Description:FIELD OF THE INVENTION
The present invention discloses a system and a method for multiple fault detection in hydraulic power systems based on a single sensor. More particularly a system and method to identify potential faults by using trained convolution neural network model. Moreover, the system and method offer a cost-effective and user-friendly solution for fault detection compared to existing methods.

BACKGROUND OF THE INVENTION
Fluid Power Systems (FPS) comprises hydraulic and pneumatic systems that utilize fluid pressure to perform work. Hydraulic systems operate under high pressures and crucial system in many industries, but wear and tear can lead to decreased efficiency and safety risks.

Hydraulic systems consist of components such as pumps, directional control valves, accumulators, and actuators, all of them are susceptible to wear and potential failure due to the continuous high-pressure operation.
To prevent system failures, monitoring and fault detection are essential to avoid issues that can lead to safety hazards and operational disruptions.

Existing methods for diagnosing faults often require multiple sensors, which is expensive and complex. Traditional monitoring systems often occur periodically rather than in real-time, emphasizing the need for continuous, intelligent diagnostic methods.

There are various patent and non-patent literature in this field of technology. Reference of one such patent application no. CN202211134708-A titled as “Fault diagnosis method for hydraulic power split continuously variable transmission based on BiCNN _ LTSM model”. This prior art discloses a fault diagnosis method for a hydraulic power split stepless gearbox. It uses a BiCNN-LSTM model to analyze oil pressure signals collected from a pressure sensor. This art involves calculating the similarity between the collected signals and a standard sample, segmenting the data into samples, labeling the samples based on fault states, dividing the data into training and testing sets, training the model with the training set, and using the trained model for fault detection. This art claimed to accurately and automatically identify the state of the hydraulic system in real time, ensuring the normal operation of the gearbox.

Another reference is made to patent application no. CN202010840213-A titled as “Electro-hydraulic servo valve fault diagnosis method and system based on transfer learning”. This prior art discloses a method and system for diagnosing faults in electro-hydraulic servo valves using transfer learning. This art involves acquiring a fault database, preprocessing the data to create a training set, establishing a fault diagnosis model using the training set and a Markov's metric transfer learning framework, and processing sample data to perform fault discrimination. This art claims that this method can effectively utilize auxiliary data to diagnose electro-hydraulic servo valve faults even with limited training data, reducing costs and achieving efficient and accurate diagnosis.

Another reference is made to patent application no. CN202310487959-A titled as “Hydraulic system fault diagnosis method and system based on diagnosis model”. This art discloses a hydraulic system fault diagnosis method and system based on a diagnosis model. This art involves acquiring actual parameters of a hydraulic system, constructing a virtual interactive environment, constructing a diagnosis model using transfer learning, collecting operation data in real time, performing fault diagnosis on the real-time state of the hydraulic system, synchronizing the diagnosis result to the virtual environment for fault location and visual feedback, and searching for a repair scheme in a knowledge base. This art claims that this method improves data quality, algorithm efficiency, working efficiency, and decision-making level of the hydraulic system and fault diagnosis system by using a visual virtual environment, real-time monitoring, and two-wheel data cleaning.

Another reference is made to non-patented document by Haohan Tao, Peng Jia, Xiangyu Wang, Liquan Wang titled as “Real-Time Fault Diagnosis for Hydraulic System Based on Multi-Sensor Convolutional Neural Network”.

This paper proposes a real-time fault diagnostic method for hydraulic systems using a multi-sensor convolutional neural network (MS-CNN). The MS-CNN incorporates feature extraction, sensor selection, and fault diagnosis into a single model. The sensor selection process is based on abstract fault-related features learned by the CNN model, which allows for the selection of sensor channels with higher-level fault-related features. This approach provides two advantages: it reduces redundant information and improves diagnostic performance, and it simplifies the model by reducing the number of sensors required, thereby reducing communication burden and computational complexity.

Recent advancements in machine learning and deep learning have significantly enhanced fault detection capabilities. Techniques such as Long Short-Term Memory (LSTM), Transformers, and Convolutional Neural Networks (CNN) have demonstrated promising results in fault classification, particularly in rotary machinery. However, monitoring hydraulic systems using single-sensor anomaly detection presents challenges due to their interconnected nature, with limited exploration of multi-component fault diagnosis.

