A switch machine oil pump credible intelligent diagnosis method and system based on conditional diffusion data enhancement

Abstract

The application discloses a switch machine oil pump credible intelligent diagnosis method and system based on conditional diffusion data enhancement, belongs to the field of railway signal equipment and industrial equipment predictive maintenance, and solves the problems of insufficient hydraulic oil pump fault samples, weak time sequence fault feature extraction capability and no reliability evaluation of diagnosis results in the prior art. The method comprises the following steps: preprocessing collected multi-source heterogeneous data of a switch machine oil pump to obtain standardized analysis sample segments; generating a conditional vector used for guiding a generation process based on a preset switch machine oil pump fault category label and a health state index code; establishing a connection between the standardized sample segments and the conditional vector to construct a real sample set; generating multi-channel synthesis by using a trained conditional diffusion model to construct an enhanced training set; extracting deep global fault features in a vibration signal to output a final feature vector; and performing fault category identification and health state evaluation on the final feature vector. The application is applied to switch machine oil pump diagnosis.

Description

Technical Field

[0001] This invention relates to the field of predictive maintenance technology for railway signaling equipment and industrial equipment, specifically to a reliable intelligent diagnostic method and system for switch machine oil pumps based on conditional diffusion data enhancement. Background Technology

[0002] With the development of railway transportation towards high speed, heavy load, and high density, higher requirements are placed on the reliability and safety of turnout switching equipment—electro-hydraulic switch machines. The hydraulic pump in the switch machine, as the core power component driving turnout switching, directly affects train operation safety. Once the pump experiences cavitation, bearing wear, or broken gear teeth, it may lead to incomplete turnout switching, locking failure, or even derailment. Therefore, achieving real-time and accurate fault prediction and health management of the switch machine's hydraulic pump has become a core requirement for railway signaling departments to ensure transportation safety and reduce maintenance costs.

[0003] Currently, intelligent diagnostic methods based on vibration signal analysis are the mainstream technical approach in this field. This method collects vibration signals during operation by placing accelerometers on the switch machine's oil pump housing and uses deep learning models to identify fault characteristics. However, single vibration signals have inherent limitations: the same vibration characteristics may originate from different fault causes (for example, cavitation and early bearing failures both generate broadband excitation and modulation sidebands), making it difficult to distinguish them based solely on vibration signals. Therefore, it is necessary to introduce multi-source sensor information such as oil pressure and temperature for auxiliary discrimination.

[0004] Although deep learning has made some progress in fault diagnosis, the following prominent shortcomings still exist in the practical application scenarios of railway switch machine oil pumps:

[0005] I. Scarcity and Imbalance of Fault Data: Equipment such as oil pumps operate under normal conditions for extended periods, resulting in extremely scarce samples of specific faults. This leads to a severe shortage of training data and a highly imbalanced class distribution. Traditional fault detection models struggle to learn the complete features of rare faults, exhibiting poor generalization ability. Furthermore, traditional oversampling or data augmentation methods (such as SMOTE) are ineffective in generating vibration signals that conform to complex physical laws.

[0006] Second, the decision-making lacks credibility assessment: Traditional fault detection models only output definite fault categories or health indices, and cannot assess the reliability or confidence level of the prediction results. When the model faces a fault mode that has never been seen before (such as a new type of cavitation) or is in a high-noise environment, it may still give a "confident" but wrong conclusion, which may cause serious consequences. This not only fails to provide effective risk warnings for operation and maintenance personnel, but also does not support subsequent active learning and continuous model optimization.

[0007] Third, poor multi-task coordination: Maintenance decisions for switch machine oil pumps require not only knowledge of the fault type but also an understanding of the equipment's remaining service life. Existing methods typically model these two tasks independently, ignoring their inherent correlation at the feature level. This not only reduces computational efficiency but may also negatively impact performance due to the learning of contradictory feature representations.

[0008] Therefore, there is an urgent need to provide a new, reliable intelligent diagnostic solution for oil pump faults that can comprehensively address the problem of multi-task collaboration. Summary of the Invention

[0009] To address the technical problems of insufficient hydraulic pump fault samples, weak ability to extract time-series fault features, and lack of reliability assessment of diagnostic results in the prior art, this invention proposes a reliable intelligent diagnostic method and system for switch machine oil pumps based on conditional diffusion data enhancement.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a reliable intelligent diagnostic method for switch machine oil pumps based on conditional diffusion data enhancement, comprising the following steps:

[0011] Step S1: Preprocess the collected multi-source heterogeneous data of the switch machine oil pump to obtain a standardized analysis sample segment; the multi-source heterogeneous data includes at least the vibration signal, temperature signal and pressure signal of the switch machine oil pump.

[0012] Step S2: Generate a condition vector to guide the generation process based on the preset fault category label and health status index code of the switch machine oil pump.

[0013] Step S3: Establish a connection between the standardized sample fragments and conditional vectors to construct a real sample set;

[0014] Step S4: Train the real sample set based on the conditional diffusion model, and embed the conditional vector into each layer of the U-Net network through the cross-attention mechanism to learn the reverse denoising process from noise to the original real sample.

