Multi-sensor fault diagnosis method based on graph regularization CNN-BiLSTM, medium and equipment

Abstract

The application discloses a kind of based on graph regularization CNN-BiLSTM multi-sensor fault diagnosis method, medium and equipment, it is related to the computer fault diagnosis system field based on specific calculation model.The method comprises the following steps: collecting the condition monitoring data of multi-sensor, and the data collected is preprocessed, to obtain training set, verification set and test set;Establish CNN-BiLSTM network, graph regularization item is added in the nearest fully connected layer of distance classifier of CNN-BiLSTM network, complete GR-CNN-BiLSTM model construction;The data in training set is used to train GR-CNN-BiLSTM model, the data in verification set is used to evaluate GR-CNN-BiLSTM model, and the network parameter when the performance of GR-CNN-BiLSTM model is optimal is obtained;The data of test set is input into the GR-CNN-BiLSTM model of optimal performance and carries out fault diagnosis, to obtain fault diagnosis result.The application improves training efficiency and diagnostic accuracy, overcome the low training efficiency of existing deep graph regularization fault diagnosis method.

Description

Technical Field

[0001] This invention relates to the field of computer fault diagnosis systems based on specific computational models, and particularly to a multi-sensor fault diagnosis method, medium, and device based on graph regularized CNN-BiLSTM. Background Technology

[0002] Fault diagnosis, as a key component of prognostics and health management (PHM), uses signal detection technology and data analysis to detect, isolate, and identify fault types, providing corresponding solutions and maintenance strategies for system repair and maintenance.

[0003] Fault diagnosis mainly consists of two parts: feature extraction and classification. Traditional fault diagnosis methods based on time-frequency signal processing rely on manual experience to extract features, which has limitations in the current context of massive, heterogeneous, rapidly generated, and sparsely valuable state monitoring big data. Deep learning, with its powerful ability to handle complex recognition tasks and feature extraction, has been widely researched and applied in equipment fault diagnosis in fields such as machinery manufacturing, aerospace, and energy and chemical engineering. Existing deep learning methods, such as convolutional neural networks (CNNs), can effectively avoid complex feature extraction and data reconstruction processes; bidirectional long short-term memory networks (BiLSTMs) process network inputs simultaneously in both directions, improving the model's prediction accuracy. Therefore, deep learning based on CNN and BiLSTM network models has been used for fault diagnosis, hereinafter referred to as: existing deep learning-based fault diagnosis methods.

[0004] Existing deep learning-based fault diagnosis methods rely on the sample independence assumption, neglecting the correlation information between samples. In reality, within a specific feature space, samples of the same class are similar and close together, while samples of different classes are significantly different and far apart. Therefore, to fully utilize the geometric structure information between samples to improve the performance of deep learning algorithms, manifold learning methods and spectral graph theory are introduced into traditional regularization methods, forming the existing deep graph regularization fault diagnosis method. This method allows data to maintain its local geometric structure in the original feature space within the new projection space; that is, data points are similar in their essential geometric structure and remain similar even when embedded or projected into the new space. However, the network nearest neighbor graph in this method is constructed in the original high-dimensional space, resulting in high computational cost, susceptibility to interference, and low training efficiency. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-sensor fault diagnosis method, medium, and device based on graph regularization CNN-BiLSTM, so as to improve the low training efficiency of existing deep graph regularization fault diagnosis methods.

[0006] To achieve the above objectives, the multi-sensor fault diagnosis method based on graph regularized CNN-BiLSTM of the present invention includes:

[0007] S1 collects status monitoring data from multiple sensors and preprocesses the collected data to obtain training set, validation set and test set;

[0008] S2, establish a CNN-BiLSTM network, add a graph regularization term to the fully connected layer of the CNN-BiLSTM network, the fully connected layer is the fully connected layer closest to the classifier, and complete the construction of the GR-CNN-BiLSTM model;

[0009] S3, train the GR-CNN-BiLSTM model using the data in the training set, evaluate the GR-CNN-BiLSTM model using the data in the validation set, and obtain the network parameters when the GR-CNN-BiLSTM model has the best performance;

[0010] S4. Input the test set data into the best-performing GR-CNN-BiLSTM model for fault diagnosis to obtain the fault diagnosis results.

[0011] Preferably, in step S1, the collected data is subjected to z-score standardization, min-max standardization, or normalization.

[0012] Preferably, step S2, establishing a CNN-BiLSTM network, specifically involves:

[0013] A1 inputs the training set into the CNN for network structure training, extracts deep feature data through convolutional layers, and then processes it through max pooling layers to obtain dimensionality-reduced deep feature data.

[0014] A2, input the dimensionality-reduced deep feature data into BiLSTM for network structure training, and then concatenate and fuse the output data obtained from BiLSTM training to obtain and learn the fused fault feature data.

[0015] A3: Input the fused fault feature data into the fully connected layer module, set a softmax classifier in the output layer of the fully connected layer module to classify the fault state and output it, update the network structure parameters, return to A1, and train the next batch of data until all the data in the training set are used for training as one iteration.

[0016] A4. After reaching the preset number of iterations, the CNN-BiLSTM network model is completed.

