A device fault diagnosis method based on an improved end-to-end ResNet-BiLSTM dual-channel model

Abstract

The present invention discloses a device fault diagnosis method based on an improved end-to-end ResNet-BiLSTM dual-channel model, specifically comprising the following steps: S1. using an acceleration sensor to collect original one-dimensional time-series fault data, which is simply normalized and used as the model input; S2. constructing an improved end-to-end ResNet-BiLSTM dual-channel fault diagnosis model, and inputting the normalized signal into the two channels of the ResNet model and the BiLSTM model; S3. constructing a 1DECANet module and connecting it in series with the ResNet model channel; S4. using a concatenate mechanism to fuse the features extracted from the two channels, and using a softmax classifier to output the diagnosis result of the device fault. The present invention relates to the construction of dual-channel models in deep learning and the technical field of industrial equipment fault diagnosis. The device fault diagnosis method based on the end-to-end dual-channel model constructs an improved end-to-end ResNet-BiLSTM dual-channel fault diagnosis model by addressing the problems of imbalance in fault diagnosis data and difficulty in extracting latent features in the industrial Internet field, effectively improving the efficiency of mechanical equipment fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical fields of dual-channel model construction in deep learning and industrial equipment fault diagnosis, and specifically to an equipment fault diagnosis method based on an improved end-to-end ResNet-BiLSTM dual-channel model. Background Art

[0002] With the rapid development of modern industry, the mechanical equipment in "smart factories" is also rapidly becoming more integrated and complex. Mechanical equipment is an indispensable component of the industrial sector. With its continuous operation, various failures are inevitable. The causes of component failures are often complex and diverse. Condition monitoring and fault diagnosis of mechanical equipment are key components of industrial fault diagnosis technology. If component failures cannot be addressed promptly and effectively, the equipment will cease to function, impacting industrial production. Therefore, mechanical equipment fault diagnosis is crucial for improving industrial production efficiency and economic benefits.

[0003] Research on mechanical equipment fault diagnosis mainly focuses on two parts: feature value extraction and state identification, which are the core parts of mechanical equipment state monitoring. During long-term operation, various sensors are connected to mechanical equipment to monitor the operating status of mechanical equipment and collect operating parameters. Among them, vibration acceleration signals are the main objects of collection and research. Deep neural networks are an effective model for solving complex system diagnosis problems. They can directly model highly nonlinear, complex, and multidimensional vibration acceleration signal data and mine the mapping relationship between data and diagnostic targets. The article [Zhang Hongbin et al., Bearing Fault Diagnosis Method Using Multi-channel Samples and Deep Convolutional Neural Networks [J]. Journal of Xi'an Jiaotong University, 2020, 54(8): 58-66.] uses continuous wavelet transform to extract the time-frequency domain features of the vibration signals of the bearings at both ends of the rotor, constructs single-channel two-dimensional graphic samples of three types of vibration signals and fuses them. The samples are input into the convolutional neural network CNN to achieve feature extraction and accurate classification of bearing fault signals. With the significant breakthroughs made by feedback neural networks in the field of natural language processing (NLP), the article [Fan Yuxue et al. Research on small sample rolling bearing fault diagnosis method based on BI-LSTM [J]. Noise and Vibration Control, 2020 (4): 108-113.] decomposes the signal by adaptive white noise modal decomposition (Complete EEMD with Adaptive Noise, CEEMDAN) and Fourier transform, and then inputs it into a bidirectional long short-term memory BiLSTM neural network to realize the fault diagnosis of high-speed train wheel bearings. The article [Tan et al. Rolling bearing fault diagnosis based on single gated unite recurrent neural networks [J]. Journal of Physics Conference Series, 2020, 1601: 042017.] designs a new single gated unite (SGD) recurrent neural network as a simplified strategy of LSTM neural network, uses wavelet packet decomposition to extract features as the input of bidirectional single gated unite to realize bearing fault diagnosis, reducing the network parameters by 36% and improving the time efficiency of fault diagnosis. The article [Chen Wei et al. Rolling bearing fault identification based on RS-LSTM [J]. Chinese Science and Technology Papers, 2018, 13(10): 1134-1141.] In order to realize the intelligent classification of vibration acceleration signals of different fault types of rolling bearings, random search (RS) is combined with LSTM neural network, and the original fault data is used as input to verify that the proposed algorithm has high generalization ability and robustness.

