A few-sample fault diagnosis method and device suitable for core components of a rail train

Abstract

The application discloses a few-sample fault diagnosis method and device suitable for core components of a rail train. Running signals of core components under current working conditions are collected from a power system of the rail train, and preprocessed running signals are obtained, and then input into a fault diagnosis model enhanced by improved DDPM model training for fault diagnosis. Since the improved DDPM model fuses a multi-scale feature extraction mechanism, simultaneously introduces a time sequence attention mechanism to establish a long-term dependence relationship, and takes a fault physical mechanism as a constraint condition, the physical rationality of generated samples is ensured. Based on the diversified and high-fidelity training samples generated, the fault diagnosis model adopts a multi-scale pyramid convolution network to perform hierarchical feature extraction, uses a time sequence attention memory network to model a fault evolution process, separates common features and private features through a feature decoupling reconstruction network, combines a graph reasoning network to mine a fault propagation path, and finally realizes accurate diagnosis under cross-working conditions.

Description

Technical Field

[0001] This application belongs to the field of fault diagnosis technology for core components of rail transit, and specifically relates to a method and device for small-sample fault diagnosis of core components of rail trains. Background Technology

[0002] In the field of rail transit, gearboxes and rolling bearings in the traction system, running gear, and transmission mechanism are core components ensuring the safe and stable operation of trains. Their operating status directly affects the reliability, safety, and operational efficiency of the entire train set and even the line. Gearboxes bear complex dynamic loads from the traction motor for extended periods, while rolling bearings are continuously subjected to high-frequency alternating stress. During long-term service, both are prone to abnormal vibration and noise due to wear, fatigue, pitting, cracks, and other faults. In severe cases, this can lead to transmission failure or even safety accidents, causing operational interruptions, soaring maintenance costs, and expanded social impact. Therefore, research on efficient, accurate, and reliable intelligent fault diagnosis technologies applicable to rail transit scenarios has extremely important engineering value and social significance.

[0003] With the development of artificial intelligence and data-driven technologies, intelligent fault diagnosis has become a research hotspot in the field of rail transit operation and maintenance. Currently, traditional methods such as Support Vector Machines (SVM), Convolutional Neural Networks (CNN), and Long Short-Term Memory Networks (LSTM) based on vibration, acoustic, or current signals have been attempted to be applied to the state recognition and fault classification of key rotating components such as locomotives and EMU trains. However, the actual operation and maintenance of rail transit systems faces the severe challenge of "scarcity of fault samples"—due to the high reliability of core train components, the actual number of collected fault samples is far less than the number of normal samples, and historical data available for the same type of fault is often only a few dozen or even a handful, forming a typical small-sample, imbalanced data scenario. To address this problem, Generative Adversarial Networks (GANs) and their variants have been introduced to expand the sample size. However, practice shows that when there are very few single-class fault samples (e.g., less than 20), such models are prone to problems such as training instability and pattern collapse, making it difficult to guarantee the quality and diversity of generated samples, thus restricting the generalization ability of the diagnostic model.

[0004] In recent years, the Denoising Diffusion Probability Model (DDPM) has shown potential in few-sample fault diagnosis due to its training stability and high-fidelity generated samples. Related research, such as the improved single-graph diffusion model by Cao Jie's team and the "DDPM+CNN" scheme proposed by Wu Jingyuan's team, has provided new ideas for the few-sample diagnosis of gearboxes and rolling bearings. However, transferring these methods to the complex operating conditions of rail transit still has significant limitations: First, the generalization ability for few samples is still insufficient; existing diffusion models experience a decline in generation quality when the sample size is extremely low (e.g., less than 10 samples), making it difficult to support joint modeling of multiple fault modes. Second, single-scale feature extraction mechanisms cannot comprehensively capture the multi-frequency and multi-modal features of faults in rotating components of rail transit. For example, local damage to gears and outer ring faults in bearings often simultaneously trigger high-frequency impacts and low-frequency modulations; traditional CNN structures have limited ability to express such complex features. Third, the utilization of temporal dynamic information is insufficient; existing methods mostly rely on static time-frequency maps as input, ignoring vibration signals. The diagnostic model suffers from several shortcomings. First, it lacks sensitivity in early identification of slowly changing faults such as progressive gear wear and bearing fatigue propagation. Second, it has weak adaptability to various operating conditions. Different operating conditions such as traction, braking, curves, and slopes cause dynamic changes in speed and load, which in turn cause the fault characteristics to shift. However, existing diagnostic models are mostly trained under fixed operating conditions and lack cross-condition adaptive and generalization mechanisms. Third, the interpretability and credibility assessment of diagnostic results are lacking. In fields like rail transit where safety requirements are extremely high, simply outputting the fault category without confidence level or uncertainty quantification makes it difficult to support risk assessment and precise intervention for operation and maintenance decisions.

[0005] In summary, existing intelligent fault diagnosis methods still have significant shortcomings in addressing the demands of limited samples, multiple scales, strong temporal sequences, variable operating conditions, and high reliability in the rail transit field, thus hindering their widespread application in practical operation and maintenance. Therefore, there is an urgent need to construct new intelligent fault diagnosis methods and systems for rail transit rotating machinery that combine high precision, strong generalization, and high reliability, in order to improve early warning and accurate diagnosis capabilities and provide key technical support for safe train operation and intelligent maintenance. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this application provides a method and apparatus for few-sample fault diagnosis of core components of railway trains. The technical problem to be solved by this application is achieved through the following technical solution:

[0007] Firstly, this application provides a method for diagnosing faults in core components of rail trains using a small sample size, including:

[0008] S100: Collect the operating signals of the core components under the current working conditions from the power system of the rail train, and preprocess the operating signals to obtain preprocessed operating signals;

[0009] S200, the preprocessed running signal is input into the trained fault diagnosis model so that the fault diagnosis model can diagnose the core component and output whether the core component has failed and the type of failure when it has failed; wherein, the fault diagnosis model uses a trained improved DDPM model to generate multiple training samples, and then uses the training samples to train a preset fault diagnosis model; the improved DDPM model integrates multi-scale features in the process of generating the multiple training samples, and introduces temporal attention to establish temporal dependencies and fault information as constraints, thereby generating multiple training samples.