[0017] The softmax classifier uses the cross-entropy loss function, L. c for:

[0018]

[0019] In equation (1), r i The true label for data sample i in the training set; y i Let be the predicted label of data sample i in the training set; N is the total number of data samples in the training set.

[0020] Preferably, the CNN-BiLSTM network includes a fully connected layer module, which comprises a first fully connected layer FC1, a second fully connected layer FC2, and an output layer arranged sequentially.

[0021] CNN-BiLSTM maps the feature information obtained from training or testing sequentially through the first fully connected layer FC1 and the second fully connected layer FC2 to the output layer.

[0022] The computational equation for each layer in the fully connected layer module is as follows:

[0023] o a =g(W a h a +b a (2)

[0024] In equation (2), o a h is the output of the a-th fully connected layer; a The input to the a-th fully connected layer; b a W represents the bias weights of the a-th fully connected layer. a is the forget gate weight of the a-th fully connected layer; g() is the ReLU function of the first fully connected layer.

[0025] Preferably, in step S2, the graph regularization term is added to the fully connected layer of the CNN-BiLSTM network, wherein the fully connected layer is the one closest to the classifier, specifically:

[0026] Based on the training set data, a nearest neighbor graph is constructed, specifically by identifying the k nearest neighbors of any data sample i and establishing connections between the k nearest neighbors of the data sample i.

[0027] Based on the nearest neighbor graph, a nearest neighbor graph regularization term L is added to the feature space of the fully connected layer closest to the classifier in the CNN-BiLSTM network. g for:

[0028]

[0029] Where i represents any data sample in the training set, j represents the data sample in the training set that is the nearest neighbor of data sample i, N is the total number of data samples in the training set, and H ij s represents the connection weight between data sample i and data sample j in the nearest neighbor graph. i s represents the feature vector corresponding to data sample i. j H represents the feature vector corresponding to data sample j. ij The calculation formula is:

[0030]

[0031] In formula (4), σ represents the average distance between all nearest neighbors in the nearest neighbor graph, and the weight between non-nearest neighbors is 0.

[0032] Preferably, in step S3, the GR-CNN-BiLSTM model is trained using data from the training set, specifically as follows:

[0033] Based on the training set, a training sample set X and a corresponding label set y after feature engineering are obtained, and then input into the GR-CNN-BiLSTM model to obtain the network parameters of the GR-CNN-BiLSTM model;

[0034] Set the hyperparameters of the GR-CNN-BiLSTM model; the hyperparameters include: learning rate, maximum number of iterations, time of introduction of graph regularization term, number of data sample batches, graph regularization term coefficient, and number of nearest neighbors;

[0035] Initialize the network parameters of the GR-CNN-BiLSTM model;

[0036] Input the b-th batch of data samples from the training sample set X into the GR-CNN-BiLSTM model, where b≥1, to obtain the predicted labels. Then, perform calculations in the pre-training and tuning stages. Based on the calculation results, use the optimizer to update the network weights and biases; thus completing the training of the GR-CNN-BiLSTM model.

[0037] Where the iteration number e < e g During the pre-training phase, the objective function L is calculated using the following formula:

[0038] L = L c (5)

[0039] When e g <e<e max At this point, during the optimization phase, the objective function L is calculated using the following formula:

[0040] L = L c +γL g (6)

[0041] In formulas (5) and (6), L represents the objective function of the GR-CNN-BiLSTM model, characterizing the difference between the expected output and the actual output; L c L represents the cross-entropy loss function; g Let γ represent the nearest neighbor graph regularization term, γ represent the weight coefficient of the graph regularization term, and e represent the number of iterations. max Indicates the maximum number of iterations; e g This indicates the time when the graph regularization term is introduced.

[0042] More preferably, the optimizer uses the Adam algorithm.

[0043] Preferably, when evaluating the GR-CNN-BiLSTM model using data from the validation set, the GR-CNN-BiLSTM model exhibits optimal performance when the evaluation parameter accuracy and / or macro f1-score are at their highest.

[0044] The beneficial effects of this invention are:

[0045] This invention presents a multi-sensor fault diagnosis method based on graph regularized CNN-BiLSTM. By constructing graph regularization terms on low-dimensional features, the network learns geometric relationship features with better clustering effect and higher recognition, thereby improving training efficiency and diagnostic accuracy and overcoming the shortcomings of low training efficiency in existing deep graph regularized fault diagnosis methods. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating a multi-sensor fault diagnosis method based on graph regularization CNN-BiLSTM.

[0047] Figure 2 This is a schematic diagram of the CNN-BiLSTM network structure;

[0048] Figure 3 This is a schematic diagram of the fault diagnosis process based on the GR-CNN-BiLSTM model;

[0049] Figure 4 This is a schematic diagram illustrating the changes in training accuracy and loss of GR-CNN-BiLSTM over epochs.

[0050] Figure 5 This is a diagram illustrating the classification accuracy and training time on the test set with different E0 values;

[0051] Figure 6 This is a schematic diagram illustrating the classification accuracy of the test set with different k values;

[0052] Figure 7 This is a schematic diagram illustrating the classification accuracy of the test set with different γ values;

[0053] Figure 8 These are schematic diagrams of the confusion matrices of test results for different model algorithms. Among them, (a) represents the confusion matrix of the KNN test results, (b) represents the confusion matrix of the GB Bayes test results, (c) represents the confusion matrix of the MLP algorithm test results, (d) represents the confusion matrix of the CNN algorithm test results, (e) represents the confusion matrix of the CNN-BiLSTM network structure test results, and (f) represents the confusion matrix of the GR-CNN-BiLSTM model algorithm test results.