[0004] The aforementioned fault diagnosis model, based on signal processing and deep learning, enables fault diagnosis and condition monitoring for mechanical equipment such as bearings and gears in the industrial sector. Different diagnostic algorithms have their own advantages and disadvantages for different scenarios and feature data. Key challenges remain: integrating the strengths of each algorithm while minimizing human intervention. Building an end-to-end diagnostic model that extracts and deeply mines features from massive amounts of feature data, thereby improving the speed and accuracy of fault diagnosis for industrial equipment, remain unresolved. Summary of the Invention

[0005] (1) Technical problems solved

[0006] To address the shortcomings of existing technologies, the present invention provides a device fault diagnosis method based on an improved end-to-end ResNet-BiLSTM dual-channel model. First, the original one-dimensional time series signal is simply normalized and directly input into the stacked BiLSTM model and the ResNet model. This method fully extracts the signal's temporal correlation characteristics and the deep implicit features of the unbalanced data, respectively. To achieve cross-channel interaction between deep and shallow data features in a deep neural network, a 1DECANet module is constructed and integrated into the ResNet feature extraction channel. Finally, the features extracted from the two channels are input into a fully connected layer, and the softmax classifier is used to accurately identify device faults.

[0007] (2) Technical solution

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: a device fault diagnosis method based on an improved end-to-end ResNet-BiLSTM dual-channel model, specifically comprising the following steps:

[0009] S1. Use an accelerometer to collect the original one-dimensional time series data of vibration acceleration, and use it as the input of the end-to-end diagnosis model after simple normalization processing;

[0010] S2. Construct a ResNet model channel to deeply mine and extract the implicit and non-obvious features of the original signal to avoid the gradient diffusion phenomenon of deep neural networks;

[0011] S3. Build the 1DECANet module and connect it in series with the ResNet model channel to perform cross-channel interaction on deep and shallow layer data features as the improved ResNet feature extraction channel;

[0012] S4. Build a stacked BiLSTM model channel to extract the temporal correlation features of the original fault data, and add a batch normalization (BN) layer and a Dropout layer to prevent overfitting of the model channel.

[0013] S5. Build an end-to-end dual-channel diagnostic model and use the concatenate mechanism to fuse the data features extracted from the improved ResNet model channel and the stacked BiLSTM model channel in the fully connected layer;

[0014] S6. Use the Focal Loss function to coordinate the imbalance of training sample data, mine the hidden features of difficult-to-classify samples, and implement parameter updates for dual-channel diagnostic model training.

[0015] S7. Use the Softmax function to classify various fault signal data to achieve accurate diagnosis of unbalanced fault signals.

[0016] Preferably, in steps S2 and S3, an improved ResNet feature extraction channel is constructed to achieve cross-channel interaction of deep and shallow feature data. The specific steps are as follows:

[0017] T1. Construct a one-dimensional ResNet neural network channel. That is, superimpose an identity mapping layer on the shallow network to extract local features of the normalized time series signal along the time axis, and perform convolution layer and identity mapping calculation according to the formula;

[0018] y lay =F(x lay ,W i )+W s x lay

[0019] F=W2σ(W,x lay )

[0020] x lay+1 =f(y lay )

[0021] where x lay represents the input of the residual module, x lay+1 represents the output of the residual module, W, W i , W2 both represent conventional one-dimensional convolution calculations, W s represents 1*1 convolution calculation, σ represents Sigmoid activation function, and f(·) represents ReLU activation function;

[0022] T2. Construct the 1DECANet module to perform deep, cross-channel mining and extraction of hidden faults in the original fault signal through a local cross-channel interaction strategy without dimensionality reduction;

[0023]

[0024]

[0025] Where C represents the number of feature data channels in the middle layer of the network, k represents the size of the one-dimensional convolution kernel calculated by the adaptive adjustment function, γ = 2, b = 1;

[0026] T3. The constructed ResNet module is fused in series with the 1DECANet module to form an improved ResNet feature extraction channel. The module is then fused with the BiLSTM feature extraction channel using the Concatenate mechanism to provide classification feature parameter basis for the end-to-end diagnostic model.

[0027] Preferably, in step S6, a Focal Loss function is used to coordinate the imbalance of fault sample data, specifically:

[0028] E1. First, add a modulation coefficient ω>0 to the standard cross entropy loss function to reduce the loss weight of easy-to-classify samples and give more weight to difficult and misclassified sample data;

[0029]

[0030]

[0031] Where L0 represents the standard cross entropy loss function, L1 represents the cross entropy loss function after adding the modulation coefficient, X represents the fault sample, n represents the total number of samples, a represents the expected output, and y represents the actual output of the neuron;

[0032] E2. Then, a balance factor λ is added to balance the uneven ratio of positive and negative samples to form the Focal Loss function.