The existing prior art discloses various systems for diagnosing faults in hydraulic systems often require multiple sensors, which is costly affair and complex to install.

The existing systems and methods do not provide real-time inventory updates or just-in-time alerts for diagnosing faults in hydraulic systems.

To address the shortcomings of existing methods, the present invention proposes a system and method for fault diagnosis in hydraulic power systems. A single sensor is used to collect data, which is then processed by a trained model to identify faults.

The present invention offers a more cost-effective and user-friendly solution compared to existing methods that require multiple sensors. This invention is trained on historical data and utilizes machine learning or deep learning techniques.

OBJECT OF THE INVENTION
To overcome the shortcomings in the existing art the main object of the present invention is to describe a system and method for identifying multiple faults in hydraulic power systems using a single sensor. The method comprises training a convolutional neural network (CNN) model on a dataset of normal and faulty system data. The trained CNN model is then used to classify new input data from the single sensor as indicative of one or more faults. The system includes a sensor module, a CNN model, and a fault identification module. The system and method offer a cost-effective and user-friendly solution for multiple fault detection compared to existing methods.

Yet another objective of the present invention is to provide a system and method for fault detection in hydraulic systems by streamlined approach which reduces complexity by using current signals instead of multiple sensors.

Yet another objective of the present invention is to provide a system and method for fault detection in hydraulic systems with fewer sensors, thereby reducing maintenance needs and minimizing downtime.

Yet another objective of the present invention is to provide a system and method for comprehensive fault identification, capable of diagnosing faults in various components of the hydraulic system.

SUMMARY OF THE INVENTION
The present invention describes a system and a method for multiple fault detection in hydraulic power systems based on a single sensor. More particularly a system and method to identify potential faults by using trained convolution neural network model. Moreover, a system and a method for multiple faut detection which is cost-effective and user-friendly solution compared to existing methods.

The current invention describes a low-cost method for monitoring hydraulic power systems that employs deep learning algorithms to detect defects in multiple components. This system eliminates the needs for sensors, making the fault diagnostic system more robust and less complex to install.

This invention addresses the limitations of multi-sensor systems by using single-sensor data, enhancing fault diagnosis efficiency. This invention methodology involves generating 2D plots from single sensor readings and applying CNNs for classification.

BRIEF DESCRIPTION OF DRAWINGS
Figure 1 depicts schematic diagram of sensors connected in line with main circuit
Figure 2(a) depicts damaged hydraulic hose.
Figure 2(b) depicts gas loaded accumulator.
Figure 2(a) depicts solenoid valve.
Figure 3 depicts finetuned transfer learning system based fault diagnosis system.
Figure 4(a) depicts J48 decision tree for accumulator fault classification.
Figure 4(b) depicts J48 decision tree for pump fault classification.
Figure 4(c) depicts J48 decision tree for stable state.
Figure 4(d) depicts J48 decision tree for vault fault classification.
Figure 5 depicts sample signal image of different fault class of component under study.
Figure 6 depicts accuracy plot of fine tuned Vgg16 with SGDM optimizer, 0.0003 learning rate, 30 epoch and 0.6 split ration for accumulator fault classification.
Figure 7 depicts loss plot of fine tuned Vgg16 with SGDM optimizer, 0.0003 learning rate, 30 epoch and 0.6 split ration for accumulator fault classification.
Figure 8 depicts confusion matrix of the fine tuned Vgg16 model for accumulator fault classification
Figure 9 depicts accuracy plot of fined tuned AlexNet model with SGDM optimizer, 0.0001 learning rate, 30 epoch and 0.8 split ratio for pump fault classification.
Figure 10 depicts loss plot of fined tuned AlexNet model with SGDM optimizer, 0.0001 learning rate, 30 epoch and 0.8 split ratio for Pump Fault classification.
Figure 11 depicts confusion matrix of fine-tuned AlexNet model for pump fault classification
Figure 12 depicts accuracy plot of fined tuned GoogLeNet model with SGDM optimizer, 0.0003 Learning rate, 30 epoch and 0.8 split ratio for stable flag conditions.
Figure 13 depicts loss plot of fined tuned GoogLeNet model with SGDM optimizer, 0.0003 learning rate, 30 epoch and 0.8 split ratio for stable flag conditions.
Figure 14 depicts confusion matrix of fine-tuned GoogLeNet model for stable flag conditions.
Figure 15 depicts accuracy plot of tuned GoogLeNet model with SGDM optimizer, 0.0003 learning rate, 20 epoch and 0.7 split ratio for valve fault classification.
Figure 16 depicts loss plot of tuned GoogLeNet model with SGDM optimizer, 0.0003 Learning rate, 20 epoch and 0.7 split ratio for valve fault classification.
Figure 17 depicts confusion matrix of the GoogLeNet model for valve fault classification.
Figure 18 depicts top performing pre-trained model specific to component fault.