[0054] Figure 9 These are two-dimensional visualization diagrams of the fully connected FC2 layer features of different deep learning models. Among them, (a) represents the two-dimensional visualization diagram of the fully connected FC2 layer features of the convolutional CNN, (b) represents the two-dimensional visualization diagram of the fully connected FC2 layer features of the LSTM, (c) represents the two-dimensional visualization diagram of the fully connected FC2 layer features of the CNN-BiLSTM network structure, and (d) represents the two-dimensional visualization diagram of the fully connected FC2 layer features of the GR-CNN-BiLSTM model.

[0055] Figure 10 It is an LSTM cell structure. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0057] A brief introduction to the background of CNN and BiLSTM is as follows:

[0058] CNN-BiLSTM combines the advantages of CNN in extracting spatial features with BiLSTM in extracting temporal features. CNN can extract deep feature information from multi-sensor data and reduce data dimensionality while preserving key features. BiLSTM extracts information from temporal data that might be overlooked by CNN from two directions. BiLSTM processes the network input simultaneously in both forward and backward directions, capturing past and future information; this enhances the model's memory of the beginning and end stages of the original input.

[0059] (I) Convolutional Neural Networks (CNNs)

[0060] A CNN is a deep feedforward neural network that incorporates convolutional computations. It has the ability to extract and combine local features from lower layers into more abstract features at higher layers. Its advantages lie in its weight-sharing structure and translation invariance. Considering an L-layer CNN structure, the convolutional feature map of the l-th layer can be represented as:

[0061]

[0062]

[0063] in, Let q be the output of the e-th feature of the l-th convolutional layer, and let q represent the current number of output features. and These correspond to the weight matrix and bias of the l-th convolutional layer, respectively, where m represents the filter index and M is the number of outputs of the convolutional layer. Taking the rectified linear unit (ReLU) activation function as an example, it obtains non-linear features, enhances the model's feature representation ability, and accelerates the convergence process of the CNN. This is the output result.

[0064] Pooling functions use the overall statistical features of the neighboring outputs at a given location to replace the network's output at that location. Introducing pooling layers between consecutive convolutional layers reduces the feature space and network parameters after convolution through non-linear downsampling, while maintaining spatial invariance. To achieve fast convergence and good generalization, the multi-sensor fault diagnosis method based on graph regularization CNN-BiLSTM described in this application employs max-pooling, selecting the maximum value among all neurons within a region.

[0065]

[0066] in, This is the output of the max pooling layer, where T and R represent the pooling layer stride and the pooling layer kernel size, respectively, and μ represents the pooling window.

[0067] (II) Bidirectional Long Short-Term Memory Neural Network (BiLSTM)

[0068] Long Short-Term Memory (LSTM) networks are a variant of Recurrent Neural Networks (RNNs). By maintaining long short-term memories in the unit states, they overcome the vanishing and exploding gradient problems during long-term sequence training. LSTMs characterize dynamic features through time series cycles and build long-term correlation models. An LSTM unit model mainly consists of a forget gate, an input gate, and an output gate to control the information flow. The LSTM unit structure is as follows: Figure 10As shown. The biggest difference between LSTM and RNN is that the input gate can temporarily store relevant information.

[0069] To obtain more information before and after each time step in a given sequence, a BiLSTM with independent forward and reverse paths was constructed. Within a specific time step, it simultaneously captures information from the past and future, enhancing the model's memory of the beginning and end stages of the original input information. The forward LSTM can discover patterns of change, while the backward LSTM can mitigate the impact of noise and smooth the prediction results.

[0070] At time step t, x t h t These represent the input state and the hidden state, f. t i t o t c t These are the forget gate, input gate, output gate, and memory unit. The forget gate determines the amount of information that passes through, the input gate determines whether new information is memorized by the cell unit, and the final output of the LSTM unit is determined by the output of the output gate and the output of the memory unit. The forward path and backward path theories are the same; this application only uses the backward path as an example, and the update equation is as follows:

[0071]

[0072] In the formula, σ represents the sigmoid function, W, V, and b are the cyclic weights of the forget gate, input, and bias in the unit, respectively, and are the model parameters to be updated.

[0073] The forget gate and the input gate can determine the next cell state c. t+1 and candidate cell state Merge into the current unit state c t The output gate can control the current cell state c. t to hidden state h t The conversion.

[0074]

[0075] In the formula, tanh is the hyperbolic tangent function.

[0076] Based on the input sequence (x) t ,…,x t The forward hidden state sequence can be obtained. Similarly, the input time series (x) is processed in reverse order. t ,…,x t It can generate backward hidden state sequences. After connecting the forward and backward hidden states, the final output sequence y of the bidirectional LSTM is... t It can be represented as U indicates transpose.