[0015] Step S5: Using the trained conditional diffusion model, generate multi-channel synthetic samples based on the specified fault category and health status to construct an enhanced training set;

[0016] Step S6: Use a time-series feature extraction model to perform block segmentation, projection coding, and Transformer coding on the multi-channel one-dimensional time-series signal in the enhanced training set to extract deep global fault features from the vibration signal and output the final feature vector.

[0017] Step S7: The extracted final feature vector is simultaneously fed into the fault classification head and the health assessment head, and the fault category probability distribution and health index prediction value are output respectively. The uncertainty of the prediction is quantified by multiple random forward propagation. The information entropy of the probability distribution of the fault category is used as the uncertainty measure, and the standard deviation of the health index prediction value is used as the uncertainty interval.

[0018] Step S8: Generate a diagnostic report based on the uncertainty measure and the uncertainty interval of the predicted health index value.

[0019] Furthermore, the preprocessing of the multi-source heterogeneous data collected from the switch machine oil pump specifically includes the following steps:

[0020] Step S11: Perform wavelet denoising on the collected multi-source heterogeneous data of the switch machine oil pump to remove power frequency interference and high frequency noise, and then perform amplitude normalization on the denoised signal.

[0021] Step S12: Use a sliding window to divide the long time-series signal after noise reduction and normalization into one-dimensional analysis sample segments of fixed length that can overlap.

[0022] Furthermore, the condition vector is a numerical representation of the fault category label and the health status index after encoding. The fault category label adopts one-hot encoding and includes at least normal, bearing outer ring fault, bearing inner ring fault and gear tooth breakage. The encoded health status index is a health degree quantification index, which is a continuous scalar with a value range of [0,1], used to characterize the degree of evolution from a healthy state to a complete failure, and is mapped according to the preset percentage of remaining service life.

[0023] Furthermore, step S4 specifically includes the following steps:

[0024] Step S41: Randomly sample real samples from the real sample set, gradually add Gaussian noise to the analysis sample segment of the real samples, and obtain pure Gaussian noise after multiple forward diffusion steps.

[0025] Step S42: Use U-Net network as the denoising network, and use the conditional vector as the key and value. Embed it into the feature map of each layer of the encoder and decoder of U-Net network through cross-attention mechanism, so that U-Net network can guide noise prediction according to the specified fault category label and health level quantification index.

[0026] Step S43: Using the mean square error between predicted noise and actual noise as the loss function, train the conditional diffusion model to learn the reverse denoising process.

[0027] Furthermore, step S5 specifically includes the following steps:

[0028] Step S51: Construct a target condition vector. The target condition vector has the same format and dimensions as the aforementioned condition vector. It includes target fault category labels and target health level quantitative indicators. This target condition vector has the same format and dimensions as the aforementioned condition vector.

[0029] Step S52: Randomly sample a pure Gaussian noise with the same dimension as the real sample as the initial state of the reverse denoising process; all elements of the pure Gaussian noise are independent of each other and all follow a standard normal distribution;

[0030] Step S53: Input pure Gaussian noise and target conditional vector into the trained U-Net network, predict and remove noise one by one, gradually restore the original real sample, and output high-quality multi-channel synthetic sample that meets the specified conditions.

[0031] Step S54: Mix the generated synthetic samples with the original real samples to form an enhanced training set.

[0032] Furthermore, step S6 specifically includes the following steps:

[0033] Step S61: Divide the augmented training set into multiple segments that may overlap along the length dimension, with each segment having a fixed length;

[0034] Step S62: Map each segment to a feature space of a specified dimension through a learnable linear layer to obtain a projection sequence, and add a learnable positional encoding matrix to preserve the temporal order;

[0035] Step S63: Process all segments using a Transformer encoder composed of multiple stacked layers. Each Transformer encoder layer contains a multi-head self-attention sub-layer and a feedforward network sub-layer. Perform residual connection and layer normalization processing on the multi-head self-attention sub-layer and the feedforward network sub-layer respectively.

[0036] Step S64: Perform global average pooling on the output features of all segments, and then fuse the multi-channel features into the final feature vector through a learnable linear layer.

[0037] Furthermore, in step S7, the fault classification head consists of a fully connected layer and a Softmax activation function, outputting the probability distribution of each fault category, and taking the category corresponding to the highest probability as the preliminary classification result;

[0038] The health assessment head consists of a fully connected layer with no special activation function, and outputs continuous scalar values ​​as preliminary predictions of health index or remaining lifespan.

[0039] Furthermore, step S8 specifically involves: inferring the confidence level of the prediction result based on the uncertainty measure; when the information entropy exceeds the preset entropy threshold, determining that the model's uncertainty regarding the prediction is too high, i.e., the confidence is too low; generating a diagnostic report based on the confidence level of the fault category and the uncertainty range of the predicted health index value; triggering a high-risk warning when the classification confidence level is lower than the preset confidence threshold or the uncertainty range exceeds the safety threshold, and automatically recording low-confidence samples for active learning to achieve continuous model evolution.

[0040] A reliable intelligent diagnostic system for switch machine oil pumps based on conditional diffusion data augmentation, which implements the above-described reliable intelligent diagnostic method for switch machine oil pumps based on conditional diffusion data augmentation, comprises:

[0041] The data preprocessing module is used to preprocess the multi-source heterogeneous data of the collected switch machine oil pump to obtain standardized analysis sample segments; based on the preset switch machine oil pump fault category labels and health status index codes, a condition vector is generated to guide the generation process; the standardized sample segments and condition vectors are linked to construct a real sample set.