DETAILED DESCRIPTION OF THE INVENTION
Some embodiments of the present disclosure, illustrating all its features, will now be discussed in detail. It must also be noted that as used herein and in the appended claims, the singular forms "a", "an" and "the" include plural references unless the context clearly dictates otherwise.

Various modifications to the embodiment will be readily apparent to those skilled in the art and the generic principles herein may be applied to other embodiments. However, one of ordinary skill in the art will readily recognize that the present disclosure including the definitions listed here below are not intended to be limited to the embodiments illustrated but is to be accorded the widest scope consistent with the principles and features described herein.

A person of ordinary skill in the art will readily ascertain that the illustrated steps detailed in the figures and here below are set out to explain the exemplary embodiments shown, and it should be anticipated that ongoing technological development will change the way functions are performed. These examples are presented herein for purposes of illustration, and not limitation. Further, the boundaries of the functional building blocks have been arbitrarily defined herein for the convenience of the description. Alternative boundaries can be defined so long as the specified functions and relationships thereof are appropriately performed. Alternatives (including equivalents, extensions, variations, deviations, etc., of those described herein) will be apparent to persons skilled in the relevant art(s) based on the teachings contained herein. Such alternatives fall within the scope and spirit of the disclosed embodiments.

Before discussing example, embodiments in more detail, it is to be noted that the drawings are to be regarded as being schematic representations and elements that are not necessarily shown to scale. Rather, the various elements are represented such that their function and general purpose becomes apparent to a person skilled in the art.

Any connection or coupling between functional blocks, devices, components, or other physical or functional units shown in the drawings or described herein may also be implemented by an indirect connection or coupling. A coupling between components may also be established over a wireless connection. Functional blocks may be implemented in hardware, firmware, software, or a combination thereof.

Further, the flowcharts provided herein, describe the operations as sequential processes. Many of the operations may be performed in parallel, concurrently, or simultaneously. In addition, the order of operations re-arranged. The processes may be terminated when their operations are completed but may also have additional steps not included in the figured. It should be noted, that in some alternative implementations, the functions/acts/ steps noted may occur out of the order noted in the figured. For example, two figures shown in succession may, in fact, be executed concurrently, or may sometimes be executed in the reverse order, depending upon the functionality/acts involved.

Further, the terms first, second etc… may be used herein to describe various elements, components, regions, layers and/or sections, these elements, components, regions, layers, and/or sections should not be limited by these terms. These terms are used only to distinguish one element, component, region, layer or section from another region, layer, or a section. Thus, a first element, component, region layer, or section discussed below could be termed a second element, component, region, layer, or section without departing form the scope of the example embodiments.

The terminology used herein is for the purpose of describing example embodiments only and is not intended to be limiting. Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which example embodiments belong. It will be further understood that terms, e.g., those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.

As used herein, the singular forms “a,” “an,” and “the,” are intended to include the plural forms as well, unless the context clearly indicates otherwise. As used herein, the terms “and/or” and “at least one of” include all combinations of one or more of the associated listed items. It will be further understood that the terms “comprises,” “comprising,” “includes,” and/or “including,” when used herein, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof.

This invention proposes a new fault diagnosis system for hydraulic power systems. It uses a single sensor to collect data, which is then processed by a trained model to identify faults. This is a more cost-effective and user-friendly solution compared to existing methods that require multiple sensors. This system trained on historical data and uses machine learning or deep learning techniques for fault detection and monitoring.
The system and method for fault detection and monitoring of hydraulic system is comprises of following structure:
• Dataset description.
• Methodology (feature extraction, sensor selection, CNN training).
• System evaluation.
• Experimental system setup.
• Result analysis of pre trained networks.
• Comparison with existing methods
• Conclusion

Dataset Description
The dataset used in this invention is sourced from the University of California (UCI) Machine Learning Repository, authored by Helwig et al. It contains signals from 15 sensors collected over a 60-second period under various fault conditions, along with data from 2 virtual sensors.