[0077] Example

[0078] The multi-sensor fault diagnosis method based on graph regularized CNN-BiLSTM described in this embodiment is as follows: Figure 1 and Figure 3 As shown, the method includes:

[0079] S1: Collect status monitoring data from multiple sensors and preprocess the collected data to obtain training and test sets;

[0080] S2, establish a CNN-BiLSTM network, add a graph regularization term to the fully connected layer of the CNN-BiLSTM network, the fully connected layer is the fully connected layer closest to the classifier, and complete the construction of the GR-CNN-BiLSTM model;

[0081] S3, evaluate the GR-CNN-BiLSTM model using data from the test set, and obtain the parameters when the GR-CNN-BiLSTM model performs optimally;

[0082] S4. Input the test set data into the GR-CNN-BiLSTM model that has been trained in S3 to perform fault diagnosis and obtain the fault diagnosis results.

[0083] More detailed explanation:

[0084] (I) On the Construction of CNN-BiLSTM Network Structure

[0085] The main challenge in classifying different fault categories lies in the nonlinearity and data uncertainty related to system changes, specifically manifested as the correlation between different features and the time-varying nature of the same feature in the time domain. Therefore, this embodiment constructs a CNN-BiLSTM network structure to nonlinearly map high-dimensional data to a low-dimensional space, including a multi-layer CNN for extracting correlations and a BiLSTM for extracting temporal correlations. The CNN-BiLSTM network structure is as follows: Figure 2 As shown.

[0086] Specifically: First, a multi-scale one-dimensional CNN containing an input layer, an output layer, and multiple hidden layers is constructed. Data from the training set is input into the CNN convolutional layers to adaptively extract deep feature data. After processing by max pooling layers, dimensionality-reduced deep feature data is obtained while preserving the main feature information.

[0087] Then, the dimensionality-reduced deep feature data is used as input to the BiLSTM layer to train the network model and learn feature information. Specifically, the output data obtained from the BiLSTM training is concatenated and fused to obtain and learn the fused fault feature data. The fully connected layers FC1 and FC2 in the fully connected layer module of the CNN-BiLSTM network structure are used to map the hidden state features to the output network layer. The computational equation for each layer in the fully connected layer module is as follows:

[0088] o a =g(W a h a +b a (6)

[0089] In equation (2), o a h is the output of the a-th fully connected layer; a The input to the a-th fully connected layer; b a W represents the bias weights of the a-th fully connected layer. a is the forget gate weight of the a-th fully connected layer; g() is the ReLU function of the first fully connected layer.

[0090] Finally, the fused fault feature data is input into the fully connected layer module. A softmax classifier is set in the output layer of the fully connected layer module to classify and output the fault state. The optimizer is used to update the network structure parameters. Then, the next batch of data is trained until all the data in the training set are used for training, which is considered one iteration. After reaching the preset number of iterations, the construction of the CNN-BiLSTM network model is completed.

[0091] The SoftMax classifier is used for learning and diagnosing fault states. The objective function of the SoftMax classifier is the cross-entropy loss function.

[0092]

[0093] In equation (7), r i The true label for data sample i in the training set; y i Let be the predicted label of data sample i in the training set; N is the total number of data samples in the training set.

[0094] In the CNN-BiLSTM network structure, forward propagation achieves non-linear mapping layer by layer, while backpropagation updates the parameters of the CNN-BiLSTM network structure to achieve model learning. By iteratively reducing the loss function between the output and the ground truth, the network parameters of each fully connected layer converge to an acceptable range, thus completing the optimization of the CNN-BiLSTM network structure.

[0095] (II) Regarding the addition of graph regularization terms to construct the GR-CNN-BiLSTM model

[0096] Spectral clustering is a highly efficient clustering method. Its main idea is to treat all data as points in space, connected by edges to form an undirected weighted graph. The weights represent the distance between two points, i.e., a similarity matrix is ​​constructed to represent the relationships between data points. Then, the graph is segmented to obtain the final clustering result.

[0097] In this embodiment, by adding graph regularization constraints to the CNN-BiLSTM network structure, the local manifold structure of the data is guaranteed based on sample similarity. Unlike the traditional approach of designing graph regularization terms in the original high-dimensional data, this embodiment designs a nearest neighbor graph in the feature space. The graph regularization term is added to the second fully connected layer FC2, which is closest to the classifier. The features output by the second fully connected layer FC2 are directly input into the classifier, so the features of the second fully connected layer FC2 are the easiest to classify, and the feature dimension of the second fully connected layer FC2 is also the lowest. Applying the graph regularization term to the second fully connected layer FC2 can not only minimize the computational cost but also improve the accuracy of graph construction.

[0098] In the feature space, the adjacency matrix H represents the nearest neighbor relationship between two samples. If s j In s i Among the k-nearest neighbors, i.e. s j ∈kNN(s i If we connect data samples i and j in the nearest neighbor graph, we can construct a weighted undirected nearest neighbor graph. The connection weight H between data sample i and data sample j in the nearest neighbor graph is... ij The Gaussian kernel function is defined as follows:

[0099]

[0100] In formula (8), σ represents the average distance between all nearest neighbors in the nearest neighbor graph, and the weight between non-nearest neighbors is 0.

