Description:Form 2
The Patent Act 1970
(39 of 1970)
&
The Patent Rules 2003

Complete Specification
(See section 10 and rule 13)

Title of the Invention:

Method and Machine Learning Based Inspection System
for
Continuous Manufacturing Process of non-fluidic Products

Applicant: ROBRO SYSTEMS PRIVATE LIMITED
Nationality: Indian
Address: 404 Commerce House
7 Race Course Road
Indore - 452003
Madhya Pradesh, India

The following specification particularly describes the invention and the manner in which it is to be perform
FIELD OF THE INVENTION
The present invention relates to an inspection system. Particularly, the invention relates to an inspection system for a continuous process. More particularly, the inspection system is machine learning based inspection system.

BACKGROUND OF THE INVENTION
Any manufacturing process is prone to defects. Defects are repetitive as well as newer in nature. Most current machines operate “mechanically”, performing repetitive actions without situational analyses of the surroundings, materials and defect locations.
In the current global competitive scenario, where defects acceptability is close to NIL or in ppm or in sigma levels, deployment of intelligent system is indispensable, particularly for high speed and high value precision manufacturing processes. Products ranging from discrete parts like consumer goods to continuous products like fabric, paper, metal sheets to fluids to bottling require intelligent defect detection means.
CN106824778 discloses an automatic machine for quality checking of incense cores. Such automatic screening is relatively easier when the quality criteria is illustratively, say, just the diameter. CN102818809 discloses a gray cloth defect on-line detecting system comprising a cross beam device, a lighting device, industrial cameras, an industrial control host machine, a display control screen and an alarm indicating module. The display control screen is connected with the industrial control host machine, diameters of the system are set through the display control screen in a setting state, and defect shapes and regions with defects can be observed in an operating state. Such system is hugely human skill and efficiency based!
The present invention recognizes that types of defects are groupable and machine learnable in order to be able to invent an intelligent and configurable system deployable for a variety of products.
SUMMARY OF INVENTION
The present invention is a configurable anomaly separation system comprising a configurable anomaly detection device around which a plurality of input sensing devices and action devices are deployed, which are diverse and specific to a prescribed manufacturing line. Action devices are different than a plurality of automation devices like motors, counters, encoders, which are integral part of any continuous manufacturing line besides the present invention. Such system incorporating the configurable anomaly detection device may be developed with a similar or dissimilar input sensing devices for another prescribed manufacturing line.
The configurable detection device deploys an inventive dual processing wherein a first processing is an approving processing or an approving neural network and a second processing is a disapproving processing or a disapproving neural network. Correspondingly, an approving processor is inputted with an acceptable product training data or an approving data while a disapproving processor is inputted with an unacceptable training data or a disapproving data. The anomaly detection device is provided with a plurality of receiving ports for the input sensors and the action devices, a dashboard and analytics.
The input sensors are inventively deployed to capture defects beyond human eye. While a contemporary RGB camera can “see” surface anomaly, a pair of RGB camera mounted mutually inclined at an intra-eye distance can “see’ a defect at a depth. Alternatively, two RGB cameras mounted opposite to one another can “see” surface defects on both sides. RGB-IR cameras and or IR cameras used in an eye-like pair can “see” an undesired inner cavity or a blow-hole. Apropos, the plurality of input sensors are deployed as a depth vision sensor assemblies having a combination of RGB cameras, RGB IR camera, and or IR camera, disposed on the same side at a mutually prescribed angle and or a mutually prescribed distance.
Important to emphasize that different placements of these sensors e.g. cameras on top and bottom or multiple-cameras etc., is also critical in creating full coverage of areas to be inspected / data to be collected.
Input sensors are inventively deployed to inspect surface cleanliness and surface readiness. Contemporary, non-invasive input sensors are integratable to sense hardness, moisture content, and other material quality details. The configurable detection device is not constrained by input sensor as long as the sensor can provide output as per corresponding industrial standard.
The input sensors feed-in prescribed manufacturing information of a running manufacturing line which is continuously compared with the acceptable/approving product data and the unacceptable/disapproving product data. Correspondingly, the action device of the anomaly separation system separates anomalous parts as per prescribed parameters, corresponding allowable values and threshold values.
The configurable detection device specific to a manufacturing site is prepared for acceptable product data and the unacceptable product data. A data acquisition is carried out either from the manufacturing site via the input sensors or from previous manufacturing data or from previous sites. The data is enriched and augmented. Illustratively, image data is non-exhaustedly moderated for contrast, reverse video, IR contents to be able to draw deeper interpretations. Thus a dataset is constructed specific to a manufacturing line. Next, an acceptability criterion/parameter setting is carried out based on expected quality requirements. Illustratively, for a cloth stitching line acceptability criterion for a thick and robust fabric like denim could be different than a silk fabric!
A trial run or a model training of the configurable detection device is then conducted to verify and validate the anomaly detection based on parameter setting. Based on loss convergence results.
Loss convergence refers to how closely the neural network’s predicted results are to the data in the dataset. A low-loss convergence means that the neural network was successful in “learning” the traits that help in identification. For a simple 2-D example, Given points (Xd, Yd) in the dataset. A predicted point (Xp, Yp)’s distance to the original point will be the loss.
The configurable system is then run for prediction of defects and consequent action by action sensors.
The process iteratively continues till a satisfactory accuracy is achieved or a stabilized production is attained.
The invention is now described as applied for a fabric cutting machine, which is a contiguous non-fluidic product line. The fabric cutting machine deploys the configurable system as per the present invention. The fabric cutting machine cuts a continuous fabric into contiguous predefined lengths after intricate inspection of weaving quality.
