Steering frame unknown fault diagnosis method based on generative model zero sample learning strategy

Abstract

The invention relates to the technical field of high-speed train bogie fault detection, and discloses a bogie unknown fault diagnosis method based on a generative model zero sample learning strategy, and the method comprises the steps: carrying out a bogie mechanical fault simulation experiment on an electric locomotive through employing multi-body dynamics analysis software SIMPACK, and collecting normal and fault data; establishing a fault attribute semantic matrix for the train operation influence factors; performing feature extraction according to the data attributes and the fault data; building an unknown fault data feature sample generation model based on a diffusion model, and performing unknown fault data feature generation by using random noise and unknown fault attribute combination; and a fault diagnosis classification model is constructed, and mapping from data features to fault attributes and then to fault categories is realized. According to the method, diagnosis of known faults and unknown faults of the high-speed train bogie is completed, the accuracy and generalization of fault diagnosis are improved, and comprehensive guarantee is provided for safe use of the high-speed train bogie.

Description

Technical Field

[0001] The present invention relates to the technical field of high-speed train bogie fault detection, and in particular to a bogie unknown fault diagnosis method based on a generative model zero-sample learning strategy. Background Art

[0002] With the acceleration of urbanization and economic globalization, high-speed trains play an increasingly important role in improving regional connectivity and promoting economic development. However, the characteristics of high-speed travel also make the reliability and safety requirements of the train system extremely high, which highlights the importance of high-speed train fault diagnosis research. As a key component of high-speed trains, bogies are responsible for the vibration reduction suspension system of the train, which directly affects the running stability and riding comfort of the train. Its failure may seriously affect the operation of the train. The key parts of the bogie include the key vibration reduction components air spring AS, the secondary lateral damper LD and the anti-snaking damper AD used to reduce the vibration between the wheel and the body. The fault diagnosis of the high-speed train bogie vibration reduction system is mainly carried out around these vibration reduction components. The air spring is a vibration reduction component that uses air as an elastic medium. Its material may degrade due to long-term use and external environmental factors, resulting in reduced elasticity, cracking and leakage; the lateral damper and anti-snaking damper are mainly used to reduce the lateral movement and snaking movement of the train body. Its components may crack and leak oil due to long-term high-speed operation, and then fail.

[0003] Existing bogie fault diagnosis methods can be roughly divided into model-based and data-based. The basic idea of ​​the former is to use mathematical models to calculate the residuals under the system operation state, and to perform fault diagnosis by analyzing these residuals. However, in practical applications, the nonlinearity and complexity of the bogie system limit the accuracy of the analytical model, which makes the model-based diagnosis method have certain limitations. The latter data-based fault diagnosis method mainly uses various sensor data collected during the operation of the system to deeply explore the equipment status information and fault characteristics implicit in the data. Through advanced data analysis and machine learning technology, fault modes can be effectively identified and potential fault trends can be predicted. At present, the field of bogie fault diagnosis is developing in a data-driven direction. Combined with advanced data analysis technology, it can effectively overcome the limitations of model-based methods and improve the accuracy and timeliness of bogie fault diagnosis.

[0004] Although data-driven methods provide many conveniences, they also face major challenges in actual working conditions. The effectiveness of data-driven methods depends largely on the quality of input data. If there are too many missing or errors in the data, this will directly affect the accuracy of fault diagnosis. For the engineering practice of high-speed train bogie faults, the data samples of normal operation are often far more than the fault samples, and the probability of a single fault is often much higher than the probability of a compound fault. However, unbalanced data has a great impact on data-driven fault diagnosis. Unbalanced data can easily cause the weight drift of the training model, causing the model to tend to predict the fault type with a large number of data samples. For different speed ranges and different fault locations, how to accurately identify the fault category will become the top priority of the problem. However, for fault types that have never appeared, how to use machine learning and deep neural networks to help improve data balance and improve the accurate identification of specific categories of fault diagnosis has become a key difficulty. Zero-shot learning strategy can help complete the task of unknown class identification. This is a strategy that uses auxiliary information to help achieve cross-modal knowledge transfer and infer unknown class information. In a general sense, the zero-shot learning strategy is to distinguish between unknown and known classes based on auxiliary information such as known class data and attributes. Zero-shot learning strategies can currently be divided into three categories: attribute prediction-based methods, embedding space learning methods, and generative model-based methods. Currently, most of the work in the field of fault diagnosis revolves around attribute prediction and embedding space. However, these two methods are easily affected by data imbalance, and bogie research problems do not require rich domain knowledge. On the contrary, there are relatively few studies on generative model-based methods that can help balance data and have good generalization and robustness performance.

