An interpretable rolling bearing fault diagnosis method

By introducing Grad-CAM++ interpretation layer in the 1D-CNN-BiLSTM model, combining convolution and BiLSTM layer, the interpretability and noise immunity of traditional rolling bearing fault diagnosis methods are solved, and high-precision fault diagnosis and classification effects are achieved .

CN114372492BActive Publication Date: 2024-11-05SOUTH CHINA UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111676747.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-31
Publication Date
2024-11-05
Estimated Expiration
2041-12-31

AI Technical Summary

Technical Problem

Traditional rolling bearing fault diagnosis methods require a lot of prior knowledge. The CNN process lacks interpretability and poor noise resistance, so it cannot effectively deal with complex operating conditions and noise influences.

Method used

Using the 1D-CNN-BiLSTM neural network model, and adding Grad-CAM++ interpretation layer on it, the model is trained to obtain a model with fault diagnosis capabilities through the combination of convolutional interpretation capabilities and BiLSTM layers, and utilize multiple one-dimensional faults. The data is trained and verified to improve the robustness and interpretability of the model.

Benefits of technology

It realizes interpretability and high accuracy of rolling bearing fault diagnosis, can effectively process one-dimensional time signals, has good noise resistance and classification effects, and significantly improves diagnostic accuracy and robustness.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114372492B_ABST
    Figure CN114372492B_ABST
Patent Text Reader

Abstract

The present invention discloses an interpretable rolling bearing fault diagnosis method, comprising the following steps: collecting one-dimensional time series signals of rolling bearings and performing sample expansion, establishing an initial 1D-CNN-BiLSTM neural network model, and adding a Grad-CAM++ interpretation layer to the neural network model to establish a neural network model with convolutional interpretation capabilities. The neural network model is trained using a variety of one-dimensional fault data to obtain a model with fault diagnosis capabilities, and then the rolling bearing is fault diagnosed using the fault diagnosis model. The present invention interprets the feature extraction process of a neural network with CNN as the basic structure, and adds BiLSTM, utilizing its characteristics of having bidirectional analysis capabilities to achieve better diagnostic accuracy, and improve the noise resistance and robustness of the fault diagnosis neural network model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of rotating machinery fault diagnosis and signal processing, and in particular relates to an interpretable rolling bearing fault diagnosis method. Background Art

[0002] Rolling bearings are widely used in various industries, such as vehicles and aerospace. The health of rolling bearings is the basic guarantee for the smooth operation of the entire equipment, so the monitoring and comprehensive evaluation of the operating status of rolling bearings are of great significance. Rolling bearings are prone to fatigue damage and performance degradation, which affects the safety and reliability of the entire system. Research on fault diagnosis and health management has important theoretical research significance and engineering application value.

[0003] There are two problems with traditional fault diagnosis methods. The first is that they require a lot of expert knowledge related to mechanical engineering and statistics. The second is that as the operating conditions of rolling bearings become more and more complex, traditional fault diagnosis methods usually do not perform very well and are not adaptive. Therefore, it is necessary to find a method based on feature learning to extract fault features and conduct diagnosis research.

[0004] With the continuous development of artificial intelligence and deep learning, convolutional neural networks have been widely used in many fields due to their advantages such as shared convolution kernels, no pressure on high-dimensional data processing, no need for prior knowledge, and automatic feature extraction. However, most CNNs are not interpretable, which leads to the process being often criticized as a "black box". In addition, when the noise is large, the diagnostic performance of conventional CNNs is not satisfactory, so it is extremely important to propose an interpretable bearing fault diagnosis method with good noise resistance.

[0005] The shortcomings of a bearing detection method based on convolutional neural network in the prior art (CN201910985498.4) are as follows:

[0006] 1. The weight is estimated by the size Z of the feature map in the derivation process of Grad-CAM, so the weight size is closely related to the size of the feature map and will affect the result. In addition, the method of the present invention cannot activate the same class that appears multiple times in the input image well.