[0010] Secondly, this application provides a few-sample fault diagnosis device suitable for core components of rail trains, comprising:

[0011] The acquisition module is configured to acquire the operating signals of the core components of the rail train under the current operating conditions from the power system, and to preprocess the operating signals to obtain preprocessed operating signals;

[0012] A fault diagnosis module is configured to input the preprocessed running signal into a trained fault diagnosis model, enabling the model to diagnose faults in the core component and output whether the core component has failed and the type of fault if it has occurred. The fault diagnosis model uses an improved DDPM model to generate multiple training samples, which are then used to train a pre-defined fault diagnosis model. The improved DDPM model integrates multi-scale features during the generation of the multiple training samples and introduces temporal attention to establish temporal dependencies, along with fault information as a constraint, thereby generating multiple training samples. Beneficial effects:

[0013] 1. This application discloses a few-sample fault diagnosis method applicable to core components of rail trains. It involves real-time acquisition of operating signals from key components in the rail train's power system under operating conditions, followed by preprocessing to generate high-quality time-series signals. These preprocessed time-series signals are input into a fault diagnosis model enhanced by an improved DDPM model. The improved DDPM model innovatively integrates a multi-scale feature extraction mechanism, enabling it to simultaneously capture the microscopic impact details, mesoscopic structural features, and macroscopic evolution trends of faults. It introduces a temporal attention mechanism to establish long-term dependencies, accurately identifying the periodic patterns and dynamic characteristics of faults. Furthermore, it uses the physical mechanism of the fault as a constraint to ensure the physical rationality of the generated samples. Based on these diverse and high-fidelity training samples, the fault diagnosis model employs a multi-scale pyramid convolutional network for hierarchical feature extraction, uses a temporal attention memory network to model the fault evolution process, separates common and private features through a feature decoupling and reconstruction network, and combines a graph reasoning network to mine fault propagation paths, ultimately achieving accurate diagnosis under various operating conditions. This solution effectively addresses the diagnostic challenges posed by the limited number of fault samples and complex operating conditions in rail trains, significantly improving the accuracy, robustness, and interpretability of fault diagnosis.

[0014] 2. This application employs an improved DDPM model as a sample augmentation model, capable of learning data distribution from a small number of real fault samples to generate a large number of high-quality, diverse synthetic training samples. The DDPM generation process is based on probability diffusion theory, generating samples through progressive denoising to ensure consistency between the generated data and the real fault distribution. Fault information is introduced as a constraint to give the generated samples physical plausibility. Through sample augmentation, this application ensures that the fault diagnosis model encounters more comprehensive fault modes during the training phase, significantly improving its generalization ability under limited sample conditions. A multi-scale feature fusion mechanism is designed in the improved DDPM to simultaneously extract high-frequency, mid-frequency, and low-frequency features. The multi-scale feature pyramid structure ensures that the generated samples contain complete information from local impact to global evolution, enabling the fault diagnosis model to learn multi-granular representations of faults. The fault diagnosis model can simultaneously identify subtle early faults and obvious severe faults, improving its sensitivity to complex fault modes of rail trains.

[0015] 3. This application introduces a temporal attention mechanism into the improved DDPM to establish long-sequence dependencies. Attention weights are dynamically assigned to the importance of different time points, focusing on key stages of fault occurrence. Temporal modeling allows generated samples to retain the periodicity, impact intervals, and evolution trajectories of faults, making the training data contain rich temporal patterns. The fault diagnosis model can learn the temporal patterns of faults, reducing false alarms and false negatives. In the sample generation process of the improved DDPM in this application, operating condition parameters are used as condition vectors to guide sample generation, enabling samples to cover multiple combinations of operating conditions and simulate the diversity of real industrial environments. The conditional generation mechanism ensures that the training data includes fault performance under changing operating conditions, allowing the fault diagnosis model to learn the correlation between operating conditions and fault characteristics.

[0016] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a few-sample fault diagnosis method for core components of rail trains provided in this application;

[0018] Figure 2 This is a schematic diagram of the fault diagnosis model provided in this application;

[0019] Figure 3 This is a schematic diagram of the structure of the improved DDPM provided in this application;

[0020] Figure 4 This is a schematic diagram of the layer structure of the improved DDPM provided in this application;

[0021] Figure 5 This is a schematic diagram of a small-sample fault diagnosis device for core components of rail trains provided in this application. Detailed Implementation

[0022] The present application will be described in further detail below with reference to specific embodiments, but the implementation of the present application is not limited thereto.

[0023] like Figure 1 As shown, this application provides a few-sample fault diagnosis method applicable to core components of rail trains, including:

[0024] S100: Collect the operating signals of the core components under the current working conditions from the power system of the rail train, and preprocess the operating signals to obtain preprocessed operating signals;

[0025] The core components include the gearbox and rolling bearings. This step can collect operating signals when both are working together or when one is operating independently. In practice, the operating signals when both are working together are generally collected. Operating signals include system parameters such as gear transmission system parameters, transmission ratio, and meshing frequency. If it is a rolling bearing, its operating signals can be local parameters such as fault frequency and film thickness ratio. When both are working together, the operating signals are generally general parameters such as vibration and temperature.