[0005] From the above background, we can clearly see that the three key points that must be solved when using data-driven methods for bogie fault diagnosis are: (1) The method must be able to help high-speed train bogies with numerous components to accurately identify unknown and known faults, replacing manual diagnosis to achieve intelligent diagnosis; (2) The method can well balance the actual engineering situation of data imbalance and have good generalization and robustness; (3) The method model needs to have both accuracy and stability, and can integrate information source data from multiple source sensors, adapt to the complexity and dynamic characteristics of the system, and ensure stable operation of the train. Summary of the invention

[0006] The purpose of the present invention is to provide a bogie unknown fault diagnosis method based on a generative model zero-sample learning strategy to complete the diagnosis of known and unknown faults of high-speed train bogies, improve the accuracy and generalization of fault diagnosis, and provide comprehensive protection for the safe use of high-speed train bogies.

[0007] In order to achieve the above purpose, the technical solutions adopted are as follows: A bogie unknown fault diagnosis method based on a generative model zero-sample learning strategy, the method comprising: Obtain train operation data; According to the fault location and running speed of the train bogie, a fault attribute semantic matrix of the train bogie is constructed, and based on the fault attribute semantic matrix, the train running data is sampled to obtain sampled data, wherein the sampled data includes known class fault data and unknown class fault data; The known class fault data and the unknown class fault data are sampled by using a residual feature extraction network to obtain known class fault data features and unknown class fault data features; Constructing an unknown fault data feature generation model based on a diffusion model, training the unknown fault data feature generation model using random noise and known class fault data features, inputting the unknown class fault data features into the trained unknown fault data feature generation model to obtain unknown class generated fault data features; A fault diagnosis classification model is constructed, and the fault diagnosis classification model is trained using the known class fault data features and the unknown class generated fault data features, and the unknown bogie fault diagnosis is realized based on the trained fault diagnosis classification model.

[0008] Furthermore, train operation data is obtained, including: Establishing a train model, wherein the train model includes a car body, two bogies and four wheel pairs, and the train model has multiple degrees of freedom; The LMA economical tread, the first track and the second track are used as model excitations, and the train operation data is obtained based on the set sampling frequency.

[0009] Furthermore, the fault attribute semantic matrix of the train bogie is classified according to the fault location and the operating speed condition. For the fault location, a fault occurs at the corresponding location and is represented by 1, while no fault occurs at the corresponding location and is represented by 0. For the operating speed condition, the train operating speed is divided into multiple non-overlapping speed intervals, and operating in the corresponding speed interval is represented by 1, while not operating in the corresponding speed interval and is represented by 0.

[0010] Furthermore, the residual feature extraction network is constructed by the following method: Building a one-dimensional convolution feature-based one-time extraction module; wherein the one-dimensional convolution feature-based one-time extraction module includes a first convolution layer and a second convolution layer, and a convolution operation is performed through the first convolution layer and the second convolution layer to extract structural features of adjacent domains; Two residual blocks are constructed. Based on the extracted structural features of the adjacent domains, the two residual blocks are used alternately twice to extract deep fault mode features and directly mapped to connect different layers of the network. An adaptive average pooling layer applied to one-dimensional signals is constructed, wherein the adaptive average pooling layer adjusts the deep fault model features inputted by the Yongjiang River to a set output size to obtain symmetrical low-dimensional data features; wherein the symmetrical low-dimensional data features include known class fault data features and unknown class fault data features.

[0011] Furthermore, the unknown fault data feature generation model based on the diffusion model includes a noise prediction model, an encoding module and a decoding module using jump connections; wherein the noise prediction model includes two embedding layers and a fusion layer, the two embedding layers are used to convert the noise step and attribute coding label features to facilitate the fusion of features, and the fusion layer is used to concatenate and fuse the features with the noise step features and the attribute coding label features to obtain the fusion features; the encoding module includes a Conv convolution layer, a ReLU activation function layer and a MaxPool pooling layer for extracting features, and the decoding module includes a ConvTrans deconvolution layer and a ReLU activation function layer for upsampling.