[0007] 2. Its noise immunity effect is unknown. Summary of the invention

[0008] The purpose of the present invention is to propose an interpretable rolling bearing fault diagnosis method in view of the shortcomings of the prior art, aiming to solve the technical problems of the traditional fault diagnosis method, such as the need for more prior knowledge, the CNN process is often criticized as a black box, and the poor noise resistance of ordinary CNN.

[0009] The present invention is achieved by at least one of the following technical solutions.

[0010] An interpretable rolling bearing fault diagnosis method comprises the following steps:

[0011] S1, collecting one-dimensional time series signals of rolling bearings and performing sample expansion;

[0012] S2, establish the initial 1D-CNN-BiLSTM neural network model;

[0013] S3, adding an explanation layer to the neural network model;

[0014] S4. Establish a neural network model with convolutional interpretation capabilities;

[0015] S5. Using a variety of one-dimensional fault data to train the neural network model to obtain a model with fault diagnosis capability;

[0016] S6. Perform fault diagnosis on the rolling bearing using the fault diagnosis model.

[0017] Furthermore, step S1 includes the following steps:

[0018] S11, collecting one-dimensional time signals of four categories, including normal, outer ring fault, inner ring fault and rolling element fault, collected by the acceleration sensor;

[0019] S12, expanding the one-dimensional time signal by overlapping sampling, and dividing the data samples into a training set and a test set;

[0020] S13, dividing the data sample into different training sets and test sets multiple times by using K-fold cross-validation, and finally taking the average result to improve the robustness of the model.

[0021] Furthermore, the structure of the initial 1D-CNN-BiLSTM neural network model includes an input layer (input), a convolutional neural network layer (CNN), a BiLSTM layer, a fully connected layer (Fully Connected) and an output layer (output):

[0022] The CNN layer includes a convolutional layer (ConV) and a pooling layer (Max Pooling). The activation function of each convolutional layer is the Relu function.

[0023] The pooling method of the pooling layer is maximum pooling;

[0024] The activation function of the output layer is the Softmax function.

[0025] Furthermore, the BiLSTM layer includes a bidirectional LSTM layer, and the composition of the LSTM layer is specifically as follows:

[0026] f t =σ(W f [h t-1 ,x t ]+b f )

[0027] i t =σ(W i [h t-1 ,x t ]+b i )

[0028] o t =σ(W o [h t-1 ,x t ]+b o )

[0029] c t =f t c t-1 +i t tanh(W c [h t-1 ,x t ]+b c )

[0030] h t =o t tanhc t

[0031] In the formula, f t For the forget gate, i t is the input gate, o t is the output gate, c t is the cell state, σ is the sigmoid activation function, W f , W i , W o , W c is the weight in the training parameters, b f 、b i 、b o 、b c is the bias term in the training parameters, h t is the state of the hidden layer at the current moment, h t-1 is the state of the hidden layer at the previous moment, x t The input at the current moment.

[0032] Furthermore, the Relu function is:

[0033]

[0034] Where f(x) is the Relu function and x is the input vector of the previous layer.

[0035] Furthermore, the Softmax activation function is:

[0036]

[0037] In the formula, z i+1 is the value output by the i+1th point in Z, C is the number of output nodes, z c is the value output by the cth point in Z, and e is a natural constant.

[0038] Furthermore, the interpretation layer is Grad-CAM++.

[0039] Further, step S3 includes the following steps:

[0040] S31. Calculate the weights of each category of each channel after the last convolutional layer of the 1D-CNN-BiLSTM neural network model:

[0041]

[0042] Where Y x is the score value of category x, i.e., its importance. is the weight of the kth feature map for category x, is the pixel value at (i, j) of the kth feature map, is the weight of category x at (i, j) of the kth feature map;

[0043] S32, generating a class activation map according to the weights:

[0044]

[0045] S33, transforming the size of the class activation map to the size of the original input image, and then overlaying it on the original image through a heat map, the size of the heat map value corresponds to the activation degree of the 1D-CNN-BiLSTM neural network model to the input signal.