[0042] The data augmentation module communicates with the data preprocessing module and is used to train the real sample set based on the conditional diffusion model. It embeds the conditional vectors into each layer of the U-Net network through the cross-attention mechanism to learn the reverse denoising process from noise to the original real sample. It generates multi-channel synthetic samples according to the specified fault category and health status to construct the augmented training set.

[0043] The temporal feature extraction module, which is connected to the data augmentation module, is used to perform block segmentation, projection coding, and Transformer coding on the multi-channel one-dimensional temporal signals in the augmentation training set using the temporal feature extraction model, so as to extract deep global fault features in the vibration signal and output the final feature vector.

[0044] The multi-task learning and uncertainty quantification module communicates with the temporal feature extraction module. It is used to simultaneously feed the extracted final feature vector to the fault classification head and the health assessment head, outputting the probability distribution of the fault category and the predicted value of the health index, respectively. It quantifies the uncertainty of the prediction through multiple random forward propagations, using the information entropy of the probability distribution of the fault category as the uncertainty measure and the standard deviation of the predicted value of the health index as the uncertainty interval.

[0045] The diagnostic report output module communicates with the multi-task learning and uncertainty quantification module to generate a diagnostic report based on the uncertainty measure and the uncertainty range of the health index prediction value.

[0046] The advantages of this invention over the prior art are as follows:

[0047] 1. The method and system of this invention generate physically meaningful and high-fidelity fault prediction values ​​through a conditional diffusion model, which effectively solves the small sample problem in the diagnosis of switch machine oil pumps, and enables the model to maintain excellent diagnostic performance even with very few real fault samples.

[0048] 2. The method and system of this invention adopt a targeted time-series feature extraction model, PatchTST, which can more accurately capture long-range periodic fault features in vibration signals, and significantly improve the accuracy of fault identification and health status assessment.

[0049] 3. The method and system of this invention introduce an uncertainty quantification mechanism to transform traditional deterministic output into a reliability assessment result with predictive confidence, providing a scientific basis for operation and maintenance decisions in critical industrial safety scenarios and having extremely high practical value. Attached Figure Description

[0050] The present invention will be further described below with reference to the accompanying drawings:

[0051] Figure 1 This is a schematic diagram of the overall system architecture and data flow of the present invention;

[0052] Figure 2 This is a schematic diagram illustrating the process of training a real sample set based on a conditional diffusion model according to the method of the present invention.

[0053] Figure 3 This is a schematic diagram of the process for generating multi-channel synthetic samples using the method of the present invention;

[0054] Figure 4 This is a schematic diagram illustrating the process of extracting deep global fault features from vibration signals using the PatchTST model in this invention. Detailed Implementation

[0055] like Figures 1 to 4 As shown, this invention provides a reliable intelligent diagnostic method for switch machine oil pumps based on conditional diffusion data enhancement, comprising the following steps:

[0056] Step S1: Preprocess the collected multi-source heterogeneous data of the switch machine oil pump to obtain standardized analysis sample segments. Multi-source heterogeneous data includes at least the vibration signal, temperature signal, and pressure signal of the switch machine oil pump;

[0057] The preprocessing of multi-source heterogeneous data collected from switch machine oil pumps specifically includes the following steps:

[0058] Step S11: Perform wavelet denoising on the collected multi-source heterogeneous data of the switch machine oil pump to remove power frequency interference and high frequency noise, and then perform amplitude normalization on the denoised signal.

[0059] Step S12: Use a sliding window to segment the long time-series signal after noise reduction and normalization into one-dimensional analysis sample segments of fixed length that can overlap. The window length is set to 1024 data points based on the sampling frequency. Each 1024 data points corresponds to 1 second of physical time, ensuring that each analysis sample segment contains the complete fault impact cycle. An overlap rate is set between adjacent windows to enhance sample diversity.

[0060] Step S2: Generate a condition vector C to guide the generation process based on the preset fault category label and health status index code of the switch machine oil pump;

[0061] The condition vector is a numerical representation of the fault category label and the health status index after encoding. The fault category label adopts one-hot encoding and includes at least normal, bearing outer ring fault, bearing inner ring fault and gear tooth breakage. The encoded health status index is a health degree quantification index, which is a continuous scalar with a value range of [0,1], used to characterize the degree of evolution from healthy state to complete failure, and is mapped according to the preset remaining useful life (RUL).

[0062] Step S3: Standardize the sample fragments Establish a connection between the conditional vector C and construct a real sample set, denoted as . ;

[0063] Specifically, each analysis sample segment obtained by sliding window segmentation uniquely corresponds to a condition vector. The condition vector consists of the fault category label and health status quantification index within the time period of the analysis sample segment. The analysis sample segment and its corresponding condition vector together constitute a real sample, and the set of all real samples constitutes the real sample set.

[0064] Step S4: Train the real sample set based on the conditional diffusion model, and embed the conditional vector into each layer of the conditional diffusion model U-Net network through the cross attention mechanism to learn the reverse denoising process from noise to the original real sample.