Approving data in the form of an image of an acceptable quality is fed into the configurable device. The image is then manipulated in terms of light, contrast, brightness, focus and other known criterion of image enhancement; and a plurality of images of correspondingly relatable acceptable quality are generated. Approving data is a benchmark data as a gross acceptability criteria based on previously acquired expertise and domain knowledge. Form the approving data, the computer program generates probable defects with or without human intervention.
Important to note that not all probable defects are generated by the computer program. We do take in marked “defect data” and use that defect data to create more variety of defects.
Approving data is generally fed only once and does NOT need to grow with in-situ learning and or day-to-day experience, unless the type of fabric changes. There will be different approving data for different types of fabric that will pass through this system.
Disapproving data in the form of a plurality of images containing a camera-visible defects of unacceptable quality are fed into the configurable device. Illustratively, each image of a different defect is then manipulated in terms of light, contrast, brightness, focus and other known criterion of image enhancement; and a plurality of additional images of correspondingly relatable unacceptable quality are derived and generated.
“Dual Data Augmentation” - creating data of disapproving quality from Data of disapproving quality, besides approving quality. Given a small quantity of data of approving or disapproving quality - we can modify and enhance that to create more data of disapproving quality. Having more data is related to better accuracy of the system since it can achieve better loss convergence. Such data extraction is via convolutional neural networks (CNN) or such equivalent neural networks.
For each disapproving data/defect, an allowable value and a threshold value is defined. Illustratively, for a fabric, a hole as a Parameter ONE may have an allowable value of 0.4mm as acceptable, but a threshold valve for rejection may be 0.7mm. A loose thread as a parameter TWO may have an allowable value of 1mm as acceptable, but a threshold value for rejection may be 3mm, a shuttle fault as a parameter THREE may have an allowable value of 0.5mm, but a threshold value for rejection may be 0.51mm.
Importantly, allowable and threshold values are moderatable or changeable depending on overall quality, cost and through-put targeted from a manufacturing line.
Disapproving data is an ever-growing data as a gross non-acceptability criteria. From the disapproving data, the computer program compares actually occurring defects primarily without human intervention. Disapproving data is generally fed repeatedly with evolving experience and expected quality; and such disapproving data grows with learning and or experience. Apropos, the neural network deployed needs to be and therefore equipped with a self-learning algorithm suited to the manufacturing line.
Importantly, the dual approach of the present invention of machine learning based on approving data AS WELL AS disapproving data implies following distinctiveness –
Approving data approach – Lesser training data executed with intense algorithm of broad objective outcome.
disapproving data approach – More and faster growing training data with narrow and unambiguous objective outcome.
Illustration – Approving data for a fabric would be a prescribed opaqueness, besides images of fabric with prescribed appearance. Opaqueness is implanted to ensure variation in fabric thickness due to thread thickness variation and weaving density variation can be captured which physical inspection may or may not be able to detect.
Corresponding disapproving data to check fabric thickness would be a combination images giving thread thickness, number of threads per unit spread, etc. Such algorithms would need lesser processing, more data and more precise camera sensor.
In other words, anomaly detection by such dual approach is complimentary and mutually exclusive.
As per the present invention the prediction is configured to be carried out in any one or all three possible ways:
By approving data, or
By disapproving data, or
By approving data AND disapproving data, or
a combination thereof.
It is easily appreciable that for a new installation of the present invention, when there is NO PRIOR manufacturing data of the new installation nor any previous relatable installations, a system as per present invention can be operational relatively faster based on the benchmark data, which is the approving data.
In case of a conflict between the approving data and the disapproving data, the system is configurable to override one over the other either manually or by a configuration setting.
The action devices in the fabric cutting machine as a preferred embodiment is a fabric cutter and a spout relieving cutter. The spout relieving cutter cuts fabric to accommodate a spout of a fabric container. After the system is trained and commissioned, a prescribed cutting length and a prescribed spout relieving location is pre-fixed. However, in the embodiment of the fabric cutter machines, as soon as the anomaly detection device detects a defect, the fabric cutter is instructed to cut the fabric before and after a zone of defect, however, only after ascertaining that the spout cutting is ON, and consequently verifying spout relieving location with respect to a defect location and any other pre-defined trimming zone. Thus,
- If the defect is within spout relieving location, then the defect is ignored since spout relieving cutter would obviate the defect;
- If the defect is in sides pre-defined for trimming, the defect is ignored since trimming would obviate the defect;
- If the defect is at other than the spout area or anywhere else, then fabric is cut before and after the defect located, as pre-programmed.
In this manner, the defective fabric is prevented from proceeding further undetected and at the same time, a wastage of full prescribed length of fabric is prevented.
Like any inspection system, there remains a possibility that the device makes an incorrect prediction. These are known as a false positive prediction and a false negative prediction. As the term suggests, a false negative prediction implies a defect is let go unchecked, which a false positive prediction implies a defect is predicted while there is none. The system as per present invention provides an observation mode wherein a manual observation is keyed in as a False Positive or a False negative prediction and such manual audit/inspection is fed for progressive machine learning. The anomaly separation system being a machine learning based invention is trained to prompt in the event the anomaly detection device is not HIGH in confidence. This allows for proactive user action and fast learning of the system wherein the system itself is able to convey areas of low-confidence and use humans-in-the-loop with domain knowledge to become more confident.
The anomaly detection device creates records or system logs, and production data for display and analytics and deep learning.
The invention is likewise applied for an incense inspection machine, which is a discrete non-fluidic product line. Another non-exhaustive application of the present invention for a continuous non-fluidic product is for an injection molded bottle cap. The present invention is limitlessly deployable for automated inspection for a continuous manufacturing processes of non-fluidic products and the process is continuously upgradable by machine learning with or without manual invention.
KWIS is a tradename of the inventive anomaly detection device, and KWIS is interchangeably used for device and system .