[0046] Furthermore, the specific steps of step S4 are:

[0047] S41, connect the interpretation layer and the BiLSTM layer after the 1D-CNN-BiLSTM neural network model;

[0048] S42, the interpretation layer is used as the middle layer of the whole model to explain the convolution process;

[0049] S43, using the BiLSTM layer as the back end of the entire model, and using the output of the last convolutional layer of CNN as the input of the back end to perform fault diagnosis.

[0050] Furthermore, the specific steps of step S5 are:

[0051] S51, classifying the samples collected in step S1 into four categories, namely normal, outer ring fault, inner ring fault and rolling element fault;

[0052] S52, dividing the fault samples into a training set and a test set, and performing training and testing using a K-fold cross validation method;

[0053] S53, using the training set as input of a neural network model with convolutional interpretation capability, performing model training, and obtaining a model with fault diagnosis capability.

[0054] Compared with the prior art, the present invention has the following advantages and effects:

[0055] (1) It is interpretable and universal. In simulation analysis and experimental verification, Grad-CAM++ was used to prove the rationality and interpretability of convolutional neural network feature extraction. The method of the present invention can be applied not only to two-dimensional images, but also to one-dimensional time signals such as rotating machinery faults, expanding the scope of application of class activation maps and having universality.

[0056] (2) Good diagnostic accuracy and classification results. Experiments were conducted using the rolling bearing fault dataset from Case Western Reserve University. The diagnostic results showed excellent clustering effects for the same category and classification effects for different categories.

[0057] (3) Good noise resistance and robustness. The diagnostic accuracy of the model described in the present invention is higher, and when the noise gradually decreases, its diagnostic accuracy converges faster. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. The drawings constitute part of this application, but are only non-limiting examples that embody the inventive concept and are not intended to make any limitation.

[0059] Figure 1 It is a flow chart of the implementation of the method of the present invention;

[0060] Figure 2 It is a neural network architecture diagram in the method of the present invention;

[0061] Figure 3a This is the Grad-CAM++ diagram of the outer ring failure of the rolling bearing experiment at a speed of 1797 r / min;

[0062] Figure 3b This is the Grad-CAM++ diagram of the inner race fault of the rolling bearing experiment at a speed of 1797 r / min;

[0063] Figure 3c This is the Grad-CAM++ diagram of the rolling element failure of the rolling bearing experiment at a speed of 1797 r / min;

[0064] Figure 3d This is the Grad-CAM++ diagram of the normal state of the rolling bearing experiment at a speed of 1797r / min;

[0065] Figure 4 It is a t-SNE analysis diagram of the fault diagnosis result at a speed of 1797 r / min in the method of the present invention;

[0066] Figure 5 It is a confusion matrix diagram of the fault diagnosis result at a speed of 1797 r / min in the method of the present invention;

[0067] Figure 6a This is the Grad-CAM++ diagram of the outer ring failure of the rolling bearing experiment at a speed of 1772 r / min;

[0068] Figure 6b This is the Grad-CAM++ diagram of the inner race fault of the rolling bearing experiment at a speed of 1772 r / min;

[0069] Figure 6c This is the Grad-CAM++ diagram of the rolling element failure of the rolling bearing experiment at a speed of 1772 r / min;

[0070] Figure 6d This is the Grad-CAM++ diagram of the normal state of the rolling bearing experiment at a speed of 1772r / min;

[0071] Figure 7 It is a t-SNE analysis diagram of the fault diagnosis result at a speed of 1772 r / min in the method of the present invention;

[0072] Figure 8 It is a confusion matrix diagram of the fault diagnosis result at a speed of 1772 r / min in the method of the present invention;

[0073] Figure 9a This is the Grad-CAM++ diagram of the outer ring failure of the rolling bearing experiment at a speed of 1750r / min;

[0074] Figure 9b This is the Grad-CAM++ diagram of the inner race fault of the rolling bearing experiment at a speed of 1750r / min;

[0075] Fig.9c This is the Grad-CAM++ diagram of the rolling element failure of the rolling bearing experiment at a speed of 1750r / min;

[0076] Figure 9dThis is the Grad-CAM++ diagram of the normal state of the rolling bearing experiment at a speed of 1750r / min;

[0077] Fig.10 It is a t-SNE analysis diagram of the fault diagnosis result at a speed of 1750 r / min in the method of the present invention;

[0078] Fig.11 It is a confusion matrix diagram of the fault diagnosis result at a speed of 1750r / min in the method of the present invention;

[0079] Fig.12 It is a diagram of the diagnostic accuracy of various network models under different signal-to-noise ratios in the method of the present invention. DETAILED DESCRIPTION

[0080] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0081] In the description of the present invention, “plurality” means two or more than two, unless otherwise clearly and specifically defined.