[0012] Furthermore, the unknown fault data feature generation model is trained using random noise and known class fault data features, and the unknown class fault data features are input into the trained unknown fault data feature generation model to obtain the unknown class generated fault data features, including: Gradually in the original data Add noise to get noisy data , and gradually transition from a clear state to a completely random noise state, forming a Markov chain:

[0013]

[0014] In the formula, Indicates State data for the time step, Indicates State data for the time step, represents a normal distribution, Indicates The noise parameter of the time step, represents the identity matrix, Indicates Time step given of The Gaussian distribution of , and the variance is , Indicates from the 1st to the All state sequences of time steps, Indicates from The entire sequence generated starts with { , , ..., } joint probability distribution, represents the index of the current time step, represents the total number of time steps, Indicates a given hour The conditional probability of make and , after reparameterization and anti-reparameterization, we get:

[0015]

[0016] In the formula, Indicates The parameters of the time step, From time step 1 to all The product of and For combination and noise The coefficient of represents the noise term sampled from the standard normal distribution; The training process learns the noise prediction function at each step. The sampling process starts with sampling from the standard Gaussian distribution, subtracts the predicted noise, and iterates continuously to finally generate the unknown class fault data features.

[0017] Furthermore, we start sampling from the standard Gaussian distribution, subtract the prediction noise, and iterate continuously to finally generate the unknown class fault data feature representation:

[0018]

[0019]

[0020]

[0021]

[0022] In the formula, Indicates that the model has parameters The joint probability distribution under covers arrive The whole process, Indicates the 0th to All state sequences of time steps, Represents the time step Status The initial probability distribution of Indicates that at time step Given In the case of The conditional probability distribution of Represents the time step Given And its prediction and variance Gaussian distribution under Indicates that based on parameters Calculation time step The mean of represents the parameters of the model, Represents the time step Parameters, Indicates that at time step The noise prediction function under Indicates that at time step The covariance matrix under Indicates that at time step The standard deviation parameter under represents the noise term sampled from a standard Gaussian distribution, Indicates that according to the time step Status The previous state generated.

[0023] Furthermore, the fault diagnosis classification model is trained using the known class fault data features and the unknown class generated fault data features, and binary cross entropy is used as a loss function.

[0024] Furthermore, the fault diagnosis classification model includes an integration layer, a fully connected layer and an attribute layer; wherein the integration layer is used to integrate the known class fault data features and the unknown class generated fault data features, and output feature information to the fully connected layer, and the fully connected layer is used to map the feature information to the attribute label space to perform data attribute prediction and fault classification, and the output of the fully connected layer is expressed as:

[0025] In the formula, Indicates The output of the layer, Indicates The input of the layer, f represents the activation function, D i 、b i Both represent learning parameters and are The weight matrices and bias terms in the layers; The attribute layer is used to convert the output of the fully connected layer into a probability value of the corresponding category.

[0026] Furthermore, the attribute layer is used to convert the output of the fully connected layer into a probability value of the corresponding category through the following formula:

[0027] In the formula, softmax represents the softmax function, which is used to convert the output of the fully connected layer into the probability of each category, softmax(y) i The output of the function The probability value of each category, Indicates The original output value of the class, Indicates The original output value of the class, Indicates Exponentiation of the raw output values ​​of the class, Indicates The index of the original output value of the class, j represents the category index, i represents the current category index, and n represents the total number of categories.

[0028] The beneficial effects of the present invention are: 1. The present invention builds a bogie-specific attribute description matrix by introducing a generalized zero-sample strategy.

[0029] 2. The present invention innovatively applies the diffusion generative model, which has significant effects in the image field, to low-dimensional data generation, which can significantly improve the adaptability and generation ability of the model in generalized zero samples, enhance the generalization ability of the model and enrich the generated data set.

[0030] 3. The present invention designs a generalized zero-sample fault diagnosis problem framework based on the conditional diffusion generative model, improves the residual network to extract the low-dimensional data features of the bogie, and completes the ternary mapping relationship of data-attribute-feature. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A flowchart of a bogie unknown fault diagnosis method based on a generative model zero-sample learning strategy according to an embodiment of the present invention is shown.

[0032] Figure 2 A schematic diagram of a high-speed train bogie fault-specific semantic matrix according to an embodiment of the present invention is shown.

[0033] Figure 3 A schematic diagram of a residual network feature extraction framework according to an embodiment of the present invention is shown.

[0034] Figure 4 A schematic diagram of the diffusion model principle according to an embodiment of the present invention is shown.

[0035] Figure 5 A schematic diagram of a diffusion model noise learning framework according to an embodiment of the present invention is shown.

[0036] Figure 6 A schematic diagram of a fault diagnosis confusion matrix of all categories according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0037] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0038] The specific implementation of the present invention is further described in detail below in conjunction with the drawings and examples.

[0039] The embodiment of the present invention provides a bogie unknown fault diagnosis method based on a generative model zero-sample learning strategy, such as Figure 1 As shown, the method comprises the following steps: S1. Obtain train operation data.