Economic Significance:

Waste reduction and time reduction is an important yardstick of economic significance. The system as per present invention facilitates faster detection and control of the cutting process when a defect is present to take action, leading to a more efficient process. The user could cut more sheets from the same roll as a result of using this system as evidenced by below analysis table, generated by the present system. On an average, the fabric saved with the installation of KWIS system for the period was 55%. The present invention thus has unprecedented scope to save on material waste, time, resources and money – directly as well as indirectly.

OBJECTIVES
To invent an inspection system that intelligently identifies defects with minimum human intervention.
To invent a system with the ability to perform cognitive tasks like learning and decision making without explicit instructions or computer program.
To invent a smart control logic that utilizes information from the above two and takes logical decisions.
To invent a system that is scalable to be deployed to a plurality of continuous manufacturing processes.
To invent a system that is scalable to a plurality of products.
To invent a system that facilitates waste reduction due to early and timely detection of defects.
To invent a system that can process multiple manufacturing alternatives based on defect locations.
To invent a system that manages defect based manufacturing processes.
To invent a system that supports decision making for routing manufacturing processes.
BRIEF DESCRIPTION OF DRAWINGS
Figure 1 is a representative diagram of the concept of the present invention.
Figure 2 is a block diagram of an anomaly detection device.
Figure 3 is a representation of incense as one of the non-fluidic product.
Figure 4 is a side representation of two cameras at a mutual angle.
Figure 5 is a perspective view of two cameras at a prescribed distance from each other.
Figure 6A-6B is a flow diagram of a method as per present invention.
Figure 7 is a flow diagram giving options of prediction as per the present invention.
Figure 8 is a perspective view of a fabric cutting machine with inventive anomaly detection device.
Figure 9A-9B is a flow diagram of the inventive system for a fabric cutting machine.
Figure 10 is an approving data of an image of acceptable quality of a fabric.
Figure 11 is an illustration of data augmentation of approving data.
Figure 12 is a disapproving data of images of unacceptable quality of the fabric.
Figure 13 is a magnified view of a disapproving data of the image of unacceptable quality of the fabric.
Figure 14 is the corresponding illustration of data augmentation of disapproving data.
Figure 14A is a process diagram based on approving as well as disapproving data for neural network.
Figure 15 is a pictorial diagram of allowable and threshold limits of defects.
Figure 16 is a perspective view of the fabric cutting machine with a spout cutter.
Figure 17 is a perspective view of a fabric container with a spout.
Figure 18 is a logic diagram illustrating intra-dependence of action devices.
Figure 19 shows various logs of the computer program of the present system.
Figure 20 shows actionables of the action devices of the fabric cutting machine.
Figure 21 is a screen of the computer program generating analyses.
Figure 22 is a flow diagram related to false decisions of the machine.
Figure 23 is an incense inspection machine using the present inventive system.
Figure 24A-24B is a flow diagram for incense inspection.
Figure 25 is an inspection set up for an injection molded part, illustrating an acceptable and an unacceptable part.
Figure 26 is a perspective view of an economical version of anomaly detection device as per present invention.