[0082] The present invention provides an explainable rolling bearing fault diagnosis method, which is described in detail below.

[0083] like Figure 1 As shown, a flow chart of an embodiment of an interpretable rolling bearing fault diagnosis method based on Grad-CAM++ and CNN-BiLSTM provided by an embodiment of the present invention includes the following steps:

[0084] S1, collecting one-dimensional time series signals of rolling bearings and performing sample expansion;

[0085] S2, establish the initial 1D-CNN-BiLSTM neural network model;

[0086] S3, adding a Grad-CAM++ interpretation layer to the neural network model;

[0087] S4. Establish a neural network model with convolutional interpretation capabilities;

[0088] S5. Using a variety of one-dimensional fault data to train the neural network model to obtain a model with fault diagnosis capability;

[0089] The data specifically includes bearing outer ring fault, inner ring fault, rolling element fault and normal data. The fault samples are divided into training set and test set, and the K-fold cross validation method is used for training and testing, with K=10, 9 samples are taken for training each time, and the remaining 1 sample is used for testing. The training set is used as the input of the 1D-CNN-BiLSTM neural network model, and each sample dimension is 2048×1. Model training is performed to obtain a model with fault diagnosis capabilities.

[0090] S6. Perform fault diagnosis on the rolling bearing using the fault diagnosis model.

[0091] The interpretable rolling bearing fault diagnosis method based on Grad-CAM++ and CNN-BiLSTM provided in the embodiment of the present invention well explains the convolution results of the convolutional neural network in the form of a visual heat map based on the Grad-CAM++ algorithm, improves the reliability of the diagnosis results through the advantages of CNN local feature extraction and BiLSTM's ability to handle nonlinear time series, and can reduce the parameters when training the network, thereby increasing the training speed. It is further compared with 1D-CNN, 1D-CNN-SVM and 1D-CNN-LSTM with the same structure and parameters as the model described in the present invention to verify the excellent noise resistance of the 1D-CNN-BiLSTM interpretable neural network.

[0092] In some embodiments of the present invention, Figure 1 As shown, S1 includes:

[0093] S11, collect one-dimensional time signals of four categories, including normal, outer ring fault, inner ring fault and rolling element fault, collected by the acceleration sensor. Here, the Case Western Reserve University rolling bearing fault data set is taken as an example, and three experiments are conducted at different speeds, namely 1797r / min, 1772r / min, and 1750r / min;

[0094] S12, expanding the one-dimensional time signal by overlapping sampling, and dividing the data samples into a training set and a test set;

[0095] S13, dividing the data sample into different training sets and test sets multiple times by using K-fold cross-validation, and finally taking the average result to improve the robustness of the model.

[0096] In some embodiments of the present invention, Figure 2As shown, the structure of the initial 1D-CNN-BiLSTM neural network model includes an input layer (input), two convolutional layers (ConV), two pooling layers (Max Pooling), a BiLSTM layer, a fully connected layer (Fully Connected) and an output layer (output):

[0097] The first convolution layer has 16 4×1 convolution kernels, and the second convolution kernel has 32 2×1 convolution kernels. The activation function of each convolution layer is the Relu function.

[0098] Specifically, the Relu function is:

[0099]

[0100] Where f(x) is the Relu function and x is the input vector of the previous layer.

[0101] The pooling method of the pooling layer is maximum pooling, the step size of the first pooling layer is 4, and the step size of the second pooling layer is 2;

[0102] The number of neurons in the first fully connected layer is 64, the number of neurons in the second fully connected layer is 4, and the dropout rate is set to 0.5 to prevent overfitting.

