DESC:FIELD OF THE INVENTION:
The present invention relates to the method of predicting the delays occurring during the manufacturing process. More particularly, the present invention relates to the method of predicting the delays occurring during the manufacturing process and the systems through data analysis making the manufacturing process smooth and quick.

BACKGROUND OF THE INVENTION:
With the industrial revolution 4.0, industries have adopted numerous modern technologies like Industrial Internet of Things IoT, Connected Systems, Predictive maintenance etc. for faster and more efficient manufacturing. In this regard, this invention deals with creating a reliable system, which could be equipped for a general setting in any manufacturing industry. The system would not only make the production process seamless and autonomous but will also provide significant improvements in many pre-existing techniques such as predictive maintenance. The aim of this invention is to combine different elements of modern day technology such as Internet of Things (IoT), Machine Learning (ML), etc. and create a mechanism that would benefit the production industry and provide increased efficiency in handling products and anomalies in the production line.

The Chinese Patent CN107045283B discloses an inference process modelling, quality prediction, and fault detection using multi-stage data separation. Process modelling techniques use a single statistical model developed from historical data of a typical process, which is used to perform quality prediction and fault detection for a number of different process states of the process. Modelling techniques determine a mean (and possibly a standard deviation) of process parameters for each of a series of product grades, production volumes, etc., compare on-line process parameter measurements to these means, and use these comparisons in a single process model to perform quality prediction and fault detection for multiple states in the process. Since only the mean and standard deviation of the process parameters of the process model are updated, quality prediction and fault detection can be performed with a single process model when the process is operating in any defined process stage or state. In addition, the sensitivity (robustness) of the process model may be corrected manually or automatically for each process parameter to adjust or adapt the model over time.

The Chinese Patent CN101140455A discloses a real-time monitoring system and a monitoring method for production processes comprise an enterprise database server, an enterprise data query terminal and a data collection terminal connected through an enterprise LAN. Wherein, the data collection terminal is composed of an address encoding unit, a collection terminal display unit, a collection terminal memory unit, a collection terminal data input device and collection terminal data I/O communication interface. The enterprise database server comprises an onsite memory unit for key standard data. In addition, the data collection terminal comprises an employee information collection terminal, an equipment information collection terminal, product selling and recovery information collection terminal and other information collection terminal required by supervision authorities. The present invention can be widely applied to safety supervision of enterprise manufacturing products involving personal safety or construction enterprises for project operation.

The Chinese Patent CN201408359Y discloses a real-time data acquisition system of industrial production line, which uses radio frequency identification technology and relates to the field of data acquisition technology; the real-time data acquisition system of industrial production line comprises a terminal reader which is used for acquiring data of radio frequency identification tag of industrial production by radio frequency identification technology, a data server for receiving the acquired data and a data acquisition computer for processing the acquired data; the terminal reader is connected with the data server, and the data server is connected with the data acquisition computer; the data can be acquired automatically by the terminal reader and using radio frequency identification technology in a real-time manner, and the data of industrial production is transmitted to the data acquisition computer in a real-time manner by the data server, a manager can acquire the information of production schedule, staff status, work efficiency of the staff and raw material loss and the like by the data acquisition computer timely, thereby arranging production reasonably and improving management efficiency.

The Japanese Patent JPS5713511A discloses an input signal and output signal of an object system are inputted to a mean value operating circuit, and a mean value, which is its output, is stored in a memory element of a signal variation extent extractor. The extractor outputs to a comparing and deciding device the extent of variation of an input signal, which is difference between the latest value of the signals and a value which has been stored in the memory element. The deciding device outputs to the extractor an off-signal and an on-signal when the extent of variation is within a present value, and when it is not within a present value, respectively. The extractor inputs a new mean value to the memory element only when the on-signal has been received. Also, the extent of variation is inputted to a simulator, an estimate of an object system output is derived, a deviation to the signal is obtained by a subtracter, and is inputted to a comparing and deciding device. In case when the deviation has exceeded a set point, the deciding device decides that a fault has occurred in the system, and display it on an indicator.

The US Patent US8571696B2 discloses methods and apparatus to predict process quality in a process control system are disclosed. A disclosed example method includes receiving process control in formation relating to a process at a first time including a first value associated with a first measured variable and a second value associated with a second measured variable, determining if a variation based on the received process control information associated with the process exceeds a threshold, if the variation exceeds the threshold, calculating a first contribution value based on a contribution of the first measured variable to the variation and a second contribution value based on a contribution of the second measured variable to the variation, determining at least one corrective action based on the first contribution value, the second contribution value, the first value, or the second value, and calculating a predicted process quality based on the at least one corrective action at a time after the first time.

D. Liebera, M.B. Stolpe, B. Konrada, J.A. Deuse and K. Morik, “Quality Prediction in Interlinked Manufacturing Processes based on Supervised & Unsupervised Machine Learning '' , 46th CIRP Conference on Manufacturing Systems, Procedia CIRP, Volume 7, Pages 193 – 198. ISSN 2212-8271, 2013. In the context of a rolling mill case study, this paper presents a methodical framework based on data mining for predicting the physical quality of intermediate products in interlinked manufacturing processes. In the first part, implemented data pre-processing and feature extraction components of the Inline Quality Prediction System are introduced. The second part shows how the combination of supervised and unsupervised data mining methods can be applied to identify most striking operational patterns, promising quality-related features and production parameters. The results indicate how sustainable and energy-efficient interlinked manufacturing processes can be achieved by the application of data mining.

M. Subramaniyana, A. Skoogha, H. Salomonssonb, P. Bangaloreb and J. Bokrantza, “A data-driven algorithm to predict throughput bottlenecks in a production system based on active periods of the machines”, Computers & Industrial Engineering, Volume 125, Pages 533-544, ISSN 0360-8352, 2018. Smart manufacturing is reshaping the manufacturing industry by boosting the integration of information and communication technologies and manufacturing process. As a result, manufacturing companies generate large volumes of machine data which can be potentially used to make data-driven operational decisions using informative computerized algorithms. In the manufacturing domain, it is well-known that the productivity of a production line is constrained by throughput bottlenecks. The operational dynamics of the production system causes the bottlenecks to shift among the production resources between the production runs. Therefore, prediction of the throughput bottlenecks of future production runs allows the production and maintenance engineers to proactively plan for resources to effectively manage the bottlenecks and achieve higher throughput. This paper proposes an active period based data-driven algorithm to predict throughput bottlenecks in the production system for the future production run from the large sets of machine data. To facilitate the prediction, we employ an auto-regressive integrated moving average (ARIMA) method to predict the active periods of the machine. The novelty of the work is the integration of ARIMA methodology with the data-driven active period technique to develop a bottleneck prediction algorithm. The proposed prediction algorithm is tested on real-world production data from an automotive production line. The bottleneck prediction algorithm is evaluated by treating it as a binary classifier problem and adapted the appropriate evaluation metrics. Furthermore, an attempt is made to determine the amount of past data needed for better forecasting the active periods.