[0040] In some embodiments, step S1 specifically includes: S11, establishing a train model, the train model includes a car body, two bogies and four wheel pairs, and the train model has multiple degrees of freedom; S12. Using the LMA economical tread, the first track and the second track as model excitations, train operation data is obtained based on a set sampling frequency.

[0041] In this embodiment, since the faulty bogie cannot run at high speed for a long time for data collection, the experimental data is collected by simulation experiment through SIMPACK. SIMPACK is a multi-body dynamics analysis software, which is widely used to simulate and analyze the dynamic performance of mechanical and electromechanical systems. The train model consists of a car body, two bogies and four wheel pairs, with 62 degrees of freedom. The nonlinearity of wheel-rail contact is fully considered. The LMA economic tread, CN60 track and Wuhan-Guangzhou line track (i.e., the first track and the second track) are set as model excitations. The sampling frequency is set to 243Hz to obtain experimental data. 220s of data of each category are collected. Considering that there is more noise in the initial stage of operation, the first 3 seconds of data are removed. 58 sensors are installed on the bogie, so the size of each 1s data information is 243*58. This simulation method can not only accurately simulate the dynamic performance of the bogie under normal and fault conditions, but also test various hypothetical fault conditions without affecting the actual operation safety and efficiency. Repeat the experiment under different operating conditions and fault scenarios to ensure that a comprehensive data set is obtained. These data are essential for subsequent analysis and training of fault diagnosis models.

[0042] S2. According to the fault location and running speed on the train bogie, a fault attribute semantic matrix of the train bogie is constructed. Based on the fault attribute semantic matrix, the train running data is sampled to obtain sampled data, and the sampled data includes known class fault data and unknown class fault data.

[0043] In this embodiment, a fault attribute semantic matrix dedicated to the high-speed train bogie is constructed according to the main fault locations and running speeds on the high-speed train bogie. The designed semantic matrix is ​​as follows: Figure 2 As shown. The table mainly classifies according to the fault location and operating speed conditions. For the fault location, a number is used to indicate whether the fault occurs or not, where 1 indicates that the corresponding location has a fault, and 0 indicates that the location has no fault. The fault may be a single component failure or a composite component failure, but since the probability of failure of all three parts is almost 0, the three-part composite fault is not studied. For the normal high-speed train operating speed range of 150km / h to 210km / h, it is evenly divided into three speed ranges, where 1 indicates running in this speed range, and 0 indicates not running in this speed range.

[0044] For the operation data obtained in step S1, if the corresponding category can be found in the fault attribute semantic matrix, it is determined to be known type fault data; if the corresponding category cannot be found in the fault attribute semantic matrix, it is determined to be unknown type fault data.

[0045] S3. Use the residual feature extraction network to sample known class fault data and unknown class fault data to obtain known class fault data features and unknown class fault data features.

[0046] In this embodiment, due to the characteristics of the diffusion model network itself, the asymmetric 243*58 size fault data is not conducive to direct generation. In order to further improve the diagnostic accuracy of the fault diagnosis framework, dig deep into the fault modal information hidden behind the data, and facilitate the diffusion model to generate data, a feature extraction network is designed to extract data fault features. The generally used layered deep convolutional network uses multiple layers of convolutional layers and pooling layers to achieve a close relationship between the extracted features and the high-level semantics of the data. Multi-layer networks bring about network degradation problems. At this time, the training effect of the shallow network is much better than that of the deep network. At the same time, the data information contained in the network decreases as the number of layers increases. Direct mapping is used to connect different layers of the network, and a residual network is introduced to solve the problem. The data is passed through the following channels: Figure 3 The network architecture shown in the figure designs two types of residual blocks. The mapping of the residual blocks effectively reduces the learning difficulty and speeds up the convergence of the model. The learning rate of the feature extraction network framework is adjusted by the Adam optimizer with an initial learning rate of 1e-3. The batch size of data training is 100 and the number of iterations is 150. The final classification network is based on the residual network, adding a fully connected layer, a linear mapping layer, and a Softmax layer for classification.

[0047] In some embodiments, the residual feature extraction network is constructed by the following steps: S31. Build a one-dimensional convolution feature extraction module. The input data size is 243 sampling frequency * 58 sensor channels. The first part of the residual feature extraction network is the convolution layer C1 and C2, which perform convolution operations to extract structural features of adjacent domains.

[0048] S32. Combining the residual connection principle, two residual blocks are used alternately twice to extract deep fault mode features, and directly mapped to connect different layers of the network to solve the network degradation problem. This step can obtain a feature map.

