A locomotive signal vehicle-mounted equipment safety fault diagnosis method and system

Abstract

The application discloses a locomotive signal vehicle-mounted equipment safety fault diagnosis method and system, the method obtains locomotive signal vehicle-mounted equipment operation data, fault event record and equipment maintenance record, slices, fills and filters the fault event time period to obtain a labeled sample and an unlabeled sample; the feature extraction network is used to obtain sample embedding features, and the labeled sample is used to construct a fault cause category prototype; in the candidate fault cause set of the unlabeled sample, reliable samples are filtered according to the nearest prototype distance and the second nearest prototype interval, pseudo labels and confidence weights are given, and the category prototype is updated by weighting; then the sample embedding features and the updated category prototype are constructed into a relationship pair feature, and the relationship discrimination module outputs a fault cause probability. The application can improve the distinguishing ability and diagnosis stability of similar fault causes, and provides equipment layer diagnosis basis for locomotive signal vehicle-mounted equipment abnormal risk identification and safety situation prompt.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance and fault diagnosis technology of railway signaling equipment, and particularly relates to a method and system for diagnosing safety faults of locomotive signaling on-board equipment. Background Technology

[0002] Locomotive signaling equipment is a crucial component in train operation control and safety assurance. Its operational status directly impacts the safe operation and anomaly risk assessment of the high-speed railway signaling control system. During on-site maintenance, equipment operating data, fault event records, and maintenance logs reflect the correlation between equipment status and fault causes, providing fundamental information for equipment safety status analysis and subsequent early warning. Current fault diagnosis typically relies on human experience, mainframe fault indications, and equipment maintenance results, which are prone to errors due to human experience and difficulty in distinguishing between similar fault phenomena.

[0003] Deep learning methods can extract features from time-series operational data, but they typically require a large number of labeled samples. For locomotive signaling onboard equipment, collecting real fault samples is difficult, the number of labeled samples confirmed by maintenance records is limited, while there are many unlabeled samples that have not been confirmed by clear maintenance records. If a conventional supervised classification model is used directly, it is easy to overfit on a small number of samples; if unlabeled samples are used without screening, noisy samples are easily introduced into the training process, causing class prototype shift.

[0004] Prototype networks can construct class centers using a small number of samples, making them suitable for small-sample classification tasks. However, traditional prototype networks typically classify based on distance metrics, which struggle to adequately characterize the nonlinear differences between similar fault causes in locomotive signal time-series data. While existing semi-supervised prototyping methods can refine class prototypes using unlabeled samples, there is still room for improvement in areas such as fault event candidate cause constraints, pseudo-label reliability screening, weighted prototype updates, and sample-prototype relationship discrimination.

[0005] Therefore, there is an urgent need for a fault diagnosis method for locomotive signal onboard equipment that can reliably utilize unlabeled samples and improve the ability to distinguish similar fault causes when there are limited labeled fault samples, so that the fault cause diagnosis results can further serve the equipment safety status assessment and abnormal risk warning. Summary of the Invention

[0006] The purpose of this invention is to provide a fault diagnosis method for JT-C series locomotive signal on-board equipment, in order to solve the problems of insufficient diagnostic accuracy and stability of existing methods when there are few tagged fault samples, untagged samples are difficult to use reliably, and similar fault causes are not easy to distinguish, and to provide a reliable basis for fault cause diagnosis for equipment safety status assessment and abnormal risk warning.

[0007] To achieve the above objectives, the present invention provides a method and system for diagnosing safety faults in locomotive signal on-board equipment.

[0008] The present invention provides a method for diagnosing safety faults in locomotive signal onboard equipment, comprising:

[0009] The system acquires the operating data, fault event records, and equipment maintenance records of the JT-C series locomotive signal onboard equipment during the fault event period, and classifies the fault event samples into labeled samples and unlabeled samples based on the equipment maintenance records.

[0010] The operational data is sliced ​​into time windows, missing data is filled in, and validity is filtered to obtain a time series sample containing continuous quantities, status quantities, lamp position information, system information, and fault indication information.

[0011] Time-series samples are input into a feature extraction network to obtain sample embedding features.