DETAILED DESCRIPTION OF THE INVENTION
The present invention shall now be described with accompanying drawings. The present invention is a configurable anomaly separation system suited to unlimited manufacturing processes and products and the description given here is one of the several embodiments. The description should therefore not be construed to limit the invention in any way whatsoever.
Figure 1, in the preferred embodiment, the present invention is the configurable anomaly separation system (100) comprising a configurable anomaly detection device (200) around which a plurality of
- input sensing devices (60) and
- action devices (90) are deployed, which are diverse and specific to a prescribed manufacturing line. Action devices (90) are different than a plurality of automation devices like motors, counters, encoders, which are integral part of any continuous manufacturing line besides the present invention.
Such system incorporating the configurable anomaly detection device (200) may be developed with a similar or dissimilar input sensing devices for another prescribed manufacturing line.
Every product has a finite number of possibilities of an acceptable product or defect-free product, while there exist relatively much larger number of possibilities of defects! Illustratively, Figure 3, for an incense (42) an acceptable product would be a near straight incense of correct cross-section with an incense coating of a correct thickness and concentricity, producing desired fragrance. On the other hand, the defects would be a camber and curve of the core and therefore or otherwise of coating, irregular incense surface, of varying adhesion, with or without localized missing or chipped coating, uncured incense coating, inappropriate mixture of ingredients, weak or broken core, inappropriate burning, a blow-hole (44), etc. Acceptability is a matter of expertise and domain while unacceptability is a manufacturing limitation!
Objectively, acceptability criteria would be acceptable and defined tolerances. Non-acceptability criteria would, on the other hand be process dependent and would therefore evolve with manufacturing process.
The present invention stems from this manufacturing acumen crisply deployed for defining and training the acceptability criterion from corresponding data to a neural network and progressively developing unacceptability criteria based on boundary conditions.
Figure 2, the configurable detection device (200) accordingly deploys an inventive dual processing with a first neural network and a second neural network, wherein a first processing is an approving processing or an approving neural network (211) and a second processing is a disapproving processing or a disapproving neural network (212). Correspondingly, an approving processor is inputted with an acceptable product training data or an approving data (201) while a disapproving processor is inputted with an unacceptable training data or a disapproving data (202). The anomaly detection device (200) is provided with a partitioned processor for the dual processing, plurality of receiving ports (206) for the input sensors (60) and the action devices (90), a dashboard and analytics (208).
The input sensors (60) are inventively deployed to capture defects beyond human eye. While a contemporary RGB camera can “see” surface anomaly, a pair of RGB camera mounted mutually inclined at an intra-eye distance can “see’ a defect at a depth. Alternatively, two RGB cameras mounted opposite to one another can “see” surface defects on both sides. RGB-IR cameras and or IR cameras used in an eye-like pair can “see” an undesired inner cavity or a blow-hole. Apropos, Figure 4,5, the plurality of input sensors (60) are deployed as a depth vision sensor assemblies having a combination of RGB cameras (53), RGB IR camera, and or IR camera, disposed on the same side at a mutually prescribed angle (51) and or a mutually prescribed distance (52).
Important to emphasize that different placements of these sensors (e.g. cameras on top and bottom or multiple-cameras etc., is also critical in creating full coverage of areas to be inspected / data to be collected.
Input sensors (60) are inventively deployed to inspect surface cleanliness and surface readiness. Contemporary, non-invasive input sensors are integratable to sense hardness, moisture content, and other material quality details. The configurable detection device (200) is not constrained by input sensor (60) as long as the sensor can provide output as per corresponding industrial standard. The configurable device (200) has a microcontroller of minimum specification: Intel i3 1.9GHz processor with 2GB RAM/equivalent - No Graphics Card.
The input sensors (60) feed-in prescribed manufacturing information of a running manufacturing line which is continuously compared with the acceptable/approving product data (201) and the unacceptable/disapproving product data (202). Correspondingly, the action device (90) of the anomaly separation system (100) separates anomalous parts as per prescribed parameters, corresponding allowable values and threshold values.
Figure 6A-6B, the configurable detection device (200) specific to a manufacturing site is prepared for acceptable product data and the unacceptable product data. A data acquisition (101) is carried out either from the manufacturing site via the input sensors (60) or from previous manufacturing data or from previous sites. The data is enriched and augmented (102). Illustratively, image data is non-exhaustedly moderated for contrast, reverse video, IR contents to be able to draw deeper interpretations. Thus a dataset is constructed (103) specific to a manufacturing line. Next, an acceptability criterion/parameter setting (104) is carried out based on expected quality requirements. Illustratively, for a cloth stitching line acceptability criterion for a thick and robust fabric like denim could be different than a silk fabric!
A trial run or a model training (105) of the configurable detection device is then conducted to verify and validate the anomaly detection based on parameter setting. Based on loss convergence results (106).
Loss convergence refers to how closely the neural network’s predicted results are to the data in the dataset. A low-loss convergence means that the neural network was successful in “learning” the traits that help in identification. For a simple 2-D example, Given points (Xd, Yd) in the dataset. A predicted point (Xp, Yp)’s distance to the original point will be the loss.
The configurable system is then run (107) for prediction (109) of defects and consequent action by action sensors (111).
NMS refers to “non-maximal suppression" which translates to combining multiple nearby defect locations together into a single and actionable defect. Say a defect has a slight hole and 2-3 stretched threads around it. Each of these will get identified as separate anomalies by the neural network - Combining those into a single result of “defect region” is carried out using the confidence values of predicted positions and their overlap.
The process iteratively continues (112) till a satisfactory accuracy is achieved or a stabilized production is attained.
Figure 8, The invention is now described as applied for a fabric cutting machine (50), which is a contiguous non-fluidic product line. Figure 9A-9B. The fabric cutting machine (50) deploys the configurable system (100) as per the present invention. The fabric cutting machine (50) cuts a continuous fabric (58) into contiguous predefined lengths after intricate inspection of weaving quality.
Approving data (201) in the form of an image (151) of an acceptable quality is fed into the configurable device (200). The image (151) is then manipulated in terms of light, contrast, brightness, focus and other known criterion of image enhancement; and a plurality of images (152) of correspondingly relatable acceptable quality are generated. Figures 10, 11. Approving data (201) is a benchmark data as a gross acceptability criteria based on previously acquired expertise and domain knowledge. Form the approving data (201), the computer program generates probable defects with or without human intervention.
Important to note that not all probable defects are generated by the computer program. We do take in marked “defect data” and use that defect data to create more variety of defects.
Approving data (201) is generally fed only once and does NOT need to grow with in-situ learning and or day-to-day experience, unless the type of fabric changes. There will be different approving data for different types of fabric that will pass through this system.
This inventive step of generating probable defects from an acceptable or approving data (201) or information can be further understood with following near similar analogy:
Approving data (201) for an English spelling defect check for say a word “compulsory” is just this word as spelled within inverted commas. All following incorrectly spelled words are implicitly generatable by the spell-checker running a random generation computer program with boundary conditions:
Compalsary, Comualsery, Compalsry, Compulsery, kompulsary
In other words, when the word file to be spell-checked contains any of the above mis-spelled words, the spell-check computer program compares the mis-spelt word with the approving data and NOT any previously stored or trained mis-spelt word!