[0049] S33. Finally, build an adaptive average pooling layer applied to one-dimensional signals to adjust the input feature map to the specified output size, which helps the output features to be symmetrical low-dimensional data features, making it easier to generate subsequent diffusion models.

[0050] S4. Construct an unknown fault data feature generation model based on a diffusion model, use random noise and known class fault data features to train the unknown fault data feature generation model, input the unknown class fault data features into the trained unknown fault data feature generation model, and obtain the unknown class generated fault data features.

[0051] In some embodiments, step S4 specifically includes the following steps: S41. Build a noise prediction model based on Unet (semantic segmentation network). First, use two embedding layers to convert the noise step and attribute encoding label features to facilitate feature fusion. Then use the fusion (Cat) layer to concatenate and fuse the features with the noise step features and attribute encoding label features as model input.

[0052] S42. Build a two-layer encoding module and decoding module with a Unet structure using jump connections. The encoding module includes a Conv convolution layer, a ReLU activation function layer, and a MaxPool pooling layer for feature extraction. The decoding module includes a ConvTrans deconvolution layer and a ReLU activation function layer for upsampling.

[0053] S43. Use random noise and the attributes and low-dimensional features of known fault classes to train a noise prediction model.

[0054] It should be noted that the known class fault data features include the attributes of the known fault class and low-dimensional features; among them, the attributes of the known fault class are determined according to the fault attribute semantic matrix, and the low-dimensional features are extracted by the residual feature extraction network.

[0055] S44: Generate unknown class fault data features using random noise and unknown class fault attributes and the trained noise prediction model.

[0056] In some embodiments, a diffusion model-based unknown fault data feature generation model is constructed by the following method: Build a diffusion model, the specific framework and working principle are as follows Figure 4 As shown in the figure. The diffusion model is a generative model that has performed well in the deep generative model, especially in image generation. Its structure mainly includes two processes, the forward diffusion process and the reverse generation process, both of which are parameterized Markov chains. The forward diffusion process gradually adds Gaussian noise to the data features until the data becomes random noise. The model gradually Add noise to the data and transform it into noise data close to Gaussian distribution through multiple steps. , the data features gradually transition from a clear state to a completely random noise state, forming a Markov chain:

[0057]

[0058] In the formula, Indicates The state data of the time step, Indicates The state data of the time step, represents a normal distribution, Indicates The noise parameter of the time step, represents the identity matrix, Indicates Time step given of The Gaussian distribution of , and the variance is , Indicates from the 1st to the All state sequences of time steps, Indicates from The entire sequence generated begins with { , ,..., } joint probability distribution, represents the index of the current time step, represents the total number of time steps, Indicates a given hour The conditional probability of .

[0059] definition and , through the reparameterization technique and anti-reparameterization, we can get:

[0060]

[0061] In the formula, Indicates The parameters of the time step, From time step 1 to all The product of and For combination and noise The coefficient of represents the noise term sampled from a standard normal distribution.

[0062] The derived formula reflects Based on the original data and random noise Linear combination for sampling, and and Also called the combination coefficient.

[0063] In contrast to the diffusion process of adding noise, the generation process (reverse) is a denoising process. The process of gradually denoising from random noise is also a Markov chain, which is a series of Gaussian distributions parameterized by neural networks.

[0064]

[0065]

[0066]

[0067]

[0068] In the above formula ,and is a parameterized Gaussian distribution, the mean and variance can be obtained by training the network Learning is to add Gaussian noise to the posterior distribution Replaced with noise prediction function , evolutionary variance No need to learn. Finally, we get:

[0069] In the formula, Indicates that the model has parameters The joint probability distribution under covers arrive The whole process, Indicates the 0th to All state sequences of time steps, Represents the time step Status The initial probability distribution of Indicates that at time step Given In the case of The conditional probability distribution of Represents the time step Given And its prediction and variance Gaussian distribution under Indicates that based on parameters Calculation time step The mean of represents the parameters of the model, Represents the time step Parameters, Indicates that at time step The noise prediction function under Indicates that at time step The covariance matrix under Indicates that at time step The standard deviation parameter under represents the noise term sampled from a standard Gaussian distribution, Indicates that according to the time step Status The previous state generated.

[0070] In summary, the training process learns the noise prediction function at each step The sampling process starts with the standard Gaussian distribution, then subtracts the prediction noise, and iterates continuously to generate new samples.

