Training method of fault diagnosis model of gearbox bearing, fault diagnosis method and device

Abstract

The invention relates to a training method of a fault diagnosis model of a gearbox bearing and a fault diagnosis method and device, and the method comprises the steps: inputting a plurality of vibration data contained in each vibration signal sample into a plurality of LSTM units, and obtaining a plurality of output results; performing weighted summation on the plurality of output results based on the plurality of attention weights; predicting a fault type corresponding to each vibration signal sample based on a weighted summation result; calculating the value of a loss function based on the predicted fault type corresponding to each vibration signal sample and the fault type label corresponding to each vibration signal sample; and adjusting the attention weight according to the value of the loss function. Therefore, by introducing the attention layer, the fault diagnosis model can be helped to better understand key parts in the vibration data (sequence data), so that the fault diagnosis model can accurately perform fault diagnosis when facing complex and variable fault data.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and more specifically, to a training method for a fault diagnosis model of a gearbox bearing, a fault diagnosis method and a device. Background Art

[0002] As a key component of rail transportation vehicles such as high-speed rail, the transmission gearbox is of vital importance for its safe service performance, so regular fault diagnosis is required. Traditional fault diagnosis methods rely on manual experience and simple signal analysis, which is difficult to meet the requirements of modern high-speed trains for high-precision fault diagnosis.

[0003] Among the related technologies, deep learning technology is mainly used for gearbox and bearing fault diagnosis, which can effectively handle vibration characteristics and has high fault diagnosis accuracy. However, although deep learning technology has shown great potential in the diagnosis of EMU gearbox and bearing faults, it still faces many challenges in practical applications. For example, the generalization ability of deep learning models in new environments is limited, and they cannot effectively cope with various complex gearbox and bearing fault diagnosis scenarios. Summary of the invention

[0004] The present disclosure provides a training method, a fault diagnosis method and a device for a gearbox bearing fault diagnosis model, so as to at least solve the problem in the above-mentioned related technologies that the generalization ability of the deep learning model in a new environment is limited and it cannot effectively cope with various complex gearbox and bearing fault diagnosis scenarios.

[0005] According to a first aspect of an embodiment of the present disclosure, a training method for a fault diagnosis model of a gearbox bearing is provided, wherein the fault diagnosis model comprises an attention layer and a long short-term memory network LSTM layer, wherein the LSTM layer comprises a plurality of LSTM units, and the training method comprises: obtaining a plurality of vibration signal samples, wherein each vibration signal sample is obtained by collecting a vibration signal of the gearbox bearing, wherein each vibration signal sample corresponds to a fault type label of the gearbox bearing, and wherein each vibration signal sample comprises a plurality of vibration data; and inputting the plurality of vibration data contained in each vibration signal sample into the plurality of LSTM units respectively, to obtain a plurality of output results h corresponding to the plurality of LSTM units one by one. i , where i>1 and i is an integer; multiple attention weights α assigned to the multiple LSTM units based on the attention layer i , for the multiple output results h iPerform weighted summation to obtain a weighted summation result; predict the fault type corresponding to each vibration signal sample based on the weighted summation result; calculate the value of the loss function based on the predicted fault type corresponding to each vibration signal sample and the fault type label corresponding to each vibration signal sample; adjust the attention weight α according to the value of the loss function i , so as to train the fault diagnosis model.

[0006] Optionally, acquiring multiple vibration signal samples includes: collecting multiple original vibration signals of the gearbox bearing; performing filtering and normalization processing on each of the multiple original vibration signals to obtain multiple vibration signal samples corresponding to the multiple original vibration signals.

[0007] Optionally, before the multiple vibration data contained in each vibration signal sample are respectively input into the multiple LSTM units, the training method further includes: dividing the multiple vibration signal samples into multiple training vibration signal samples and multiple verification vibration signal samples; the multiple vibration data contained in each vibration signal sample are respectively input into the multiple LSTM units, including: inputting the multiple training vibration data contained in each training vibration signal sample in the multiple training vibration signal samples into the multiple LSTM units, and obtaining multiple output results h corresponding to the multiple LSTM units one by one. i The training method further comprises: inputting a plurality of verification vibration data contained in each of the plurality of verification vibration signal samples into the trained fault diagnosis model to obtain a predicted fault type corresponding to each of the verification vibration signal samples; and adjusting the depth, number of layers, and attention weight α of the LSTM layer when the predicted fault type of at least one verification vibration signal sample in the plurality of verification vibration signal samples is inconsistent with the fault type label corresponding to the at least one verification vibration signal sample. i At least one of the above is used to optimize and adjust the trained fault diagnosis model.

[0008] Optionally, the attention weight α i Calculated by the following formula:

[0009] α(s t-1 ,h i )=v a T tanh(U a h j +W a s t-1 )

[0010]

[0011] Among them, s t-1 is the hidden state of the fault diagnosis model at time (t-1), h i is the output result of the i-th LSTM unit in the multiple LSTM units, h j is the output result of the jth LSTM unit in the multiple LSTM units, v a , U a , W a are weight matrices, α t，i is the attention weight of the i-th LSTM unit in the multiple LSTM units at time t, and N is the number of the multiple LSTM units. According to a second aspect of an embodiment of the present disclosure, a fault diagnosis method for a gearbox bearing is provided, comprising: obtaining a target vibration signal of the gearbox bearing, wherein the target vibration signal comprises a plurality of target vibration data; and inputting the plurality of target vibration data contained in the target vibration signal into a fault diagnosis model trained according to the training method of the present disclosure, respectively, to predict the fault type corresponding to the target vibration signal.