[0012] Initial category prototypes for each fault cause category are constructed based on the sample embedding features of labeled samples.

[0013] For unlabeled samples, a set of candidate fault causes is determined based on their corresponding fault events, and the distance between the sample embedding features and the class prototype of the unlabeled sample is calculated only within the set of candidate fault causes.

[0014] When an unlabeled sample simultaneously satisfies both the nearest prototype distance condition and the second nearest prototype interval condition, it is assigned a pseudo-label, and the pseudo-label confidence weight is determined based on the nearest prototype distance and the second nearest prototype interval.

[0015] The category prototype is updated by weighting the initial category prototype and the unlabeled samples with pseudo-labels, resulting in the updated category prototype.

[0016] The sample embedding features of the sample to be diagnosed and the updated category prototypes in the corresponding candidate fault cause set are constructed as sample-prototype relationship pair features. The relationship discrimination module is input to obtain the relationship score, and the relationship score of the candidate fault cause category is normalized. The fault cause diagnosis result is output based on the normalized relationship score.

[0017] Furthermore, the operating data includes the status of machine A, the status of machine B, the fault code of machine A, the fault code of machine B, the system, the lamp position, the power supply voltage, the amplitude, the carrier frequency, the low frequency, and the 50Hz power frequency signal.

[0018] Among them, the status of machine A, the status of machine B, and the system are used as binary features, the fault codes of machine A, the fault codes of machine B, and the lamp positions are used as discrete features, and the power supply voltage, amplitude, carrier frequency, low frequency, and 50Hz power frequency signal are used as continuous features.

[0019] The tagged samples are fault event slices with equipment maintenance records confirming the cause of the fault.

[0020] Unlabeled samples are fault event slices that do not have a clear correspondence with maintenance records but have fault event records.

[0021] The set of candidate fault causes is determined based on equipment maintenance rules and expert knowledge.

[0022] Furthermore, the time window slicing includes: sampling operational data within the time period of a fault event at a fixed time length, generating one or more slices for each fault event. When data records are missing at certain time points within a slice, forward padding is used to fill in the gaps. If the number of valid record points within a slice is less than a preset number, the slice is discarded.

[0023] The fixed time length is 60 seconds. Each fault event generates a maximum of 5 slices. When there are fewer than 45 valid record points in a slice, the slice is discarded.

[0024] Furthermore, the feature extraction network includes a tensor quantization module, a local temporal feature extraction and channel weight calibration module, a global temporal dependency construction module, and a pooling output module.

[0025] The tensor quantization module converts continuous quantities, state quantities, lamp position information, standard information, and fault indication information into multivariate temporal tensors. The local temporal feature extraction and channel weight calibration module extracts local temporal variations and enhances key feature channels. The global temporal dependency construction module models global correlations between different time locations. The pooling output module outputs the embedded features of the samples.

[0026] The local temporal feature extraction and channel weight calibration module includes a two-level one-dimensional convolution module and a compression excitation module, while the global temporal dependency construction module includes a position encoding module and a Transformer encoder.

[0027] The kernel size in the two-level one-dimensional convolutional module is 3, the number of output channels in the first convolutional module is 64, the number of output channels in the second convolutional module is 96, the Transformer encoder has 2 layers, the number of self-attention heads is 4, and the hidden layer dimension of the feedforward network is 192.

[0028] Furthermore, the feature extraction network is trained based on an "N-way K-shot" meta-task. When training with an "N-way K-shot" meta-task, each meta-task includes a support set and a query set. Support set class prototypes are calculated based on the embedding features of the support set samples, and class probabilities are calculated based on the distance between the query set sample embedding features and the support set class prototypes. The feature extraction network is then updated using the loss function corresponding to the query set labels. Here, a "9-way 5-shot" meta-task is used to train the feature extraction network.

[0029] Furthermore, the initial category prototype is the mean of the embedded features of labeled samples of the same failure cause category.

[0030] Furthermore, the nearest prototype distance condition is that the distance from an unlabeled sample to the nearest category prototype is less than the category adaptive threshold determined by the distance distribution from the labeled sample of the corresponding fault cause category to the category prototype.