[0065] Step S4 specifically includes the following steps:

[0066] Step S41: Randomly sample real samples from the real sample set, and gradually add Gaussian noise to the analysis sample segment of the real samples. After multiple forward diffusion steps, pure Gaussian noise is obtained. ;

[0067] The forward diffusion process is defined as: from the initial real sample Starting with the standardized analysis sample segment, Gaussian noise is gradually added, after... After the step, pure Gaussian noise is obtained. Each step of the noise addition process is described by a conditional probability distribution:

[0068] ;

[0069] In the formula, Indicates the first The analyzed sample segment after adding noise; Indicates the first The analyzed sample segment after adding noise; These are preset noise scheduling parameters used to control the intensity of noise added at each step; It is an identity matrix, and its dimensions are the same as those of the analysis sample segment. The dimensions are the same; Indicates the first Analysis sample segment after adding noise Follow the mean variance is Gaussian distribution; Indicates that in the known number of... Noise-added analysis sample segment Under the conditions, the first Analysis sample segment after adding noise The probability distribution it follows describes the pattern of noise addition during forward diffusion. In the forward diffusion process, the actual noise added at each step is... , and After linear combination, we get During the training phase, the neural network needs to learn from noisy samples. The noise was predicted in the middle. ,therefore It is considered a true label during the training process;

[0070] Step S42: Use U-Net network as the denoising network, and use the conditional vector as the key and value. Embed it into the feature map of each layer of the encoder and decoder of U-Net network through cross-attention mechanism, so that U-Net network can guide noise prediction according to the specified fault category label and health level quantification index.

[0071] U-Net is a neural network with an encoder-decoder structure and skip connections, capable of effectively fusing multi-scale features. It maintains consistency in the size of input and output features while modeling both global semantics and local details. In conditional diffusion models, the task of U-Net is to learn to analyze sample segments from arbitrary intermediate points. Predict the added noise .

[0072] Step S43: Using the mean square error between predicted noise and actual noise as the loss function, train the conditional diffusion model to learn the reverse denoising process.

[0073] The goal of the inverse denoising process is to learn the conditional probability distribution. The expression for the conditional probability of inverse denoising is:

[0074] ;

[0075] In the formula, Indicates the first Analysis sample segment after adding noise Follow the mean variance is Gaussian distribution; Represents the predicted covariance matrix; This indicates that during the reverse denoising process, guided by the conditional vector C, the analysis starts from the current noisy sample segment. Restore a cleaner sample segment from the previous step. The probability distribution. This process is iteratively executed until... The final generated It possesses the attributes specified by condition vector C (i.e., fault category labels and health status quantification indicators).

[0076] loss function The expression is:

[0077] ;

[0078] In the formula, This represents the expectation, where t is the time step for random sampling; These are the original, real samples; Added real Gaussian noise; The first prediction for the U-Net network Real sample after adding noise (step 1) Analysis sample segment after adding noise The noise of the U-Net network and its corresponding condition vector C is used to accurately predict the noise added at each step, thus learning the reverse denoising process.

[0079] Step S5: Using the trained conditional diffusion model, starting with pure Gaussian noise, perform reverse denoising iterations under the guidance of the conditional vector to gradually recover multi-channel synthetic samples of fault category labels and health level quantification indicators, and construct an enhanced training set.

[0080] Step S5 specifically includes the following steps:

[0081] Step S51: Construct the target condition vector Target condition vector It has the same format and dimensions as the aforementioned condition vector C; including target fault category labels and target health level quantitative indicators, this target condition vector has the same format and dimensions as the aforementioned condition vector;

[0082] Specifically, it constructs a target condition vector based on the user-specified sample generation conditions. (e.g., the fault category is "bearing inner ring fault" and the health level quantification index is 0.7); the target condition vector includes the target fault category label and the target health level quantification index. The target condition vector has the same format and dimensions as the aforementioned condition vector to ensure the controllability of the generation process.

[0083] Step S52: Randomly sample pure Gaussian noise with the same dimension as the real sample. As the initial state for the reverse denoising process; pure Gaussian noise. All elements are independent of each other and all follow a standard normal distribution N(0,1);

[0084] Step S53: Combine the pure Gaussian noise and the target condition vector Input the trained U-Net network, predict and remove noise successively, repeat this process T times, and gradually restore the original real sample. , , ..., Output high-quality, multi-channel synthetic samples that meet specified conditions. ;

[0085] Step S54: Mix the generated synthetic samples with the original real samples to form a large-scale, class-balanced quality-enhanced training set. Each sample in the enhanced training set is a multi-channel one-dimensional time-series signal; its morphology is [ , ],in Number of channels ( (These correspond to vibration signal, temperature signal, and pressure signal, respectively). The sequence length is given.

[0086] Step S6: Use a time-series feature extraction model (such as the PatchTST model) to perform block segmentation, projection coding, and Transformer coding on the multi-channel one-dimensional time-series signals in the enhanced training set to extract deep global fault features from the vibration signals.