[0026] This application can convert operating signals, such as vibration signals, into time-frequency diagrams through continuous wavelet transform (CWT), and then input them into the model for fault diagnosis.

[0027] S200, the preprocessed running signal is input into the trained fault diagnosis model so that the fault diagnosis model can diagnose the core component and output whether the core component has failed and the type of failure when it has failed; wherein, the fault diagnosis model uses a trained improved DDPM model to generate multiple training samples, and then uses the training samples to train a preset fault diagnosis model; the improved DDPM model integrates multi-scale features in the process of generating the multiple training samples, and introduces temporal attention to establish temporal dependencies and fault information as constraints, thereby generating multiple training samples.

[0028] refer to Figure 2 As shown, the fault diagnosis model of this application includes an input layer, a multi-scale pyramid convolutional network, a temporal attention memory network, a feature decoupling and reconstruction network, a graph reasoning network, and a perceptual classification network connected in sequence. The input layer is used to input a preprocessed signal. The multi-scale pyramid convolutional network uses three branches to extract features from the preprocessed signal, obtaining feature maps for each branch, and adaptively fuses all feature maps to obtain an adaptive fused feature sequence. The temporal attention memory network performs temporal processing on the adaptive fused feature sequence, introducing historical fault time sequences as memory vectors during the processing to obtain a temporal feature sequence. The feature decoupling and reconstruction network decouples the temporal feature sequence to obtain a decoupled feature sequence. The graph reasoning network performs graph relationship reasoning on the decoupled feature sequence to obtain global features. The perceptual classification network quantizes and classifies the global features to obtain a final diagnostic result indicating whether a core component has failed and the type of failure when it has occurred.

[0029] refer to Figure 2As shown, the multi-scale pyramid convolutional network of this application includes three parallel branch convolutional layers and a weight learning network, wherein the three branch convolutional layers are high-frequency convolutional layers, mid-frequency convolutional layers, and low-frequency convolutional layers, respectively; the temporal attention memory network includes a sequentially connected sliding window processing module, a relative position encoding module, a multi-head self-attention mechanism module, and a memory module; the feature decoupling reconstruction network includes three parallel feature encoders and a reconstruction decoder; the outputs of the three parallel feature encoders are all connected to the input of the reconstruction decoder, wherein the three parallel feature encoders are a common feature encoder, a private feature encoder, and a noisy feature encoder, respectively; the graph reasoning network includes a sequentially connected graph construction module, a graph convolutional network, a graph attention module, and an average pooling layer; and the perceptual classification network includes an evidence network and an output layer.

[0030] The fault diagnosis model proposed in this application solves the diagnostic challenge of track faults under conditions of limited samples and multiple operating conditions through a multi-module collaborative deep architecture. This fault diagnosis model employs a multi-scale pyramid convolutional network to simultaneously capture the microscopic impact details, mesoscopic structural features, and macroscopic trend evolution of faults, overcoming the limitations of single-scale analysis. Its temporal attention memory network introduces historical fault patterns as memory vectors, establishing long-range dependencies and accurately identifying the periodic patterns and dynamic evolution characteristics of faults. The feature decoupling and reconstruction network effectively extracts the essential attributes of faults by separating common features, private features, and noise components, significantly improving the model's cross-operating-condition generalization ability. The graph reasoning network further mines the topological relationships between fault features, simulating the propagation path of faults between components, enhancing the interpretability of diagnostic results. Finally, a perceptual classification network achieves accurate classification of uncertainty perception. This layer-by-layer progressive feature learning mechanism enables the model to maintain high accuracy and robustness even under conditions of limited samples, providing reliable assurance for the intelligent operation and maintenance of critical equipment such as gearboxes and rolling bearings.

[0031] In one specific embodiment of this application,

[0032] The high-frequency convolutional layer is used to extract detailed features from the preprocessed signal to obtain a first feature map;

[0033] The intermediate frequency convolutional layer is used to extract intermediate frequency features from the preprocessed signal to obtain a second feature map;

[0034] The low-frequency convolutional layer is used to extract global features from the preprocessed signal to obtain a third feature map;

[0035] The weight learning network is used to assign adaptive weights to the detailed features, the mid-frequency features and the global features according to the fault type, and to fuse the first feature map, the second feature map and the third feature map based on the adaptive weights to obtain an adaptive fused feature sequence.

[0036] It should be noted that the multi-scale pyramid convolutional network of this application resembles a pyramid, containing three parallel convolutional branches, each using convolutional kernels of different sizes to capture features at different scales. Branch 1 has a 3x1 kernel, Branch 2 has a 5x1 kernel, and Branch 3 has a 7x1 kernel. Each branch is followed by an activation function (such as ReLU) and a pooling layer (such as max pooling). The weight learning network fuses the outputs of multiple branches along the channel dimension. The multi-scale feature extraction module of this application simultaneously extracts features from high-frequency, mid-frequency, and low-frequency signals to ensure that fault information across different frequency ranges is captured.

[0037] The sliding window processing module is used to divide the adaptive fusion feature sequence into multiple sub-sequences with partial overlap using a sliding window;

[0038] The relative position encoding module performs position encoding on the sequence within each window of the sliding window;

[0039] The memory module is used to store historical fault sequences and introduces them as memory vectors into the multi-head self-attention mechanism module.