A.I. Maniu, and V. Gh. Voda, Gh, "Prediction Based On Time Series. Applications In Quality Control", Journal for Economic Forecasting, Institute for Economic Forecasting, Vol. 0(1), pages 70-80, March 2010. A prediction model based on time series involving EWMA type approach. After a brief historical sketch and a short presentation of the GLM - General Linear Model we construct the predictor, which is an average exponentially weighted depending on previous and current values of the series. The last paragraph is dedicated to an analogy with SPC - Statistical Process Control and possible applications are emphasized. Open theoretical problems are discussed also.

Kim J., Jeong K., Choi H., Seo K. “GAN-Based Anomaly Detection In Imbalance Problems”. In: Bartoli A., Fusiello A. (eds) Computer Vision – ECCV 2020 Workshops. ECCV 2020. Lecture Notes in Computer Science, Vol. 12540. Springer, Cham, 2020. Imbalance pre one of the key issues that affect the performance greatly. Our focus in this work is to address an imbalance problem arising from defect detection in industrial inspections, including the different number of defect and non-defect dataset, the gap of distribution among defect classes, and various sizes of defects. To this end, we adopt the anomaly detection method that is to identify unusual patterns to address such challenging problems. Especially generative adversarial network (GAN) and autoencoder-based approaches have shown to be effective in this field. In this work (1) we propose a novel GAN-based anomaly detection model which consists of an auto encoder as the generator and two separate discriminators for each of normal and anomaly input; and (2) we also explore a way to effectively optimize our model by proposing new loss functions: Patch loss and Anomaly adversarial loss, and further combining them to jointly train the model. In our experiment, we evaluate our model on conventional benchmark datasets such as MNIST, Fashion MNIST, CIFAR 10/100 data as well as on real-world industrial dataset – smartphone case defects. Finally, experimental results demonstrate the effectiveness of our approach by showing the results of outperforming the current State-Of-The-Art approaches in terms of the average area under the ROC curve (AUROC).

Bergmann, P., Batzner, K., Fauser, M. et al. The MVTec Anomaly Detection Dataset: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection. Int J Comput Vis 129, Pages 1038–1059 (2021). The detection of anomalous structures in natural image data is of utmost importance for numerous tasks in the field of computer vision. The development of methods for unsupervised anomaly detection requires data on which to train and evaluate new approaches and ideas. The MVTec introduce anomaly detection dataset containing 5354 high-resolution color images of different object and texture categories. It contains normal, i.e., defect-free images intended for training and images with anomalies intended for testing. The anomalies manifest themselves in the form of over 70 different types of defects such as scratches, dents, contaminations, and various structural changes. In addition, we provide pixel-precise ground truth annotations for all anomalies. We conduct a thorough evaluation of current state-of-the-art unsupervised anomaly detection methods based on deep architectures such as convolutional auto encoders, generative adversarial networks, and feature descriptors using pertained convolutional neural networks, as well as classical computer vision methods. We highlight the advantages and disadvantages of multiple performance metrics as well as threshold estimation techniques. This benchmark indicates that methods that leverage descriptors of pretrained networks outperform all other approaches and deep-learning-based generative models show considerable room for improvement.

O.B. Sezer, M. U. Gudelek, A.M. Ozbayoglu, “Financial time series forecasting with deep learning: A systematic literature review: 2005–2019”, Applied Soft Computing, Volume 90, 106181, ISSN 1568-4946, 2020. Financial time series forecasting is undoubtedly the top choice of computational intelligence for finance researchers in both academia and the finance industry due to its broad implementation areas and substantial impact. Machine Learning (ML) researchers have created various models, and a vast number of studies have been published accordingly. As such, a significant number of surveys exist covering ML studies on financial time series forecasting. Lately, Deep Learning (DL) models have appeared within the field, with results that significantly outperform their traditional ML counterparts. Even though there is a growing interest in developing models for financial time series forecasting, there is a lack of review papers that solely focus on DL for finance. Hence, the motivation of this paper is to provide a comprehensive literature review of DL studies on financial time series forecasting implementation. We not only categorized the studies according to their intended forecasting implementation areas, such as index, forex, and commodity forecasting, but we also grouped them based on their DL model choices, such as Convolutional Neural Networks (CNNs), Deep Belief Networks (DBNs), and Long-Short Term Memory (LSTM). We also tried to envision the future of the field by highlighting its possible setbacks and opportunities for the benefit of interested researchers.

Sak, Hasim / Senior, Andrew / Beaufays, Françoise "Long short-term memory recurrent neural network architectures for large scale acoustic modeling", In INTERSPEECH, 338-342, 2014. ong Short-Term Memory (LSTM) is a recurrent neural network (RNN) architecture that has been designed to address the vanishing and exploding gradient problems of conventional RNNs. Unlike feedforward neural networks, RNNs have cyclic connections making them powerful for modeling sequences. They have been successfully used for sequence labeling and sequence prediction tasks, such as handwriting recognition, language modeling, phonetic labeling of acoustic frames. However, in contrast to the deep neural networks, the use of RNNs in speech recognition has been limited to phone recognition in small scale tasks. In this paper, we present novel LSTM based RNN architectures which make more effective use of model parameters to train acoustic models for large vocabulary speech recognition. We train and compare LSTM, RNN and DNN models at various numbers of parameters and configurations. We show that LSTM models converge quickly and give state of the art speech recognition performance for relatively small sized models.

SUMMARY OF THE INVENTION:
The principal object of the present invention is a process to detect the anomalies occurring in the industries during the manufacturing process of the goods/products and to predict future delays in manufacturing and production process of the goods/products in the industries and detect the anomalies occurring during the manufacturing process thereof. The present invention is a method of predicting delays in the manufacturing process by detecting the anomalies in the products or product parts comprises recipe block to signify processes and their parameters to make decisions and steps to manufacture the product; the said recipe block accompany with predefined rules and set parameters of production process; the process block gives an unique ID and a set ideal time of completion to each stage of production process; the data acquisition block acquires data from each stage of production where the nature of data includes machine parameters or the parameters of the product itself, useful for the inspection and anomaly detection during the production process; the anomaly detection block detects the anomalies and forms the basis of rejection and recycling of products; the anomaly decision block decides whether the product is to be send for recycling or rejection and whether the product is recoverable or not is decided by the recovery block; the non-recoverable products or products parts are sent for recycling; the rejection block discards the products or product parts completely in damaged or irreparable condition; the action block sends the products or product parts for further production process.

Another object of the present invention is to acquire data from each stage of production and feed the data to the data-sorting block for sorting the data; the ideal parameters are recorded and sorted by the standard value block; the ideal time value block acquires the ideal benchmark timing for process and the ideal rate value block acquires the rate of production; the readings of the time and production rate are recorded at the observed value block where the readings of the time and production rate are later fed to the time delay prediction system as input; the recorded time value records actual error of timing observed by difference between observed timings and ideal timings; the recorded rate value block records error of rate of production observed by difference between observed rate of production and ideal rate of production; the errors in time and production rate are analyzed as a Time Series; the time series collects and visualizes the data in a meaningful way; the time series forecasting makes future predictions of delay in the production pipeline, the predictions facilitates applications like predictive maintenance; the time series forecasting detects the anomalies and takes required action.