The signal acquisition setup is illustrated in Figure 1 as a schematic diagram. The frequency at which the sensor data is acquired is detailed in Table 1. The setup includes essential components:
• Pump (MP1)
• Accumulators (A1-A4) [31]
• Directional control valve (V10) [32]
• Pressure relief valve (V11), which acts as an actuator under steady load conditions to simulate the real-time functioning of hydraulic system.

Simulate specific fault conditions produced by these components, valves (V1-V9) were manually operated and connected in line to the main circuit. These valves were systematically controlled to replicate different fault scenarios. Additionally, a total of 17 sensors were connected in line to the main circuit board. Data from these sensors were collected during various fault events to create the dataset.

Table 1 sensor details
S.no Sensor No of sensor Abbreviation of the sensor Frequency
1 Pressure sensor 6 Ps1-Ps6 100Hz
2 Energy /Watt meter 1 Eps1 100Hz
3 Flow Sensor 2 Fs1-Fs2 10Hz
4 Temperature sensor 5 Ts1-Ts5 1Hz
5 Accelerometer/Vibration Sensor 1 Vs1 1Hz
6 Efficiency Factor 1 Se 1Hz
7 Virtual sensor-cooling power 1 Cp 1Hz
8 Virtual sensor-cooling efficiency 1 Ce 1Hz

For this study, three major components prone to failure in hydraulic systems are specifically examine, namely, hydraulic hoses/pumps, accumulators, and valves. These components are illustrated in Figures 2a, 2b, and 2c. The several types of faults associated with each component are listed in Table 2. The dataset for the accumulator includes four distinct types of faults, though the number of examples for each type is not balanced. There are a total of 2,205 examples for the accumulator, with the same number of examples for the other components.

Table 2 fault details and description
Component Number of class Descriptions Abbreviation of class used Instances/ samples
Accumulator 4 Optimal pressure Acc 130 599
Slightly reduced pressure Acc 115 399
Severely reduced pressure Acc 100 399
Close to total failure Acc 90 808
Pump / hose 3 No leakage Pump 0 1221
Weak leakage Pump 1 492
Severe leakage Pump 2 492
Stable flag 2 Conditions were stable Stable 1449
Unstable Un stable 756
Valve 4 Optimal switching
Behaviour Valve 100 1125
Small lag Valve 90 360
Severe lag Valve 80 360
Close to total failure Valve 73 360

Methodology
This invention methodology comprises of:
- Feature Extraction.
- Sensor selection.
- Convolution neural network training
- Image generation and preprocessing,
- Transfer learning and model selection,
- Evaluation.

Feature Extraction
In this invention crucial part is extraction of statistical features from the sensor data for machine learning in fault detection system. Statistical features were extracted using the Microsoft Data Analysis Toolbox and Visual Basic for Applications (VBA). Statistical features used as input for the J48 decision tree to identify crucial features. The feature selection process involves extracting valuable information from raw data to facilitate machine classification across different classes.

The present inventive system discloses a total of 12 statistical features which were extracted from each sensor's data: mean, median, mode, standard deviation, sum, standard error, sample variance, skewness, kurtosis, maximum, minimum, and range. These features were calculated from 17 sensor readings, resulting in a total of 204 features (17 sensors × 12 statistical features).

Each statistical feature is named to indicate its source i.e., "Eps1_AM" denotes the arithmetic mean from Energy Sensor 1, while "Fs1_SM" represents the sum of values from Flow Sensor 1, as described in Table 2. This compiled dataset is then used as input for the J48 decision tree algorithm to conduct a feature selection study across the four different components.

Sensor selection.
To selects most informative sensors for fault detection, a dataset containing 204 features is fed into the J48 decision tree algorithm for all four classification problems, using default parameters of a confidence factor of 0.25 and a minimum number of objects set to 2.