[0040] The multi-head self-attention mechanism module is used to perform self-attention weights and output a temporal feature sequence based on the self-attention weights;

[0041] It should be noted that this application segments long sequences into overlapping or non-overlapping local windows, reducing computational complexity and enhancing local feature extraction. The relative position encoding module of this application provides the attention mechanism with information on the relative distances between elements in the sequence, compensating for the position insensitivity of the self-attention mechanism itself. This module does not focus on the absolute coordinates of each position, but rather on the relative distance between any two elements in the sequence, resulting in stronger generalization ability. This can be implemented using a set of learnable bias matrices, which are lookup tables based on the relative position indices between elements. The multi-head self-attention mechanism module of this application primarily evaluates the importance of features at different time steps, ensuring that each element (time step) in the sequence can pay attention to all other elements in the sequence, capturing long-distance dependencies. Furthermore, the memory module of this application stores historical fault sequences and uses them as a memory vector to introduce the historical fault sequences into the multi-head self-attention mechanism module.

[0042] The three parallel feature encoders are used to map the temporal feature sequence to their own space to obtain their respective features;

[0043] The reconstruction decoder is used to concatenate the corresponding features and decouple the original features based on the concatenated features;

[0044] It should be noted that the common feature encoder extracts shared, general equipment operation features across different fault modes. These features are independent of specific fault types but reflect the basic operating state and common patterns of the equipment, such as basic vibration characteristics and operating parameters. This application uses a shallow neural network layer, focusing on global statistical features and low-frequency components to ensure the universality and stability of the extracted features. The private feature encoder extracts unique features specific to a certain fault type. These features are fault-sensitive and can clearly distinguish different fault modes. For example, the impact frequency feature unique to bearing outer ring faults and the modulation feature unique to gear tooth breakage. This ensures that each fault type has its unique feature representation, and that these features have maximum discriminative power across different fault types. The noise feature encoder separates interference information and random fluctuations irrelevant to fault diagnosis, including measurement noise, environmental interference, random fluctuations, and other irrelevant information. This application can achieve this through residual learning. The reconstruction decoder receives the outputs from the common feature encoder and the private feature encoder as input. This application does not directly use noise features for reconstruction, but forces the network to learn an effective representation after denoising. The goal is to reconstruct the original input features or some transformed version thereof as accurately as possible.

[0045] The graph construction module is used to construct a feature relationship graph based on the decoupled original features and their respective corresponding features.

[0046] The graph convolutional network is used to propagate and aggregate node information in the feature relationship graph to obtain node features enhanced by graph reasoning;

[0047] The graph attention module adjusts the weights of neighboring nodes during propagation and aggregation;

[0048] The average pooling layer is used to perform average pooling on the node features of the inference enhancement to obtain global features;

[0049] It's important to note that the graph construction module aims to transform the input feature vectors into a graph structure, i.e., constructing graph nodes and edges. Each node represents a feature vector or a sample, but in fault diagnosis, the feature vector at each time step or for each sample is typically considered a node. Edge construction is usually based on the similarity between nodes or predefined rules, such as k-nearest neighbors or fully connected edges. Graph Convolutional Networks (GCNs) are neural networks used for graph-structured data, updating node representations by aggregating information from neighboring nodes. The output of a GCN is a node feature matrix H and an adjacency matrix A. Through multiple graph attention layers, each node updates its features based on the attention weights of its neighboring nodes. Multi-head attention can be used to stabilize the learning process, resulting in the node feature matrix after attention-weighted aggregation. Average pooling layers are used to aggregate node-level features into graph-level features. By averaging the features of all nodes, a global graph representation vector is obtained. This vector contains information about the entire graph and can be used for subsequent classification tasks.

[0050] The evidence network is used to classify each fault category based on the global features to obtain an evidence vector for each fault category, and to calculate the expected probability of each fault category based on the evidence vector and the Dirichlet distribution parameters.

[0051] The output layer is used to determine the fault category based on the desired probability of each fault category.

[0052] It's important to note that the goal of an evidence network is to generate evidence for each category. Evidence can be understood as the degree to which a sample belongs to a certain category. This evidence is typically non-negative because negative values ​​have no intuitive interpretation. In evidence theory, evidence is represented by a function called the base probability assignment. For classification problems, a base probability mass is assigned to each category. The evidence network outputs an evidence value for each category, and then uses these evidence values ​​to calculate the probability of belonging to each class and the overall uncertainty. Evidence networks typically consist of several fully connected layers, with the final step using a non-negative activation function to ensure that the evidence values ​​are non-negative.

[0053] Specifically, we assume that the probability distribution of fault categories follows a multinomial distribution. In the Bayesian framework, to infer the probability vector of the multinomial distribution, a Dirichlet distribution is introduced as the prior distribution. The Dirichlet distribution is defined by a set of parameters (called prior parameters), which express the initial belief about the probability of each fault category before seeing any evidence. For example, if all prior parameters are equal, it means that all fault categories are initially considered to have the same probability of occurrence. When actual evidence is collected, this evidence is formalized as an evidence vector. Each element in the evidence vector corresponds to a fault category, and its value represents the strength or frequency of evidence for that category. Using Bayes' theorem, the prior distribution is combined with the evidence to obtain the posterior distribution. Since the Dirichlet distribution is the conjugate prior of the multinomial distribution, this means that the posterior distribution is also a Dirichlet distribution, and its parameters are equal to the sum of the prior parameters and the evidence vector. This property simplifies the calculation: the parameters of the posterior Dirichlet distribution can be obtained simply by adding the prior parameters and the evidence vector element by element. To obtain the expected probability, or hope probability, of each fault category, we calculate the expected value of the posterior Dirichlet distribution. For a Dirichlet distribution, each component of its expected probability vector is equal to its posterior parameter divided by the sum of all posterior parameters. This normalization operation ensures that the sum of the expected probabilities of all categories is 1, thus forming an efficient probability distribution.