Another object of the present invention is to collect the production time error and production rate error from the data acquisition block where the production time error and production rate error are fed as input to the time delay prediction algorithm; the Nonparametric Time Series Forecasting is applied on the acquired data of production time error and production rate error; the application of Nonparametric Time Series on acquire data is done using Recurrent Neural Network (RNN), also known as Long Short Term Memory Cells (LSTM); the neural network undergoes the training procedure where the trained model is capable of making predictions on any future delays; the LSTM model predicts the delay in future along with error in Production rate.

Another object of the present invention is to acquire images of the products or product parts for detection of anomalies; the acquired images are processed; the trained model identifies the anomalies from the acquired images; the anomaly block decides from the processed images that whether the product or product part is anomalous or not; sometimes the faulty products are treated as good; it is necessary to control the number of such faulty products; the faulty products are controlled by focusing on the minute features of the images by identifying the Region of Interest (RoI); RoI is a smaller region within the image canvas, which is identified by the time prediction delay algorithm; the knowledge base of the anomalies is maintained; knowledge base contains the data of the previously occurred anomalies in product or product part; the RoI (Region of Interest) is fed as an input to the knowledge base to match the anomalies in the product; the knowledge base does not miss the anomalous condition of the product or product part by updating all the instances and prevents the production rate from being hampered; the instance block decides that if the product or product part is anomalous and conveys the product or product part for recycling or rejection and if the product or product part is non-anomalous, the product or product part is send for further production process.

Yet another object of the present invention is to detect the anomalies in the products or product parts causing delay in the production of the goods and maintain proper flow of production of the goods especially in the industries manufacturing Fast Moving Consumer Goods.

Another object of the present invention is to find out the cause of defects in the products or product parts while manufacturing and utilize the resources in a better manner.

Still another object of the present invention is a method to detect anomalies of the machines at different stages of manufacturing by real time monitoring, outputs important information about the manufacturing process, and analyze the performance of manufacturing process and machines.

Another object of the present invention is to extract the data of defects in products or product parts at different stages of production process using Digital Shadow, which provides important information regarding the defects in the products or product parts caused during the manufacturing process.

Another object of the present invention is to accumulate the data of defects in components/parts and command the system/user to send the particular component/part for the recycling purpose, if the defect is not found, commands the system/user to send the component/part to further stage in production process and discard the component/part, which is defective.

Yet another object of the present invention is to accumulate timing data from the different stages and calculate the error timing occurring in the manufacturing of the products/goods and predict the delay in the manufacturing of the products/goods.

BRIEF DESCRIPTION OF DRAWINGS:
Figure 1:- Illustrates the Process Flow Diagram.
Figure 2:- Illustrates flow diagram of data acquisition block.
Figure 3:- Illustrates flow diagram of series forecasting.
Figure 4:- Illustrates flow diagram of Anomaly detection.
Figure 5:- UC1: Time Series representation of Delays In conveyor Belt
Figure 6:- UC2: Magnified View of Time Series representation of Delays In conveyor Belt
Figure 7:- UC3: Time Series representation of Delays In Conveyor Belt with Anomaly Detection Time.
Figure 8:- UC4 Magnified Time Series representation of Delays In Conveyor Belt with Anomaly Detection Time
Figure 9:- UC5 Sample Images
Figure 10:- UC6 Periodic Circular Trajectory of robotic hand for data generation
Figure 11:-UC7: Time Series Forecasting data of robotic hand Time Cycle
Figure 12:- UC8: True and Predicted Time Series Forecasting data in terms of slopes

DETAILED DESCRIPTION OF THE INVENTION:
The present invention relates to the method of predicting upcoming delays in manufacturing process and system thereof. The present invention is a method, which detects the anomalies in the manufacturing process, calculates the delay occurring in the manufacturing process resulting from the anomalies or the failures in machines or component parts of machines and allows the user to repair the anomalies. The anomaly data is accumulated at each stage of manufacturing making it easier for the user to identify anomaly at a particular stage and helps to take quick action for repairing the anomaly. The present invention allows the user to identify whether the defect is repairable or not. If the defect is repairable, then the machine or component part of the machine is sent for recycling process and if the defect is non-repairable, the machine or component part of the machine is discarded. If there is no defect in the machine or component part of machine, then the machine or component part of machine is send back to production line for manufacturing process.

The manufacturing process flow describes how the application works in the manufacturing setup. The blocks in Fig. 1 represent the different tasks carried out along with their significance.

100) The process is initiated as a pipeline. The raw materials and other pre-processed materials are loaded in a synchronized manner, to mark the commencement of the production.

200) The product recipe block signifies the processes and their parameters to make decisions and steps to manufacture the product. This block has certain predefined rules. These rules help in the decision making process and identification of the right process for a production pipeline is carried out. All the steps are carried out at the action phase with either computational or manual inspection techniques.

300) The process pipeline is the combination of several independent and interlinked processes. Every process is given an ID and every process has an ideal time of completion which is fixed (Tj). Over the given time (Tj), the rate of production is also fixed (Qj) for a given (Tj) . This block takes into account the sequential and parallel processes as it is typically found in industry.

400) Data Acquisition is done from each individual process in the process pipelines. The nature of data would include machinery parameters or the parameters of the product itself. This data is useful for the inspection and anomaly detection phase of the manufacturing process. The nature of collected data varies according to the type of product or process. The classes of data include but are not limited to: images, video feed or raw numerical data.

500) After the end of manufacturing procedures the product undergoes an inspection to judge the quality and the quantity of the product that will be packaged in the forthcoming step. The anomaly detection block uses different techniques to find the possible anomalies in the product. This block forms the basis of rejection and recycling of products. In case of no anomalies, the product is forwarded for further processing (e.g. - Packaging, Manual Inspection etc.)

600) The anomaly decision block identifies the anomalies in the product. If the product is recoverable, the process product to the recovery block to identify the possible components that can be recovered. Otherwise, the product moves onto the action block for further processing.

700) This block identifies whether the components of the product can be recovered. This stage is very useful in order to minimise wastage in a plant. This would ensure that monetary losses due to wastage are significantly reduced with the recovery of components.

800) The spare parts or components In case of non recoverable products of the product would be recycled. This would reduce the demand for the raw materials which in-turn would result in an increase in efficiency of the usage of the raw materials. Therefore at this step, there would be a feedback to the product whenever the spare parts of the product are recyclable.

900) This block is used to discard the products or parts of the product which are completely damaged during the production process. This block ensures the rejection of the product if it is in a non-recyclable or non- reusable state.

1000) The action block is a descriptor for the further processing that a product undergoes. This processing can be in the form of packaging or maybe a manual quality check or inspection. This block also carries out the finishing task manually for some products, as required.

1100) At this stage the pipeline comes after coming across many inspection stages to an end; and the finished product is packaged and ready for dispatch. This is the final step of implementation in the proposed manufacturing process. It is viable in most cases as more quantities are packages in boxes. Therefore, products wait till specified quantity arrives to the same stage

The idea and novelty of the Production pipeline is the ‘Future delay prediction’ which involves different sub-processes. Once the data is acquired from the processes, the next step would be the utilisation of this data and analyze it using a single algorithm or abstract individual processes. Fig. 2 explains how this data will be analysed and utilised for the prediction applications. Each block from Fig. 2 is explained in detail below:

410) The sorting block uses the data acquired from the various processes. Every process tends to have benchmark data and the observed data. This block sorts the benchmark data and the observed data from the array of data of various processes that has been acquired.