[0087] Specifically, this involves inputting multi-channel one-dimensional timing signals in batches. , , That is, augmenting the training set. , B represents the batch size. For the number of channels, For sequence length, For size A real matrix;

[0088] Step S6 specifically includes the following steps:

[0089] Step S61, Blocking: Divide the augmented training set into multiple segments that may overlap along the length dimension, with each segment having a fixed length;

[0090] In one specific embodiment, the augmented training set is divided along the length dimension into There are 10 potentially overlapping segments, each segment having a length of 1. Each segment is of size [size missing]. real matrix The step size of the starting position of adjacent segments is S. The expression for each fragment is:

[0091] ;

[0092] In the formula, This indicates that all batches and channel dimensions are retrieved;

[0093] In the first From each segment, take the sequence length dimension from... arrive The range.

[0094] Total number of segments Determined by the following formula:

[0095] ;

[0096] when When there is overlap between adjacent segments, the robustness of the conditional diffusion model to local features can be enhanced.

[0097] Step S62, Projection and Encoding: Map each segment to a feature space of a specified dimension through a learnable linear layer to obtain a projection sequence, and add a learnable positional encoding matrix to preserve the temporal order;

[0098] In one embodiment, each length is The fragments are mapped through a learnable linear layer to... dimensional feature space, Let the embedding dimension be ; let the projection matrix be . , bias is For any fragment After projection, we get:

[0099] ;

[0100] For the first Embedding vectors of each segment; For the first The original signal values ​​of each segment;

[0101] The projection sequence is obtained after parallel processing. To preserve the temporal order of the segments, a learnable positional encoding matrix is ​​added. :

[0102] ;

[0103] In the formula, The features after adding position encoding are used as input to the subsequent Transformer encoder;

[0104] To indicate the first The projection vectors of all batches and channels of each segment;

[0105] For the position encoding matrix of the first Line, that is, with the first The position encoding vectors corresponding to each segment;

[0106] Step S63, Transformer Encoding: All segments are processed using a Transformer encoder composed of multiple stacked layers. Each encoder layer contains a multi-head self-attention sublayer and a feedforward network sublayer, and residual connections and layer normalization are performed on these two sublayers respectively. For the first... layer Its input is the output of the previous layer. The calculation process is as follows:

[0107] Multi-head self-attention quantum layer:

[0108] ;

[0109] Feedforward sublayer:

[0110] ;

[0111] In the formula:

[0112] This is a multi-head self-attention function whose inputs are query, key, and value, all of which are... This indicates that attention modeling is performed on the enhanced training set itself;

[0113] For the first The intermediate output obtained after passing through the multi-head self-attention sublayer and residual connections, and then undergoing layer normalization, will be used as the input to the feedforward network sublayer.

[0114] It is a feedforward network, consisting of two fully connected layers and a non-linear activation function (ReLU) to independently transform the features at each location;

[0115] This is a layer normalization operation used to normalize features and stabilize the training process.

[0116] For the first The feature matrix output by the layer Transformer encoder;

[0117] For the first The feature matrix output by the Transformer encoder layer is also the first layer's feature matrix. The input feature matrix of the layer Transformer encoder is of size . real matrix ;when hour, ,in The embedded sequence after projection and position encoding;

[0118] Step S64, Pooling and Output: Global average pooling is performed on the output features of all segments, and then multi-channel features are fused into a final feature vector through a learnable linear layer. This final feature vector includes key information for fault detection and health status assessment.

[0119] In one embodiment, a preliminary aggregated feature vector is obtained by performing global average pooling on the output features of all segments. Each channel corresponds to a feature representation of a sensor signal (vibration signal, temperature signal, and pressure signal are collectively referred to as sensing signals). The expression for the initially aggregated feature vector is: ;

[0120] In the formula, Indicates passing through all After the first layer of Transformer encoding, the third... The encoded characteristics of each segment across all batches and all channels;

[0121] To facilitate subsequent multi-task processing, the feature vectors from multiple channels need to be initially aggregated. The features are then fused into a single feature vector. Specifically, a learnable linear layer is used for fusion to obtain the final feature vector. :

[0122] ;

[0123] In the formula, This represents the operation of flattening a multidimensional tensor into a one-dimensional vector.

[0124] The flattened feature matrix;

[0125] This is the weight matrix of the fusion layer;

[0126] This is the bias vector of the fusion layer;

[0127] For the final feature vector Dimensions.

[0128] Final feature vector It includes all the key information used to determine faults and health status, which will serve as input for subsequent multi-task learning modules.

[0129] Compared to traditional Transformer encoders, the PatchTST model architecture used in this invention has the following three advantages:

[0130] Firstly, through block division, the length of the input sequence is reduced from the original length. Compress to the number of fragments Accordingly, the core computational complexity will be reduced from Down to This significantly improves the training and inference efficiency of the conditional diffusion model;

[0131] Secondly, by using local segments as the basic processing unit, feature extraction can be performed while preserving the details of the original signal waveform. Compared with point-based modeling that directly processes the global sequence, it is more suitable for capturing transient impact modes in vibration signals.

[0132] Third, by modeling the long-range dependencies between segments through a self-attention mechanism, the field of view limitation of local windows is avoided, thereby effectively identifying the periodic fault characteristics in the switch machine oil pump and improving diagnostic accuracy.