[0101] Based on the training set data, a nearest neighbor graph is constructed, specifically: k nearest neighbors of any data sample i are identified, and connections are established between these k nearest neighbors; based on the nearest neighbor graph, a nearest neighbor graph regularization term L is added to the feature space of the fully connected layer closest to the classifier in the CNN-BiLSTM network. g for:

[0102]

[0103] In formula (9), i represents any data sample in the training set, j represents the data sample in the training set that is the nearest neighbor of data sample i, N is the total number of data samples in the training set, and H ijs represents the connection weight between data sample i and data sample j in the nearest neighbor graph. i s represents the feature vector corresponding to data sample i. j This represents the feature vector corresponding to data sample j.

[0104] Using the backpropagation gradient descent algorithm to iteratively improve L g Minimize the distance between nearest neighbors so that the GR-CNN-BiLSTM model can learn features with good clustering effect and high recognition, thus achieving efficient fault classification.

[0105] (III) Training the GR-CNN-BiLSTM model using a two-stage training method

[0106] In this specific example, due to insufficient training of the CNN-BiLSTM network structure, it is difficult to distinguish high-order features, making it difficult to correctly construct the nearest neighbor graph. In addition, introducing the graph regularization term too early will increase the complexity and computation time of the GR-CNN-BiLSTM model. The training algorithm shown in Table 1 includes two stages of training the GR-CNN-BiLSTM model. The network characterization model's expected output and the objective function L of the actual output differ in each stage.

[0107] (1) Pre-training phase. During iteration e... <e g At that time, L only contains the loss function L of the classifier. c .

[0108] L = L c (10)

[0109] (2) Optimization phase. At iteration number e... g <e<e max At that time, the network objective function L also includes the classifier's loss function L. c And graph regularization term L g The classifier loss learns significant low-dimensional features of the data, while the graph regularization term makes the learned features more discriminative.

[0110] L = L c +γL g (11)

[0111] In formulas (10) and (11), L represents the objective function of the GR-CNN-BiLSTM model, characterizing the difference between the expected output and the actual output; L c L represents the objective function of the classifier, in this embodiment. c L is the cross-entropy loss function; g Let γ represent the nearest neighbor graph regularization term, γ represent the weight coefficient of the graph regularization term, and e represent the number of iterations. max Indicates the maximum number of iterations; eg The timing of introducing graph regularization terms needs to be determined and adjusted manually.

[0112] The deep learning model is considered to have optimal performance when any one or more of its evaluation parameters—Accuracy, Precision, Recall, and macro f1-score—are optimal. In this embodiment, when evaluating the GR-CNN-BiLSTM model using data from the validation set, the GR-CNN-BiLSTM model is considered to have optimal performance when the accuracy and / or macro f1-score of the evaluation parameters are at their highest.

[0113] Table 1 Two-stage training algorithm for GR-CNN-BiLSTM model

[0114]

[0115] The model parameter set of the trained GR-CNN-BiLSTM model includes the network weights and biases of each layer.

[0116] To better illustrate the effectiveness and beneficial effects of the multi-sensor fault diagnosis method based on graph regularized CNN-BiLSTM claimed in this embodiment, the effectiveness of the method described in this embodiment is verified by specific experiments and model comparisons.

[0117] Experimental verification and model comparison

[0118] 1.1 Description of the ultrasonic motor dataset

[0119] Experiments were conducted to verify the effectiveness of the multi-sensor fault diagnosis method based on graph regularization CNN-BiLSTM described in this embodiment.

[0120] The dataset includes five signals—driving voltage, driving current, driving frequency, lone electrode feedback voltage, and internal temperature—for an ultrasonic motor under normal conditions and four fault states: cracked piezoelectric ceramic sheet, worn friction sheet, loose adhesive layer, and broken elastomer teeth. Training and test sets are constructed based on the collected multi-sensor data, ensuring complete independence between the two sets, each containing different fault modes. Each sample in both sets contains data from the five signal channels of the sensors, with 18,652 samples in each set. The goal is to train the GR-CNN-BiLSTM model using the training set and then validate its performance using the test set.

[0121] Table 2. Description of normal and fault datasets for 8 motors

[0122]

[0123] The raw data collected by multiple sensors is difficult to use directly for fault diagnosis due to its high sampling frequency and high noise. Therefore, z-score is used to normalize the raw data collected by multiple sensors to eliminate the scaling effect and ensure effective comparison between different fault characteristics.

[0124]

[0125] In equation (12), x i The original data, The characteristic mean, σ x The characteristic standard deviation, x i_norm Z-score represents the deviation from the mean in units of standard deviation.