[0033]

[0034] Preferably, in step S1, an acceleration sensor is used to collect the original one-dimensional time series data of vibration acceleration, and after simple normalization processing, the data is used as the input of the end-to-end diagnostic model, which specifically includes the following steps:

[0035] P1. Use vibration acceleration sensors to collect fault time series data under different working conditions, and cut the collected time series data according to the sliding window size to generate fault diagnosis sample data As shown in the following formula:

[0036]

[0037] The sliding window length is l sw , where nl sw ×i≥l sw ; i=1,2,3..., the sample sequence of each time point is expressed as in are all d-dimensional vectors;

[0038] P2. Perform simple normalization on the original time-series vibration acceleration sample data and directly use it as the input data of the end-to-end fault diagnosis model X inputs =[X (1) ,X (2) ,...,X (n-1) ,X (n) ], as shown below:

[0039]

[0040] Preferably, in step S5, an end-to-end dual-channel diagnostic model is built, and the data features extracted from the improved ResNet model channel and the stacked BiLSTM model channel are fused using a concatenate mechanism in the fully connected layer, specifically:

[0041] A1 and Channel1 are set as the improved ResNet neural network model, which is composed of a ResNet module and a 1DECNet module in series. The output of the ResNet module serves as the input of the 1DECANet module.

[0042] A2 and Channel2 are stacked bidirectional long short-term memory (Stacked Bi-LSTM) neural network models. By adjusting the number of memory cells and the number of neural network layers in the BiLSTM network, we can extract signal features of different dimensions.

[0043] The two feature extraction channels A3, Channel1 and Channel2 are expanded by the Flatten mechanism and then fused by the Concatenate mechanism.

[0044] (3) Beneficial effects

[0045] This invention provides a device fault diagnosis method based on an improved end-to-end ResNet-BiLSTM dual-channel model. Compared with the existing technology, it has the following advantages:

[0046] (1) The device fault diagnosis method of the end-to-end dual-channel model addresses the existing problems of imbalanced fault diagnosis sample data, time series characteristics, and difficulty in extracting hidden faults in the industrial Internet field. The present invention improves the three processes of end-to-end model construction, data feature extraction and mining, and model parameter updating based on ResNet neural network and BiLSTM neural network. The improved deep learning algorithm of the present invention is called a fault diagnosis method based on an improved end-to-end ResNet-BiLSTM dual-channel diagnosis model. Compared with traditional fault diagnosis methods, this method not only adds an attention mechanism to the feature extraction channel, but also uses the Focal Loss loss function to update the dual-channel diagnosis model parameters to cope with unbalanced sample data, thereby solving the problem of mechanical equipment fault diagnosis in the industrial field.

[0047] (2) Compared with the traditional diagnosis model based on signal processing and deep learning, the equipment fault diagnosis method of the end-to-end dual-channel model: 1) The original vibration acceleration data is only simply normalized and directly input into the end-to-end diagnosis model, which maximizes the retention of the original characteristics of the time series data and avoids the subsequent chain reaction caused by the poor effect of signal preprocessing technology; 2) An improved end-to-end ResNet-BiLSTM dual-channel model fault diagnosis method is proposed to fully mine the implicit and difficult-to-identify features in unbalanced data and the contextual information of time series data; 3) Dropout layers and batch normalization layers are added to the dual-channel model to standardize feature data and avoid overfitting of the model; 4) The channel attention mechanism is introduced, and the 1DECNet module is constructed to embed the residual network channel to avoid the influence of the dimensionality reduction of the intermediate layer data on feature extraction, and at the same time realize the cross-channel interaction of deep and shallow layer data features in the deep neural network.