[0012] According to three aspects of the embodiments of the present disclosure, a training device for a fault diagnosis model of a gearbox bearing is provided, wherein the fault diagnosis model comprises an attention layer and a long short-term memory network LSTM layer, wherein the LSTM layer comprises a plurality of LSTM units, and the training device comprises: a sample acquisition module, configured to acquire a plurality of vibration signal samples, wherein each vibration signal sample is obtained by collecting a vibration signal of the gearbox bearing, wherein each vibration signal sample corresponds to a fault type label of the gearbox bearing, and wherein each vibration signal sample comprises a plurality of vibration data; and a vibration signal sample input module, configured to input the plurality of vibration data contained in each vibration signal sample into the plurality of LSTM units respectively, to obtain a plurality of output results h corresponding to the plurality of LSTM units one by one. i , where i>1 and i is an integer; a weighted summation module configured to allocate multiple attention weights α to the multiple LSTM units based on the attention layer i , for the multiple output results h i The weighted summation is performed to obtain a weighted summation result; a fault type prediction module is configured to predict the fault type corresponding to each vibration signal sample based on the weighted summation result; a loss function calculation module is configured to calculate the value of the loss function based on the predicted fault type corresponding to each vibration signal sample and the fault type label corresponding to each vibration signal sample; a training module is configured to adjust the attention weight α according to the value of the loss function i , so as to train the fault diagnosis model.

[0013] Optionally, the sample acquisition module is configured to: collect multiple original vibration signals of the gearbox bearing; perform filtering and normalization processing on each of the multiple original vibration signals to obtain multiple vibration signal samples corresponding to the multiple original vibration signals.

[0014] Optionally, the training device further comprises: a division module configured to divide the plurality of vibration signal samples into a plurality of training vibration signal samples and a plurality of verification vibration signal samples; the vibration signal sample input module is configured to: input a plurality of training vibration data contained in each of the plurality of training vibration signal samples into the plurality of LSTM units respectively, and obtain a plurality of output results h corresponding to the plurality of LSTM units one by one i ; The training device also includes: a verification vibration signal sample input module, which is configured to input the multiple verification vibration data contained in each verification vibration signal sample of the multiple verification vibration signal samples into the trained fault diagnosis model to obtain the predicted fault type corresponding to each verification vibration signal sample; an optimization adjustment module, which is configured to adjust the depth, number of layers, and attention weight α of the LSTM layer when the predicted fault type of at least one verification vibration signal sample in the multiple verification vibration signal samples is inconsistent with the fault type label corresponding to the at least one verification vibration signal sample. i At least one of the above is used to optimize and adjust the trained fault diagnosis model.

[0015] Optionally, the attention weight α i Calculated by the following formula:

[0016] α(s t-1 ,h i )=v a T tanh(U a h j +W a s t-1 )

[0017]

[0018] Among them, s t-1 is the hidden state of the fault diagnosis model at time (t-1), h i is the output result of the i-th LSTM unit in the multiple LSTM units, h j is the output result of the jth LSTM unit in the multiple LSTM units, v a , U a , W a are weight matrices, α t，iis the attention weight of the i-th LSTM unit in the multiple LSTM units at time t, and N is the number of the multiple LSTM units.

[0019] According to four aspects of the embodiments of the present disclosure, a fault diagnosis device for a gearbox bearing is provided, comprising: a vibration signal acquisition module, configured to acquire a target vibration signal of the gearbox bearing, wherein the target vibration signal includes a plurality of target vibration data; and a fault prediction module, configured to input the plurality of target vibration data included in the target vibration signal into a fault diagnosis model trained according to the training method of the present disclosure, so as to predict a fault type corresponding to the target vibration signal.

[0020] The technical solution provided by the embodiments of the present disclosure brings at least the following beneficial effects:

[0021] In the present disclosure, by introducing the attention layer, the fault diagnosis model can be helped to better understand the key parts of the vibration data (sequence data), so that the fault diagnosis model can accurately perform fault diagnosis when facing complex and changeable fault data. That is, the fault diagnosis model provided by the present disclosure can effectively cope with various complex gearbox bearing fault diagnosis scenarios, and the generalization ability and robustness of the fault diagnosis model are good.

[0022] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute improper limitations on the present disclosure.

[0024] Figure 1 is a flow chart illustrating a method for training a fault diagnosis model of a gearbox bearing according to an exemplary embodiment of the present disclosure;

[0025] Figure 2 is a structural diagram showing a fault diagnosis model based on CNN-LSTM-Self-Attention according to an exemplary embodiment of the present disclosure;

[0026] Figure 3 is a network architecture diagram showing a bidirectional LSTM according to an exemplary embodiment of the present disclosure;

[0027] Figure 4 is a schematic diagram showing a model architecture of an attention mechanism according to an exemplary embodiment of the present disclosure;

[0028] Figure 5is a flow chart showing a fault diagnosis method of a gearbox bearing according to an exemplary embodiment of the present disclosure;

[0029] Figure 6 is a block diagram illustrating a training device for a fault diagnosis model of a gearbox bearing according to an exemplary embodiment of the present disclosure;

[0030] Figure 7 is a block diagram showing a fault diagnosis device for a gearbox bearing according to an exemplary embodiment of the present disclosure;

[0031] Figure 8 is a block diagram illustrating an electronic device according to an exemplary embodiment of the present disclosure. DETAILED DESCRIPTION

[0032] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings.

[0033] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The implementation methods described in the following examples do not represent all implementation methods consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the attached claims.