420) The standard parameters that are recorded for a particular sub-process and are sorted into this module. These parameters are dependent on the industry specification and the kind of product that is being manufactured. Ideal Parameters are constant for a process at any point of production. These parameters in general consist of the time for production and rate of production metrics.

421) This block acquires Ideal benchmark timing for process completion, as set by the manufacturer. Industrial processes are cyclic and each cycle has a time of completion. This benchmark time data varies according to the product type. Some products have an objective towards higher production rate and others need higher precision. Therefore, depending on the requirement of production, the manufacturers would have the flexibility to change the time of completion of a particular process.

422) This block acquires Ideal benchmark for rate of production which was set by the manufacturer. This value depends upon the state of machinery and how fast the machine is able to manufacture the product. Every industry has a goal towards its production. This is mainly seen in the FMCG industry where productivity is very huge with respect to its demand. Therefore, it is important to take note of the rate of products being produced.

430) Observed Parameters are the recorded readings of time and production rate, which are obtained from the machinery involved in the process or from the process itself. The observed timings and rate of production are recorded here in this stage and later fed to our time delay prediction system.

431) Actual data points collected from 430 are in real time during the manufacturing process. These readings are continuous with respect to the timestamps or Point of time collected. Consider a process which needs to put labels on boxes. For each box, time taken might not be the same. This may be due to faults in machines. Nevertheless this affects the entire productivity.

432) This is the observed rate at which the production is currently taking place. Production rate is the ratio between the number of products being produced by the process to the number of products coming into the processes. For instance, production rate is assumed as the process that can take 5 products in a particular time but forwards only 4 due to machine faults. Ideally this value should be 1 because incoming products do not get delayed midway. In the observed scenario, some products may be lost due to anomalies. Ultimate aim is to minimize the anomalies by figuring them out priorly.

440) Every process will have observed parameters and there is no guarantee that both observed and Ideal may not be equal. The time error is obtained by referencing the observed timings with the ideal timings. The mathematical aspect of calculating the error may vary according to the complexity of process and the complexity of data. The error helps in figuring out the convergence or divergence between ideal and observed parameters.

450) The error rate is obtained by referencing the observed rate with the ideal rate. Calculating this error is tedious because of multiple factors that affect the rate of production. The ideal readings may also vary depending on the physical and technical situation. Machines frequently cause failure due to Internal or External faults. The delay might occur even if faults have no presence because in interlinked processes, one process could affect others too. Therefore resulting in delay of pipeline.

460) Both the errors in time and rate are analysed as a Time Series. The most important use of studying time series is that it helps us to predict the future behaviour of the variable based on past experience. It is helpful for business planning as it helps in comparing the actual current performance with the expected one. The purpose of time series here is to help in collecting and visualizing data in a meaningful way. This helps in making further predictions by using Time Series forecasting techniques. Time series forecasting can make future predictions of delay in the production pipeline, these predictions will facilitate applications like predictive maintenance.

Anomalies have a huge role in adding delays to processes. The presence of frequent anomalies adds up extra time and negatively diverges from the ideal rate of production. Fig. 3 explains the process of anomaly detection and required action after detection.

461) This block collects both time and rate errors with respect to their timestamps and makes it into a suitable form of data which requires further processing. The converted data is fed as an input to the prediction algorithm. This time data is extracted from the block (400). The time series data is often visualized graphically for an easy to understand representation. This is done by comparing the true data with its predicted ones.

462) On the data, Nonparametric Time Series Forecasting is applied using techniques such as but not limited to: Recurrent Neural Network (RNN), Long Short Term Memory Cells (LSTM). An LSTM module has a cell state and three gates which provides them with the power to selectively learn, unlearn or retain information from each of the units. The cell state in LSTM helps the information to flow through the units without being altered by allowing only a few linear interactions. Each unit has an input, output and a forget gate (Cache gate) which can add or remove or clean the information to the cell state. The forget gate decides which information from the previous cell state should be forgotten and uses a sigmoid function. The input gate controls the information flow to the current cell state using a pointwise multiplication operation of ‘sigmoid’ and ‘tanh’ respectively. Finally, the output gate decides which information should be passed on to the next hidden state. In LSTMs, The data is time flagged and each cell's output depends on the cumulative output of all the previous cells (i.e the past values in time). To generalise the algorithm, the output of a cell at tn depends on the output of tn-1, tn-2 ,.............., t0

463) Once the neural network undergoes the training procedure, a trained model is obtained which is capable of making predictions on any future delays. The trained model is updated with the new data stream. Continuous update to data is important for algorithms to be synchronised. It is important to keep your model updated for effective and true predictions.

464) Using the LSTM model, the delay in future can be predicted along with error in Production rate. This use case of our model will assist in applications such as predictive maintenance.

510) Visual Inspection methods for anomaly detection are very popular in today's research. Visual inspection is used in manufacturing for quality or defect assessment. In non-production environments, it can be used to determine whether the features indicative of a target are present and prevent potential negative impacts. One of the most important steps is Image Acquisition. The images are acquired individually, marked with the same ID as the product. This will make the further processing of the product easier.

520) The image acquired is to be processed. There are a significant number of techniques available. However, most of them have a class imbalance problem i.e.; they require a significant number of classes of data. Use of GAN (General Adversarial Networks) has good results for imbalance data, but the techniques are not specific /designed for anomaly detection. Hence the processed image forms the basis for our anomaly detection, A trained model is able to identify the faults and anomalies in the product image.

530) The anomaly decision block helps in deciding whether the product is anomalous or not. This decision is made from the result of 520. 520 outputs RoI (Region of Interest) for every image based on the training set but labeling is done here. If output is anomalous then the product is pushed to the reject unit, otherwise it proceeds for further processing.

540) False Positives (faulty product but treated as Good) are inevitable in the production line, but it is critical to minimize their number. In order to tackle the scenario of false negatives, the focus is on the minute features of the product image. These features are extracted by identifying the RoI. Image segmentation is the best technique for RoI extraction.

550) A knowledge base of possible anomalies is maintained. Knowledge base contains the data of most possible anomalies that occurred previously. Each individual RoI is fed as an input to this knowledge base to match for any anomaly in the product. This knowledge base is expected to be updated at all instances, so that no anomalous condition is missed and the production sequence does not get hampered.

560) The instance block takes a decision of forwarding to rejection or that requires further inspection. It is the last decision block and any further selections or findings should follow the same flow chart repetitively irrespective of product or its nature.

570) The product could be sent in for rejection or recycling depending on the nature of the anomaly. It would depend upon the severity of the fault, as described in the knowledge base.

580) During the course of anomaly detection, there may be instances of false negatives and false positives. In the case of the FMCG industry, False negatives are mostly ignored but dealing with False positives is a critical aspect. This is the reason for using a two-step verification process for the detected anomaly. In the first step, the anomaly detection is done for the entire acquired image. In the second step, there is a comparison between the acquired image and the knowledge base which was previously collected. Doing this, one can come to a better prediction of finding defects in products. Once it is verified that there is no anomaly in the product, the product is cleared for further processing.