[0054] In one specific embodiment of this application, the process of generating a pre-defined fault diagnosis model using the trained improved DDPM model, comprising multiple training samples, and then using the training samples to train the pre-defined fault diagnosis model includes:

[0055] a1. Obtain historical fault data and corresponding fault tags, and preprocess the historical fault data to obtain initial samples;

[0056] b1, extract the fault conditions from the historical fault data and randomly select a time step t, and determine the noise at the current time step according to the predefined noise scheduling;

[0057] c1, the initial sample, fault condition and noise of the current time step are input into the trained improved DDPM model so that the output is multiple training samples corresponding to a single initial sample; in the process of generating the multiple training samples, the improved DDPM model integrates multi-scale features and introduces temporal attention to establish temporal dependencies and fault information as constraints.

[0058] refer to Figure 3As shown, the improved DDPM model includes a sequentially connected conditional encoding module, a multi-scale diffusion model, and an output layer. The conditional encoding module encodes the conditional information into conditional vectors using an encoder, and merges all conditional vectors into a single comprehensive conditional vector. The multi-scale diffusion model extracts multiple scale features from the initial samples and, guided by the comprehensive conditional vector, generates a noisy fault signal. It then searches for similar noise from a database based on the comprehensive conditional vector. Multiple fault samples conforming to real faults are generated based on the noisy fault signal and similar noise. The output layer outputs these multiple fault samples conforming to real faults.

[0059] It should be noted that the condition encoding module of this application consists of an embedding layer and a Transformer encoder, which can encode discrete or continuous condition information (such as fault type labels and equipment operating conditions) into dense vector representations and use a self-attention mechanism to capture the correlation within the condition information, thereby outputting a comprehensive condition vector c as a guiding signal for the generation process.

[0060] refer to Figure 3 and Figure 4 The multi-scale diffusion model includes an encoder, a temporal attention bottleneck layer, a decoder, a multi-scale fusion layer, and an adaptive noise prediction network. The encoder extracts multi-scale features from initial samples under the influence of the comprehensive conditional vector and generates multi-scale feature maps. The temporal attention bottleneck layer reassembles the feature maps into a sequence under the influence of the conditional vector, capturing long-range dependencies through a multi-head self-attention mechanism and preserving temporal information using relative position encoding. The decoder gradually increases the feature map size under the influence of the conditional vector, and each upsampling block is connected to the corresponding multi-scale feature map of the encoder via skip connections, thereby fusing all multi-scale feature maps to obtain fused features. These fused features are then used to generate a noisy fault signal. The adaptive noise prediction network searches for similar noise in a database based on the noisy fault signal and the comprehensive conditional vector to obtain predicted noise. Based on the noisy fault signal and the predicted noise, it generates multiple fault samples that conform to the actual fault.

[0061] It should be noted that the multi-scale diffusion model in this application adopts a U-net hierarchical structure, progressively downsampling to extract multi-scale features, and then upsampling and fusing the features. The encoder is responsible for extracting multi-scale features from the input noisy data. It gradually reduces the sequence length through a series of downsampling operations while increasing the number of channels, thereby capturing contextual information at different scales. The encoder typically consists of multiple downsampling blocks, each containing a convolutional layer, a normalization layer, and an activation function. Downsampling is performed using one-dimensional convolutional layers, and residual connections are used after each downsampling block to mitigate gradient vanishing. The features extracted at each scale are preserved for use in skip connections in the decoder. The temporal attention bottleneck layer is located between the encoder and decoder and is responsible for capturing long-term dependencies in the time series. Since the sequence length is shortened after downsampling by the encoder, computational attention is more efficient. This layer models the relationships between time steps through a self-attention mechanism, enhancing the model's ability to capture global temporal patterns. A self-attention mechanism (such as multi-head self-attention in Transformer) is used to compute the representation of each time step. Since the sequence length is shortened by downsampling, the cost of computational attention is reduced, and positional encoding can also be added to preserve temporal order information. The decoder upsamples the features extracted by the encoder and the output of the temporal attention bottleneck layer, gradually restoring the sequence length while reducing the number of channels. The decoder uses features from different scales of the encoder to supplement detailed information through skip connections, thereby accurately reconstructing the denoised data. The decoder uses one-dimensional transposed convolutions or interpolation upsampling to increase the sequence length. Each upsampling block fuses features from the encoder at the corresponding scale through skip connections; each block typically contains convolutions, normalization, and activation functions. The multi-scale fusion layer integrates features from different scales to obtain a richer representation. The multi-scale fusion layer fuses features from different scales of the encoder, the output of the temporal attention bottleneck layer, and intermediate features from the decoder, combining low-level details and high-level semantics. A pyramid structure is used to upsample features from different scales to the same resolution before fusion. The fusion method can be weighted summation, concatenation followed by convolution, or adaptive fusion using an attention mechanism. The adaptive noise prediction network consists of an adaptive fusion layer and a noise prediction head, used to output predicted noise.

[0062] d1, the training samples are combined into a training set, and the training set is used to iteratively train the preset fault diagnosis model to obtain the trained fault diagnosis model.