The outputs from the J48 decision tree are then analysed to determine which features are essential for fault classification. the rank of each feature is assessed in the decision tree by employing a drop-down approach In this method, the root node (the topmost feature) of the decision tree is considered the most significant. The significance levels decrease progressively from the root node to the leaves and from left to right across the tree. The decision trees used to extract the significance of features are presented in Figures 4a-4d for the accumulator, pump, stable state, and valve, respectively.

The significance levels of each feature in the classification study are identified and presented in Table 4. The features listed in this table represent the optimal number necessary for achieving maximum classification accuracy with a minimum number of features.

Table 3 highlights the significant contributions of sensors and their features to fault classification, with a particular focus on individual components. The features in Table 3 are optimized to yield maximum accuracy while using fewer features. For instance, in accumulator fault classification, utilizing the eight features listed in the accumulator column achieves a classification accuracy of 95.10%, compared to an accuracy of 95.14% when using the original 204 features. Similarly, for stable state classification, the model requires only ten features out of the available 204 to achieve an optimal classification accuracy of 97.23%. Additionally, for other fault classifications, the performance of the J48 model improves with the selected features, as indicated in Table 3.

Table 3 feature ranking based on J48 decision tree and its performance.
Rank Accumulator Pump Stable flag Valve
1 Se AM Fs1 MD Ps2 KT Ps2 SK
2 Fs 1SK Se MD Eps KT Eps KT
3 Ps5 RG Eps1 MM Fs1 MD Ps2 KT
4 Cp MD Ps3 MD Ps1 SM TS1 SD
5 Fs2 AM Ps1 RMS Ts2 SK Ts1 SV
6 Eps1 AM Fs1 SD Fs1 SD Eps SK
7 Ps3SD Ps5 SV Eps1 RG Ps1 RMS
8 Ts4 MX Eps1 MX Ts1 MD Ts4 RG
9 Cp MO
10 Ps2 MD
J48 Decision tree -Accuracy (%)
With the selected features 95.10 99.31 97.23 99.00
With all 204 features 95.14 99.27 96.73 98.54

Analysis of Table 4 provides valuable insights for fault detection. Among the 17 sensors, Eps1 and Fs1 consistently contribute significantly to fault detection across all classification tasks. However, the FS1 sensor is particularly significant in the cases of accumulator and pump fault classification, while it shows no significance in valve fault classification and less significance than Eps1 in stable state classification. Furthermore, Eps1 captures data at a much higher frequency (100 Hz) compared to the FS1 sensor. This increased sampling rate translates to richer and more detailed information within the signals, which is especially valuable for accurate classification, such as when generating images from the sensor data for deep learning-based classification.

Given these findings, Eps1 is chosen for further investigation using a transfer learning approach in the fault detection process. This strategic selection ensures that the model receives the essential data required for deep learning techniques, emphasizing the ongoing significance of Eps1 in enhancing fault diagnosis accuracy. Additionally, it is worth noting that the dataset considered for the study used valve V11 instead of an actuator, with its output connected to a flow control valve. This may not accurately reflect real-time applications.

Table 4 characteristics features of adopted pre-trained networks
Network Number of Layers Input Size Learnable Parameters Unique Features Complexity
AlexNet 8 227x227 60.00 Uses nolinear rely activation function Low
VGG16 16 224x224 137.00 Uses small size kernels High
Google net 22 224x224 7.10 Replace large filters with small Low

Convolution neural network training
- Image generation and preprocessing
Signals from the Eps1 sensor are processed using deep neural networks to manage the high dimensionality of the data and to learn features from sequential time series data. During data collection, the data acquisition (DAQ) system is configured to gather data for one minute at a rate of 100 Hz, resulting in a sample length of 6,000. This data is stored as a series of signal files. These signal files are then used to generate line charts with a MATLAB program, where the X-axis represents the sample length, the Y-axis represents power in watts (W). Ensure consistency in the generated images, the limits of both axes are kept constant: the X-axis ranges from 0 to 6,000 (iterations), while the Y-axis ranges from 2,200 to 3,000 watts. This process of generating colour images is repeated for all fault classes considered in the study. Additionally, to maintain uniformity, 300 images per class are generated for each component fault study. An illustrative sample image generated through the MATLAB program is shown in Figure 5. Figure 5 showcases the line chart plotted in MATLAB, representing different fault states of the components analysed for the fault classification study. The extracted images serve as inputs for training the model. The images generated by the MATLAB program have a size of 875 x 656 pixels, which is incompatible with the pre-trained network inputs. Therefore, these images are preprocessed by resizing to dimensions of 227 x 227 pixels for AlexNet, and to 224 x 224 pixels for VGG16 and GoogleNet.