[0126] 1.2 Model Parameter Settings and Sensitivity Analysis

[0127] The preprocessed 5D multi-sensor signals were input into the GR-CNN-BiLSTM model. The network hyperparameters were set as follows: CNN batch size of 149, number of kernels of 32, kernel size of 2, stride of 2, and pooling window size of 2; the number of units in the hidden layers of the BiLSTM was set to 128, and the number of units in the fully connected layers FC1 and FC2 were set to 64 and 32 respectively; the classifier output classes were 5. During the training of the GR-CNN-BiLSTM model, the Adam algorithm was used, with an initial learning rate of 0.001 and weight decay set to e^(-1 / 2). -5 One-hot encoding is used to encode the classification results. Therefore, the outputs for no fault, cracked piezoelectric ceramic sheet, worn friction sheet, loose adhesive layer, and broken elastomer tooth state are [1,0,0,0,0], [0,1,0,0,0], [0,0,1,0,0], [0,0,0,1,0], and [0,0,0,0,1], respectively. The impact of the maximum number of iterations (epoch) parameter on the accuracy and loss of the training set in the GR-CNN-BiLSTM model is as follows: Figure 4 As shown, when the maximum number of iterations (epochs) is 120, the training accuracy and loss tend to stabilize. Overtraining can lead to model overfitting. Therefore, the maximum number of iterations (epochs) is set to 120 to iteratively optimize and update the model parameters, minimizing the loss function.

[0128] The nearest neighbor graph regularization term was introduced at time E0, set to 60, with the number of nearest neighbors k set to 3, and the graph regularization coefficient γ set to 0.01. Next, the sensitivity of key parameters affecting network performance, such as E0, k, and γ, was analyzed using hierarchical five-fold cross-validation. To save training time, the sample size in the following analysis was one-tenth of the original sample size, i.e., 1865 samples.

[0129] 1.2.1 Analysis on Determining the Timing of the Introduction of the Graph Regularization Term

[0130] When γ = 0.01 and k = 3, under different E0 conditions, the mean diagnostic accuracy and training time of graph-regularized CNN-BiLSTM on the test set are as follows: Figure 5 As shown, it can be observed that as E0 increases, the training time gradually shortens, and the classification accuracy first increases and then decreases, reaching its maximum when E0 is 60. This is because the original samples exhibit severe aliasing in the data space, making them difficult to distinguish. Introducing graph regularization too early results in poor accuracy of the constructed nearest neighbor graph, causing the model to learn incorrect geometric relationship features. Conversely, introducing graph regularization too late leads to poor parameter optimization due to approaching the iteration termination number. To achieve a balance between accuracy and training efficiency, an E0 value of 60 was chosen.

[0131] 1.2.2 Analysis of Determining the Number of Nearest Neighbors

[0132] If the number of nearest neighbors, k, is too large, incorrect nearest neighbors may be selected, affecting the accuracy of the nearest neighbor graph construction. Since the training batch size is 149, with an average of 30 samples per class, the value of k should be less than 30. Figure 6 The figure shows the mean diagnostic accuracy and training time of GR-CNN-BiLSTM on the test set under different k conditions when E0=60 and γ=0.01. It can be found that as k increases, the classification accuracy first improves, reaching a maximum at k=3, and then gradually decreases. Therefore, a value of k=3 is selected.

[0133] 1.2.3 Analysis of the Determination of Graph Regularization Coefficient

[0134] If the graph regularization coefficient γ is too large, it may affect the learning process of the classifier; if it is too small, it will not be able to learn the geometric structure information between samples to improve model performance. When E0 = 60 and k = 3, the mean diagnostic accuracy of the GR-CNN-BiLSTM model on the test set under different γ conditions is as follows: Figure 7 As shown, it can be observed that as γ increases, the classification accuracy first improves, reaching its maximum value when k is 0.01, and then gradually decreases. Therefore, a value of γ of 0.01 was selected.

[0135] 1.3 Model Evaluation

[0136] Based on the aforementioned dataset, the performance of the multi-sensor fault diagnosis method based on the GR-CNN-BiLSTM model described in this embodiment is compared with that of several commonly used fault diagnosis methods. In this embodiment, accuracy, precision, recall, and F1-score were used to evaluate model performance. In addition to commonly used metrics representing the classification accuracy of each fault type, such as average classification accuracy, precision, and recall, the macro F1-score, which is unaffected by data imbalance, was selected to compare the multi-classification performance of different models. The macro F1-score is calculated by averaging the F1-scores (see the formula below). The performance of the multi-sensor fault diagnosis method based on the GR-CNN-BiLSTM model was evaluated, and the performance of the method described in this embodiment was compared with that of commonly used fault diagnosis algorithms.

[0137]

[0138] In formula (13), v represents the v-th type of fault. In this experiment, v = 1, 2, 3, 4, 5, which represent the normal state of the ultrasonic motor, the cracking fault of the electro-ceramic sheet, the wear fault of the friction sheet, the loosening fault of the adhesive layer, and the broken tooth fault of the elastomer.

[0139] The graph regularization term in traditional deep graph regularization networks is constructed in the original data space, while the graph regularization term in the multi-sensor fault diagnosis method based on the GR-CNN-BiLSTM model described in this embodiment is constructed in the feature space. Since the dimension (64) of the feature space of the second fully connected layer FC2 is much lower than the dimension of the original training dataset (74610 = 14922 * 5), the GR-CNN-BiLSTM model described in this embodiment obviously has higher training efficiency.