Figures 12, 13, 14, disapproving data (202) in the form of a plurality of images (155A, 155B, 155C, 155D, 155E, 155F, 155G, 155H, 155J) containing a camera-visible defects of unacceptable quality are fed into the configurable device (200). Illustratively, each image (155) of a different defect is then manipulated in terms of light, contrast, brightness, focus and other known criterion of image enhancement; and a plurality of additional images (156) of correspondingly relatable unacceptable quality are derived and generated.
Figure 14A, termed as “Dual Data Augmentation” (157) - creating data of disapproving quality from Data of disapproving quality, besides approving quality. Given a small quantity of data of approving or disapproving quality - we can modify and enhance that to create more data of disapproving quality. Having more data is related to better accuracy of the system since it can achieve better loss convergence. Such data extraction is via convolutional neural networks (CNN) or such equivalent neural networks.
Figure 15, for each disapproving data/defect, an allowable value and a threshold value is defined. Illustratively, for a fabric, a hole as a Parameter ONE (161) may have an allowable value of 0.4mm as acceptable, but a threshold valve for rejection may be 0.7mm. A loose thread as a parameter TWO (162) may have an allowable value of 1mm as acceptable, but a threshold value for rejection may be 3mm, a shuttle fault as a parameter THREE (163) may have an allowable value of 0.5mm, but a threshold valve for rejection may be 0.51mm.
Importantly, allowable and threshold values are moderatable or changeable depending on overall quality, cost and through-put targeted from a manufacturing line.
Disapproving data is an ever-growing data as a gross non-acceptability criteria. From the disapproving data, the computer program compares actually occurring defects primarily without human intervention. Disapproving data is generally fed repeatedly with evolving experience and expected quality; and such disapproving data grows with learning and or experience. Apropos, the neural network deployed needs to be and therefore equipped with a self-learning algorithm suited to the manufacturing line.
Importantly, the dual approach of the present invention of machine learning based on approving data AS WELL AS disapproving data implies following distinctiveness –
Approving data approach – Lesser training data executed with intense algorithm of broad objective outcome.
Disapproving data approach – More and faster growing training data with narrow and unambiguous objective outcome.
Illustration – Approving data for a fabric would be a prescribed opaqueness, besides images of fabric with prescribed appearance. Opaqueness is implanted to ensure variation in fabric thickness due to thread thickness variation and weaving density variation can be captured which physical inspection may or may not be able to detect.
Corresponding disapproving data to check fabric thickness would be a combination images giving thread thickness, number of threads per unit spread, etc. Such algorithms would need lesser processing, more data and more precise camera sensor.
In other words, anomaly detection by such dual approach is complimentary and mutually exclusive.
Figure 7, as per the present invention the prediction (109) is configured to be carried out in any one or all three possible ways:
By approving data (201), or
By disapproving data (202), or
By approving data (201) AND disapproving data (202), or
a combination thereof.
It is easily appreciable that for a new installation of the present invention, when there is NO PRIOR manufacturing data of the new installation nor any previous relatable installations, a system as per present invention can be operational relatively faster based on the benchmark data, which is the approving data.
In case of a conflict between the approving data and the disapproving data, the system is configurable to override one over the other either manually or by a configuration setting.
The action devices (90) in the fabric cutting machine (50) as a preferred embodiment is a fabric cutter (54) and a spout relieving cutter (55). The spout relieving cutter (55) cuts fabric (58) to accommodate a spout (59A) of a fabric container (59), Figures 16, 17. After the system is trained and commissioned, a prescribed cutting length and a prescribed spout relieving location is pre-fixed. However, in the embodiment of the fabric cutter machines (50), as soon as the anomaly detection device (200) detects a defect, the fabric cutter is instructed to cut the fabric before (181) and after (182) a zone (185) of defect, however, only after ascertaining that the spout cutting is ON (120), and consequently verifying spout relieving location with respect to a defect location and any other pre-defined trimming zone. Thus, Figures 18, 20,
- If the defect is within spout relieving location (121), then the defect is ignored since spout relieving cutter would obviate the defect;
- If the defect is in sides pre-defined for trimming (124), the defect is ignored since trimming would obviate the defect;
- If the defect is at other than the spout area (122) or anywhere else (123), then fabric (58) is cut before (181) and after (182) the defect located, as pre-programmed.
In this manner, the defective fabric is prevented from proceeding further undetected and at the same time, a wastage of full prescribed length of fabric is prevented.
In other words, the present invention is NOT merely a defect detector and remover. The system (100) recognizes that every defect has a different location and thus needs to be assessed with respect to all concerned action devices (90), which ought to work in co-ordination with one another or intra-dependently, lest the system may end up generating excessive or unreal “waste”. The system (100) as per the present invention creates different cutting plans, causes alternate cut length, block the spout cutter and create intelligent cutting plans when the system (100) is installed on a plurality of machines.
In other words, the action devices (90) are configurable to act independently or mutually dependently, which means intra-dependently.
It is to be noted that automation devices like motor (56) and encoder (57) which are essential part of such fabric cutting machine (50) are instructed by the computer program to operate in synchronism with and to support inventive operation of the action devices (90) and are not detailed in order to maintain clarity of the essence of the invention.
Figure 22, it is to be noted that like any inspection system, there remains a possibility that the device makes an incorrect prediction. These are known as a false positive prediction (167) and a false negative prediction (166). As the term suggests, a false negative prediction (166) implies a defect is let go unchecked, which a false positive prediction (167) implies a defect is predicted while there is none. The system as per present invention provides an observation mode wherein a manual observation is keyed in as a False Positive or a False negative prediction and such manual audit/inspection is fed for progressive machine learning. The anomaly separation system (100) being a machine learning based invention is trained to prompt in the event the anomaly detection device (200) is not HIGH in confidence. This allows for proactive user action and fast learning of the system wherein the system itself is able to convey areas of low-confidence and use humans-in-the-loop with domain knowledge to become more confident.
Figure 19, 21, the anomaly detection device (200) creates records or system logs (158), and production data (159) for display and analytics and deep learning.
A preferred embodiment of core cognitive algorithms is as follows, not to be construed to limit the present invention in any manner:
1 // class to handle all image grabbing from camera
2 class CameraImageGrabber
3 {
4 // functions
5 void initialize();
6 void startGrabbingImages(); // starts a thread to send images
7 };
8
9 // class that handles the training of the neural network
10 class NeuralNetwork
11 {
12 // function definitons
13 void updateNNWithNewInput(new_image, manual_marking)
14 {
15 // data augmentation
16
vector new_images_to_train = createDataAugmentationImages(new_ima
ge, new_boxes);
17 reTrainNN(new_images_to_train);
18 }
19 // this identifies all the disapproving data (defects)
20 vector identifyObjectsUsingNN(current_image); // uses NN to predict and returns objects
21 };
22 // class to handle all defect identification scenarios
23 class DefectIdentifier
24 {
25 NeuralNetwork nn;
26 bool defect_present;
27 void processAndPublishImage(cv::Mat current_image) // gets called everytime a new image is available.
28 {
29 current_image = preProcess(current_image);
30 current_objects = nn.identifyObjectsUsingNN(current_image);
31 // varying levels of strictness of inspection are stored in recipe
32
shortlisted_and_merged_objects = useRecipeToShortlistAndMerge(current_obj
ects, current_recipe, defect_present);
33 }
34
vector useRecipeToShortlistAndMerge(current_objects, current_recipe, defect_ present)
35 {
36 for(box in current_objects)
37 {
38 if(checkSizeAndNumberOfCount() &&
39 checkTypeAndLocation())
40 {
41 update(current_objects);
42 defect_present = true;
43 }
44 }
45 return current_objects;
46 }
47 };
48