Furthermore, a balanced dataset approach is implemented by maintaining 300 images per fault class, regardless of the component considered. This ensures uniformity while proposing a single CNN model for use in the fault diagnosis system across different components of the hydraulic system.

- Transfer learning and model selection
Pre-trained Convolutional Neural Networks (CNNs) play a crucial role in Deep Neural Network (DNN) architectures, particularly within the realm of image recognition. transfer learning emerges as a valuable strategy for fault diagnosis studies, especially when resources like datasets and computational power are limited. The present invention discloses pre-trained networks such as AlexNet, VGG16, and GoogleNet are utilized, as they have proven effective in handling smaller datasets and optimizing computational resources. Table 4 describes the characteristic features of the chosen pre-trained networks.

System Evaluation
The performance of the transfer learning model is evaluated using various metrics, with a focus on accuracy and training time. Given that the models are trained using a balanced dataset, both accuracy and efficiency are crucial. Additional metrics, such as precision, recall, and F1 score, are also considered, particularly when the model's accuracy is not 100%. This evaluation provides insights into the model's performance, especially for individual fault classes. Detailed calculations of accuracy, precision, recall, and F1 score are provided.

Experimental system setup
The system consists of an Intel Xeon W-1350 CPU, 64GB of RAM, a Radeon™ PRO WX 5100 graphics card, and 1TB of storage, running MATLAB R2022b. For fault classification, the system utilized four datasets representing components: accumulator, valve, pump, and stable state.

These datasets were balanced, with 900 images for the pump and 600 for the stable state, and 300 images per class, resulting into a total of 1,200 images for the accumulator and valve fault classification studies. To further enhance the neural network, the hyperparameter values listed in Table 6 were used to train the transfer learning model relevant to the dataset, and the best configurations for each component were selected for real-time application.

Table 5 Parameter Consider for Training Pertained Network
Parameter setting Value range
Optimizer SGDM*, Adam & RMS prop
Learning rate 0.0003*, 0.0001 &0.001
Epochs 30*, 20 & 10
Split ratio 0.8*, 0.7, 0.6

Result analysis of pre trained networks.
The analysis begins by establishing a baseline, where pre-trained models (AlexNet, VGG16, and GoogleNet) are trained using their default parameters as shown in see Table 5. These include the optimizer (SGDM), learning rate, epoch count, and train-test split ratio, which separates the data for training and evaluation.

The learning rate is fine-tuned using the specified optimizer, and models are trained with learning rates different from the default value (0.0003) to identify the best results. Similarly, the epoch count, which indicates how many times the model sees the entire training data, is modified. Models are trained using 20 and 10 epochs, alongside the default of 30 epochs, selecting the count that yields optimal results. The train-test split ratio, which determines the proportion of data used for training versus testing, is optimized using ratios of 0.7, 0.6, and the standard 0.8, selecting the one that provides the best overall performance.

After tuning these hyperparameters for each pre-trained network (AlexNet, VGG16, and GoogleNet), their performance is evaluated to determine the most effective model for identifying faults in hydraulic system components.

Analysis of pre trained network for accumulator fault classification.
The choice of optimizer significantly impacts the performance of fine-tuned pre-trained models. Adam was the best choice for AlexNet and GoogleNet, achieving accuracies of 94.6% and 95.42%, respectively. VGG16 performed best with the default SGDM optimizer, reaching an accuracy of 96.67%. VGG16 excelled in parameter tuning, achieving a classification accuracy of 97.50% with a combination of hyperparameters: SGDM optimizer, a learning rate of 0.0003, 30 epochs, and a 0.6 train-test split ratio, with a training time of 14 minutes and 3 seconds. This model is a reliable tool for classifying accumulator faults, as shown in Figures 6, 7, and 8, which display the accuracy, loss curve during training, and confusion matrix. The model's precision for the "Acc100" class is lower at 0.92, likely due to misclassifications, but the overall classification accuracy remains impressive at 97.50%. Performance metrics are detailed in Table 7.