[0103] The number of neurons in the output layer is consistent with the number of fault types, and the activation function is the Softmax function.

[0104] Specifically, the Softmax function is:

[0105]

[0106] In the formula, z i+1 is the value output by the i+1th point in Z, C is the number of output nodes, z c is the value output by the cth point in Z, and e is a natural constant.

[0107] Further, in some embodiments of the present invention, the Grad-CAM++ interpretation layer is added to the neural network model to obtain a model with convolutional interpretation capability, the result of which is shown in FIG3 , S3 includes:

[0108] S31. Calculate the weights of each channel for each category after the convolution layer, as shown below:

[0109]

[0110] Where Y x is the score value of category x, i.e., its importance. is the weight of the kth feature map for category x, is the pixel value at (i, j) of the kth feature map, is the weight of category x at (i, j) of the kth feature map.

[0111] S32, the calculation method of the generated class activation map is as follows:

[0112]

[0113] S33, the class activation map needs to be resized to the size of the original input image, and then overlaid on the original image through a heat map, and the value of the heat map corresponds to the activation degree of the convolutional neural network at this position.

[0114] Three embodiments are selected for illustration. Embodiment 1 is the case when the rotation speed is equal to 1797 r / min, Embodiment 2 is the case when the rotation speed is equal to 1772 r / min, and Embodiment 3 is the case when the rotation speed is equal to 1750 r / min. The three embodiments are analyzed together as follows.

[0115] Rolling bearing outer ring failure bearing such as Figure 3a , Figure 6a , Figure 9a As shown in the figure, the 1D-CNN-BiLSTM neural network has a greater activation degree for the impact part of the input signal, and the convolution layer of the neural network is responsible for extracting fault features in this experiment, which means that the network has a greater weight on this impact part and pays more attention to it.

[0116] Bearings with inner ring failures Figure 3b , Figure 6b , Figure 9b As shown, similar to the case of outer ring failure, the neural network also has a greater degree of activation for the impact part of the input signal, indicating that the network has a greater weight on this part.

[0117] Rolling element failure of bearings such as Figure 3c , Figure 6c , Fig.9c As shown in the figure, unlike the inner and outer ring faults, the impact part of the time domain signal of the rolling element fault sample is unevenly distributed, and there is no particularly obvious impact area, but it is different from the signal of the normal bearing. Figure 3d , Figure 6d , Figure 9d In comparison, it still shows that the neural network selectively activates such signals, and the weight of the activated area is larger.

[0118] The time domain signal of a normal bearing is evenly distributed, such as Figure 3d , Figure 6d , Figure 9dAs shown in Figure 2, similarly, the activated parts of its Grad-CAM++ graph are also scattered in various areas, that is, the weights of each part are similar, so it can indicate the state of a normal bearing.

[0119] Specifically, the auxiliary visualization method Figure 4 , Figure 7 and Fig.10 The fault diagnosis results were analyzed and the fault diagnosis accuracy was over 98.5%, which clearly shows the superiority of the method of the present invention. Figure 5 , Figure 8 and Fig.11 This is a t-SNE dimension reduction visualization diagram. Figure 5 For example, in category 3, some samples are classified in category 2 and category 4, and in category 4, a small number of samples are classified in category 2, so the diagnosis result does not reach 100%. However, according to the results of dimensionality reduction visualization, it can be seen that the rolling bearing fault data set of Case Western Reserve University is used for experiments. The diagnosis results and t-SNE analysis show that the clustering effect of the method of the present invention for the same category and the classification effect for different categories are excellent. The network of the present invention can cluster samples of the same category well, and can also classify different samples well.

[0120] Furthermore, it is compared with 1D-CNN, 1D-CNN-SVM and 1D-CNN-LSTM with the same structure and parameters as the model of the present invention, and the results are shown in Figures 6, Fig.12 As shown, the results show that under the same noise conditions, the diagnostic accuracy of the model of the present invention is higher, and when the noise gradually decreases, its diagnostic accuracy converges faster, verifying the excellent noise resistance of the 1D-CNN-BiLSTM interpretable neural network.