[0034] It should be noted that the phrase "at least one of the items" in the present disclosure includes three types of parallel situations: "any one of the items", "a combination of any number of the items", and "all of the items". For example, "including at least one of A and B" includes the following three parallel situations: (1) including A; (2) including B; (3) including A and B. Another example is "executing at least one of step 1 and step 2" which means the following three parallel situations: (1) executing step 1; (2) executing step 2; (3) executing step 1 and step 2.

[0035] As a key component of rail transportation vehicles such as high-speed rail, the transmission gearbox is of vital importance for its safe service performance, so regular fault diagnosis is required. Traditional diagnostic methods rely on manual experience and simple signal analysis, which is difficult to meet the requirements of modern high-speed trains for high-precision fault diagnosis. With the development of signal detection technology, computer technology, digital signal processing technology and artificial intelligence technology, it is particularly important to use more advanced methods for fault diagnosis.

[0036] Deep learning technology has gradually attracted attention in intelligent fault diagnosis. It can effectively process complex vibration characteristics and improve the accuracy of fault diagnosis. Data-driven methods can automatically extract features, avoiding the tediousness of manually extracting feature information, and greatly improving the effectiveness of fault classification systems. In addition, deep learning methods are widely used in the field of mechanical fault diagnosis, significantly improving the efficiency and accuracy of gearbox fault diagnosis, and becoming an important means to ensure the service life of gearboxes.

[0037] Although deep learning technology has shown great potential in fault diagnosis of gearbox bearings of EMUs, it still faces many challenges in practical applications. First, it is difficult and costly to obtain sufficient labeled data, especially data of specific fault types. In addition, the relative scarcity of fault instances leads to an imbalance in the ratio of positive and negative samples in the training data set. Second, data quality is crucial to the model effect. Any noise in acquisition or preprocessing may affect model learning, and the accuracy of data annotation will directly affect model performance. In addition, the complexity and difficulty of interpreting deep learning models become disadvantages in industrial applications that require clear fault analysis. At the same time, training these models requires a lot of computing resources, which is an insurmountable obstacle for small enterprises and research institutions with limited resources; in applications that require instant feedback, the computational delay and lightweight requirements of the model are particularly prominent; the lack of unified standards also increases the difficulty of model transplantation and interoperability; although deep learning models perform well on specific data sets, their generalization ability is limited when faced with new environments or unseen data.

[0038] In addition, most current studies rely on single time series data, which limits the comprehensive understanding of fault characteristics and may lead to model overfitting, affecting the accuracy of prediction. Secondly, existing methods often lack in-depth and comprehensive analysis of fault diagnosis, but rely more on local perspectives. Prediction models designed for a single data source may not be applicable in multimodal information scenarios, resulting in reduced prediction performance.

[0039] In order to solve the above problems existing in the related art, the training method, fault diagnosis method and device of the gearbox bearing fault diagnosis model provided by the present disclosure can help the fault diagnosis model better understand the key parts of the vibration data (sequence data) by introducing the attention layer, so that the fault diagnosis model can accurately perform fault diagnosis when facing complex and changeable fault data. That is, the fault diagnosis model provided by the present disclosure can effectively cope with various complex gearbox bearing fault diagnosis scenarios, and the generalization ability and robustness of the fault diagnosis model are good.

[0040] Figure 11 is a flowchart showing a method for training a fault diagnosis model for a gearbox bearing according to an exemplary embodiment of the present disclosure. The fault diagnosis model may include an attention layer and a long short-term memory network LSTM layer, and the LSTM layer may include multiple LSTM units.

[0041] Reference Figure 1 In step 101, a plurality of vibration signal samples may be obtained, wherein each vibration signal sample may be obtained by collecting a vibration signal of a gearbox bearing, each vibration signal sample may correspond to a fault type label of the gearbox bearing, and each vibration signal sample may include a plurality of vibration data: (x0, x1, x2, x3, ..., x t ).

[0042] Specifically, a high-precision sensor can be used to collect vibration signals of the EMU transmission gearbox under different working conditions for data collection. The data collection may include but is not limited to: no fault, rolling element pitting, rolling element failure, outer ring pitting, outer ring failure, rolling element and outer ring pitting, rolling element and outer ring failure, etc. Then, the collected data can be converted into a corresponding text format.

[0043] Furthermore, the gearbox and its small end bearing of the EMU can be selected as the research object. By collecting and processing the vibration data during operation, the faulty bearing of the gearbox can be accurately identified, and the location selection of the vibration measurement point can be reasonably guided. Accelerometers can be set on the bearing seats of the small gear end and the large gear end of the gearbox to measure the vibration signals in the vertical (Z), lateral (Y) and axial (X) directions. The test conditions can involve a variety of speed conditions. For example, the speed can be gradually increased from zero to the maximum operating speed of 6070rpm, and the vibration characteristics can be recorded at these speeds. In addition, the test can also simulate a variety of conditions such as no fault state, outer ring fault, rolling element fault, and simultaneous fault of the outer ring and rolling element to collect comprehensive vibration data.

[0044] According to an exemplary embodiment of the present disclosure, multiple original vibration signals of a gearbox bearing may be collected, and then filtering and normalizing may be performed on each of the multiple original vibration signals to obtain multiple vibration signal samples corresponding to the multiple original vibration signals.

[0045] Specifically, a large number of signal data sets of gearbox bearings can be obtained and data preprocessing can be performed. In order to ensure the integrity of the time series data during preprocessing, missing values ​​must be filled before further analysis. In addition, the scales of different dimensional features of the original data set may be inconsistent. In order to convert them into data with consistent scales, the digital features need to be standardized. Then, the preprocessed data set can be classified and divided into training sets, validation sets, and test sets according to the proportion. In this way, by performing preprocessing operations such as filtering and normalization on the collected data, the data quality and model training effect can be improved.