Table 6 performance of pre trained networks for accumulator fault classification.

Table 7 performance of fine tuned Vgg16 for accumulator fault classification

.
Analysis of pre trained network for Pump Fault Classification
The AlexNet model performed well with default parameters, except for the learning rate, achieving an accuracy of 97.8% with a training time of 14 minutes and 7 seconds. Figures 9, 10, and 11 illustrate the model's accuracy, loss curve, and confusion matrix. The AlexNet model's average performance for pump fault classification exceeds 97%, demonstrating the effectiveness and robustness of the sensor data, ensuring high accuracy and F-measure.

Table 8 performance of pretrained network for pump fault classification

Table 9 performance of fine tuned AlexNet model for pump fault classification

Analysis of pre trained network for Stable State Condition Classification
In analyzing the impact of hyperparameters on pre-trained networks, the GoogleNet model achieved 100% accuracy in fault classification, while AlexNet and VGG16 achieved 98.30% and 98.33%, respectively. The proposed GoogleNet model is suitable for determining hydraulic system stability. Figures 12, 13, and 14 present the model's accuracy, loss curve, and confusion matrix of the fine-tuned model with optimal hyperparameters.

Table 10 performance of pre trained networks for stable flag classification.

Analysis of pre trained network for Valve Fault Classification
The pre-trained network is fine-tuned to adapt to the dataset for classifying faults in the valve (V10) of the hydraulic system. The hyperparameter values used are highlighted in Table 11. Most models achieved maximum classification accuracy of 100%. However, among the three pre-trained networks, GoogleNet took significantly less time to train (225 seconds) compared to the others, which is crucial for real-time applications in dynamic environments. The energy meter readings show significant variations for different valve fault conditions, forming distinct patterns that enhance the model's classification accuracy. The model's accuracy, loss curve, and confusion matrix for the fine-tuned GoogleNet model with optimal hyperparameters are presented in Figures 15, 16, and 17.

Table 11 performance of pre trained networks for valve fault classification.

Analysis of pre trained network for Fault Classification Across Multiple Components
This analysis evaluates the performance of pre-trained networks across all hydraulic components considered in this study. Table 12 presents the network parameters that achieved the highest accuracy in fault classification, along with their corresponding training times. VGG16 achieved the highest average fault classification accuracy of 98.13%, with an average training time of 8.4 minutes. GoogleNet achieved an average accuracy of 98.03%, just 0.1% lower than VGG16 but with a training time approximately 50% shorter. Due to its efficient training speed, GoogleNet is preferred for real-time applications, despite the minor decrease in accuracy.

The recommended models for fault classification in hydraulic system components are as follows: VGG16 for accumulator faults (97.5% accuracy), AlexNet for pump leakage detection (97.8% accuracy), and GoogleNet for valve and stable state fault classification (100% accuracy). These models show great potential for real-time applications in fault classification within hydraulic systems.

Table 12 performance comparison with respect to fine pre tuned model

Comparison of Proposed Model with State-of-the-Art Literature
The proposed method is compared with existing literature using the Helwig dataset, training models with signals from all 17 sensors. This transfer learning model addresses the gap by utilizing CNNs with colour images.

Table 13 provides an overview of studies using the Helwig dataset for fault classification, including features, models, and the best-reported accuracies.

While the consistent use of datasets with signals from all 17 sensors is noted, variations in model performance depending on the fault classification tasks are also evident. For example, an ANN model by Helwig et al. achieved 80% accuracy in pump fault classification using statistical and FFT features, while Quatrini et al. achieved 99.89% accuracy with a combination of slope of the line features and statistical features. This highlights the importance of feature selection and meticulous feature extraction, which require considerable time and domain expertise. The performance of the same model varies for different component fault classifications, underscoring the necessity of domain knowledge in training machine learning models.

The proposed methodology for multi-component fault diagnosis employs a single sensor input and a CNN to achieve classification accuracies exceeding 90%. This innovative approach, utilizing 50% or less of the typically required data, emphasizes the need for further research in testing such methods across various components for fault classification.

Table 13 comparison of proposed methodology with existing literature.