[0133] Step S7: The extracted final feature vector is simultaneously fed into the fault classification head and the health assessment head, and the fault category probability distribution and health index prediction value are output respectively. In the inference stage, Monte Carlo Dropout is used to perform multiple forward propagations to calculate the information entropy of the fault category probability distribution as an uncertainty measure and to calculate the standard deviation of the health index prediction value as an uncertainty interval.

[0134] Specifically:

[0135] The fault classification head consists of a fully connected layer and a Softmax activation function, which outputs the probability distribution of each fault category and takes the category with the highest probability as the preliminary classification result.

[0136] The fault classification head is used to classify the final feature vector. Mapped to Probability distribution of each fault category The above values ​​are used to determine the initial classification result, and the fault category with the highest probability is selected. The probability distribution of the fault categories is expressed as follows:

[0137] ;

[0138] In the formula, This is the weight matrix of the classification heads. for Dimensions This is the bias vector for the classification head;

[0139] The health assessment head consists of a fully connected layer with no special activation function, and outputs continuous scalar values ​​as preliminary predictions of health index or remaining lifespan.

[0140] The health assessment head is used to convert the final feature vector Mapped to a continuous scalar value This represents the initial prediction of the health index or remaining useful life (RUL). The initial prediction is a single prediction result without uncertainty quantification and will serve as the basis for subsequent Monte Carlo sampling; during training, the mean squared error (MSE) is used as the loss function for this health index prediction.

[0141] The Monte Carlo Dropout method is used for uncertainty quantification. This involves maintaining the activation of the Dropout layer during the model inference phase and simulating the posterior distribution of model parameters through multiple random forward propagations to estimate the confidence level of the prediction results. This method provides an uncertainty measure for fault classification and health index assessment, enabling the system to have risk perception capabilities. Specifically, it includes the following sub-steps:

[0142] First, perform multiple forward propagations: after model training is complete, maintain the activation of the Dropout layer during the inference and prediction phase; for the same input sample... ,conduct The forward propagation yields the predicted value after forward propagation, thus obtaining the _th_ prediction. The predicted value obtained after the second forward propagation:

[0143] ;

[0144] In the formula, This represents a trained neural network with parameters θ and the Dropout layer remaining on.

[0145] Dropout represents a random discarding operation, where different neurons are randomly discarded during each forward propagation.

[0146] For the first The predicted values ​​obtained from the second forward propagation (the probability distribution of the fault category is output as a probability vector, and the predicted value of the health index is output as a scalar value).

[0147] The number of Monte Carlo samplings; due to the randomness of Dropout, the predicted values ​​obtained after each forward propagation will differ slightly.

[0148] Then, perform uncertainty calculations on the probability distribution of fault categories: for the probability distribution of fault categories, perform statistical analysis. The probability distribution obtained from the first prediction is used to calculate the average probability as the final prediction value, and the entropy of this probability distribution is calculated as a measure of uncertainty; the higher the entropy, the more uncertain the model. The uncertainty measure of the probability distribution of the fault category is defined as:

[0149] ;

[0150] In the formula, Total number of fault categories; No. In the second forward propagation, the model predicts the probability that a sample belongs to the c-th class;

[0151] for In the second prediction The average probability of the class is denoted as ;

[0152] The information entropy for prediction is used to measure uncertainty; the larger the entropy value, the more uncertain the model is.

[0153] Then, the uncertainty calculation of the health index prediction value is performed: for the health index prediction value, statistics are... The predicted health index The mean of the predicted value is calculated as the final predicted value, and the standard deviation is calculated as the uncertainty interval. The uncertainty interval of the predicted health index is expressed as follows:

[0154] ;

[0155] In the formula, for The average of the previous predictions will be used as the final prediction.

[0156] for The standard deviation of the forecast represents the uncertainty of the forecast;

[0157] The final output format is (For example: Remaining lifetime RUL is 120 ± 15 hours); Physical meaning: The smaller, the more... The more concentrated the prediction results are, the more confident the model is in the prediction; The larger, the more The more dispersed the prediction results, the more uncertain the model is about that prediction; when If the preset threshold is exceeded, an alert will be triggered, and manual review is recommended.

[0158] Step S8: Based on the uncertainty measure, the confidence level of the prediction result is inferred. When the information entropy exceeds the preset entropy threshold, it is determined that the uncertainty of the model for the prediction is too high, that is, the confidence is too low. A diagnostic report is generated based on the confidence of the fault category and the uncertainty range of the predicted health index value. When the classification confidence is lower than the preset confidence threshold or the uncertainty range exceeds the safety threshold, a high-risk warning is triggered, and low-confidence samples are automatically recorded for active learning to achieve continuous model evolution.

[0159] The final diagnostic report includes the most likely fault type and its prediction confidence, the remaining useful life prediction and its uncertainty range; when the classification confidence is lower than the preset threshold or the regression uncertainty range exceeds the safety threshold, a high-risk warning is triggered and manual review is recommended; all low-confidence samples are automatically recorded for subsequent active learning processes to guide the labeling of the most valuable data and achieve continuous model evolution.

[0160] This invention provides a reliable intelligent diagnostic system for switch machine oil pumps based on conditional diffusion data augmentation, which implements the aforementioned reliable intelligent diagnostic method for switch machine oil pumps based on conditional diffusion data augmentation. The system comprises:

[0161] The data preprocessing module is used to preprocess the multi-source heterogeneous data of the collected switch machine oil pump to obtain standardized analysis sample segments; based on the preset switch machine oil pump fault category labels and health status index codes, a condition vector is generated to guide the generation process; the standardized sample segments and condition vectors are linked to construct a real sample set.