49
50
51
52
53

// class to do traditional matching operation with average image to identify anomalous areas
class AnomalyDetector
{
bool is_training_mode_on;
cv::Mat master_average_image; // This contains all the "approving" data (OK average image)
54 cv::Mat updateAverageMasterImage(current_image);
55 vector matchWithMasterImage(current_image);
56 };

1 int main()
2 {
3 CameraImageGrabber imageGrabber;
4 DefectIdentifier defectIdentifier;
5 AnomalyDetector anomalyDetector;
6 imageGrabber.initialize();
7 imageGrabber.startGrabbingImages();
8 if(first_fifteen_meters)
9 {
10 anomalyDetector.is_training_mode_on = true;
11
anomalyDetector.updateAverageMasterImage(defectIdentifier.c urrent_image);
12 }
13
anomaly_boxes = anomalyDetector.matchWithMasterImage(current_im
age);
14
defect_boxes = defectIdentifier.identifyObjectsUsingNN(current_
image);
15
16 if(diff(defect_boxes, anomaly_boxes))
17 {
18 updateNNWithNewInput(current_image, diff_boxes);
19 if(manual_input_required)
20 {
21 getManualMarkingsfromUser();
22 }
23 }
24 if(defectIdentifier.defect_found)
25 {
26 updateCuttingPlan(current_defect);
27 issueStoppingCommandToMachine(distance_to_stop);
28 }
29 spin(); // Blocking call so other operations run in parallel
30 }