[0031] The category adaptive threshold is the quantile threshold of the distance distribution from the labeled sample to the category prototype in the corresponding fault cause category, which is the 90th quantile threshold here.

[0032] The second-nearest prototype interval condition is that the ratio of the difference between the second-nearest category prototype distance and the nearest category prototype distance to the nearest category prototype distance is not less than a set minimum interval ratio threshold.

[0033] The confidence weight of the pseudo-label is determined based on the nearest neighbor score and the interval score.

[0034] Among them, the nearest neighbor score increases as the nearest prototype distance decreases relative to the class adaptive threshold, and the interval score increases as the distance ratio between the second nearest class prototype distance and the nearest class prototype distance increases.

[0035] Furthermore, the weighted update of the category prototype includes: setting the weight of the unlabeled sample assigned a pseudo-label to its pseudo-label confidence weight, and performing a weighted update using the embedded features of the initial category prototype and the pseudo-label sample to obtain the updated category prototype.

[0036] Furthermore, the sample-prototype relationship features include the sample embedding features of the sample to be diagnosed, the absolute difference between the sample embedding features of the sample to be diagnosed and the class prototype, the Euclidean distance between the sample embedding features of the sample to be diagnosed and the class prototype, and the inner product between the sample embedding features of the sample to be diagnosed and the class prototype.

[0037] Furthermore, the relationship discrimination module is a multilayer perceptron. The multilayer perceptron outputs a relationship score between a sample and its corresponding category prototype based on the sample-prototype relationship. During training, samples are paired with their true category or pseudo-label category prototypes to form positive relationship pairs, and samples are paired with prototypes of other categories or other candidate categories to form negative relationship pairs. During diagnosis, the relationship scores of the candidate fault cause categories corresponding to the sample to be diagnosed are normalized before outputting the probability of the fault cause. The multilayer perceptron includes two hidden layers.

[0038] The present invention provides a safety fault diagnosis system for locomotive signal on-board equipment, comprising:

[0039] The data acquisition module is used to acquire the operating data, fault event records, and equipment maintenance records of the locomotive signal onboard equipment during the fault event period.

[0040] The sample construction module is used to slice the runtime data into time windows, fill in missing data, and filter validity, and to divide the data into labeled and unlabeled samples based on equipment maintenance records.

[0041] The feature extraction module is used to input time-series samples into the feature extraction network to obtain sample embedded features.

[0042] The prototype building module is used to build initial category prototypes based on the sample embedding features of labeled samples.

[0043] The pseudo-label screening module is used to screen reliable unlabeled samples from the candidate fault cause set of unlabeled samples based on the nearest prototype distance condition and the second nearest prototype interval condition, and to determine the pseudo-label and pseudo-label confidence weight.

[0044] The prototype update module is used to perform weighted updates on category prototypes based on labeled samples and pseudo-labeled samples.

[0045] The relation discrimination module is used to construct sample-prototype relation pair features and output relation scores and fault cause diagnosis results.

[0046] The beneficial technical effects of this invention compared to the prior art are as follows:

[0047] (1) The present invention transforms the running process into a fixed-length multivariate time series sample by time window slicing, forward filling and multi-source locomotive signal feature tensor quantization;

[0048] (2) This invention narrows the category range of pseudo-label allocation and diagnostic output by using a candidate fault cause set, thereby reducing interference from irrelevant categories;

[0049] (3) The present invention filters unlabeled samples by using the nearest prototype distance condition and the second nearest prototype interval condition, thereby reducing prototype shift caused by low confidence samples;

[0050] (4) This invention uses confidence weights to participate in category prototype updates, enabling reliable unlabeled samples to participate in category center estimation;

[0051] (5) The present invention outputs the probability of candidate fault causes by using the sample-prototype relationship to distinguish features and relationships, thereby improving the ability to distinguish similar fault causes.

[0052] (6) This invention provides support for the safety status assessment, abnormal risk identification and operation and maintenance early warning of locomotive signal on-board equipment by outputting fault cause probability and confidence information. Attached Figure Description

[0053] Figure 1 The overall flowchart of the locomotive signal on-board equipment safety fault diagnosis method provided in the embodiments of the present invention is shown.