[0046] Figure 2 is a structural diagram showing a fault diagnosis model based on CNN-LSTM-Self-Attention according to an exemplary embodiment of the present disclosure. Figure 2 , the convolutional layer and the bidirectional LSTM layer can be integrated into a complete network, and the attention layer, the fully connected layer, and the output layer can be added after the bidirectional LSTM layer to handle multi-classification problems. The network can be trained using a training dataset to optimize the network parameters, that is, the loss value and accuracy can be monitored during the training process, and the hyperparameters can be adjusted to improve the model performance.

[0047] Figure 3 is a network architecture diagram showing a bidirectional LSTM according to an exemplary embodiment of the present disclosure. Figure 3 , the convolutional neural network structure in the time feature extraction path can adopt a bidirectional LSTM network structure. The bidirectional LSTM structure can effectively capture the continuity of the vibration signal and build dependencies. Specifically, the forward LSTM (forward LSTM) is used to capture the information flow from the past to the present in the time series, that is, the forward dependency features of the sequence can be extracted; the reverse LSTM (backward LSTM) is used to capture the reverse information flow from the future to the present, that is, the backward dependency features of the sequence can be extracted. The hidden layer update state of the forward LSTM, backward LSTM and the final output process of BiLSTM can be expressed by the following formula:

[0048] A i =f1(w1x i +w2A i-1 )

[0049] B i =f2(w3x i +w5B i+1 )

[0050] Y i =f3(w4A i +w6B i )

[0051] Among them, f1, f2, and f3 are activation functions between different layers. By combining the information in the above two directions, BiLSTM can more comprehensively understand the characteristic connection of data in time series, thereby improving the model's utilization of data and prediction accuracy.

[0052] Reference Figure 2 In the present disclosure, the CNN-inceptionv2 structure can be used to extract preprocessed time series data and output a time feature set. The CNN-Inception V2 model uses the Inception module, which can capture features of different scales by stacking multiple parallel convolution kernel operations and pooling operations of different sizes. In addition, by using two 3x3 convolution kernels, not only can the depth of the model be increased, but also its feature expression ability can be enhanced, and at the same time, the computational efficiency can be improved. In addition, 1x1 convolution can be used to reduce the dimension of the feature, and nonlinear transformation can be introduced to enhance its expression ability.

[0053] In step 102, the multiple vibration data contained in each vibration signal sample can be input into multiple LSTM units respectively to obtain multiple output results h corresponding to the multiple LSTM units. i , where i>1 and i is an integer.

[0054] Specifically, each LSTM unit can usually be composed of the following three parts: forget gate, input gate and output gate.

[0055] The forget gate selects the information to be discarded from the cell state at time (t-1). The forget gate reads the hidden layer state h at the previous moment. (t-1) and the current input sequence x (t) , can generate an output value between 0 and 1, where 1 can represent completely retaining the information and 0 can represent completely discarding the information.

[0056]

[0057] Among them, f (t) represents the output of the forget gate at the current time t, W f and b f They represent the weight matrix and bias term of the forget gate respectively, and σ represents the sigmoid activation function.

[0058] The main function of the input gate is to (t) Determine the information that needs to be stored in the neuron. First, the candidate memory cell state at time t can be generated through the tanh layer: Then, the cell state can be updated by combining the output of the forget gate to obtain a new cell state: C (t)The specific operation process of the input gate can be as follows:

[0059] i (t) =σ(W i ·[h (t-1) , x (t) ]+b i )

[0060]

[0061] Among them, i (t) is the output state of the input gate at time t, which determines the input x (t) Update cell state C (t) The degree of influence of W i and b i are the weight and bias of the input gate respectively; W c and b c is the weight matrix and bias for generating candidate memory units; tanh is the hyperbolic tangent activation function; Represents bit-wise multiplication of elements.

[0062] The output gate is mainly used to extract and output key information from the current cell state. First, the σ layer can be used to determine which neuron states need to be output. Then, these neuron states can be processed by the tanh layer and multiplied by the σ output to generate the final output value h. (t) , and this value can be used as the hidden layer input for the next time step. The calculation process of the output gate can be expressed by the following formula:

[0063] o (t) =σ(W0·[h (t-1) ,x (t) ]+b0)

[0064] Among them, (t) is the output state of the output gate at time t; W0 and b0 are the weight matrix and bias term of the output gate respectively.

[0065] In step 103, multiple attention weights α assigned to multiple LSTM units one by one based on the attention layer can be i , for multiple output results h i Perform weighted summation to obtain a weighted summation result.

[0066] Figure 4 is a schematic diagram showing a model architecture of an attention mechanism according to an exemplary embodiment of the present disclosure. Figure 4, after the convolutional network and the bidirectional long short-term memory recurrent network, the attention mechanism can be introduced, which can be used to generalize the output layer of the hidden layer. Specifically, the Attention model can be introduced between the hidden layer and the output layer. The attention mechanism considers the input of all positions at a certain moment, and can assign an attention weight to each feature, and then train it. It is more suitable for time series prediction problems.

[0067] The attention mechanism calculation formula can be as follows:

[0068]

[0069] Among them, k can be the number of LSTM units, α i The attention weight assigned to the i-th LSTM unit, h i It can be the output result of the i-th LSTM unit.