[0121] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. An interpretable rolling bearing fault diagnosis method, characterized in that: The following steps are involved: S1. Collecting one-dimensional time series signals of rolling bearings and performing sample expansion, including the following steps: S11, collecting one-dimensional time signals of four categories, including normal, outer ring fault, inner ring fault and rolling element fault, collected by the acceleration sensor; S12, expanding the one-dimensional time series signal by using overlapping sampling, and dividing the data samples into a training set and a test set; S13, dividing the data sample into different training sets and test sets multiple times by using a K-fold cross-validation method, and finally taking the average result to improve the robustness of the model; S2. Establish an initial 1D-CNN-BiLSTM neural network model; the structure of the initial 1D-CNN-BiLSTM neural network model includes an input layer, a convolutional neural network layer (CNN), a BiLSTM layer, a fully connected layer and an output layer: The CNN layer includes convolutional layer and pooling layer. The activation function of each convolutional layer is the Relu function. The pooling method of the pooling layer is maximum pooling; The activation function of the output layer is a Softmax function; S3, adding an interpretation layer to the neural network model, wherein the interpretation layer is Grad-CAM++, comprising the following steps: S31. Calculate the weights of each category of each channel after the last convolutional layer of the 1D-CNN-BiLSTM neural network model: Where Y x is the score value of category x, i.e., its importance. is the weight of the kth feature map for category x, is the pixel value at (i, j) of the kth feature map, is the weight of category x at (i, j) of the kth feature map; S32, generating a class activation map according to the weights: S33, transforming the size of the class activation map to the size of the original input image, and then overlaying it on the original image through a heat map, wherein the value of the heat map corresponds to the activation degree of the input signal by the 1D-CNN-BiLSTM neural network model; S4. Establish a neural network model with convolutional interpretation capabilities. The specific steps are as follows: S41, connect the interpretation layer and the BiLSTM layer after the 1D-CNN-BiLSTM neural network model; S42, the interpretation layer is used as the middle layer of the whole model to explain the convolution process; S43, using the BiLSTM layer as the back end of the entire model, and using the output of the last convolutional layer of CNN as the input of the back end to perform fault diagnosis; S5, using a variety of one-dimensional fault data to train the neural network model to obtain a fault diagnosis model; S6. Perform fault diagnosis on the rolling bearing using the fault diagnosis model.

2. The interpretable rolling bearing fault diagnosis method according to claim 1, characterized in that: The BiLSTM layer includes a bidirectional LSTM layer, and the composition of the LSTM layer is specifically as follows: f t =σ(W f [h t-1 ,x t ]+b f ) i t =σ(W i [h t-1 ,x t ]+b i ) the t =σ(W o [h t-1 ,x t ]+b o ) c t =f t c t-1 +i t tanh(W c [h t-1 ,x t ]+b c ) h t =oh t thank you t In the formula, f t For the forget gate, i t is the input gate, o t is the output gate, c t is the cell state, σ is the sigmoid activation function, W f , W i , W o , W c is the weight in the training parameters, b f , b i , b o , b c is the bias term in the training parameters, h t is the state of the hidden layer at the current moment, h t-1 is the state of the hidden layer at the previous moment, x t The input at the current moment.

3. The interpretable rolling bearing fault diagnosis method according to claim 1, characterized in that: The Softmax function is: In the formula, z i+1 is the value output by the i+1th point in Z, C is the number of output nodes, z c is the value output by the cth point in Z, and e is a natural constant.

4. An interpretable rolling bearing fault diagnosis method according to any one of claims 1 to 3, characterized in that: The specific steps of step S5 are: S51, classifying the samples collected in step S1 into four categories, namely normal, outer ring fault, inner ring fault and rolling element fault; S52, dividing the fault samples into a training set and a test set, and performing training and testing using a K-fold cross validation method; S53, using the training set as input of a neural network model with convolutional interpretation capability, performing model training, and obtaining a fault diagnosis model.