[0054] Figure 2 This is a schematic diagram of sample construction provided in an embodiment of the present invention.

[0055] Figure 3 This is a schematic diagram of the feature extraction network structure provided in an embodiment of the present invention.

[0056] Figure 4 This is a schematic diagram illustrating pseudo-label filtering and prototype updating provided in an embodiment of the present invention.

[0057] Figure 5 This is a schematic diagram of the relationship discrimination and diagnosis output provided in an embodiment of the present invention. Detailed Implementation

[0058] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0059] The flowchart of a safety fault diagnosis method for locomotive signal on-board equipment of the present invention is as follows: Figure 1 As shown, it specifically includes:

[0060] Step 1: Obtain the operating data, fault event records, and equipment maintenance records of the JT-C series locomotive signal onboard equipment during the fault event period, and divide the samples into labeled samples and unlabeled samples based on the maintenance records.

[0061] Step 2: Perform time window slicing, forward filling, and validity filtering on the running data to generate a time series sample containing continuous quantities, status quantities, lamp position information, system information, and fault indication information.

[0062] Step 3: Input the temporal samples into the feature extraction network to obtain the sample embedding features. The feature extraction network processes the temporal samples in the following order: tensor quantization, local temporal feature extraction and channel weight calibration, global temporal dependency construction, and pooling output.

[0063] Step 4: Construct an initial category prototype for the fault cause category based on the embedded features of the training set samples with labels.

[0064] Step 5: For unlabeled samples, determine the candidate fault cause set based on their fault events, and calculate the distance between the unlabeled sample and the category prototype within the candidate fault cause set; when the unlabeled sample meets the nearest prototype distance condition and the second nearest prototype interval condition, assign a pseudo-label to the unlabeled sample, and determine the pseudo-label confidence weight based on the proximity degree and the interval degree.

[0065] Step 6: Based on the initial category prototype and the embedded features of the pseudo-label samples, perform a weighted update on the category prototype to obtain the updated category prototype.

[0066] Step 7: Construct sample-prototype relationship pair features by combining the embedded features of the sample to be diagnosed with the updated category prototypes in the corresponding candidate fault cause set. Input the relationship discrimination module to obtain the relationship score between the sample and each candidate fault cause category prototype, normalize the candidate category relationship score, and output the fault cause diagnosis result based on the normalization result.

[0067] In actual execution, the data acquisition phase outputs operational data, fault event records, and maintenance records within the fault event time period; the sample construction phase outputs labeled and unlabeled samples; the feature extraction phase outputs sample embedding features; the prototype construction phase outputs initial category prototypes; the unlabeled sample screening phase outputs pseudo-labeled samples and their confidence weights; the prototype update phase outputs updated category prototypes; and the relationship discrimination phase outputs the relationship scores and normalized probabilities of each candidate fault cause category. The outputs of each of these phases serve as inputs for the next phase, forming a closed-loop processing mechanism from field data to probabilistic fault cause diagnosis results. These diagnostic results can be used as input information for equipment safety status analysis and anomaly risk alerts.

[0068] like Figure 2 As shown, the data acquisition phase involves acquiring operational data, fault event records, and equipment maintenance records from the locomotive signaling onboard equipment. Operational data includes the status of locomotive A and locomotive B, fault codes for locomotive A and B, signal type, lamp position, power supply voltage, amplitude, carrier frequency, low-frequency signal, and 50Hz power frequency signal. Fault event records are used to determine the time period of fault events and the range of candidate fault causes, while equipment maintenance records are used to determine whether fault events have a clear fault cause label.

[0069] In one embodiment of the present invention, the sample construction module samples data within a fault event time period using a fixed time window of 60 seconds. For fault events with a long duration, a maximum of five 60-second slices are generated for each fault event. If no data is recorded within certain seconds, forward padding is used to fill in the gaps; if a slice contains fewer than 45 valid record points, the slice is discarded. Through the above processing, each slice forms a time-series sample with a fixed time step. Fault event slices with a clear correspondence to maintenance records are designated as labeled samples, while slices without a clear correspondence to maintenance records but containing fault event information are designated as unlabeled samples.