[0140] 1.3.1 Comparison of diagnostic results with commonly used fault diagnosis models

[0141] To verify the diagnostic performance of the multi-sensor fault diagnosis method based on the GR-CNN-BiLSTM model described in this embodiment, Table 3 shows the comparison results of the GR-CNN-BiLSTM model with six other commonly used models. It can be found that the diagnostic accuracies of k-nearest neighbors (KNN), Gaussian naive Bayes (GB Bayes), and multilayer perceptron (MLP) learning algorithms are 0.8779, 0.8259, and 0.8852, respectively, which are inferior to the GR-CNN-BiLSTM model in this embodiment. These three learning algorithms suffer from problems such as reliance on manual feature extraction, overfitting, and parameter selection issues. Compared with shallow machine learning methods, the GR-CNN-BiLSTM model in this embodiment significantly improves fault diagnosis performance due to its powerful feature extraction capabilities, confirming the necessity of establishing a deep network structure. Furthermore, the three deep learning models in this embodiment—GR-CNN-BiLSTM, CNN, and CNN-BiLSTM—were compared. With identical parameter settings, the CNN and CNN-BiLSTM models achieved accuracies of 0.9158 and 0.9389, respectively, while the GR-CNN-BiLSTM model achieved an accuracy of 0.9491. The GR-CNN-BiLSTM model also performed best in terms of precision, recall, and macro f1-score. This is because CNN can learn the short-term correlations of multiple signals at each step, while BiLSTM can learn the long-term correlations of sequences in both forward and backward directions. By simultaneously learning spatial and temporal features, the macro f1-score of CNN-BiLSTM was improved by 3.75% compared to CNN; by introducing a graph regularization term, the macro f1-score of GR-CNN-BiLSTM was improved by 1.47% compared to CNN-BiLSTM. Figure 8 The confusion matrix of the test results of different models is given. It can be found that the GR-CNN-BiLSTM model has the best classification effect. Except for some misclassification of fault 1 and fault 4, it can almost completely and correctly classify the fault types in the test samples.

[0142] Table 3 Evaluation results based on different fault diagnosis classification models

[0143]

[0144] To intuitively verify the effectiveness of the multi-sensor fault diagnosis method based on GR-CNN-BiLSTM described in this embodiment, the t-SNE algorithm is used to visualize the features of the fully connected FC2 layer of different deep learning models in two dimensions, such as... Figure 9 As shown, the GR-CNN-BiLSTM model can distinguish the five states, including normal and different fault types, quite well, while the other three deep learning models all exhibit overlapping boundaries in the data distribution.

[0145] 1.3.2 Comparison of diagnostic results between multi-sensor data and single-sensor data

[0146] To verify the necessity of fusing multi-sensor data in fault diagnosis methods, the impact of input samples from multi-sensor signals and single-sensor signals on the diagnostic results was analyzed. Fault diagnosis was performed using each of the five signals individually, and the results were compared with those using a combination of all five signals. The results show that the model parameters of GR-CNN-BiLSTM remained consistent throughout.

[0147] Table 4 Comparison of fault diagnosis results based on multi-sensor and single-sensor data

[0148]

[0149] Table 4 shows the comparison results. It can be seen that a satisfactory diagnostic result cannot be achieved based on any single signal, while combining multi-sensor signals can improve the diagnostic effect. When using all five signals, all evaluation indicators reached their optimal levels. This result verifies that fault diagnosis based on single-channel sensor data is insufficient to fully reflect the health status of the system, while fusing multi-sensor data and expanding the feature space can significantly improve the fault diagnosis effect.

[0150] By adopting the above-disclosed technical solution of this invention, the following beneficial effects are obtained: The multi-sensor fault diagnosis method based on GR-CNN-BiLSTM described in this application can simultaneously possess the ability of CNN to extract spatial correlation of multi-dimensional features and BiLSTM to extract temporal feature information under the condition of multi-sensor state monitoring data; it also adopts the nearest neighbor graph regularization method to mine data geometric relationship features in low-dimensional features and constructs a graph regularization deep learning framework; furthermore, a two-stage model training algorithm is designed to achieve efficient fault identification and classification on the basis of balancing training time and diagnostic accuracy. The main conclusions drawn from the experimental verification and analysis are as follows: (1) The classification performance of the GR-CNN-BiLSTM model on which the method described in this application depends is better than the shallow machine learning methods kNN, Bayes, and MLP, as well as the deep learning methods CNN and CNN-BiLSTM; (2) Compared with the input samples of single sensor signals, the method described in this application can improve the accuracy of the fault diagnosis model by fusing multi-sensor data; (3) The method described in this application overcomes the disadvantage of low training efficiency of traditional deep graph regularization networks. By constructing graph regularization terms on low-dimensional features, the network learns geometric relationship features with better clustering effect and higher recognition, thereby improving training efficiency and diagnostic accuracy.