[0063] In one specific embodiment of this application, before the improved DDPM model generates multiple training samples, the few-sample fault diagnosis method further includes:

[0064] a2, an improved DDPM model is obtained by improving the DDPM model;

[0065] b2, Train the improved DDPM model to obtain an improved DDPM model for generating samples; the training process is as follows:

[0066] a3, Input a single initial sample into the condition injection module to obtain the comprehensive condition vector of the single initial sample;

[0067] b3, randomly select a time step t, and determine the noise at the current time step according to the predefined noise scheduling;

[0068] c3. The integrated condition vector, the initial sample, and the sampling noise at the current time step are input into the multi-scale diffusion model to extract multiple scale features of the initial sample. Guided by the integrated condition vector, the noise is added to the initial sample to obtain an initial sample containing noise. Based on the initial sample containing noise and the integrated condition vector, similar noise is searched from the database to obtain the predicted noise.

[0069] d3, calculate the loss using the predicted noise and the sampled noise, and adjust the internal parameters of the improved DDPM using the loss;

[0070] e3, repeat the initial sample selection, and repeat the process from b3 to d3 until the training is complete;

[0071] f3. After training, the improved DDPM model is evaluated using evaluation metrics, and the improved DDPM model that meets the evaluation criteria is selected as the completed DDPM model.

[0072] This application may use one or more evaluation indicators disclosed in the relevant technology for evaluation, and this application does not impose any restrictions.

[0073] In one specific embodiment of this application, S200 includes:

[0074] S210, The preprocessed running signal is input into the fault diagnosis model so that it can extract features from multiple branches to obtain the feature map of each branch;

[0075] S220, Adaptively fuse all the feature maps to obtain an adaptive fused feature map;

[0076] S230, perform time-series processing on the adaptive fusion features, and introduce historical fault time series as memory vectors during the processing to obtain time-series feature sequences;

[0077] S240, Decouple the time-series feature sequence to obtain a decoupled feature sequence;

[0078] S250, perform graph relation reasoning on the decoupled feature sequence to obtain global features, and perform uncertainty quantification classification on the global features to obtain the final diagnostic result of whether the core component has failed and the type of failure when it has failed.

[0079] It should be noted that the fault diagnosis model outputs a probability distribution of fault types. The category with the highest probability is selected. If this category is a fault, then it can be determined that the core component has failed. In this case, the fault type of this category is the fault type of the core component.

[0080] Secondly, refer to Figure 5 This application provides a few-sample fault diagnosis device suitable for core components of rail trains, comprising:

[0081] The acquisition module 501 is configured to acquire the operating signals of the core components under the current working conditions from the transmission system in real time, and to preprocess the operating signals to obtain preprocessed operating signals.

[0082] The fault diagnosis module 502 is configured to input the preprocessed running signal into the trained fault diagnosis model, so that the fault diagnosis model can diagnose the core component and output whether the core component has failed and the type of failure when it has failed; wherein, the fault diagnosis model uses an improved DDPM model to generate multiple training samples, and then uses the training samples to train a preset fault diagnosis model; the improved DDPM model integrates multi-scale features in the process of generating the multiple training samples, and introduces temporal attention to establish temporal dependencies and fault information as constraints, thereby generating multiple training samples.

[0083] To verify the effectiveness of this application, the following simulation is performed:

[0084] MATLAB Simulink was used to build a system-level model of the gearbox and bearings in the traction system of a rail train, simulating their dynamic behavior under different operating conditions. ROMAX Designer, a professional gearbox analysis tool, was then used for analysis, providing detailed gear meshing and bearing contact analysis to generate gearbox fault data covering common fault types such as gear wear, tooth breakage, and pitting. Vibration signals containing specific fault characteristics were synthesized for algorithm verification. Simulation results are shown in Tables 1 and 2.

[0085] Table 1

[0086]

[0087] Table 2

[0088]

[0089] Table 1 presents the performance indicators for gearbox fault diagnosis. As can be seen from Table 1, this application demonstrates excellent overall performance in gearbox fault diagnosis, with a weighted average accuracy of 95.3% and an F1 score of 94.9%, reflecting the model's accurate fault identification and classification capabilities. Particularly in the diagnosis of severe faults (such as multiple broken teeth and severe wear), the accuracy exceeds 97%, with a false alarm rate of less than 1%, ensuring high reliability. It also maintains high sensitivity for early minor faults (minor wear 89.3%, localized pitting 92.5%), and the detection delay significantly decreases with the severity of the fault (from 15.2 samples to 4.2 samples), verifying the model's progressive capture capability during fault evolution. The balanced and excellent performance across all indicators fully demonstrates the high accuracy and strong robustness of this method in engineering practice.

[0090] Table 2 shows the predictive performance of different models. The fault diagnosis model trained with the improved DDPM generated samples in this application demonstrates significant advantages, achieving diagnoses of 96.8% and 95.9% in gearbox and bearing diagnostics, respectively, comprehensively outperforming traditional methods (such as frequency domain analysis at 82.5%) and mainstream deep learning models (such as one-dimensional CNN at 93.5%). Although the training time is slightly longer (18.2 hours), the inference time remains efficient (12.5 ms) and also possesses high interpretability (8.5 / 10), significantly outperforming models with similar parameter sizes but low interpretability (such as DBN at only 5.8 points). This design, while maintaining high accuracy in deep learning, solves the few-sample problem through sample generation augmentation, achieving an optimal balance between accuracy and interpretability, and providing a reliable guarantee for industrial applications.

[0091] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of this application and should not be construed as limiting the specific implementation of this application to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of this application, and all such modifications or substitutions should be considered within the scope of protection of this application.