[0071] The unknown fault data feature generation model based on the diffusion model is essentially a noise learning model. It is based on the Unet architecture. Unet is an effective convolutional neural network that is often used in the field of image segmentation. Its characteristics are that it can effectively capture and utilize multi-scale context information through the symmetrical structure of the encoder and decoder, plus the jump connection. The specific network design is as follows Figure 5 As shown in the framework. In this embodiment, the network is used to generate symmetric low-dimensional fault data features. The input of the model is processed by the embedding layer, which includes two parts: one is the encoding of the noise adding step, and the other is the encoding of the fault attribute. The noise adding step encoding is used to represent the collective specific stage in the diffusion process. Each noise adding step corresponds to a time point in the diffusion model, which identifies the degree of transition from the noise-free state to the full noise state. This encoding helps the model understand the noise level of the current data so that the denoising process can be adjusted appropriately. The attribute encoding defines the specific characteristics of the fault, such as the fault location and running speed. These attribute auxiliary information is used to guide the addition of noise and the recovery of characteristics of different fault types to ensure that the generated data meets the expectations of the fault characteristics. These encoded features are spliced ​​through the Cat layer to ensure that information from different sources can be effectively integrated in the model.

[0072] In the encoding module of the model learning phase, the convolution layer (Conv) and the pooling layer (MaxPool) work together to extract the deep features of the input data. The ReLU activation function provides nonlinear processing in this process, helping the model capture more complex data patterns. Next, the data flows to the decoding module, which includes a deconvolution layer (ConvTrans) to upsample the feature map to a higher resolution. The ReLU activation function is also used in this process to ensure the nonlinear expression of the features. The values ​​are determined according to the linear range between (0.00001, 0.005), and the total number of noise addition steps is The loss function is set to 50, and the mean square error (MSE) is used as the loss function to measure the difference between the predicted noise and the actual noise at each step. The optimizer uses the stochastic gradient descent (SGD) method with periodic decay. A total of 150,000 training iterations are set. The initial learning rate is 0.01, and the learning rate is adjusted by multiplying it by 0.9 every 5000 epochs.

[0073] Through such an encoding-decoding process, Unet is able to reconstruct fault features with specific attributes from preliminary noise data. The training of this process relies on the attributes of known fault classes and low-dimensional feature data. After training, the model can accept new random noise inputs, combined with the attribute labels of unknown fault classes, to generate corresponding fault data. These synthetic data not only enhance the diversity of the training set, but also help the model learn how to accurately diagnose faults in unseen fault situations, effectively expanding the capabilities of the fault diagnosis system so that it can handle a wider range of fault scenarios, thereby improving the overall performance and reliability of the system.

[0074] S5. Construct a fault diagnosis classification model, use the known class fault data features and the unknown class generated fault data features to train the fault diagnosis classification model, and realize the unknown fault diagnosis of the bogie based on the trained fault diagnosis classification model.

[0075] In this embodiment, the generalized zero-shot learning problem is converted into a supervised classification problem, a fault diagnosis classification model is built, and the fault data features generated by the unknown class are used for training with the real data features of the known class. The low-dimensional features are converted into one-dimensional features using the integration (Flatten) layer, and finally the extracted feature information is converted into the attribute label space through the fully connected layer to complete the data classification. The final output data is converted into the probability value of the corresponding category through softmax to predict the fault class.

[0076] Specifically, based on the fault diagnosis classification model, the fault features of the unknown fault class generated by the generation model are combined with the fault features of the known fault class to perform attribute prediction and category training on the final fault. The fault diagnosis classification model first integrates the features extracted by the model through the Flatten layer, and maps the extracted feature information to the attribute label space through the fully connected layer to perform data attribute prediction and fault classification. Its output can be expressed as:

[0077] In the formula, Indicates The output of the layer, Indicates The input of the layer, f represents the activation function, D i 、b i Both represent learning parameters and are The weight matrices and bias terms in the layer.

[0078] The last layer of the model is a Dense layer with 6 neurons, and uses the sigmoid activation function to output 6 values, each corresponding to the predicted value of an attribute. The range of these 6 output values ​​is between 0 and 1, and the binary cross entropy is used as the loss function, which is suitable for multi-label classification tasks or binary classification tasks of multiple attributes. The Adam optimizer is used for optimization to evaluate the accuracy, and the training is performed for 100 rounds and the monitoring loss is set to trigger a 10-fold reduction in the learning rate adjustment if there is no improvement within 5 training rounds.