As shown in Figure 1 of the present invention, it relates to the process flow diagram of time prediction delay system 50. The process flow diagram represents the different tasks carried out along with their significance. The process begins as a pipeline. The manufacturing process begins from the starting point 100 by loading the raw materials. The product manufacturing parameters are set at the product recipe block 200 allowing the user to decide the steps to manufacture goods. The steps at the product recipe block 200 are carried out at the action phase with either computational or manual inspection techniques. The processes block 300 are divided into interdependent and interlinked processes. All the processes block 300 possess their own time of completion based on which the production rate is also fixed. Data is accumulated from each stage of the process at the data acquisition block 400. The data obtained at data acquisition block 400 includes machinery parameters or the parameters of the product itself. The data collected helps to find out the defect, inspect and repair the same. The classes of data include but are not limited to images, video feed or raw numerical data. The product is inspected after the manufacturing process where the quality of the product is checked. Different techniques are used to detect the anomalies at the anomaly detection block 500. At the anomaly detection block 500, it is decided that whether the product/good, machine or machine component is to be send for recycling process or rejected. If no anomalies are detected at the anomaly detection block 500, then the product/goods, machine or component part of machine is sent back to further manufacturing. The decision of whether the product is anomalous or not is decided at the anomaly decision block 600. If the product is anomalous, the process product moves to the recovery block to identify the components that can be recovered. If there is no anomaly, the product moves onto the action block for further processing. The recoverable block 700 identifies whether the product of component is recoverable or not. This stage is very useful in order to minimize wastage in a plant. This would ensure that monetary losses due to wastage are significantly reduced with the recovery of components. In cases where the products are not recoverable, the component parts of the products are recycled at spare part recycling block 800, which would decrease the demand of raw materials resulting in increase in efficiency in use of raw materials. The parts or products completely damaged during the production process are discarded at the rejection block 900. The rejection block 900 ensures that rejection is done only if no part of the product is in a recyclable or reusable state. After the products or parts are discarded, the products that are recycled or defect is repaired are sent to the action block 1000 where the products or parts are sent for further production process. Once the final product is packaged and dispatched, the production line ends at the stop block 1100.

As shown in Figure 2 of the present invention, it relates to flow diagram of data acquisition block 400, explains how the acquired data is analyzed and utilized for the prediction of delay in production process. The data acquired from each stage of production is gathered at the data-sorting block 410. The data-sorting block 410 gathers the benchmark data and observed data. The data-sorting block 410 sorts the benchmark data and the observed data from the array of data of various processes that has been acquired. The standard parameters that are recorded for a particular sub-process and are sorted into standard value block 420, these parameters are dependent on the industry specification and the kind of product that is being manufactured. Ideal Parameters are usually constant for a process at any point of production. These parameters usually consist of the time for production and rate of production metrics. The ideal benchmark timings of completion of process is acquired in the ideal time value block 421 as set by the manufacturer. The ideal benchmark of rate of production set by the manufacturer is acquired by the ideal rate value block 422. The data acquired by the ideal rate value 422 depends on the speed, efficiency and time taken by the machine for the production process. The recorded readings of timing and production are recorded in the observed value block 430. The recorded readings from observed value block 430 is sent to the time delay prediction system 50. The real time recordings of production are recorded at the recorded time value block 431. The observed rate of production are recorded at recorded rate value block 432. The time error between the observed value and recorded value are calculated by the time error block 440 making it easier to find out the duration of difference between the observed value of recordings and ideal value of recordings. The difference of observed rate and ideal rate is obtained at the rate error block 450. The time series forecasting block 460 analyzes the errors in time and rate, which helps in predicting the delay in manufacturing.

As shown in Figure 3 of the present invention, it relates to the flow diagram of time series forecasting 460 wherein the anomalies are detected and required action is taken. The time data block 461 collects the data from data acquisition block 400 and the collected data is sent to the prediction algorithm. The time series data is often visualized graphically for an easy to understand representation. The Nonparametric Time Series Forecasting is applied on data, which is done using Recurrent Neural Network (RNN), also known as Long Short Term Memory Cells (LSTM). The LSTM block 462 has a cell state and three gates, which provides them with the power to selectively learn, unlearn or retain information from each of the units. The cell state in LSTM helps the information to flow through the units without being altered by allowing only a few linear interactions. Each unit has an input, output and a forget gate which can add or remove the information to the cell state. The forget gate decides which information from the previous cell state should be forgotten for which it uses a sigmoid function. The input gate controls the information flow to the current cell state using a pointwise multiplication operation of ‘sigmoid’ and ‘tanh’ respectively. Finally, the output gate decides which information should be passed on to the next hidden state. A training model 463 is obtained after the neural network undergoes the training procedure, which is capable of making predictions on any future delays. The trained model 463 can be re-trained with the new data stream. The delay prediction block 464 predicts the delay along with the error in production rate by using LSTM 462.

As shown in Figure 4 of the present invention, it relates to the flow diagram of anomaly detection 500. Image acquisition block 510 is very important process for the detection of anomalies. It becomes easier for the user to identify the anomalies through images. The images are acquired individually, marked with the same ID as the product. This will make the further processing of the product easier. The images acquired from the image acquisition block 510 are processed at the process image block 520. The anomaly decision block 530 decided from the results of process image block 520 that whether the product is anomalous or not. False Positives (faulty product but treated as Good) are inevitable in the production line, but it is critical to minimize their number. In order to tackle the scenario of false negatives, the focus is on the minute features of the product image. These features are extracted by identifying the RoI (Region of Interest) algorithm 540. The data of anomalies occurred previously is maintained by knowledge base of anomalies 550. The RoI 540 is fed as an input to the knowledge base of anomalies 550 to match for any anomaly in the product. This knowledge base is expected to be updated at all instances, so that no anomalous condition is missed and the production sequence does not get hampered. Whether the product is anomalous or not and if yes, then whether the product should go for rejection or recycling and if the product has no anomalies then the product should be send further for production process, is decided by the instances block 560. The product is sent for rejection or recycling through rejection/recycling block 570 depending upon the severity of anomaly. Two step verification is done for detection of anomaly. In the first step, anomaly detection is done for the entire image acquired. In the second step, there is a comparison between image acquired from the image acquisition block 510 and knowledge base for anomalies block 550, which was previously collected. Upon doing this, one can come to a better explanation of finding defects in products. Once it is verified that there is no anomaly in the product, the product is cleared for further processing through clearance block 580.

One of the preferred embodiments of the present invention is to provide a time prediction delay system 50, which allows the user to identify the defects in the product or product parts and help in predicting delay resulting in timely repairing of components and maintain the flow of production.

One of the preferred embodiments of the present invention is to find out the causes of failures or defects in the products or product parts, repair or recycle the product or product parts and utilize the resources efficiently.

One of the preferred embodiments of the present invention is to find out the anomalies at each stage of production by calculating difference between observed values and ideal values of time and production.

One of the preferred embodiments of the present invention is to acquire the images of the products or product parts through image acquisition block 510 and send the products or product parts for rejection or recycling through rejection/recycling block 570.

One of the preferred embodiments of the present invention is to detect the anomaly in product or part of product and decide whether the product or product part should be sent for rejection or recycling and if there are no anomalies in the product or product part, then send the product or product part for further manufacturing process.