[0070] According to an exemplary embodiment of the present disclosure, the attention weight α i It can be calculated by the following formula:

[0071] a(s t-1 , h i)=v a T tanh(U a h j +W a s t-1 )

[0072]

[0073] Among them, s t-1 is the hidden state of the fault diagnosis model at time (t-1), h i is the output result of the i-th LSTM unit among multiple LSTM units, h j is the output result of the jth LSTM unit among multiple LSTM units, v a , U a , W a are weight matrices, α t，i is the attention weight of the i-th LSTM unit in multiple LSTM units at time t, and N is the number of multiple LSTM units.

[0074] In step 104, the fault type corresponding to each vibration signal sample can be predicted based on the weighted summation result. Specifically, the weighted summation result can be input into the Softmax layer for full connection calculation to obtain the final fault type.

[0075] In step 105 , the value of the loss function may be calculated based on the predicted fault type corresponding to each vibration signal sample and the fault type label corresponding to each vibration signal sample.

[0076] In step 106, the attention weight α can be adjusted according to the value of the loss function i , so as to train the fault diagnosis model.

[0077] According to an exemplary embodiment of the present disclosure, before the multiple vibration data contained in each vibration signal sample is respectively input into the multiple LSTM units, that is, before step 102, the multiple vibration signal samples can also be divided into multiple training vibration signal samples and multiple verification vibration signal samples. In this way, in step 102, the multiple training vibration data contained in each training vibration signal sample in the multiple training vibration signal samples can be respectively input into the multiple LSTM units to obtain multiple output results h corresponding to the multiple LSTM units one by one. i .

[0078] Accordingly, the multiple verification vibration data contained in each of the multiple verification vibration signal samples can also be input into the trained fault diagnosis model to obtain the predicted fault type corresponding to each verification vibration signal sample. In the case where the predicted fault type of at least one verification vibration signal sample in the multiple verification vibration signal samples is inconsistent with the fault type label corresponding to at least one verification vibration signal sample, the depth, number of layers, and attention weight α of the LSTM layer can be adjusted. i At least one of the above is used to optimize and adjust the trained fault diagnosis model.

[0079] Specifically, the trained fault diagnosis model can be verified on the validation set to evaluate the generalization ability and performance of the model. For example, the depth or number of layers of the convolutional neural network and the long short-term memory network can be adjusted based on the validation results, and the aforementioned attention weight α can also be adjusted. i Modifications are made until the model reaches the expected performance level. In addition, when the performance of the model on the validation set meets the preset standards, the test set can also be used to conduct a comprehensive performance evaluation of the verified fault diagnosis model.

[0080] It should be noted that, in the present disclosure, the trained gearbox bearing fault diagnosis model based on CNN-LSTM-Self-Attention can be used for both gearbox fault diagnosis and bearing fault diagnosis.

[0081] The EMU gearbox bearing fault diagnosis model based on CNN-LSTM-Self-Attention provided by the present disclosure can not only effectively improve the accuracy of fault diagnosis by integrating feature extraction, attention mechanism and prediction model, but also alleviate the various limitations encountered by traditional deep learning models in industrial applications to a certain extent. Specifically, the method disclosed in the present disclosure enhances the ability to capture complex patterns in vibration signals and improves the processing effect of time series data by integrating convolutional neural networks (CNN), long short-term memory networks (LSTM) and self-attention mechanisms (Self-Attention). In addition, through data preprocessing and enhancement technology, data quality is improved and the problem of data imbalance is solved.

[0082] In addition, in terms of feature extraction of time series data, the present invention adopts a CNN-inceptionv2 model with multi-branch convolution kernels, which improves feature expression capabilities by increasing model depth, improves computational efficiency, and enables the model to more effectively extract features from time series data, thereby improving the prediction performance of the model; by utilizing the attention mechanism, the fault diagnosis model provided by the present invention can focus more on the extraction of key information, and can fully consider the relationship between data, thereby improving the prediction performance and accuracy of the model; the prediction module can use a BiLSTM model to capture long-term dependencies, and can integrate local features of CNN to enhance the prediction performance of the model.

[0083] Figure 5 is a flow chart illustrating a fault diagnosis method of a gearbox bearing according to an exemplary embodiment of the present disclosure.

[0084] Reference Figure 5 In step 501, a target vibration signal of a gearbox bearing may be acquired, wherein the target vibration signal may include a plurality of target vibration data.

[0085] In step 502, a plurality of target vibration data included in the target vibration signal may be respectively input into the fault diagnosis model trained according to the training method of the present disclosure to predict the fault type corresponding to the target vibration signal.

[0086] It should be noted that, compared with the related art, the present disclosure has the following advantages:

[0087] 1. Local feature extraction: Through the convolutional neural network (CNN) layer, the local features of the gearbox vibration signal can be effectively extracted, and then subtle fault signals can be captured; the self-attention model can enhance the focus on key features and improve the model's sensitivity to important information, so as to better understand complex fault modes.

[0088] 2. Improved generalization ability: LSTM networks may have poor generalization ability when processing hidden layer outputs, resulting in poor performance of the model on new data. This disclosure introduces a self-attention model to help the model better understand the key parts of sequence data, so that the model can accurately diagnose complex and changeable fault data, thereby improving the generalization ability and robustness of the model.

[0089] 3. Multi-dimensional diagnosis: Combining the CNN-LSTM-Self-Attention model architecture, the present disclosure can not only capture the local features of the gearbox vibration signal, but also use the LSTM network's ability to process time series data and the self-attention mechanism to strengthen key features, thereby achieving in-depth, multi-dimensional diagnosis of gearbox faults. Through comprehensive analysis of multimodal features, the present disclosure can more comprehensively identify and diagnose various fault modes of gearboxes.