[0079] Finally, the attribute layer is converted into the probability value of the corresponding category by softmax, and its expression is:

[0080] In the formula, softmax represents the softmax function, which is used to convert the output of the fully connected layer into the probability of each category, softmax(y) i The output of the function The probability value of each category, Indicates The original output value of the class, Indicates The original output value of the class, Indicates Exponentiation of the raw output values ​​of the class, Indicates The index of the original output value of the class, j represents the category index, i represents the current category index, and n represents the total number of categories.

[0081] The confusion matrix is ​​one of the important tools for evaluating the performance of classification models. It is a square matrix that shows the comparison between the predicted results of the classification model and the actual results. Each row of the matrix represents the actual category, and each column represents the predicted category. Through the confusion matrix, we can understand the classification of the model in each category and identify the advantages and disadvantages of the model. Four groups were randomly selected as unknown faults, and the others were known faults. The diagnostic accuracy of known class faults was higher than 95%, and the accuracy of unknown class known faults was more than 70%. The harmonic mean of the two exceeded 80%. The specific diagnosis results are shown in Figure 6 .

[0082] In summary, the overall solution of the present invention has a complete diagnostic process and provides an effective method for detecting unknown faults of high-speed train bogies.

[0083] In view of the incomplete data of unknown faults, a dedicated bogie fault attribute knowledge base is constructed, combining the advantages of the diffusion generative model network and the generalized zero-shot strategy based on the generative model, deeply mining the hidden fault mode, and using the deep neural network to learn the mapping relationship between fault data, fault characteristics, fault attributes, and fault categories, taking into account accuracy, generalization, and robustness. The visualization results of the confusion matrix show that the designed diagnostic network framework is very sensitive to the fault characteristics of the bogie, and can effectively distinguish the known and unknown faults of the high-speed train bogie, meeting the diagnostic requirements.

[0084] The above implementation modes are only used to illustrate the present invention, but not to limit the present invention. Ordinary technicians in the relevant technical field can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions also belong to the scope of the present invention. The patent protection scope of the present invention should be defined by the claims.

Claims

1. A bogie unknown fault diagnosis method based on a generative model zero-shot learning strategy, characterized in that: The method comprises: Obtain train operation data; According to the fault location and running speed of the train bogie, a fault attribute semantic matrix of the train bogie is constructed, and based on the fault attribute semantic matrix, the train running data is sampled to obtain sampled data, wherein the sampled data includes known class fault data and unknown class fault data; The known class fault data and the unknown class fault data are sampled by using a residual feature extraction network to obtain known class fault data features and unknown class fault data features; Constructing an unknown fault data feature generation model based on a diffusion model, training the unknown fault data feature generation model using random noise and known class fault data features, inputting the unknown class fault data features into the trained unknown fault data feature generation model to obtain unknown class generated fault data features; A fault diagnosis classification model is constructed, and the fault diagnosis classification model is trained using the known class fault data features and the unknown class generated fault data features, and the unknown bogie fault diagnosis is realized based on the trained fault diagnosis classification model.

2. The bogie unknown fault diagnosis method based on the generative model zero-sample learning strategy according to claim 1 is characterized in that: Obtain train operation data, including: Establishing a train model, wherein the train model includes a car body, two bogies and four wheel pairs, and the train model has multiple degrees of freedom; The LMA economical tread, the first track and the second track are used as model excitations, and the train operation data is obtained based on the set sampling frequency.

3. The bogie unknown fault diagnosis method based on the generative model zero-sample learning strategy according to claim 1 is characterized in that: The fault attribute semantic matrix of the train bogie is classified according to the fault location and the operating speed condition. For the fault location, a fault at the corresponding location is represented by 1, and a no fault at the corresponding location is represented by 0. For the operating speed condition, the train operating speed is divided into multiple non-overlapping speed intervals. Running in the corresponding speed interval is represented by 1, and not running in the corresponding speed interval is represented by 0.

4. The bogie unknown fault diagnosis method based on the generative model zero-sample learning strategy according to claim 1, characterized in that: The residual feature extraction network is constructed by the following method: Building a one-dimensional convolution feature-based one-time extraction module; wherein the one-dimensional convolution feature-based one-time extraction module includes a first convolution layer and a second convolution layer, and a convolution operation is performed through the first convolution layer and the second convolution layer to extract structural features of adjacent domains; Two residual blocks are constructed. Based on the extracted structural features of the adjacent domains, the two residual blocks are used alternately twice to extract deep fault mode features and directly mapped to connect different layers of the network. An adaptive average pooling layer applied to one-dimensional signals is constructed, wherein the adaptive average pooling layer adjusts the deep fault model features inputted by the Yongjiang River to a set output size to obtain symmetrical low-dimensional data features; wherein the symmetrical low-dimensional data features include known class fault data features and unknown class fault data features.