Experiment 1:
The experiment to validate process pipeline, by testing the effects of time delay on the operations of conventional industrial machinery. The expriment primarily focus on the complete behaviour of the conveyor belt, ranging from mechanical aspects to possible anomalous behaviour (caused due to unknown characteristics). The conveyor belt plays an important role in transportation of products from one stage to another. The multiple variants offer solutions for a wider range of applications. For our experimental set-up, the possible choices of conveyor belt are as follows (but not limited to):
a. Roller Bed Conveyor Belt
b. Flat Belt Conveyors
c. Modular Belt Conveyors
d. Cleated Belt Conveyors
e. Curved Belt Conveyors
The output of the use-case will forecast the future delays in the functioning of the conveyor belt. The focal point will be the time delay only and not the cause behind the time delay.
Parameters:
Vital details of the experimental set-up have been furnished below, to understand the complete apparatus and its functionality.

The length of the conveyor belt is 10 feet. It moves at a speed of 65 feet per minute. Hence, from the relation, velocity = distance / time; the product will cover the belt distance in approximately 9 seconds (This is the ideal time requirement for a product to traverse from one end to another). The conveyor belt is equipped with a vision-based anomaly detection module whose anomaly determination time is ideally 1 second (Theoretical time complexity). Taking into consideration, the delay caused due to fetch, compute and release of data through the cloud. The total time required for anomaly determination will fall into the range of 2 to 2.5 seconds. Therefore, the total time of traversing for a product on the conveyor belt is: 9 + (2 to 2.5) = (11 to 11.5) seconds approximately.

Detailed Working:
The industry incorporates PID (Proportional Integral Derivative) controllers to maintain the parameters of any given machinery. Since, it is practically impossible to converge to a given parameter with 0% error, the ideal parameters are not achieved. The same applies for the conveyor belt. As an effort to achieve the ideal parameters (as discussed above), the observed parameters in the form of time and rate are acquired and used to predict the future behaviour, presenting a scope for taking preventive measures. Even though our primary focus is on delay prediction, We would like to illustrate a few possible reasons for this increased time of production. This will provide the much needed correlation for the delay (We don’t endorse the reasons as the primary reason for the caused delays). These reasons include, but not limited to:
a. Mechanical defects in conveyer belt
b. Excess weight on Conveyor belt
c. Irregularities in power supply
d. Low performance of the Anomaly Detection module

The data collection is done at an interval of 100 millisecond and relayed to the remote storage facility (e.g. - cloud). The data acquired from the conveyor belt is in Time Series format; as can be seen in Fig. 2. With the predicted delays, the average error in operations of the conveyor belt can be minimised. The overview of the prediction process can be referred from Fig. 3.

A. Data Generation:
Generally, conveyor belts are meant to operate continuously without halts. This might result in gradual reduction in performance. In order to predict this behaviour, historical working data would be trained alongside careful inspection. Consider a conveyor carrying sets of packaged boxes. The objective is to check whether the quality of boxes meets the product standards (as per manufacturer) and at the same time move them from one section to another. Therefore doing multiple tasks simultaneously is of greater significance. The inability to perform tasks simultaneously results in increased time delays. Before any inspection, the stand alone behaviour of the conveyor belt must be monitored. Similarly, a vision-based inspection unit, typically consisting of a camera and a microcontroller to capture images and process them locally or over the cloud, should be put in place which will capture and inspect the product quality.

To begin with, synthetic (computer program generated) data generation is done and the same data is utilised for prediction purposes. The primary reason to generate a synthetic data set is the unavailability of publicly accessible data sets (of real time conveyor belt parameters). As per literature, anomalies are the result of irregularities in the manufacturing process. Hence, generating an irregularity is an easy task through simulation. In the particular use case, the functionality of every parameter can be determined from the time required for each process. Hence, any deflection in ideal time (specified by manufacturer) will in turn increase delays. Hence, adding suitable noise for ideal parameters will lead to fault generation. In order to inspect and predict the future behavior, both scenarios of - presence of anomaly and absence of anomaly are investigated and presented below:

1. A simulator tool known as ‘V-Rep’ is used. V-Rep assists in setting up the optimal environment for the simulation.
2. In order to introduce the delays which synthetically slow down the Conveyor Belt, the input parameters of the conveyor are modeled aligning to the real world scenario. This is done by introducing two reference points onto the conveyor belt: one that references the movement of the product and other being stationary (end point).
3. Initially, both points coincide and the time cycle is measured when the reference point on the conveyor belt moves around and converges again to the end point.
4. The speed of the conveyor belt and the time of acquisition are managed in V-rep while Gamma noise function is used to add faults.
5. At the edge of the conveyor, a camera is set up which internally triggers an anomaly detection mechanism.
6. This runtime is added to the Time cycle of the conveyor belt to get the total time cycle. The total time cycle is the sum of Time Cycle without Anomaly Detection Time + Runtime of code for anomaly detection.

After data generation, the data is stored for a specified time frame. The stored data (approximately 25 minutes) is then sent to the time series forecasting mechanism. The details pertaining to the output of the time series mechanism is explained in Fig. 3. In the environment, several physical properties are abstracted into a single variable of time. For scenarios where the goal is to predict the delay after X amount of time has elapsed (Our prediction set point, X = 0.25 hour) the process flow will initiate a prediction window of 0.25 hour. Notably, X is the ideal time of production through which delays shall be compared. Hence, the value of X is tuned as per requirement. For a smaller prediction window size (X), the precision of prediction is higher compared to a larger window. It is conclusive of the fact that the amount of training data remains the same but the prediction horizon expands thus hampering the precision. The anomaly detection module from Fig. 4 gives the decision on whether the product is being manufactured correctly; or it would move towards the process of rejection/ recycling.

Results:
In accordance with the simulator, results are classified into 3 scenarios.
1. Scenario-1: It represents delay prediction for the conveyor belt.
2. Scenario-2: It represents delays caused when an anomaly is detected in a unit.
3. Scenario-3: It represents prediction of rapid changes and its analysis with true data.

Scenario 1:
1. The data collected earlier is trained for 1500 seconds and predictions for the forthcoming 1000 seconds are shown in Fig. UC1.
2. Variation from True and Predicted values is presented in a magnified view in Fig. UC2. This shows how prediction and true data converges or diverges.
3. Mean Square error (MSE) is taken as a performance metric and the goal is to optimize MSE. MSE is defined as Mean or Average of the square of the difference between actual and estimated values.
4. Since the values are estimated, MSE calculates errors and also puts great emphasis on the large errors that can make a great impact.
5. Generally MSEs are focused on future predictions, which in turn help in decision making through estimations.
6. In the case of Zero error, the MSE is an ideal situation. Though it is highly improbable to find it, this method gives satisfiable outcomes that can be used for forecasting.
n
MSE = [ (1/n)? (Xobs,i -Xmodel,i)² ]
i=1
7. The Ideal graph resembles a rectangular wave with unequal widths. Width of each rectangular pulse defines the time taken by one cycle.
8. Data collection is discrete which makes graphs easily identifiable and see the contrast between the forecasted and ideal graphs.
Observations:
1. It is observed from both Fig. UC1 and its magnified view of Fig. UC2 that true values and predicted ones are very similar. This implies that through the use of the proposed mechanism, the true values can be predicted. Further, the predicted value gives the result with slight delay which can be attributed to the additional time required during the training and testing process.
2. The calculated MSE from the simulated results is equal to 3.25, compared to classical models where MSE ~ 6 are considered to be a good fit, our method improves the error defiance capability by almost 45%
3. Notably, the predicted values display a low magnitude of divergence even after a lack of random data change points.