[0048] (3) The equipment fault diagnosis method of the end-to-end dual-channel model was applied to the faulty bearing dataset of Case Western Reserve University in the United States. Through experimental analysis, the effectiveness of the proposed improved end-to-end ResNet-BiLSTM dual-channel model fault diagnosis method was verified. The diagnostic accuracy on 11 different imbalance ratio test sets reached a minimum of 97.67% and a maximum diagnostic accuracy of 99.52%. Compared with the ResNet-BiLSTM dual-channel model without the 1DECANet module, the model convergence speed was significantly improved, and the minimum model loss value was reduced to 0.0711, which can achieve accurate fault diagnosis of bearing components of mechanical equipment in the industrial field. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a framework diagram of the end-to-end neural network model of the present invention;

[0050] Figure 2This is a schematic diagram of the residual module ResNet of the present invention;

[0051] Figure 3 1DECANet schematic diagram of the present invention;

[0052] Figure 4 This is the improved ResNet feature extraction channel framework diagram of the present invention;

[0053] Figure 5 This is a diagram of the stacked BiLSTM feature extraction channel framework of the present invention;

[0054] Figure 6 This is the overall framework diagram of the improved end-to-end ResNet-BiLSTM dual-channel model of the present invention;

[0055] Figure 7 This is the dual-channel model training and diagnosis flow chart proposed by the present invention;

[0056] Figure 8 This is a visualization diagram of the vibration acceleration signals of the normal and nine types of faults of the rolling bearing of the present invention;

[0057] Figure 9 This is a visualization diagram of the input layer of the dual-channel model proposed in this invention;

[0058] Figure 10 This is the visualization of the fusion layer T-SNE of the present invention;

[0059] Figure 11 This is a curve diagram of the diagnostic accuracy of the dual-channel model proposed in the present invention;

[0060] Figure 12 A graph showing a decrease in loss value according to the present invention;

[0061] Figure 13 This is the confusion matrix diagram of the 9 types of fault diagnosis classification results of the present invention;

[0062] Figure 14 Comparison of model training convergence before and after the incorporation of 1DECANet into the present invention. DETAILED DESCRIPTION

[0063] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only 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 making creative efforts are within the scope of protection of the present invention.

[0064] See also Figure 1-14The embodiment of the present invention provides a technical solution: a device fault diagnosis method based on an improved end-to-end ResNet-BiLSTM dual-channel model, with industrial machinery equipment fault diagnosis as the carrier, and ResNet and BiLSTM algorithms as the main algorithm framework. Figure 1 As shown, it specifically includes the following steps:

[0065] S1. Use an accelerometer to collect the original one-dimensional time series data of vibration acceleration, and use it as the input of the end-to-end diagnosis model after simple normalization processing.

[0066] 1-1) Use vibration acceleration sensors to collect fault time series data under different working conditions, and cut the collected time series data according to the sliding window size to generate fault diagnosis sample data As shown in the following formula:

[0067]

[0068] The sliding window length is l sw , where nl sw ×i≥l sw ; i=1,2,3..., the sample sequence of each time point is expressed as in are all d-dimensional vectors.

[0069] 1-2) The original time-series vibration acceleration sample data is simply normalized and directly used as the input data of the end-to-end fault diagnosis model. The framework structure of the end-to-end neural network model is as follows: Figure 1 As shown, data X inputs =[X (1) ,X (2) ,...,X (n-1) ,X (n) ], as shown below:

[0070]

[0071] S2. Construct a ResNet model channel to deeply mine and extract the implicit and non-obvious features of the original signal to avoid the gradient diffusion phenomenon of deep neural networks.

[0072] The main idea of ​​the ResNet module is to superimpose the identity mapping layer (IdentityMappings) on the shallow network to avoid the problem of network degradation as the depth increases. The residual module principle of the ResNet neural network is as follows Figure 2 As shown. When the identity mapping function H(x lay )=x layIt is relatively difficult, so the residual function is introduced as F(x lay )=H(x lay )-x lay , when F(x lay )→0, it is an identity mapping, where x lay Represents the input of the residual module. The shortcut performs the same mapping, where the addition of F(x) to x is the element-by-element addition process, and the formula is as follows:

[0073] y lay =F(x lay ,W i )+x lay

[0074] F=W2σ(W,x lay )

[0075] where x lay Represents the input of the residual module, W, W i , W2 both represent one-dimensional convolution calculations, and σ represents the Sigmoid activation function.

[0076] When F(x) and x have different dimensions, a linear mapping needs to be performed on x to match the dimensions. There are two common solutions for linear mapping: (1) directly increasing the dimensions through zero padding, and (2) projecting the result into a new space by multiplying the W matrix. This can be achieved by performing a 1x1 convolution and varying the number of filters in the 1x1 convolution. This paper uses the second solution to calculate the output of the residual module, using the following formula:

[0077] y lay =F(x lay ,W i )+W s x lay

[0078] x lay+1 =f(y lay )

[0079] Where W s Indicates 1*1 convolution calculation, x lay+1 is the output of the residual module, and f(·) is the ReLU activation function.