[0162] The data augmentation module communicates with the data preprocessing module and is used to train the real sample set based on the conditional diffusion model. It embeds the conditional vectors into each layer of the U-Net network through the cross-attention mechanism to learn the reverse denoising process from noise to the original real sample. It generates multi-channel synthetic samples according to the specified fault category and health status to construct the augmented training set.

[0163] The temporal feature extraction module, which is connected to the data augmentation module, is used to perform block segmentation, projection coding, and Transformer coding on the multi-channel one-dimensional temporal signals in the augmentation training set using the temporal feature extraction model, so as to extract deep global fault features in the vibration signal and output the final feature vector.

[0164] The multi-task learning and uncertainty quantification module communicates with the temporal feature extraction module. It is used to simultaneously feed the extracted final feature vector to the fault classification head and the health assessment head, outputting the probability distribution of the fault category and the predicted value of the health index, respectively. It quantifies the uncertainty of the prediction through multiple random forward propagations, using the information entropy of the probability distribution of the fault category as the uncertainty measure and the standard deviation of the predicted value of the health index as the uncertainty interval.

[0165] The diagnostic report output module communicates with the multi-task learning and uncertainty quantification module to generate a diagnostic report based on the uncertainty measure and the uncertainty range of the health index prediction value.

[0166] Regarding the specific structure of this invention, it should be noted that the connection relationships between the various component modules used in this invention are definite and achievable. Except as specifically described in the embodiments, their specific connection relationships can bring about corresponding technical effects and solve the technical problems proposed by this invention without relying on the execution of corresponding software programs. The models of the components, modules, and specific components appearing in this invention, the connection methods between them, and the conventional usage methods and expected technical effects brought about by the above technical features, unless specifically described, are all publicly disclosed content in patents, journal articles, technical manuals, technical dictionaries, and textbooks that can be obtained by those skilled in the art before the application date, or belong to conventional technology, common knowledge, and other existing technologies in this field. There is no need to elaborate, which makes the technical solution provided in this case clear, complete, and achievable, and can reproduce or obtain corresponding physical products based on this technical means.

[0167] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A reliable intelligent diagnostic method for switch machine oil pumps based on conditional diffusion data augmentation, characterized in that, Includes the following steps: Step S1: Preprocess the collected multi-source heterogeneous data of the switch machine oil pump to obtain a standardized analysis sample segment; the multi-source heterogeneous data includes at least the vibration signal, temperature signal and pressure signal of the switch machine oil pump. Step S2: Generate a condition vector to guide the generation process based on the preset fault category label and health status index code of the switch machine oil pump. Step S3: Establish a connection between the standardized sample fragments and conditional vectors to construct a real sample set; Step S4: Train the real sample set based on the conditional diffusion model, and embed the conditional vector into each layer of the U-Net network through the cross-attention mechanism to learn the reverse denoising process from noise to the original real sample. Step S5: Using the trained conditional diffusion model, generate multi-channel synthetic samples based on the specified fault category and health status to construct an enhanced training set; Step S6: Use a time-series feature extraction model to perform block segmentation, projection coding, and Transformer coding on the multi-channel one-dimensional time-series signal in the enhanced training set to extract deep global fault features from the vibration signal and output the final feature vector. Step S7: The extracted final feature vector is simultaneously fed into the fault classification head and the health assessment head, and the fault category probability distribution and health index prediction value are output respectively. The uncertainty of the prediction is quantified by multiple random forward propagation. The information entropy of the probability distribution of the fault category is used as the uncertainty measure, and the standard deviation of the health index prediction value is used as the uncertainty interval. Step S8: Generate a diagnostic report based on the uncertainty measure and the uncertainty interval of the predicted health index value.

2. The reliable intelligent diagnostic method for switch machine oil pumps based on conditional diffusion data enhancement according to claim 1, characterized in that, The preprocessing of multi-source heterogeneous data collected from switch machine oil pumps specifically includes the following steps: Step S11: Perform wavelet denoising on the collected multi-source heterogeneous data of the switch machine oil pump to remove power frequency interference and high frequency noise, and then perform amplitude normalization on the denoised signal. Step S12: Use a sliding window to divide the long time-series signal after noise reduction and normalization into one-dimensional analysis sample segments of fixed length that can overlap.

3. The reliable intelligent diagnostic method for switch machine oil pumps based on conditional diffusion data enhancement according to claim 1, characterized in that, The condition vector is a numerical representation of the fault category label and the health status index after encoding. The fault category label adopts one-hot encoding and includes at least normal, bearing outer ring fault, bearing inner ring fault and gear tooth breakage. The encoded health status index is a health degree quantification index, which is a continuous scalar with a value range of [0,1], used to characterize the degree of evolution from a healthy state to a complete failure, and is mapped according to the preset percentage of remaining service life.