Conclusions
This invention utilized an open-source dataset to monitor hydraulic systems for potential issues by extracting key information from sensor data. It employed machine learning and deep learning techniques to detect anomalies in three different components, as well as in the stable state of the system. A notable distinction of this invention is that its use of convolutional neural networks (CNNs) with colour images, a novel approach not previously explored. The present invention outperformed existing studies, operating efficiently with just one sensor and a limited amount of data, thus enhancing cost-effectiveness.

ADVANTAGES OVER EXISTING SYSTEMS
• Reduced Cost and Increased Efficiency: Utilizing a single sensor simplifies installation and maintenance, leading to lower costs and improved operational efficiency.
• Streamlined Data Acquisition: Collecting data from a single sensor minimizes complexity and reduces the potential for inconsistencies.
• Enhanced Fault Detection and System Health: Machine learning and deep learning algorithms can identify subtle anomalies and predict potential faults, improving overall system health.
• Improved Diagnostic Accuracy: The trained model offers more accurate diagnoses compared to traditional methods, enhancing reliability.
• Broader Fault Coverage: The system is capable of detecting a wider range of faults, including those not directly monitored by the sensor.
• Non-Invasive Approach: The system requires no physical modifications to the hydraulic system, making it easier to implement.
• Real-Time Monitoring: Continuous monitoring facilitates early fault detection and enables preventative maintenance, enhancing system longevity.

POTENTIAL COMMERCIAL APPLICATIONS
• Marine Vessels: Early detection of faults in critical components, such as steering and transmission systems, enhances safety and reliability while reducing downtime.
• Automation Industries: Timely fault detection ensures the smooth operation of vital hydraulic components, minimizing disruptions in manufacturing processes.

Mobile Equipment Operators: The system's cost-effectiveness and user-friendliness make it ideal for various mobile equipment that rely on hydraulics, including construction machinery and agricultural equipment.
, C , Claims:WE CLAIM:
1. A system for fault detection in a hydraulic power system comprising:
a. at least one sensor module for collecting data from the hydraulic power system;
b. at least one convolutional neural network (CNN) model for processing the collected data; and
c. atleast one fault identification module for identifying potential faults based on the output of the CNN model;
wherein the system can detect multiple faults in hydraulic systems based on a single sensor.

2. The system as claimed in claim 1, wherein the sensor module comprises a single sensor for collecting data.

3. The system as claimed in claim 1, wherein the CNN model is trained on a dataset of normal and faulty data from the hydraulic power system.

4. The system as claimed in claim 1, wherein the CNN model is a pre-trained CNN model.

5. The system as claimed in claim 4, wherein the pre-trained CNN model is selected from the group consisting of AlexNet, VGG16, and GoogleNet.

6. The system as claimed in claim 1, wherein the fault identification module identifies multiple faults based on the output of the CNN model.

7. The system as claimed in claim 1, wherein the fault identification module identifies faults in at least one of the following components of the hydraulic power system: an accumulator, a pump, a valve, and a stable state.

8. The system as claimed in claim 1, wherein the CNN model is trained on color images generated from the sensor data.

9. The system as claimed in claim 1, wherein the fault identification module is configured to provide real-time fault detection.

10. A method for fault detection in a hydraulic power system comprising:
a. collecting data from the hydraulic power system using a sensor module;
b. b. processing the collected data using a convolutional neural network (CNN) model; and
c. identifying potential faults based on the output of the CNN model;
wherein the method provides real-time fault detection.

11. The method as claimed in claim 11, wherein the sensor module comprises a single sensor.

12. The method as claimed in claim 10, wherein the CNN model is trained on a dataset of normal and faulty data from the hydraulic power system.

13. The method as claimed in claim 10, wherein the CNN model is a pre-trained CNN model.

14. The method as claimed in claim 14, wherein the pre-trained CNN model is selected from the group consisting of AlexNet, VGG16, and GoogleNet.

15. The method as claimed in claim 11, wherein the method identifies multiple faults in the hydraulic power system.

16. The method as claimed in claim 11, wherein the method identifies faults in at least one of the following components of the hydraulic power system: an accumulator, a pump, a valve, and a stable state.

17. The method as claimed in claim 11, wherein the CNN model is trained on color images generated from the sensor data.