Claims

1. A method for diagnosing faults in core components of rail trains using a small sample size, characterized in that, include: S100: The system collects operating signals of core components under current operating conditions from the power system of the railcar, and preprocesses these signals to obtain preprocessed operating signals. The core components include a gearbox and rolling bearings. If the core component is a gearbox, the operating signals are gear transmission system parameters, transmission ratio, and meshing frequency. If the core component is a rolling bearing, the operating signals are fault frequency and film thickness ratio. If the core components are a gearbox and rolling bearings, the operating signals are vibration signals and temperature. S200, the preprocessed running signal is input into the trained fault diagnosis model so that the fault diagnosis model can diagnose the core component and output whether the core component has failed and the type of failure when it has failed; wherein, the fault diagnosis model uses a trained improved DDPM model to generate multiple training samples, and then uses the training samples to train a preset fault diagnosis model; the improved DDPM model integrates multi-scale features in the process of generating the multiple training samples, and introduces temporal attention to establish temporal dependencies and fault information as constraints, thereby generating multiple training samples; The fault diagnosis model includes an input layer, a multi-scale pyramid convolutional network, a temporal attention memory network, a feature decoupling and reconstruction network, a graph reasoning network, and a perceptual classification network connected in sequence. The improved DDPM model includes a sequentially connected conditional encoding module, a multi-scale diffusion model, and an output layer. The conditional encoding module encodes conditional information into conditional vectors using an encoder, and merges all conditional vectors into a single comprehensive conditional vector. The multi-scale diffusion model extracts multiple scale features from the initial samples and, guided by the comprehensive conditional vector, generates a noisy fault signal. Based on the comprehensive conditional vector, it searches for similar noise in a database. Multiple fault samples conforming to real faults are generated based on the noisy fault signal and similar noise. The output layer outputs these multiple fault samples conforming to real faults.

2. The method for few-sample fault diagnosis of core components of rail trains according to claim 1, characterized in that, The input layer is used to input the preprocessed signal; the multi-scale pyramid convolutional network is used to extract features from the preprocessed signal using three branches to obtain the feature map of each branch, and adaptively fuses the feature maps of all branches to obtain an adaptive fused feature sequence. The temporal attention memory network is used to perform temporal processing on the adaptive fusion feature sequence, and introduces historical fault time sequences as memory vectors during the processing to obtain a temporal feature sequence; the feature decoupling reconstruction network is used to decouple the temporal feature sequence to obtain a decoupled feature sequence; the graph reasoning network is used to perform graph relationship reasoning on the decoupled feature sequence to obtain global features; the perceptual classification network is used to quantize and classify the global features to obtain the final diagnostic result of whether the core component has failed and the type of failure when it has failed.

3. The method for few-sample fault diagnosis of core components of rail trains according to claim 1, characterized in that, The multi-scale pyramid convolutional network includes three parallel branch convolutional layers and a weight learning network, wherein the three branch convolutional layers are high-frequency convolutional layers, mid-frequency convolutional layers, and low-frequency convolutional layers, respectively; the temporal attention memory network includes a sequentially connected sliding window processing module, a relative position encoding module, a multi-head self-attention mechanism module, and a memory module; the feature decoupling reconstruction network includes three parallel feature encoders and a reconstruction decoder; the outputs of the three parallel feature encoders are all connected to the input of the reconstruction decoder, wherein the three parallel feature encoders are a common feature encoder, a private feature encoder, and a noisy feature encoder, respectively; The graph reasoning network includes a sequentially connected graph construction module, a graph convolutional network, a graph attention module, and an average pooling layer; the perceptual classification network includes an evidence network and an output layer.

4. The method for few-sample fault diagnosis of core components of rail trains according to claim 3, characterized in that, The high-frequency convolutional layer is used to extract detailed features from the preprocessed signal to obtain a first feature map; The intermediate frequency convolutional layer is used to extract intermediate frequency features from the preprocessed signal to obtain a second feature map; The low-frequency convolutional layer is used to extract global features from the preprocessed signal to obtain a third feature map; The weight learning network is used to assign adaptive weights to the detailed features, the mid-frequency features and the global features according to the fault type, and to fuse the first feature map, the second feature map and the third feature map based on the adaptive weights to obtain an adaptive fused feature sequence. The sliding window processing module is used to divide the adaptive fusion feature sequence into multiple sub-sequences with partial overlap using a sliding window; The relative position encoding module performs position encoding on the sequence within each window of the sliding window; The memory module is used to store historical fault sequences and introduces them as memory vectors into the multi-head self-attention mechanism module. The multi-head self-attention mechanism module is used to perform self-attention weights and output a temporal feature sequence based on the self-attention weights; The three parallel feature encoders are used to map the temporal feature sequence to their own space to obtain their respective features; The reconstruction decoder is used to concatenate the corresponding features and decouple the original features based on the concatenated features; The graph construction module is used to construct a feature relationship graph based on the decoupled original features and their respective corresponding features. The graph convolutional network is used to propagate and aggregate node information in the feature relationship graph to obtain node features enhanced by graph reasoning; The graph attention module adjusts the weights of neighboring nodes during propagation and aggregation; The average pooling layer is used to perform average pooling on the node features of the inference enhancement to obtain global features; The evidence network is used to classify based on the global features, obtain an evidence vector for each fault category, and calculate the expected probability for each fault category based on the evidence vector and the Dirichlet distribution parameters. The output layer is used to determine the fault category based on the desired probability of each fault category.