[0151] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multi-sensor fault diagnosis method based on graph regularized CNN-BiLSTM, characterized in that, The method includes: S1 collects status monitoring data from multiple sensors and preprocesses the collected data to obtain training set, validation set and test set; S2, establish a CNN-BiLSTM network, add a graph regularization term to the fully connected layer of the CNN-BiLSTM network, the fully connected layer is the fully connected layer closest to the classifier, and complete the construction of the GR-CNN-BiLSTM model; S3, train the GR-CNN-BiLSTM model using the data in the training set, evaluate the GR-CNN-BiLSTM model using the data in the validation set, and obtain the network parameters when the GR-CNN-BiLSTM model has the best performance; S4. Input the test set data into the best-performing GR-CNN-BiLSTM model for fault diagnosis to obtain the fault diagnosis results. Step S2, establish the CNN-BiLSTM network, specifically as follows: A1 inputs the training set into the CNN for network structure training, extracts deep feature data through convolutional layers, and then processes it through max pooling layers to obtain dimensionality-reduced deep feature data. A2, input the dimensionality-reduced deep feature data into BiLSTM for network structure training, and then concatenate and fuse the output data obtained from BiLSTM training to obtain and learn the fused fault feature data. A3: Input the fused fault feature data into the fully connected layer module, set a softmax classifier in the output layer of the fully connected layer module to classify the fault state and output it, update the network structure parameters, return to A1, and train the next batch of data until all the data in the training set are used for training as one iteration. A4. After reaching the preset number of iterations, the CNN-BiLSTM network model is completed. The softmax classifier uses the cross-entropy loss function, L. c for: In equation (1), r i Let be the true label of data sample i in the training set; yi be the predicted label of data sample i in the training set; N be the total number of data samples in the training set; The CNN-BiLSTM network is equipped with a fully connected layer module, which includes a first fully connected layer FC1, a second fully connected layer FC2, and an output layer arranged sequentially. CNN-BiLSTM maps the feature information obtained from training or testing sequentially through the first fully connected layer FC1 and the second fully connected layer FC2 to the output layer. The computational equation for each layer in the fully connected layer module is as follows: o a =g(W a h a +b a ) (2) In equation (2), o a h is the output of the a-th fully connected layer; a The input to the a-th fully connected layer; b a W represents the bias weights of the a-th fully connected layer. a g is the forget gate weight of the a-th fully connected layer; g() is the ReLU function of the first fully connected layer; In step S2, a graph regularization term is added to the fully connected layer of the CNN-BiLSTM network. This fully connected layer is the one closest to the classifier. Specifically: Based on the training set data, a nearest neighbor graph is constructed, specifically by identifying the k nearest neighbors of any data sample i and establishing connections between the k nearest neighbors of the data sample i. Based on the nearest neighbor graph, a nearest neighbor graph regularization term L is added to the feature space of the fully connected layer closest to the classifier in the CNN-BiLSTM network. g for: Where i represents any data sample in the training set, j represents the data sample in the training set that is the nearest neighbor of data sample i, N is the total number of data samples in the training set, and H ij s represents the connection weight between data sample i and data sample j in the nearest neighbor graph. i s represents the feature vector corresponding to data sample i. j H represents the feature vector corresponding to data sample j. ij The calculation formula is: In formula (4), σ represents the average distance between all nearest neighbors in the nearest neighbor graph, and the weight between non-nearest neighbors is 0.

2. The multi-sensor fault diagnosis method based on graph regularized CNN-BiLSTM according to claim 1, characterized in that, In step S1, the collected data is subjected to z-score standardization, min-max standardization, or normalization.

3. The multi-sensor fault diagnosis method based on graph regularized CNN-BiLSTM according to claim 1, characterized in that, Step S3 involves training the GR-CNN-BiLSTM model using data from the training set, specifically as follows: Based on the training set, a training sample set X and a corresponding label set y after feature engineering are obtained, and then input into the GR-CNN-BiLSTM model to obtain the network parameters of the GR-CNN-BiLSTM model; Set the hyperparameters of the GR-CNN-BiLSTM model; the hyperparameters include: learning rate, maximum number of iterations, time of introduction of graph regularization term, number of data sample batches, graph regularization term coefficient, and number of nearest neighbors; Initialize the network parameters of the GR-CNN-BiLSTM model; Input the b-th batch of data samples from the training sample set X into the GR-CNN-BiLSTM model, where b≥1, to obtain the predicted labels. Then, perform calculations in the pre-training and tuning stages. Based on the calculation results, use the optimizer to update the network weights and biases; thus completing the training of the GR-CNN-BiLSTM model. Where, when the iteration number e <e g During the pre-training phase, the objective function L is calculated using the following formula: L=L c (5) When e g <e<e max At this point, during the optimization phase, the objective function L is calculated using the following formula: L=L c +γL g (6) In formulas (5) and (6), L represents the objective function of the GR-CNN-BiLSTM model, characterizing the difference between the expected output and the actual output; L c L represents the cross-entropy loss function; g Let γ represent the nearest neighbor graph regularization term, γ represent the weight coefficient of the graph regularization term, and e represent the number of iterations. max Indicates the maximum number of iterations; e g This indicates the time when the graph regularization term is introduced.

4. The multi-sensor fault diagnosis method based on graph regularized CNN-BiLSTM according to claim 3, characterized in that, The optimizer used is the Adam algorithm.

5. The multi-sensor fault diagnosis method based on graph regularized CNN-BiLSTM according to claim 1, characterized in that, When evaluating the GR-CNN-BiLSTM model using data from the validation set, the GR-CNN-BiLSTM model exhibits optimal performance when the evaluation parameter accuracy and / or macrof1-score are at their highest.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the multi-sensor fault diagnosis method based on graph regularized CNN-BiLSTM as described in any one of claims 1-5.

7. An electronic device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the multi-sensor fault diagnosis method based on graph regularization CNN-BiLSTM as described in any one of claims 1-5.

Images