[0080] S3: Construct a 1DECANet module and connect it in series with the ResNet model channel to perform cross-channel interaction on deep and shallow layer data features as the improved ResNet feature extraction channel.

[0081] 2-1) Constructing the 1DECANet module: First, use one-dimensional global average pooling to obtain a 1*1*C feature map;

[0082] 2-2) Then, the convolution kernel size k is calculated through an adaptive adjustment function. Since the number of channels in the middle layer of the network is generally an exponential of 2, the adaptive convolution kernel size k is calculated as follows:

[0083]

[0084]

[0085] Where C represents the number of data channels in the middle layer, γ = 2, and b = 1.

[0086] 2-3) The convolution kernel size k of the one-dimensional convolution is adaptively adjusted with the help of a function, so that layers with a larger number of channels can interact more across channels. Applying k to the one-dimensional convolution calculation, the weight of each channel is obtained through the activation function Sigmoid, and finally the information interaction between channels is completed. The principle of the constructed 1DECANet module is as follows Figure 3 As shown;

[0087] 2-4) The constructed 1DECANet module is connected in series with the ResNet feature extraction module, where the output of the ResNet module is used as the input of the 1DECANet module. The 1DECANet+ResNet modules formed in series together constitute the improved ResNet feature extraction channel, as shown in Figure 4 shown.

[0088] S4. Build a stacked BiLSTM model channel to extract the temporal correlation features of the original fault data, and add batch normalization (BN) and dropout layers to prevent overfitting of the model channel.

[0089] The LSTM unit cell calculation steps are as follows:

[0090] f t =σ(W t [h t-1 ,x t ]+b t )

[0091] i t =σ(W i [h t-1 ,x t ]+b t )

[0092]

[0093]

[0094] o t =σ(W o [h t-1 ,x t ]+b o )

[0095] H t =o t tanhC t

[0096] Among them, f,i,o represent the calculation results of the forget gate, input gate, and output gate respectively. is the value to be updated for the LSTM cell state, C t is the updated cell state value, H t is the final output value of the LSTM cell unit, W and b represent the weight matrix and bias vector involved in the training of the memory cell unit respectively; σ and tanh represent the Sigmoid function and the hyperbolic tangent function respectively, acting as activation functions of different "gate" structures.

[0097] To achieve contextual judgment of signal data, the present invention uses the hidden layer of the BiLSTM neural network to implement two processes: forward calculation and reverse calculation, which can provide data context information for the network. The update process of BiLSTM is shown in the following formula:

[0098]

[0099]

[0100]

[0101] Among them LSTM + (·), LSTM - (·) are all LSTM cell unit operations mentioned above, W hy and W h ' y are the weight values ​​of the BiLSTM forward calculation layer and the reverse calculation layer respectively; b y is the bias vector of the output layer. In the bidirectional LSTM calculation process, the context information of the data is weighted and learned more data features than the traditional LSTM neural network, such as Figure 5 shown.

[0102] S5. Build an end-to-end dual-channel diagnostic model and use the concatenate mechanism to fuse the data features extracted from the improved ResNet model channel and the stacked BiLSTM model channel in the fully connected layer.

[0103] The framework structure of the improved end-to-end ResNet-BiLSTM dual-channel model is as follows Figure 6 As shown in the figure, the features extracted by combining the two channels are fused by the concatenate mechanism after passing through the fully connected (FC) layer, providing more classification feature parameter basis for the input of the classification layer.

[0104] S6. Use the Focal Loss function to coordinate the imbalance of training sample data, mine the implicit features of difficult-to-distinguish samples, and realize parameter update of dual-channel diagnosis model training.

[0105] The traditional cross-loss function is as follows:

[0106]

[0107] Where X represents the fault sample, n represents the total number of samples, a represents the expected output, and y represents the actual output of the neuron. After calculation by the activation function, its value is between 0 and 1.

[0108] The Focal Loss loss function adds a modulation coefficient ω>0 to the standard cross entropy loss function to reduce the loss weight of easy-to-classify samples and give greater weight to difficult and misclassified sample data. The formula is as follows:

[0109]

[0110] On this basis, a balance factor λ is added to balance the uneven ratio of positive and negative samples. The formula is as follows:

[0111]

[0112] S7: Softmax function is used to classify various fault signal data to achieve accurate diagnosis of unbalanced fault signals.