[0070] In one embodiment of the present invention, the tensor module converts the status of machine A, the status of machine B, and the system into binary features; converts the lamp positions into lamp position embedded features; splits the fault codes of machine A and machine B into two fault code fields and converts them into fault code embedded features; and treats the power supply voltage, amplitude, carrier frequency, low frequency, and 50Hz power frequency signal as continuous features and performs normalization processing. The above features are concatenated step-by-step to form a multivariate time-series tensor.

[0071] In one embodiment of the present invention, fault events include host malfunction events, light outage or multiple light outage events, coil malfunction events, uplink / downlink switch malfunction events, and no TAX box information events; fault cause categories include I-SZ or II-SZ cable faults, uplink / downlink circuit open circuits, host-TAX box communication faults, host board or connection board faults, ground signal malfunctions, light display short circuits, power board faults, coil or coil-to-host cable faults, and recording board faults. The candidate fault cause set can be determined based on the correspondence between fault events and fault causes.

[0072] In one embodiment of the present invention, the correspondence between fault events and candidate fault cause sets can be determined as follows: the candidate fault cause set corresponding to host abnormal events, light-out or multi-light events, and no TAX box information events includes I-SZ or II-SZ cable faults, uplink / downlink circuit open circuits, host-TAX box communication faults, host board or connection board faults, ground signal abnormalities, light display short circuits, power board faults, coil or coil-to-host cable faults, and recording board faults; the candidate fault cause set corresponding to uplink / downlink switch abnormal events includes I-SZ or II-SZ cable faults, uplink / downlink circuit open circuits, and host board or connection board faults; the candidate fault cause set corresponding to coil abnormal events includes ground signal abnormalities and coil or coil-to-host cable faults.

[0073] In one embodiment of the present invention, the range of possible fault causes corresponding to different fault events can be determined by expert knowledge and equipment maintenance rules, forming a candidate fault cause set. This candidate fault cause set is used for subsequent pseudo-label screening of unlabeled samples and relationship determination of samples to be diagnosed, avoiding irrelevant categories from participating in the competition.

[0074] like Figure 3 As shown, the feature extraction network first converts multi-source machine information data into multivariate temporal tensors through a tensor quantization module. Subsequently, the multivariate temporal tensors are passed through two levels of one-dimensional convolution, batch normalization, activation function, and compression activation module to extract local temporal features and assign weights to key channels. Then, positional encoding is added and input into the Transformer encoder to construct global temporal dependencies through a self-attention mechanism. Finally, the pooled output samples embed the features.

[0075] In one embodiment of the present invention, the first one-dimensional convolutional module has 64 output channels, the second one-dimensional convolutional module has 96 output channels, and the convolutional kernel size is 3; the Transformer encoder has 2 layers, the number of self-attention heads is 4, and the hidden layer dimension of the feedforward network is 192. The above parameters are preferred embodiments, and can be adjusted according to the data scale and device computing power without departing from the technical concept of the present invention.

[0076] In one embodiment of the present invention, the feature extraction network is trained using a few-shot meta-task approach. Each meta-task includes a support set and a query set. For samples in the support set belonging to the same fault cause category, the mean value of their embedded features is calculated to obtain the support set category prototype. For query set samples, the category probability is obtained based on the distance between the query set sample's embedded features and the support set category prototype, and the feature extraction network is updated based on the query set label. Preferably, in an implementation scenario with nine fault cause categories, the feature extraction network is trained using a "9-way 5-shot" meta-task.

[0077] In one embodiment of the present invention, the initial category prototype can be obtained by averaging the embedded features of labeled training set samples of the same fault cause category; when screening unlabeled samples, only the distance between the unlabeled sample and the category prototype in its candidate fault cause set is calculated.

[0078] In one embodiment of the present invention, if the distance from an unlabeled sample to its nearest class prototype is less than the class adaptive threshold of that class, and the distance ratio between the distance to the second nearest class prototype and the distance to the nearest class prototype is not less than a set minimum distance ratio threshold, then the unlabeled sample is considered a reliable unlabeled sample, and the class corresponding to the nearest class prototype is used as its pseudo-label. The class adaptive threshold is taken as the 90th percentile of the distance distribution from the labeled samples of that class to the class prototype.