[0090] 4. Real-time and accuracy: The fault diagnosis model provided by the present disclosure has high real-time performance and can quickly diagnose the fault of the EMU transmission gearbox, meeting the online monitoring needs of the industrial site. The fault diagnosis model provided by the present disclosure has a significantly higher fault identification accuracy on multiple gearbox fault data sets than traditional methods, especially under complex working conditions and with large noise interference.

[0091] Figure 6 is a block diagram illustrating a training device 600 of a fault diagnosis model of a gearbox bearing according to an exemplary embodiment of the present disclosure.

[0092] Reference Figure 6 The training device 600 of the fault diagnosis model of the gearbox bearing may include a sample acquisition module 601, a vibration signal sample input module 602, a weighted summation module 603, a fault type prediction module 604, a loss function calculation module 605 and a training module 606.

[0093] The sample acquisition module 601 can acquire multiple vibration signal samples, wherein each vibration signal sample can be obtained by collecting the vibration signal of the gearbox bearing, each vibration signal sample can correspond to a fault type label of the gearbox bearing, and each vibration signal sample can include multiple vibration data: (x0, x1, x2, x3, ..., x t ).

[0094] According to an exemplary embodiment of the present disclosure, the sample acquisition module 601 can collect multiple original vibration signals of the gearbox bearing. Then, the sample acquisition module 601 can perform filtering and normalization processing on each of the multiple original vibration signals to obtain multiple vibration signal samples corresponding to the multiple original vibration signals.

[0095] Specifically, a large number of signal data sets of gearbox bearings can be obtained and data preprocessing can be performed. In order to ensure the integrity of the time series data during preprocessing, missing values ​​must be filled before further analysis. In addition, the scales of different dimensional features of the original data set may be inconsistent. In order to convert them into data with consistent scales, the digital features need to be standardized. Then, the preprocessed data set can be classified and divided into training sets, validation sets, and test sets according to the proportion. In this way, by performing preprocessing operations such as filtering and normalization on the collected data, the data quality and model training effect can be improved.

[0096] The vibration signal sample input module 602 can input multiple vibration data contained in each vibration signal sample into multiple LSTM units respectively to obtain multiple output results h corresponding to the multiple LSTM units. i , where i>1 and i is an integer.

[0097] The weighted summation module 603 can be based on the multiple attention weights α assigned to the multiple LSTM units by the attention layer. i , for multiple output results h i Perform weighted summation to obtain a weighted summation result.

[0098] According to an exemplary embodiment of the present disclosure, the attention weight α i It can be calculated by the following formula:

[0099] α( s t -1 ,h i )=v a T tanh(U a h y +W a s t-1 )

[0100]

[0101] Among them, s t-1 is the hidden state of the fault diagnosis model at time (t-1), h i is the output result of the i-th LSTM unit among multiple LSTM units, h j is the output result of the jth LSTM unit among multiple LSTM units, v a , U a , W a are weight matrices, α t，i is the attention weight of the i-th LSTM unit in multiple LSTM units at time t, and N is the number of multiple LSTM units.

[0102] The fault type prediction module 604 can predict the fault type corresponding to each vibration signal sample based on the weighted summation result. Specifically, the weighted summation result can be input into the Softmax layer for full connection calculation to obtain the final fault type.

[0103] The loss function calculation module 605 may calculate the value of the loss function based on the predicted fault type corresponding to each vibration signal sample and the fault type label corresponding to each vibration signal sample.

[0104] The training module 606 can adjust the attention weight α according to the value of the loss function i , so as to train the fault diagnosis model.

[0105] According to an exemplary embodiment of the present disclosure, the training device 600 for the fault diagnosis model of the gearbox bearing may further include a division module, a verification vibration signal sample input module and an optimization adjustment module.

[0106] The division module can divide the multiple vibration signal samples into multiple training vibration signal samples and multiple verification vibration signal samples. The multiple training vibration data contained in each training vibration signal sample in the multiple training vibration signal samples can be input into multiple LSTM units respectively to obtain multiple output results h corresponding to the multiple LSTM units one by one. i The verification vibration signal sample input module can also input multiple verification vibration data contained in each verification vibration signal sample of the multiple verification vibration signal samples into the trained fault diagnosis model to obtain the predicted fault type corresponding to each verification vibration signal sample.

[0107] When the predicted fault type of at least one verification vibration signal sample among multiple verification vibration signal samples is inconsistent with the fault type label corresponding to at least one verification vibration signal sample, the optimization adjustment module can adjust the depth, number of layers, and attention weight α of the LSTM layer. i At least one of the above is used to optimize and adjust the trained fault diagnosis model.

[0108] Figure 7 is a block diagram illustrating a fault diagnosis apparatus 700 for a gearbox bearing according to an exemplary embodiment of the present disclosure.

[0109] Reference Figure 7 The gearbox bearing fault diagnosis device 700 may include a vibration signal acquisition module 701 and a fault prediction module 702 .

[0110] The vibration signal acquisition module 701 can acquire a target vibration signal of a gearbox bearing, wherein the target vibration signal can include a plurality of target vibration data.

[0111] The fault prediction module 702 may input a plurality of target vibration data included in the target vibration signal into the fault diagnosis model trained according to the training method of the present disclosure, so as to predict the fault type corresponding to the target vibration signal.

[0112] Figure 8 is a block diagram illustrating an electronic device 800 according to an exemplary embodiment of the present disclosure.

[0113] Reference Figure 8 The electronic device 800 includes at least one memory 801 and at least one processor 802, wherein the at least one memory 801 stores instructions, and when the instructions are executed by the at least one processor 802, a training method for a fault diagnosis model of a gearbox bearing or a fault diagnosis method for a gearbox bearing according to an exemplary embodiment of the present disclosure is executed.