[0113] The classification layer takes the sample feature vector after dual-channel fusion as input and uses the Softmax classifier to obtain the probability distribution of the fault vibration acceleration time series signal belonging to different categories, as shown in the following formula:

[0114]

[0115] Where m represents the number of units in the output layer of the classification layer, that is, the number of fault signal categories.

[0116] Based on the above steps, the present invention effectively solves the problem of mechanical equipment fault diagnosis. First, the method performs simple normalization on the raw vibration acceleration signal as input data for the end-to-end diagnostic model. Second, to address the temporal nature of unbalanced fault signals and the difficulty in extracting implicit features, a ResNet model (Channel 1) and a stacked BiLSTM model (Channel 2) are constructed, each integrating a 1DECANet module, to perform in-depth, cross-channel data feature mining. Finally, the data features extracted from both channels are fused and accurately classified at the classification layer. This method significantly improves the accuracy of mechanical equipment fault diagnosis and effectively accelerates the convergence of model training.

[0117] The present invention is based on the experimental verification of the fault diagnosis method of the improved end-to-end ResNet-BiLSTM dual-channel model:

[0118] 1. Data description

[0119] The experimental data was derived from a rolling bearing fault vibration dataset maintained by the Case Western Reserve University (CWRU) Bearing Data Center. SKF 6205-2RS JEM deep groove ball bearings and SKF 6203-2RS JEM deep groove ball bearings were installed at the drive end and fan end, respectively. Using electrospark machining (EDM), single-point damage of varying severity was applied to the inner raceway, outer raceway, and rolling element of the rolling bearings at the drive and fan ends, respectively. Fault diameters were set to 0.007, 0.014, and 0.021 inches, respectively, and fault depths to 0.011, 0.050, and 0.150 inches, respectively. Vibration data was collected at 12 kHz and 48 kHz using vibration accelerometers placed at the drive end, fan end, and base, respectively. The present invention conducts fault diagnosis experiments based on a sampling frequency of 12 kHz for nine types of faults, including three different fault diameters at three different fault locations on the inner ring, outer ring, and rolling element of the driving end. Each type of fault sample is divided into data samples with different imbalance ratios using a time step of 100 data points. The experimental sample information of the nine types of bearing faults is shown in Table 1:

[0120] Table 1 Experimental sample information of nine types of bearing faults (fault diameter unit: feet)

[0121]

[0122] The vibration acceleration data of the rolling bearing under normal conditions and nine different fault types are visualized in a specified range of 1000 time steps, as shown in the following example: Figure 8As shown in the figure, there are differences in amplitude and period between the vibration signals of the nine different fault types, and there is weak periodicity between the vibration acceleration signals. Their hidden characteristics are difficult to diagnose and identify directly through data distribution patterns.

[0123] 2. Improved end-to-end ResNet-BiLSTM dual-channel model structure parameters

[0124] For the improved end-to-end ResNet-BiLSTM dual-channel fault diagnosis model, the original unbalanced time series data is simply normalized and directly input into the model. The overall fit of the end-to-end model's hidden layers is utilized to deeply mine and extract the implicit features of the unbalanced sample data. The parameters of the end-to-end classification model are shown in Table 2. The stacked BiLSTM channel fully extracts contextual information from the time series signal, while the ResNet channel fully extracts implicit and non-obvious features in the local space of the unbalanced data and enables cross-channel communication between shallow-layer data and intermediate-layer feature data. The improved ResNet channel is embedded in the 1DECANet module to prevent the impact of dimensionality reduction of the intermediate-layer feature data on feature extraction. This improves the overall diagnostic accuracy of the model while slightly increasing the model's computational parameters.

[0125] Table 2 End-to-end classification model parameters

[0126]

[0127]

[0128] 3. Improved end-to-end ResNet-BiLSTM dual-channel model fault diagnosis effect

[0129] The normalized fault signals and corresponding category labels of the original one-dimensional time series signals are randomly divided into training and test sets according to a ratio of 7:3. The training set is further randomly divided into training and verification data according to a ratio of 9:1 for evaluating the current training results of the model. The proposed dual-channel model training and diagnosis process is as follows: Figure 7 The model training parameters are shown in Table 3.