[0079] like Figure 4 As shown, in the pseudo-label screening and prototype update stages, each unlabeled sample is first input into the feature extraction network to obtain embedded features, and then the candidate fault cause set is determined based on the fault event to which the unlabeled sample belongs. When performing pseudo-label judgment on unlabeled samples, the distance between the unlabeled sample and each category prototype is calculated only within its candidate fault cause set, without including irrelevant fault cause categories in the competition.

[0080] In one embodiment of the present invention, the distance results within the candidate fault cause set are sorted to determine the nearest category prototype distance. Distance to the next nearest class prototype If an unlabeled sample simultaneously satisfies both the nearest prototype distance condition and the second nearest prototype interval condition, then the fault cause category corresponding to the nearest category prototype is used as the pseudo-label for the unlabeled sample; otherwise, the unlabeled sample is not used for category prototype updating. The above two judgment conditions can be expressed as follows:

[0081]

[0082]

[0083] Among them, the category adaptive distance threshold The distance distribution from the labeled samples corresponding to the fault cause category to the category prototype is determined, preferably by the 90th percentile; The minimum distance ratio threshold represents the ratio between the second-best and best distances. Used to control the degree of separation between the second nearest class prototype and the nearest class prototype. This prevents the denominator from being zero in the formula. The nearest prototype distance condition is used to determine whether an unlabeled sample is close enough to the nearest prototype in its candidate category, and the second nearest prototype interval condition is used to determine whether the unlabeled sample is sufficiently distinguishable from other candidate prototypes.

[0084] In one embodiment of the present invention, the confidence weight of the pseudo-label is further determined for the unlabeled samples that have passed the screening. The confidence weight is determined jointly by proximity and distance: the closer an unlabeled sample is to its pseudo-label category prototype, and the more significant the distance between it and other candidate category prototypes, the greater the weight of the unlabeled sample in subsequent prototype updates. The weight calculation relationship can be expressed as follows:

[0085]

[0086] Using the above methods, pseudo-labeled samples with higher reliability have a greater impact on category prototype updates, while unlabeled samples that are close to the boundary or whose candidate categories are not clearly distinguishable are filtered out or have their impact reduced, thereby reducing prototype shifts caused by erroneous pseudo-labels.

[0087] In one embodiment of the present invention, the prototype update module sets the weights of the selected pseudo-label samples as their pseudo-label confidence weights, and performs a weighted average of the embedded features of samples under the same fault cause category to obtain the updated category prototype. The update relationship can be represented as follows:

[0088]

[0089] in, For category Updated category prototype, For category The initial category prototype, For categories in the training set The number of samples, For pseudo-label samples belonging to category The sample data. This processing enables unlabeled samples to supplement class center information under reliability constraints and suppresses the influence of low-confidence samples on class prototypes.

[0090] like Figure 5As shown, in the relationship discrimination stage, the relationship discrimination module constructs a sample-prototype relationship pair feature by combining the embedded features of the sample to be diagnosed with the updated category prototypes in the candidate fault cause set. This relationship pair feature includes the sample embedded features, the absolute difference between the sample embedded features and the category prototypes, the Euclidean distance between them, and the inner product between them, which are used to describe the matching relationship from multiple perspectives such as sample self-representation, channel difference, distance metric, and directional similarity.

[0091] In one embodiment of the present invention, the construction relationship of the sample-prototype relationship with respect to features can be expressed as follows:

[0092]

[0093] The sample-prototype relationship pair features are obtained by concatenating multiple matching information sets mentioned above. The relationship discrimination module is a multilayer perceptron, which takes the sample-prototype relationship pair features as input and outputs the relationship score between the sample and its corresponding category prototype. This relationship score reflects the probability that the sample to be diagnosed belongs to the corresponding candidate fault cause category.