5. The bogie unknown fault diagnosis method based on the generative model zero-sample learning strategy according to claim 1, characterized in that: The unknown fault data feature generation model based on the diffusion model includes a noise prediction model, an encoding module using jump connections, and a decoding module; wherein the noise prediction model includes two embedding layers and a fusion layer, the two embedding layers are used to convert the noise step and attribute coding label features to facilitate the fusion of features, and the fusion layer is used to concatenate and fuse the features with the noise step features and the attribute coding label features to obtain the fusion features; the encoding module includes a Conv convolution layer, a ReLU activation function layer, and a MaxPool pooling layer for extracting features, and the decoding module includes a ConvTrans deconvolution layer and a ReLU activation function layer for upsampling.

6. The bogie unknown fault diagnosis method based on the generative model zero-sample learning strategy according to claim 5, characterized in that: The unknown fault data feature generation model is trained using random noise and known class fault data features, and the unknown class fault data features are input into the trained unknown fault data feature generation model to obtain the unknown class generated fault data features, including: Gradually in the original data Add noise to get noisy data , and gradually transition from a clear state to a completely random noise state, forming a Markov chain: In the formula, Indicates State data for the time step, Indicates State data for the time step, represents a normal distribution, Indicates The noise parameter of the time step, represents the identity matrix, Indicates Time step given of The Gaussian distribution of , and the variance is , Indicates from the 1st to the All state sequences of time steps, Indicates from The entire sequence generated starts with { , , ..., } joint probability distribution, represents the index of the current time step, represents the total number of time steps, Indicates a given hour The conditional probability of make and , after reparameterization and anti-reparameterization, we get: In the formula, Indicates The parameters of the time step, From time step 1 to all The product of and For combination and noise The coefficient of represents the noise term sampled from the standard normal distribution; The training process learns the noise prediction function at each step. The sampling process starts with sampling from the standard Gaussian distribution, subtracts the predicted noise, and iterates continuously to finally generate the unknown class fault data features.

7. The bogie unknown fault diagnosis method based on the generative model zero-sample learning strategy according to claim 6, characterized in that: Sampling starts from the standard Gaussian distribution, subtracting the prediction noise, and iterating continuously to finally generate the unknown class to generate the fault data feature representation: In the formula, Indicates that the model has parameters The joint probability distribution under covers arrive The whole process, Indicates the 0th to All state sequences of time steps, Represents the time step Status The initial probability distribution of Indicates that at time step Given In the case of The conditional probability distribution of Represents the time step Given And its prediction and variance Gaussian distribution under Indicates that based on parameters Calculation time step The mean of represents the parameters of the model, Represents the time step Parameters, Indicates that at time step The noise prediction function under Indicates that at time step The covariance matrix under Indicates that at time step The standard deviation parameter under represents the noise term sampled from a standard Gaussian distribution, Indicates that according to the time step Status The previous state generated.

8. The bogie unknown fault diagnosis method based on the generative model zero-sample learning strategy according to claim 1, characterized in that: The fault diagnosis classification model is trained using the known class fault data features and the unknown class generated fault data features, and binary cross entropy is used as a loss function.

9. The bogie unknown fault diagnosis method based on the generative model zero-sample learning strategy according to claim 1, characterized in that: The fault diagnosis classification model includes an integration layer, a fully connected layer and an attribute layer; wherein the integration layer is used to integrate the known class fault data features and the unknown class generated fault data features, and output feature information to the fully connected layer, and the fully connected layer is used to map the feature information to the attribute label space to perform data attribute prediction and fault classification. The output of the fully connected layer is expressed as: In the formula, Indicates The output of the layer, Indicates The input of the layer, f represents the activation function, D i , b i Both represent learning parameters and are The weight matrices and bias terms in the layers; The attribute layer is used to convert the output of the fully connected layer into a probability value of the corresponding category.

10. The bogie unknown fault diagnosis method based on generative model zero-sample learning strategy according to claim 1, characterized in that: The attribute layer is used to convert the output of the fully connected layer into the probability value of the corresponding category through the following formula: In the formula, softmax represents the softmax function, which is used to convert the output of the fully connected layer into the probability of each category, softmax(y) i The output of the function The probability value of each category, Indicates The original output value of the class, Indicates The original output value of the class, Indicates Exponentiation of the raw output values ​​of the class, Indicates The index of the original output value of the class, j represents the category index, i represents the current category index, and n represents the total number of categories.

Images