[0130] Table 3 Description of model training process

[0131]

[0132] To verify the effectiveness of the improved end-to-end ResNet-BiLSTM diagnostic model for implicit feature extraction from imbalanced data, this section constructs 11 imbalanced sample datasets and compares the diagnostic accuracy and test set loss values ​​on these datasets. The test set data is then fed into the best-trained model for fault diagnosis. As shown in Table 4, the improved end-to-end ResNet-BiLSTM dual-channel diagnostic model achieves a minimum diagnostic accuracy of 97.67% on the imbalanced test set.

[0133] Table 4 Diagnosis results of unbalanced dataset

[0134]

[0135]

[0136] In order to more intuitively demonstrate the effectiveness of the improved end-to-end ResNet-BiLSTM dual-channel fault diagnosis model in extracting implicit features from unbalanced data, the input layer and fusion layer of the end-to-end model are visualized using the T-SNE algorithm. The test sample set numbered 6 is selected for the experiment. Figure 9 The visualization results of the input layer of the mid-end-to-end model show that the 9 different fault feature data are randomly distributed and unbalanced. The number of samples with fault label 6 is much larger than the number of other fault samples. Figure 10 The end-to-end model's fusion layer shows low overlap between the nine different fault feature data types, demonstrating that the model significantly distinguishes the different implicit features of imbalanced sample data. This further confirms that the improved end-to-end ResNet-BiLSTM dual-channel diagnostic model can effectively and fully extract the implicit features of unbalanced samples, ultimately achieving accurate classification of unbalanced fault samples.

[0137] The improved end-to-end ResNet-BiLSTM dual-channel diagnosis model has a ResNet channel with a layer of ResNet modules, that is, only one cross-channel communication of feature data is achieved and then connected in series with 1DECANet module. The end-to-end model shows the best average diagnosis accuracy. The model training accuracy and loss value decrease curve are shown in Figure 2. Figure 10 The loss value drop curve shows that the model has basically converged when the training batch reaches about 20 times. The diagnosis results of the model on the test set are displayed using the confusion matrix, as shown in Figure 11 As shown in FIG, the average value of the diagnosis results of the nine types of faults is taken as the overall diagnosis performance of the model. The average diagnosis result of the above 11 types of imbalance fault sample test set reaches 98.74%.

[0138] In order to further verify the role of the 1DECANet module in the training convergence process of the improved end-to-end ResNet-BiLSTM dual-channel fault diagnosis model, an ablation experiment was conducted on the model training loss value before and after adding the 1DECANet module. Figure 12 The results show that the model loss value decreases faster after embedding the 1DECANet module. This efficient channel attention mechanism achieves lower loss values ​​and faster convergence speed while adding a small number of parameters. This means that the model can achieve higher diagnostic accuracy without significantly increasing the model complexity by incorporating the 1DECANet module.

[0139] 4. Conclusion

[0140] To address the challenges of latent feature extraction due to the lack of raw fault signal data collected under complex operating conditions and the imbalance of different fault samples, this paper proposes an improved end-to-end ResNet-BiLSTM dual-channel fault diagnosis model. By introducing a ResNet residual module to enable cross-channel communication of feature data and deep latent feature mining, a 1DECANet module is constructed and embedded in the ResNet feature extraction channel to prevent the dimensionality reduction of feature data from affecting feature extraction during training. This module also avoids the dramatic increase in model parameters and slow model convergence caused by conventional complex attention mechanisms. To verify the effectiveness of the proposed model in extracting latent features from unbalanced data, the paper constructs various unbalanced bearing sample sets of varying proportions and inputs them into the end-to-end diagnosis model for fault diagnosis. The Focal Loss function is used as the model convergence factor. Experimental results demonstrate that the improved end-to-end ResNet-BiLSTM dual-channel diagnosis model can effectively handle a variety of unbalanced sample datasets and effectively extract latent features from different types of unbalanced datasets.

[0141] Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0142] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprise," "include," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0143] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A device fault diagnosis method based on an improved end-to-end ResNet-BiLSTM dual-channel model, characterized by: The specific steps include: S1. Use an accelerometer to collect the original one-dimensional time series data of vibration acceleration, and use it as the input of the end-to-end diagnosis model after simple normalization processing; S2. Construct a ResNet model channel to deeply mine and extract the implicit and non-obvious features of the original signal; S3. Build the 1DECANet module and connect it in series with the ResNet model channel to perform cross-channel interaction on deep and shallow layer data features as the improved ResNet feature extraction channel; S4. Build a stacked BiLSTM model channel to extract the temporal correlation features of the original fault data, and add batch normalization and dropout layers to prevent overfitting of the model channel. S5. Build an end-to-end dual-channel diagnostic model and use the concatenate mechanism to fuse the data features extracted from the improved ResNet model channel and the stacked BiLSTM model channel in the fully connected layer; S6. Use the Focal Loss function to coordinate the imbalance of training sample data, mine the hidden features of difficult-to-classify samples, and implement parameter updates for dual-channel diagnostic model training. S7. Use the Softmax function to classify various fault signal data to achieve accurate diagnosis of unbalanced fault signals.