[0114] As an example, the electronic device 800 may be a PC, a tablet device, a personal digital assistant, a smart phone, or other device capable of executing the above instructions. Here, the electronic device 800 is not necessarily a single electronic device, but may also be any device or circuit capable of executing the above instructions (or instruction sets) individually or in combination. The electronic device 800 may also be part of an integrated control system or system manager, or may be configured as a portable electronic device interconnected with a local or remote (e.g., via wireless transmission) interface.

[0115] In the electronic device 800, the processor 802 may include a central processing unit (CPU), a graphics processing unit (GPU), a programmable logic device, a dedicated processor system, a microcontroller or a microprocessor. As an example and not limitation, the processor may also include an analog processor, a digital processor, a microprocessor, a multi-core processor, a processor array, a network processor, etc.

[0116] The processor 802 may execute instructions or codes stored in the memory 801, wherein the memory 801 may also store data. Instructions and data may also be sent and received over a network via a network interface device, wherein the network interface device may employ any known transmission protocol.

[0117] The memory 801 may be integrated with the processor 802, for example, by placing RAM or flash memory within an integrated circuit microprocessor or the like. In addition, the memory 801 may include a separate device, such as an external disk drive, a storage array, or any other storage device that can be used by a database system. The memory 801 and the processor 802 may be operatively coupled, or may communicate with each other, such as through an I / O port, a network connection, etc., so that the processor 802 can read files stored in the memory.

[0118] In addition, the electronic device 800 may also include a video display (such as a liquid crystal display) and a user interaction interface (such as a keyboard, a mouse, a touch input device, etc.) All components of the electronic device 800 may be connected to each other via a bus and / or a network.

[0119] According to an exemplary embodiment of the present disclosure, a computer-readable storage medium may also be provided. When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the above-mentioned training method of the fault diagnosis model of the gearbox bearing or the fault diagnosis method of the gearbox bearing. Examples of computer-readable storage media here include: read-only memory (ROM), random access programmable read-only memory (PROM), electrically erasable programmable read-only memory (EEPROM), random access memory (RAM), dynamic random access memory (DRAM), static random access memory (SRAM), flash memory, non-volatile memory, CD-ROM, CD-R, CD+R, CD-RW, CD+RW, DVD-ROM, DVD-R, DVD+R, DVD-RW, DVD+RW, DVD-RAM, BD-ROM, BD-R, BD-R LTH, BD-RE, Blu-ray or optical disk storage, hard disk drive (HDD), solid state drive (SSD), card storage (such as, multimedia card, secure digital (SD) card or extreme digital (XD) card), magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid state disk and any other device, any other device is configured to store computer programs and any associated data, data files and data structures in a non-transitory manner and provide the computer programs and any associated data, data files and data structures to a processor or computer so that the processor or computer can execute the computer program. The computer program in the above-mentioned computer-readable storage medium can be run in an environment deployed in a computer device such as a client, a host, an agent device, a server, etc. In addition, in one example, the computer program and any associated data, data files and data structures are distributed on a networked computer system, so that the computer program and any associated data, data files and data structures are stored, accessed and executed in a distributed manner by one or more processors or computers.

[0120] According to an exemplary embodiment of the present disclosure, a computer program product may also be provided, including a computer program, which, when executed by a processor, implements the training method of the fault diagnosis model of the gearbox bearing or the fault diagnosis method of the gearbox bearing according to the present disclosure.

[0121] According to the training method, fault diagnosis method and device of the gearbox bearing fault diagnosis model disclosed in the present invention, by introducing the attention layer, the fault diagnosis model can be helped to better understand the key parts of the vibration data (sequence data), so that the fault diagnosis model can accurately perform fault diagnosis when facing complex and changeable fault data. That is, the fault diagnosis model provided by the present invention can effectively cope with various complex gearbox bearing fault diagnosis scenarios, and the generalization ability and robustness of the fault diagnosis model are good.

[0122] According to an exemplary embodiment of the present disclosure, by performing preprocessing operations such as filtering and normalization on the collected data, the data quality and the model training effect can be improved.

[0123] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. The present disclosure is intended to cover any variations, uses or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The description and examples are to be considered exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0124] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A training method for a fault diagnosis model of a gearbox bearing, characterized in that: The fault diagnosis model comprises an attention layer and a long short-term memory network LSTM layer, the LSTM layer comprises a plurality of LSTM units, and the training method comprises: Acquire a plurality of vibration signal samples, wherein each vibration signal sample is obtained by collecting a vibration signal of the gearbox bearing, each vibration signal sample corresponds to a fault type label of the gearbox bearing, and each vibration signal sample contains a plurality of vibration data; Input the multiple vibration data contained in each vibration signal sample into the multiple LSTM units respectively, and obtain multiple output results h corresponding to the multiple LSTM units one by one i , where i>1 and i is an integer; Based on the multiple attention weights α assigned to the multiple LSTM units one by one by the attention layer i , for the multiple output results h i Perform weighted summation to obtain a weighted summation result; Predicting the fault type corresponding to each vibration signal sample based on the weighted summation result; Calculating a value of a loss function based on the predicted fault type corresponding to each vibration signal sample and the fault type label corresponding to each vibration signal sample; By adjusting the attention weight α according to the value of the loss function i , so as to train the fault diagnosis model.

2. The training method according to claim 1, characterized in that: The obtaining of a plurality of vibration signal samples comprises: Collecting a plurality of original vibration signals of the gearbox bearing; Filtering and normalizing are performed on each of the multiple original vibration signals to obtain multiple vibration signal samples corresponding to the multiple original vibration signals.