[0094] During training of the relationship discrimination module, samples are paired with their true or pseudo-label class prototypes to form positive relationship pairs, and samples are paired with other candidate class prototypes to form negative relationship pairs. The relationship discrimination module is then updated based on the loss function corresponding to the relationship scores. During diagnosis, the sample to be diagnosed is paired with each updated class prototype in its candidate fault cause set to obtain multiple relationship scores. These relationship scores are then normalized, and the probability of each candidate fault cause class is output. The class with the highest probability is taken as the diagnosis result.

[0095] The present invention provides a safety fault diagnosis system for locomotive signal on-board equipment, comprising:

[0096] The data acquisition module is used to acquire the operating data, fault event records, and equipment maintenance records of the locomotive signal onboard equipment during the fault event period.

[0097] The sample construction module is used to slice the runtime data into time windows, fill in missing data, and filter validity, and to divide the data into labeled and unlabeled samples based on equipment maintenance records.

[0098] The feature extraction module is used to input time-series samples into the feature extraction network to obtain sample embedded features.

[0099] The prototype building module is used to build initial category prototypes based on the sample embedding features of labeled samples.

[0100] The pseudo-label screening module is used to screen reliable unlabeled samples from the candidate fault cause set of unlabeled samples based on the nearest prototype distance condition and the second nearest prototype interval condition, and to determine the pseudo-label and pseudo-label confidence weight.

[0101] The prototype update module is used to perform weighted updates on category prototypes based on labeled samples and pseudo-labeled samples.

[0102] The relation discrimination module is used to construct sample-prototype relation pair features and output relation scores and fault cause diagnosis results.

[0103] In one embodiment of the present invention, the method of the present invention can be executed by an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it implements the method of the present invention.

[0104] The specific window length, number of network layers, number of channels, threshold, hidden layer dimension, and training parameters described in the above embodiments are preferred settings for ease of explanation. Those skilled in the art can make equivalent substitutions or parameter adjustments without changing the core technical solution of the present invention.

Claims

1. A method for diagnosing safety faults in locomotive signaling onboard equipment, characterized in that, include: The system acquires the operating data, fault event records, and equipment maintenance records of the JT-C series locomotive signal onboard equipment during the fault event period, and classifies the fault event samples into labeled samples and unlabeled samples based on the equipment maintenance records. The running data is sliced ​​into time windows, missing data is filled in, and validity is filtered to obtain a time series sample containing continuous quantities, status quantities, lamp position information, system information, and fault indication information; The time-series samples are input into the feature extraction network to obtain the sample embedding features; Initial category prototypes for each fault cause category are constructed based on the sample embedding features of labeled samples; For unlabeled samples, a set of candidate fault causes is determined based on their corresponding fault events, and the distance between the sample embedding features and the class prototype of the unlabeled sample is calculated only within the set of candidate fault causes. When an unlabeled sample simultaneously satisfies both the nearest prototype distance condition and the second nearest prototype interval condition, it is assigned a pseudo-label, and the pseudo-label confidence weight is determined based on the nearest prototype distance and the second nearest prototype interval. The category prototype is updated by weighting the initial category prototype and the unlabeled samples with pseudo-labels to obtain the updated category prototype; The sample embedding features of the sample to be diagnosed and the updated category prototypes in the corresponding candidate fault cause set are constructed as sample-prototype relationship pair features. The relationship discrimination module is input to obtain the relationship score, and the relationship score of the candidate fault cause category is normalized. The fault cause diagnosis result is output based on the normalized relationship score.

2. The method for diagnosing safety faults in locomotive signaling onboard equipment according to claim 1, characterized in that, The operating data includes the status of machine A, the status of machine B, the fault code of machine A, the fault code of machine B, the system, the lamp position, the power supply voltage, the amplitude, the carrier frequency, the low frequency, and the 50Hz power frequency signal. Among them, the status of machine A, the status of machine B, and the system are used as binary features, the fault codes of machine A, the fault codes of machine B, and the lamp positions are used as discrete features, and the power supply voltage, amplitude, carrier frequency, low frequency, and 50Hz power frequency signal are used as continuous features. The tagged sample is a slice of fault events with equipment maintenance records confirming the cause of the fault; The unlabeled samples are fault event slices that do not have a clear correspondence with maintenance records but have fault event records.