The invention is likewise applied for an incense inspection machine (40), which is a discrete non-fluidic product line. The incense inspection machine (40) inspects each and every incense as per a trained neural network as per approving data (201) and or as per disapproving data (202). The input sensors are RGB cameras (53), IR and eye-like cameras with requisite lighting. A stick rotator (43) rotates each incense stick by 180 degrees so that the input sensors (60) can capture the compete surface of the incense (42). The action sensors (90) are a pneumatic nozzle (41) which pushes away the defective incenses. Besides, the incense inspection machine (40) has a motor (46) and other automation devices, Figure 23, 24A, 24B.
Another non-exhaustive application of the present invention for a continuous non-fluidic product is for an injection molded bottle cap (71), Figure 25. Injection molded bottle caps (71) of current times have complex geometries including split threads, that make inspection challenging, but an intelligent inspection system can address this by using advanced imaging techniques and machine learning algorithms to detect defects in even the latent construction. The system can be designed to perform cognitive tasks like learning and decision making without explicit instructions or computer programs. This can enable the system to adapt to changes in the manufacturing process. A smart control logic can be created that utilizes information from the inspection system and performs logical decisions to address defects. Here the input sensors are a camera and the output action device is a mechanical means like a blow nozzle connected via a pressured supply for rejecting the defected pieces. The method of approving and disapproving data collection is applicable in a similar manner here too.
The present invention is limitlessly deployable for automated inspection for a continuous manufacturing processes of non-fluidic products and the process is continuously upgradable by machine learning with or without manual invention.
Grossly, the method revolves around the steps of:
a. Collecting an approving details set of the non-fluidic product, for expediting initial installation and commissioning, and subsequently for optimizing quantum of anomalous non-fluidic product. Approving details is generally a crisp defect free detailing of a product like – an unleaky and smooth folding umbrella.
b. Collecting an disapproving details set of the non-fluidic product. Continuing with previous example, a disapproving details would be a long list of possible defects in the umbrella – loose handle, missing spokes, folding button too hard, etc.
c. In-situ comparing a prescribed quantum of the non-fluidic product with the approving details and or the disapproving details
d. Segregating an anomalous part of the non-fluidic product
e. Continuously refining the approving details and the disapproving details
f. Continuously reducing a false approving and a false disapproving outcome by a manual inspection
wherein the approving details comprise cognitive acceptance details derived from previous knowledge, potential improvisation thereon and corresponding data, while the disapproving details comprise actual non-acceptance details and corresponding data.
The acceptance details for fabric, incense, and an injection molded cap, for example could be just “free from visual and functional defect” along with a picture or two of the corresponding defect free product(s). On the other hand, the corresponding disapproving details would be several clear pictures showing holes, blow holes, scratches, tears, distortions, etc. The acceptance details therefore are broad, thoughtful and therefore cognitive – more algorithmic and less data intensive, while the non-acceptance details would be the other way! In other words, the method of automated anomaly detection of a continuous production of a discrete or a contiguous non-fluidic product as is such that the approving details comprise a higher order cognitive acceptance details and a lower order data while the disapproving details comprise a lower order cognitive non-acceptance details and a higher order data.
To execute above inventive method, computer program instructs the configurable anomaly separation system (100) through a user interface residing interactively on the industrial computer and the configurable anomaly detection device (200), configured to carry out the following steps:
a. Acquisitioning (101) approving data (201) for an approving neural network (211),
b. Acquisitioning (101) disapproving data (202) for a disapproving neural network (212),
c. Augmenting and enriching approving data and disapproving data,
d. Training (105) approving neural network (211) and disapproving neural network (212),
e. Setting parameter values (104),
f. Running (107) a continuous production line with prescribed action devices (90) and automation devices,
g. Receiving manufacturing input (108) from input sensors (60), interpreting defects through the neural network (211, 212) with reference to the parameter values,
h. Generating statistics for analysis and deep learning, and or
i. Supervising manually for false approving and false disapproving and further machine learning of neural networks (211, 212);
wherein the approving neural network is inputted with a first cognitive algorithm of acceptance derived from previous knowledge and potential improvisation thereon and corresponding data, while the disapproving neural network is inputted with a second cognitive algorithm of non-acceptance derived from actual non-acceptance instances and corresponding data. The configurable anomaly detection system is such that the first cognitive algorithm is a higher order cognitive algorithm with a lower order data, while the second cognitive algorithm is relatively a lower order cognitive algorithm with a relatively higher order data.
The anomaly separation system (100) as per present invention caters to a plurality of manufacturing lines running different recipes and cross-learns.
While above process steps primarily lists defect based manufacturing process control, the present invention includes manufacturing control for conventional manufacturing process as well.
The configurable anomaly separation system (100) as per present invention is retrofittable on the existing conventional continuous manufacturing lines as the anomaly detection device (200) is configurable with pre-existing action devices (90).
KWIS is a tradename of the inventive anomaly detection device (200), and KWIS is interchangeably used for device (200) and system (100).

Economic Significance:

At the heart of the present invention is the anomaly detection device (200). Being an industrial product, the device needs to be robust and conducive or immune to varying industrial conditions and abuses. The device (200) comprises a compact enclosure of ingress protection IP52 and higher. It is well understood that environment conditions of different places and countries is different and so are economic requirements, which the present invention specifically addresses, with a particular applicability to India. See Figure 26 which is a stripped down economical version of the product as per present invention with uncompromised suitability for Indian environment and dust levels.
The device (200), accordingly, is a specific electronic hardware of a cost effective bill of material for target electrical and electronic parameters as tabulated here below:

Waste reduction and time reduction is another yardstick of economic significance. The system as per present invention facilitates faster detection and control of the cutting process when a defect is present to take action, leading to a more efficient process. The user could cut more sheets from the same roll as a result of using this system as evidenced by below analysis table, generated by the present system.

On an average, the fabric saved with the installation of KWIS system for the period was 55%.

Illustration:
A flexible intermediate bulk container (FIBC), jumbo, bulk bag, super sack, big bag, or tonne bag is an industrial container made of flexible fabric that is designed for storing and transporting dry, flowable products, such as sand, fertilizer, and granules of plastic.
As per the Textile Association, FIBC production in India is recorded as 306,996 MT in 2021. The food-grade FIBC production was nearly 28% of the total production of FIBC in India. The total export sales of FIBC from India increased 3 times over the past decade and reached US$ 708.48 Million from 2020 to 2021.
Considering FIBC industry standard of 5 % waste generation of total production. 15349 MT was the waste produced in India in the year 2020 -2021 by the Industry.
The present invention has demonstrated its capability to reduce such waste to HALF, thereby reduces carbon print as well, yet another yardstick of recognized economic significance.

The present invention thus has unprecedented scope to save on material waste, time, resources and money – directly as well as indirectly.
, C , Claims:WE CLAIM:
1. A method of automated anomaly detection of a continuous production of a discrete or a contiguous non-fluidic product, the method comprising the steps of:
a. Collecting an approving details set of the non-fluidic product,
b. Collecting a disapproving details set of the non-fluidic product,
c. In-situ comparing a prescribed quantum of the non-fluidic product with the approving details and or the disapproving details,
d. Segregating an anomalous part of the non-fluidic product,
e. Continuously refining the approving details and the disapproving details,
f. Continuously reducing a false approving and a false disapproving outcome by utilizing manual intervention, and
the approving details set expediting an initial production of an anomaly-free initial produce; the method reducing a waste measure of anomalous non-fluidic product of a stabilized production line to a half of industrial acceptance.
2. The method of automated anomaly detection of a continuous production of a discrete or a contiguous non-fluidic product as claimed in claim 1, wherein the approving details comprise cognitive acceptance details derived from previous knowledge, potential improvisation thereon and corresponding data, while the disapproving details comprise actual non-acceptance details and corresponding data.
3. The method of automated anomaly detection of a continuous production of a discrete or a contiguous non-fluidic product as claimed in claim 2, wherein the approving details comprise a higher order cognitive acceptance details and a lower order data while the disapproving details comprise a lower order cognitive non-acceptance details and a higher order data.