3. The training method according to claim 1, characterized in that: Before inputting the plurality of vibration data contained in each vibration signal sample into the plurality of LSTM units respectively, the training method further comprises: dividing the plurality of vibration signal samples into a plurality of training vibration signal samples and a plurality of verification vibration signal samples; The step of inputting the plurality of vibration data contained in each vibration signal sample into the plurality of LSTM units respectively comprises: Inputting the plurality of training vibration data contained in each of the plurality of training vibration signal samples into the plurality of LSTM units respectively, and obtaining a plurality of output results corresponding to the plurality of LSTM units one by one h i; The training method further comprises: Inputting a plurality of verification vibration data contained in each of the plurality of verification vibration signal samples into the trained fault diagnosis model to obtain a predicted fault type corresponding to each verification vibration signal sample; When the predicted fault type of at least one verification vibration signal sample among the multiple verification vibration signal samples is inconsistent with the fault type label corresponding to the at least one verification vibration signal sample, the depth, number of layers, and attention weight α of the LSTM layer are adjusted. i At least one of the above is used to optimize and adjust the trained fault diagnosis model.

4. The training method according to claim 1, characterized in that: The attention weight α i Calculated by the following formula: α(s t-1 ,h i )=v a T tanh(U a h j +W a s t-1 ) Among them, s t-1 is the hidden state of the fault diagnosis model at time (t-1), h i is the output result of the i-th LSTM unit in the multiple LSTM units, h j is the output result of the jth LSTM unit in the multiple LSTM units, v a , U a , W a are weight matrices, α t，i is the attention weight of the i-th LSTM unit in the multiple LSTM units at time t, and N is the number of the multiple LSTM units.

5. A gearbox bearing fault diagnosis method, characterized in that: include: Acquire a target vibration signal of the gearbox bearing, wherein the target vibration signal includes a plurality of target vibration data; The multiple target vibration data included in the target vibration signal are respectively input into the fault diagnosis model trained according to any one of the training methods in claims 1 to 4 to predict the fault type corresponding to the target vibration signal.

6. A training device for a fault diagnosis model of a gearbox bearing, characterized in that: The fault diagnosis model comprises an attention layer and a long short-term memory network LSTM layer, the LSTM layer comprises a plurality of LSTM units, and the training device comprises: A sample acquisition module is configured to acquire a plurality of vibration signal samples, wherein each vibration signal sample is acquired by collecting a vibration signal of the gearbox bearing, each vibration signal sample corresponds to a fault type label of the gearbox bearing, and each vibration signal sample includes a plurality of vibration data; The vibration signal sample input module is configured to input the multiple vibration data contained in each vibration signal sample into the multiple LSTM units respectively, and obtain multiple output results h corresponding to the multiple LSTM units one by one i , where i>1 and i is an integer; A weighted summation module is configured to allocate multiple attention weights α to the multiple LSTM units based on the attention layer. i , for the multiple output results h i Perform weighted summation to obtain a weighted summation result; A fault type prediction module is configured to predict the fault type corresponding to each vibration signal sample based on the weighted sum result; A loss function calculation module, configured to calculate a value of a loss function based on the predicted fault type corresponding to each vibration signal sample and the fault type label corresponding to each vibration signal sample; A training module is configured to adjust the attention weight α according to the value of the loss function i , so as to train the fault diagnosis model.

7. The training device according to claim 6, characterized in that The sample acquisition module is configured as follows: Collecting a plurality of original vibration signals of the gearbox bearing; Filtering and normalizing are performed on each of the multiple original vibration signals to obtain multiple vibration signal samples corresponding to the multiple original vibration signals.

8. The training device according to claim 6, characterized in that The training device also includes: a division module, configured to divide the plurality of vibration signal samples into a plurality of training vibration signal samples and a plurality of verification vibration signal samples; The vibration signal sample input module is configured as follows: Inputting the plurality of training vibration data contained in each of the plurality of training vibration signal samples into the plurality of LSTM units respectively, and obtaining a plurality of output results corresponding to the plurality of LSTM units one by one h i; The training device also includes: A verification vibration signal sample input module is configured to input a plurality of verification vibration data contained in each verification vibration signal sample of the plurality of verification vibration signal samples into a trained fault diagnosis model to obtain a predicted fault type corresponding to each verification vibration signal sample; The optimization and adjustment module is configured to adjust the depth, number of layers, and attention weight α of the LSTM layer when the predicted fault type of at least one verification vibration signal sample among the multiple verification vibration signal samples is inconsistent with the fault type label corresponding to the at least one verification vibration signal sample. i At least one of the above is used to optimize and adjust the trained fault diagnosis model.

9. The training device according to claim 6, characterized in that The attention weight α i Calculated by the following formula: α(s t-1 ,h i )=v a T tanh(U a h j +W a s t-1 ) Among them, s t-1 is the hidden state of the fault diagnosis model at time (t-1), h i is the output result of the i-th LSTM unit in the multiple LSTM units, h j is the jth LSTM in the multiple LSTM units 单 The output result of the element, v a , U a , W a are weight matrices, α t，i is the attention weight of the i-th LSTM unit in the multiple LSTM units at time t, and N is the number of the multiple LSTM units.

10. A gearbox bearing fault diagnosis device, characterized in that: include: a vibration signal acquisition module, configured to acquire a target vibration signal of the gearbox bearing, wherein the target vibration signal includes a plurality of target vibration data; The fault prediction module is configured to input the multiple target vibration data contained in the target vibration signal into the fault diagnosis model trained according to any one of the training methods in claims 1 to 4, so as to predict the fault type corresponding to the target vibration signal.

Images