2. The device fault diagnosis method based on the improved end-to-end ResNet-BiLSTM dual-channel model according to claim 1 is characterized in that: In steps S2 and S3, an improved ResNet feature extraction channel is constructed to achieve cross-channel interaction of deep and shallow feature data. The specific steps are as follows: T1. Construct a one-dimensional ResNet neural network channel. That is, superimpose an identity mapping layer on the shallow network to extract local features of the normalized time series signal along the time axis, and perform convolution layer and identity mapping calculation according to the formula; y lay =F(x lay ,W i )+W s x lay F=W2σ(W,x lay ) x lay+1 =f(y lay ) where x lay represents the input of the residual module, x lay+1 represents the output of the residual module, W, W i , W2 both represent conventional one-dimensional convolution calculations, W s represents 1*1 convolution calculation, σ represents Sigmoid activation function, and f(·) represents ReLU activation function; T2. Construct the 1DECANet module to perform deep, cross-channel mining and extraction of hidden faults in the original fault signal through a local cross-channel interaction strategy without dimensionality reduction; Where C represents the number of feature data channels in the middle layer of the network, k represents the size of the one-dimensional convolution kernel calculated by the adaptive adjustment function, γ = 2, b = 1; T3. The constructed ResNet module is fused in series with the 1DECANet module to form an improved ResNet feature extraction channel, and the concatenation mechanism is used to fuse them in the BiLSTM feature extraction channel to provide classification feature parameter basis for the end-to-end diagnosis model.

3. The device fault diagnosis method based on the improved end-to-end ResNet-BiLSTM dual-channel model according to claim 1 is characterized in that: In step S6, the Focal Loss function is used to coordinate the imbalance of fault sample data, specifically: E1. First, add a modulation coefficient ω>0 to the standard cross entropy loss function to reduce the loss weight of easy-to-classify samples and give more weight to difficult and misclassified sample data; Where L0 represents the standard cross entropy loss function, L1 represents the cross entropy loss function after adding the modulation coefficient, X represents the fault sample, n represents the total number of samples, a represents the expected output, and y represents the actual output of the neuron; E2. Then, on this basis, a balance factor λ is added to balance the uneven ratio of positive and negative samples to form the FocalLoss loss function.

4. The device fault diagnosis method based on the improved end-to-end ResNet-BiLSTM dual-channel model according to claim 1 is characterized in that: In step S1, an acceleration sensor is used to collect the original one-dimensional time series data of vibration acceleration, and after simple normalization processing, the data is used as the input of the end-to-end diagnosis model, which specifically includes the following steps: P1. Use vibration acceleration sensors to collect fault time series data under different working conditions, and cut the collected time series data according to the sliding window size to generate fault diagnosis sample data As shown in the following formula: The sliding window length is l sw , where nl sw ×i≥l sw ; i=1,2,3..., the sample sequence of each time point is expressed as in are all d-dimensional vectors; P2. Perform simple normalization on the original time-series vibration acceleration sample data and directly use it as the input data of the end-to-end fault diagnosis model X inputs =[X (1) ,X (2) ,...,X (n-1) ,X (n) ], as shown below:

5. The device fault diagnosis method based on the improved end-to-end ResNet-BiLSTM dual-channel model according to claim 1 is characterized in that: In step S5, an end-to-end dual-channel diagnostic model is built, and the data features extracted from the improved ResNet model channel and the stacked BiLSTM model channel are fused using the Concatenate mechanism at the fully connected layer, specifically: A1 and Channel1 are set as the improved ResNet neural network model, which is composed of a ResNet module and a 1DECNet module in series. The output of the ResNet module serves as the input of the 1DECANet module. A2 and Channel2 are stacked bidirectional long short-term memory neural network models. The number of memory units in the BiLSTM network and the number of neural network layers are adjusted to extract signal features of different dimensions. The two feature extraction channels A3, Channel1 and Channel2 are expanded by the Flatten mechanism and then fused by the Concatenate mechanism.

Images