4. The reliable intelligent diagnostic method for switch machine oil pumps based on conditional diffusion data enhancement according to claim 3, characterized in that, Step S4 specifically includes the following steps: Step S41: Randomly sample real samples from the real sample set, gradually add Gaussian noise to the analysis sample segment of the real samples, and obtain pure Gaussian noise after multiple forward diffusion steps. Step S42: Use U-Net network as the denoising network, and use the conditional vector as the key and value. Embed it into the feature map of each layer of the encoder and decoder of U-Net network through cross-attention mechanism, so that U-Net network can guide noise prediction according to the specified fault category label and health level quantification index. Step S43: Using the mean square error between predicted noise and actual noise as the loss function, train the conditional diffusion model to learn the reverse denoising process.

5. The reliable intelligent diagnostic method for switch machine oil pumps based on conditional diffusion data enhancement according to claim 4, characterized in that, Step S5 specifically includes the following steps: Step S51: Construct a target condition vector. The target condition vector has the same format and dimensions as the aforementioned condition vector. It includes target fault category labels and target health level quantitative indicators. This target condition vector has the same format and dimensions as the aforementioned condition vector. Step S52: Randomly sample a pure Gaussian noise with the same dimension as the real sample as the initial state of the reverse denoising process; all elements of the pure Gaussian noise are independent of each other and all follow a standard normal distribution; Step S53: Input pure Gaussian noise and target conditional vector into the trained U-Net network, predict and remove noise one by one, gradually restore the original real sample, and output high-quality multi-channel synthetic sample that meets the specified conditions. Step S54: Mix the generated synthetic samples with the original real samples to form an enhanced training set.

6. The reliable intelligent diagnostic method for switch machine oil pumps based on conditional diffusion data enhancement according to claim 5, characterized in that, Step S6 specifically includes the following steps: Step S61: Divide the augmented training set into multiple segments that may overlap along the length dimension, with each segment having a fixed length; Step S62: Map each segment to a feature space of a specified dimension through a learnable linear layer to obtain a projection sequence, and add a learnable positional encoding matrix to preserve the temporal order; Step S63: Process all segments using a Transformer encoder composed of multiple stacked layers. Each Transformer encoder layer contains a multi-head self-attention sub-layer and a feedforward network sub-layer. Perform residual connection and layer normalization processing on the multi-head self-attention sub-layer and the feedforward network sub-layer respectively. Step S64: Perform global average pooling on the output features of all segments, and then fuse the multi-channel features into the final feature vector through a learnable linear layer.

7. The reliable intelligent diagnostic method for switch machine oil pumps based on conditional diffusion data enhancement according to claim 6, characterized in that, In step S7, The fault classification head consists of a fully connected layer and a Softmax activation function, which outputs the probability distribution of each fault category and takes the category with the highest probability as the preliminary classification result. The health assessment head consists of a fully connected layer with no special activation function, and outputs continuous scalar values ​​as preliminary predictions of health index or remaining lifespan.

8. The reliable intelligent diagnostic method for switch machine oil pumps based on conditional diffusion data enhancement according to claim 7, characterized in that, Step S8 specifically involves inferring the confidence level of the prediction result based on the uncertainty measure. When the information entropy exceeds the preset entropy threshold, it is determined that the model's uncertainty regarding the prediction is too high, i.e., the confidence level is too low. A diagnostic report is generated based on the confidence level of the fault category and the uncertainty range of the predicted health index value. When the classification confidence level is lower than the preset confidence threshold or the uncertainty range exceeds the safety threshold, a high-risk warning is triggered, and low-confidence samples are automatically recorded for active learning to achieve continuous model evolution.

9. A reliable intelligent diagnostic system for a switch machine oil pump based on conditional diffusion data augmentation, implementing the single-based conditional diffusion data augmentation reliable intelligent diagnostic method for switch machine oil pumps as described in any one of claims 1-8, characterized in that, include: The data preprocessing module is used to preprocess the multi-source heterogeneous data of the collected switch machine oil pump to obtain standardized analysis sample segments. A condition vector is generated to guide the generation process based on preset fault category labels and health status index codes for switch machine oil pumps. Establish a connection between standardized sample fragments and conditional vectors to construct a real sample set; The data augmentation module communicates with the data preprocessing module and is used to train the real sample set based on the conditional diffusion model. It also embeds the conditional vectors into each layer of the U-Net network through the cross-attention mechanism to learn the reverse denoising process from noise to the original real sample. Multi-channel synthetic samples are generated based on specified fault categories and health states to construct an enhanced training set; The temporal feature extraction module, which is connected to the data augmentation module, is used to perform block segmentation, projection coding, and Transformer coding on the multi-channel one-dimensional temporal signals in the augmentation training set using the temporal feature extraction model, so as to extract deep global fault features in the vibration signal and output the final feature vector. The multi-task learning and uncertainty quantification module communicates with the temporal feature extraction module. It is used to simultaneously feed the extracted final feature vector to the fault classification head and the health assessment head, outputting the probability distribution of the fault category and the predicted value of the health index, respectively. It quantifies the uncertainty of the prediction through multiple random forward propagations, using the information entropy of the probability distribution of the fault category as the uncertainty measure and the standard deviation of the predicted value of the health index as the uncertainty interval. The diagnostic report output module communicates with the multi-task learning and uncertainty quantification module to generate a diagnostic report based on the uncertainty measure and the uncertainty range of the health index prediction value.

Images