4. A configurable anomaly separation system (100) comprising an industrial computer; an anomaly detection device (200) having a programmable logic controller, a power supply (204), a cloud module (209); a plurality of input sensors (60); a plurality of action devices (90); characterized in that:
• the anomaly detection device (200) has a partitioned processor for the dual processing,
• the anomaly detection device (200) has a dual neural network, a first neural network is an approving neural network (211) and a second neural network is a disapproving neural network (212),
• the anomaly detection device (200) correspondingly has the approving neural network (211) inputted with an acceptable product training data, the disapproving neural network (212) inputted with an unacceptable training data,
• the anomaly detection device (200) having a plurality of receiving ports (206) for the plurality of input sensors (60) and the plurality of action devices (90), a dashboard and analytics (208),
• the plurality of action devices (90) intra-dependently reducing quantum of anomalous non-fluidic product,
• the configurable anomaly separation system (100) having a plurality of input sensors (60) detecting anomaly as per a neural training,
• the configurable anomaly separation system (100) having manual supervision to detect a false positive and a false negative anomaly detection generating a new training data, and
• the configurable anomaly separation system (100) iteratively trained to include the new training data in the dual neural network;
the configurable anomaly separation system (100) reducing waste generation to a HALF of industry acceptance.
5. The configurable anomaly detection system (100) as claimed in claim 4, wherein the first neural network (211) is inputted with a first cognitive algorithm of acceptance derived from previous knowledge and potential improvisation thereon and corresponding data, while the second neural network (212) is inputted with a second cognitive algorithm of non-acceptance derived from actual non-acceptance instances and corresponding data.
6. The configurable anomaly detection system (100) as claimed in claim 5, wherein the first cognitive algorithm is a higher order cognitive algorithm with a lower order data, while the second cognitive algorithm is a lower order cognitive algorithm with a higher order data.
7. The configurable anomaly detection system (100) as claimed in claim 4, wherein the plurality of input sensors (60) devices are two RGB cameras (53) disposed on opposite sides of the non-fluidic product.
8. The configurable anomaly detection system (100) as claimed in claim 4, wherein the plurality of input sensors (60) are depth vision sensor assemblies having a combination of RGB cameras (53), RGB IR camera, and or IR camera, disposed on the same side at a mutually prescribed angle (51).
9. The configurable anomaly detection system (100) as claimed in claim 4, wherein the plurality of input sensors (60) are depth vision sensor assemblies having a combination of RGB cameras (53), RGB IR camera, and or IR camera, disposed on the same side at a mutually prescribed distance (52).
10. The configurable anomaly detection system (100) as claimed in claim 4, wherein the plurality of input sensors (60) are surface readiness and surface cleanliness sensors.
11. The configurable anomaly detection system (100) as claimed in claim 4, wherein the plurality of input sensors (60) are position sensors, laser profilers, distance measurement devices, and or torque measurement devices.

12. The configurable anomaly separation system (100) as claimed in claim 4, wherein the configurable anomaly separation system (200) is deployed on a fabric cutting machine (50), and wherein the action devices (90) is a fabric cutter (54) and a spout relieving cutter (55) intra-dependently reducing quantum of a defective fabric for a continuous production line of a contiguous non-fluidic product.
13. The configurable anomaly separation system (100) as claimed in claim 4, wherein the configurable anomaly system (200) is deployed on an incense inspection machine (40) wherein the action device (90) is a pneumatic nozzle (41) removing a defective incense for a continuous production line of the discrete non-fluidic product.
14. The configurable anomaly separation system (100) as claimed in claim 13, wherein the configurable anomaly separation system (100) has an incense stick rotator (43) to enable input sensors (60) capture a complete surface of the incense (42).
15. The configurable anomaly separation system (100) as claimed in claim 4, wherein the configurable anomaly system (200) is deployed on an injection molding inspection machine (70) wherein the action device (90) is a mechanical means for removing a defective molded part for a continuous production line of the discrete non-fluidic product.
16. A computer program to instruct a configurable anomaly separation system (100) through a user interface residing interactively on an industrial computer and a configurable anomaly detection device (200), configured to carry out the following steps, as configured:
a. Acquisitioning (101) approving data (201) for an approving neural network (211),
b. Acquisitioning (101) disapproving data (202) for a disapproving neural network (212),
c. Augmenting and enriching approving data and disapproving data,
d. Training (105) approving neural network (211) and disapproving neural network (212),
e. Setting parameter values (104),
f. Running (107) a continuous production line with prescribed action devices (90) and automation devices,
g. Receiving manufacturing input (108) from input sensors (60),
h. Predicting defects through the neural network (211, 212) with reference to the parameter values,
i. Generating statistics for analysis and deep learning, and or
j. Supervising manually for false positive prediction (166) and false negative prediction (167) and further machine learning of neural networks (211, 212).
17. The computer program as claimed in claim 16, wherein the approving neural network is inputted with a first cognitive algorithm of acceptance derived from previous knowledge and potential improvisation thereon and corresponding data, while the disapproving neural network is inputted with a second cognitive algorithm of non-acceptance derived from actual non-acceptance instances and corresponding data.
18. The computer program as claimed in claim 16, wherein the predicting is by one of the approving data (201), the disapproving data (202), or the approving data (201) and the disapproving data (202).