Scenario 2 of Experiment 1:
1. The run time delay caused by remotely deploying the visual inspection mechanism on the cloud infrastructure is added upon the operation time of scenario 1.
2. When the product is in the vicinity of the camera, the camera captures the image of the product. The obtained image is processed and output is generated in accordance with Fig. 4 of the flow chart.
3. It is to be noted that both the quality and reliability are equally important and need to be balanced. In order to maintain this balance, a two-step inspection mechanism is introduced.
4. The first step of the two-step rectification is to analyze the captured image. Further, if the output is anomalous, the unit proceeds to the rejection block. The second step is followed only when the unit is detected to be non anomalous.
5. In the 2nd step, the image segmentation methods are used to identify the Region of Interest (RoI) before using the classification techniques. This helps in giving a reliable decision, thereby increasing the accuracy. Significantly, this results in a reduction of the false positives, as specified in Fig. 4 of flow chart.
6. The run-time of the anomaly detection mechanism is added to the time of operation calculated in Use Case-1. This gives us the complete functional time as seen in conveyor belts used in industries.
7. Notably, with the increased operation time (Time delay due to cloud deployment + Regular operation time) in presence of anomalies, the predicted values initially diverge with true ones. With sufficient training, the resultant divergences are found to be tracing with the true values. This can be seen in Fig. UC3 which is further magnified and shown in in Fig. UC4. The Fig. UC4 clearly shows the variation of true values and the same being traced in the prediction with latency.
8. The divergence between true and predicted values is due to irregularities in code compilation time. This irregularity is not approximated with any probabilistic functions considering the randomness in network throughput (To and fro movement of data through cloud).
9. It can be observed from Fig. UC4 that the average offset is approximately equal to ± 0.7 sec. {Offset = |Avg(True)| - |Avg( Pred)| }. ( || -> Modulus )
10. The positive part of the offset indicates the increasing trend in the predicted graph while the negative part indicates the decreasing trend in offset.
Analysis:
a. Three different techniques (GANomaly, CAE:VAE and Skip-GANomaly) were investigated to analyze the performance.
b. The anomalies happening more frequently with higher probability) are grouped together and a separate class is created.
c. MVTEC benchmark data set is used for all observations.
d. The accuracies were tested for “all texture classes”, “all object classes” and for “all classes”; for both before segmentation and ‘after segmentation’.
e. The accuracies are presented as follows. It can be observed that for all 3 techniques, the accuracy improves considerably after segmentation. This shows the benefit obtained through segmentation. Further, the accuracy is the highest when it is for “all object classes”
f. Fig UC5 has 5 sub-figures. The Fig.a, Fig b, Fig c, Fig e are the sample images tested and their output is shown. In Fig d and Fig e. Segmentation removed the maximum anomalous region which might give less RoI area. This eventually becomes a false positive. This problem can be solved by increasing the accuracy, which can be done if the training data is large enough.
Class Description GANomaly
(Technique 1) CAE:VAE
(Technique 2) Skip-GANomaly
(Technique 3)
Before Segmentation After Segmentation Before Segmentation After Segmentation Before Segmentation After Segmentation
All Texture Classes 72.8% 74.8% 74.7% 77.7% 73.6% 77.6%
All Object Classes 85.6% 84.6% 86.8% 88.8% 86.4% 86.8%
All Classes 79.2% 80% 81% 83.5% 80% 82%

Experiment 2: Robotic Arm for Manufacturing
Background: Today, in industry, the use of Robots has replaced people in executing tedious, recurring production activities thereby, increasing productivity, efficiency and reducing the cost. A significant aspect of Robotics is the use of the Robotic arm. A Robot with “N” degrees of freedom would have N links (typically, steel sections) and N - 1 joints. This Robotic arm could be used for numerous manufacturing activities, ranging from preparation of food items, pharmaceuticals, bottling, manufacturing of FMCG (fast moving consumer goods), distilleries, vehicles, etc. Mostly, the Robotic arm is used for picking and packing from the assembly lines. Further, assume a Robotic arm needs to perform an activity such as Drawing a circular trajectory periodically. Each cycle has a time of completion and a constant rate of completing the task. Ideally the rate of operation is specified by the manufacturer though in practise, it is not always required that the ideal time cycle would be equal to the actual observed time; as described in the flowchart in Fig. 2. The anomaly in the Robot-based manufacturing process is very less; and would arise primarily due to either: Environmental effects, Internal Kinematics or the error in Sensors data. For each of the anomalies there needs to be an inspection process which takes an additional amount of time on each cycle of task.

Hence, the ideal time of completion equals the time taken by the Robotic arm to move the product from Point A to B around a circular trajectory without any defects in the robot. On the other hand, the Observed Time of Completion = Ideal Time of completion + Time taken by Anomalies. Notably, the Observed Time Cycle is always more than or equal to the ideal Production Time Cycle. Importantly, the time cycle considered can be referred to as Round Trip Time.

Data Collection:
Generally, robotic arms have predefined trajectories. They follow certain predefined rules and typically, have no decision making capabilities in the dynamic operation environment. Most of the industries use/aim to use some kind of robotic arms. The performance of the robotic arm can be made adaptive by being able to bring-in intelligence which would enable making changes/decisions in the dynamic environment. This would in-turn assist in identifying anomalies as traditional rules would not have information about faults and anomalies in the process.

A simple way to add intelligence is to train the robots for different scenarios and get a lot of data. Though it makes the development computation heavy, it would guide the robot to follow a certain trajectory pattern in taking decisions. At the same time, this additional processing adds some time. Delays in both case studies represent the existence of anomalies in machinery. Therefore predicting before the occurrence of the failure/potential latency helps to decrease the delays. The process of operation is followed from Fig. 2 of the flowchart. Therefore, the time series data based prediction is given more importance than any other forms. The technique used is as shown in Case-Study 1; and is also represented in Fig. 3 in the drawings/flowcharts.

An industrial robotic arm with 3 DoF is created which has a high repeatability, i.e., there is no randomness involved in the operation. This is shown in Fig. UC6. Repeated circular trajectories were performed and the round trip time (RTT) of the entire system was collected. In the case of circular trajectory, the ideal or acceptable RTT is in the range of 2.75 to 3.25 seconds. But the observed RTT was slightly higher. This is mainly because of the faults, occurring either internally or externally. Hence, RTT is used to figure out the faults in the machines.