5. The method for few-sample fault diagnosis of core components of rail trains according to claim 1, characterized in that, The process of generating multiple training samples using the improved DDPM model after training, and then using the training samples to train a preset fault diagnosis model includes: a1. Obtain historical fault data and corresponding fault tags, and preprocess the historical fault data to obtain initial samples; b1, extract the fault conditions from the historical fault data and randomly select a time step t, and determine the noise at the current time step according to the predefined noise scheduling; c1, the initial sample, fault condition and noise of the current time step are input into the trained improved DDPM model so that the output is multiple training samples corresponding to a single initial sample; in the process of generating the multiple training samples, the improved DDPM model integrates multi-scale features and introduces temporal attention to establish temporal dependencies and fault information as constraints. d1, the training samples are combined into a training set, and the training set is used to iteratively train the preset fault diagnosis model to obtain the trained fault diagnosis model.

6. The method for few-sample fault diagnosis of core components of rail trains according to claim 1, characterized in that, The few-sample fault diagnosis method further includes the following steps before the improved DDPM model generates multiple training samples: a2, an improved DDPM model is obtained by improving the DDPM model; b2, Train the improved DDPM model to obtain an improved DDPM model for generating samples; the training process is as follows: a3, Input a single initial sample into the conditional encoding module to obtain the comprehensive conditional vector of the single initial sample; b3, randomly select a time step t, and determine the noise at the current time step according to the predefined noise scheduling; c3. The integrated condition vector, the initial sample, and the sampling noise at the current time step are input into the multi-scale diffusion model to extract multiple scale features of the initial sample. Guided by the integrated condition vector, the noise is added to the initial sample to obtain an initial sample containing noise. Based on the initial sample containing noise and the integrated condition vector, similar noise is searched from the database to obtain the predicted noise. d3, calculate the loss using the predicted noise and the sampled noise, and adjust the internal parameters of the improved DDPM using the loss; e3, repeat the initial sample selection, and repeat the process from b3 to d3 until the training is complete; f3. After training, the improved DDPM model is evaluated using evaluation metrics, and the improved DDPM model that meets the evaluation criteria is selected as the completed DDPM model.

7. The method for few-sample fault diagnosis of core components of rail trains according to claim 1, characterized in that, The multi-scale diffusion model includes an encoder, a temporal attention bottleneck layer, a decoder, a multi-scale fusion layer, and an adaptive noise prediction network. The encoder is used to extract multi-scale features of the initial samples under the action of the comprehensive conditional vector and generate multi-scale feature maps. The temporal attention bottleneck layer is used to reassemble the feature maps into a sequence under the action of the conditional vector, and capture long-range dependencies through a multi-head self-attention mechanism and retain temporal information using relative position encoding. The decoder is used to gradually increase the feature map size under the action of the conditional vector, and each upsampling block is connected to the corresponding multi-scale feature map of the encoder in a skip connection, thereby fusing all multi-scale feature maps to obtain fused features. The fused features are used to generate a fault signal containing noise. The adaptive noise prediction network is used to search for similar noise in the database based on the fault signal containing noise and the comprehensive conditional vector to obtain predicted noise, and generate multiple fault samples that conform to the real fault based on the fault signal containing noise and the predicted noise.

8. The method for few-sample fault diagnosis of core components of rail trains according to claim 1, characterized in that, S200 includes: S210, The preprocessed running signal is input into the fault diagnosis model so that it can extract features from multiple branches to obtain the feature map of each branch; S220, adaptively fuse the feature maps of all branches to obtain an adaptive fused feature map; S230, perform time-series processing on the adaptive fusion features, and introduce historical fault time series as memory vectors during the processing to obtain time-series feature sequences; S240, Decouple the time-series feature sequence to obtain a decoupled feature sequence; S250, perform graph relation reasoning on the decoupled feature sequence to obtain global features, and perform uncertainty quantification classification on the global features to obtain the final diagnostic result of whether the core component has failed and the type of failure when it has failed.

9. A few-sample fault diagnosis device suitable for core components of rail trains, characterized in that, include: The acquisition module is configured to acquire operating signals of core components from the power system of the railcar under current operating conditions, and preprocess the operating signals to obtain preprocessed operating signals. The core components include a gearbox and rolling bearings. If the core component is a gearbox, the operating signals are gear transmission system parameters, transmission ratio, and meshing frequency. If the core component is a rolling bearing, the operating signals are fault frequency and film thickness ratio. If the core component is both a gearbox and rolling bearings, the operating signals are vibration signals and temperature. The fault diagnosis module is configured to input the preprocessed running signal into the trained fault diagnosis model, so that the fault diagnosis model can diagnose the core component and output whether the core component has failed and the type of failure when it has failed. The fault diagnosis model uses an improved DDPM model to generate multiple training samples, and then uses these training samples to train a preset fault diagnosis model. The improved DDPM model integrates multi-scale features in the process of generating the multiple training samples, and introduces temporal attention to establish temporal dependencies and fault information as constraints, thereby generating multiple training samples. The fault diagnosis model includes an input layer, a multi-scale pyramid convolutional network, a temporal attention memory network, a feature decoupling and reconstruction network, a graph reasoning network, and a perceptual classification network connected in sequence. The improved DDPM model includes a sequentially connected conditional encoding module, a multi-scale diffusion model, and an output layer. The conditional encoding module encodes conditional information into conditional vectors using an encoder, and merges all conditional vectors into a single comprehensive conditional vector. The multi-scale diffusion model extracts multiple scale features from the initial samples and, guided by the comprehensive conditional vector, generates a noisy fault signal. Based on the comprehensive conditional vector, it searches for similar noise in a database. Multiple fault samples conforming to real faults are generated based on the noisy fault signal and similar noise. The output layer outputs these multiple fault samples conforming to real faults.

Images