Analysis and Results:
1. Data is collected for 1500 seconds and trained using Time Series Forecasting algorithms specified in Fig. UC7.
2. The data is collected and predictions are made simultaneously based on the data. The algorithm is constantly updated with the newly obtained data.
3. Both true and predicted data have substantial correlation or Less Error which suggests that the model is working efficiently. This is shown in Fig. UC7 where true and predicted are ideally the same.
4. An MSE of 1.4 which validates the point that Time Series Forecasting algorithm successfully predicts in future Round Trip Time.
5. Tracing in graphs displays the changes being predicted with good accuracy. If tracing has less divergence then the algorithm is predicting correct values.
6. Predicted values trigger the same changes whenever Ideal or True Value changes even when the time ( Y-Axis ) is changing at a faster rate.
7. Fig UC8 is the Integral Representation of Fig. UC7. Providing more information in terms of subtle changes of slope of functions. It is observed in Fig. UC8, that both predicted and true values have similar slopes which suggests performance of the algorithm is optimum.

Hence, taking the measures of ideal and observed metrics and feeding the data into Fig. 4 of flow charts gives future delays as shown in Fig. UC7. Before feeding the data, manufacturers need to set a prediction horizon. The Prediction Horizon implies how far is the mechanism able to forecast at the moment. When the prediction horizon is similar to the delay between ideal and predicted time, manufacturer learns how to control the system more rapidly and achieve better performance

The Prediction Horizon tells how far ahead the model predicts the future. When the Prediction Horizon is well matched to the lag between input and output, the user learns how to control the system more rapidly, and achieve better performance. On increasing the Prediction Horizon, the output performance may be reduced. This is observed in Case Study-1, Scenario-2. Therefore Prediction Horizon should be tuned properly. Finally, after finding the delays from Prediction Horizon, one can conclude whether the Robotic Arm has any defects in its working. This is done when the prediction horizon value diverges or deviates from the ideal case.
Advancements
This work brings in a mechanism for predictive maintenance that can be used in manufacturing industries, especially where the rate of production is very high. This is because, in the FMCG (fast moving consumer goods) manufacturing industries, the rate of production is very high. Hence, the machines cannot be shut-down as and when required. This invention proposes a dynamic predictive maintenance process and makes use of novel technologies based on time series forecasting mechanism on the data collected and the pattern observed to detect the anomalies.
,CLAIMS:CLAIMS:
We Claim,

1) A method of predicting delays in the manufacturing process comprises:
a) manufacturing process loading the raw materials at starting point 100;
b) the product manufacturing parameters are set at the product recipe block 200 and processes block 300;
c) each processes block 300 possess time of completion based on which the production rate is also fixed;
d) data is accumulated from each stage of the process at the data acquisition block 400 of process parameters, machine parameters and product parameters;
e) determining variation in data based on the received process parameters, machine parameters and product parameters at the anomaly detection block 500, based on a set ideal time of completion to each stage of production process it is decided at the anomaly decision block 600 that whether the product, machine or machine component is to be send for recycling process or rejected;
f) determining variation in data the product is anomalous, the process product moves to the recovery block 700 to identify the components that can be recovered, the products are recycled or defect is repaired are sent to the action block 1000 where the products or parts are sent for further production process;
g) where the products are not recoverable, the component parts of the products are recycled at spare part recycling block 800; the parts or products completely damaged during the production process are discarded at the rejection block 900 and ensures that rejection is done only if no part of the product is in a recyclable or reusable state;

2) The method of predicting delays in the manufacturing process as claimed in claim 1 wherein, the data include but are not limited to images, video feed, raw numerical data, process parameter data, product parameter data.

3) A method of predicting delays in the manufacturing process the data acquisition block 400 data acquired comprises:
a) from each stage of production is gathered at the data-sorting block 410 the data-sorting block 410 gathers the benchmark data and observed data and sorts from the array of data of various processes that has been acquired;
b) the standard parameters for a particular sub-process are sorted into standard value block 420;
c) the ideal benchmark timings of completion of process is acquired in the ideal time value block 421 as set by the manufacturing process;
d) the ideal benchmark of rate of production set by the manufacturing process is acquired by the ideal rate value block 422;
e) the recorded readings of timing and production are recorded in the observed value block 430;
f) the recorded readings from observed value block 430 is sent to the time delay prediction system 460;
g) the real time recordings of production are recorded at the recorded time value block 431;
h) the observed rate of production are recorded at recorded rate value block 432;
i) the time error between the observed value and recorded value are calculated by the time error block 440 making it easier to find out the duration of difference between the observed value of recordings and ideal value of recordings;
j) the difference of observed rate and ideal rate is obtained at the rate error block 450;
k) the time series forecasting block 460 analyzes the errors in time and rate, which helps in predicting the delay in manufacturing.

4) The method of predicting delays in the manufacturing process as claimed in claim 3 wherein, standard parameters for a particular sub-process are dependent on the industry specification and the kind of product that is being manufactured.

5) The method of predicting delays in the manufacturing process as claimed in claim 3 wherein, Ideal Parameters are usually constant for a process at any point of production.
6) The method of predicting delays in the manufacturing process as claimed in claim 3 wherein, Ideal Parameters consist of the time for production and rate of production metrics.

7) The method of predicting delays in the manufacturing process as claimed in claim 3 wherein, the data acquired by the ideal rate value 422 depends on the speed, efficiency and time taken by the machine for the production process.

8) A method of predicting delays in the manufacturing process the time series forecasting 460 comprises:
a) The time data block 461 collects the data from data acquisition block 400 and the collected data is processed for determining if a variation;
b) The Nonparametric Time Series Forecasting is applied on data, which is done using Long Short Term Memory cells (LSTM), an advancement of the RNN;
c) The block 462 has a cell state and three gates, which provides them with the power to selectively learn, unlearn or retain information from each of the units;
d) a training model 463 is obtained after the neural network undergoes the training procedure, which is capable of making predictions on any future delays;
e) the trained model 463 can be re-trained with the new data stream;
a) the delay prediction block 464 predicts the delay along with the error in production rate by using LSTM 462;
b) The time series data is visualized graphically,

9) A system of predicting delays in the manufacturing process the anomaly detection 500 comprises:
a) Image acquisition block 510 for the detection of anomalies the images are acquired individually, marked with the same ID as the product;
b) The images acquired from the image acquisition block 510 are processed at the process image block 520;
c) The anomaly decision block 530 decided from the results of process image block 520 that whether the product is anomalous or not;
d) The data of anomalies are stored in knowledge base of anomalies 550;
e) The RoI 540 is fed as an input to the stored data from knowledge base of anomalies 550 to match for any anomaly in the product;
f) The data of anomalies is updated at all instances, so that no anomalous condition is missed and the production sequence is not hampered;
g) determining variation in the product for rejection or recycling
h) the product has no anomalies then the product is sent further for production process, is decided by the instances block 560;
i) the product is sent for rejection or recycling through recycling block 570 depending upon the severity of anomaly;
j) Two step verification is done for detection of anomaly, In the first step, anomaly detection is done for the entire image acquired, In the second step, there is a comparison between image acquired from the image acquisition block 510 and knowledge base for anomalies block 550, which was previously collected;
k) verified product there is no anomaly in the product, the product is cleared for further processing through clearance block 580

10) The system of predicting delays in the manufacturing process the anomaly detection 500 as claimed in claim 9 wherein, False Positives (faulty product but treated as Good) are inevitable in the production line, which are tackle by focus is on the minute features of the product image extracted by identifying the RoI (Region of Interest) 540.

Dated this 08th Dec, 2021.