3. The method for diagnosing safety faults in locomotive signaling onboard equipment according to claim 1, characterized in that, The time window slicing includes: sampling the operational data within the time period of the fault event at a fixed time length, generating one or more slices for each fault event; filling in missing data records at some time points within a slice using a forward filling method; and discarding a slice when the number of valid record points within the slice is less than a preset number. The fixed time length is 60 seconds. Each fault event generates a maximum of 5 slices. When there are fewer than 45 valid record points in a slice, the slice is discarded.

4. The method for diagnosing safety faults in locomotive signaling onboard equipment according to claim 1, characterized in that, The feature extraction network includes a tensor module, a local temporal feature extraction and channel weight calibration module, a global temporal dependency construction module, and a pooling output module. The tensor quantization module converts continuous quantities, state quantities, lamp position information, standard information, and fault indication information into multivariable temporal tensors; the local temporal feature extraction and channel weight calibration module extracts local temporal changes and enhances key feature channels; the global temporal dependency construction module models the global correlation between different time positions; and the pooling output module outputs the embedded features of the samples. The local temporal feature extraction and channel weight calibration module includes a two-level one-dimensional convolution module and a compression excitation module, and the global temporal dependency construction module includes a position encoding module and a Transformer encoder. The two-level one-dimensional convolutional module has a kernel size of 3, the first convolutional module has 64 output channels, the second convolutional module has 96 output channels, the Transformer encoder has 2 layers, the number of self-attention heads is 4, and the hidden layer dimension of the feedforward network is 192.

5. The method for diagnosing safety faults in locomotive signaling onboard equipment according to claim 4, characterized in that, The feature extraction network is trained based on the "N-way K-shot" meta-task. When training with the "N-way K-shot" meta-task, each meta-task includes a support set and a query set. The support set category prototype is calculated based on the embedded features of the support set samples, and the category probability is calculated based on the distance between the embedded features of the query set samples and the support set category prototype. The feature extraction network is updated with the loss function corresponding to the query set label. The feature extraction network is trained using a "9-way 5-shot" meta-task.

6. The method for diagnosing safety faults in locomotive signaling onboard equipment according to claim 1, characterized in that, The nearest prototype distance condition is that the distance from an unlabeled sample to the nearest category prototype is less than the category adaptive threshold determined by the distance distribution from the labeled sample of the corresponding fault cause category to the category prototype. The category adaptive threshold is the quantile threshold of the distance distribution from the labeled sample to the category prototype in the corresponding fault cause category, which is the 90th quantile threshold here; The second nearest prototype spacing condition is that the ratio of the difference between the second nearest category prototype distance and the nearest category prototype distance to the nearest category prototype distance is not less than a set minimum spacing ratio threshold. The pseudo-label confidence weight is determined based on the nearest neighbor score and the interval score; Among them, the nearest neighbor score increases as the nearest prototype distance decreases relative to the class adaptive threshold, and the interval score increases as the distance ratio between the second nearest class prototype distance and the nearest class prototype distance increases.

7. The method for diagnosing safety faults in locomotive signaling onboard equipment according to claim 1, characterized in that, The weighted update of the category prototype includes: setting the weight of the unlabeled sample with the pseudo-label to its pseudo-label confidence weight, and performing a weighted update using the embedded features of the initial category prototype and the pseudo-label sample to obtain the updated category prototype.

8. The method for diagnosing safety faults in locomotive signaling onboard equipment according to claim 1, characterized in that, The sample-prototype relationship features include the sample embedding features of the sample to be diagnosed, the absolute difference between the sample embedding features of the sample to be diagnosed and the class prototype, the Euclidean distance between the sample embedding features of the sample to be diagnosed and the class prototype, and the inner product between the sample embedding features of the sample to be diagnosed and the class prototype.

9. The method for diagnosing safety faults in locomotive signaling onboard equipment according to claim 1, characterized in that, The relationship discrimination module is a multilayer perceptron. The multilayer perceptron outputs a relationship score between the sample and the corresponding category prototype based on the sample-prototype relationship. During training, the sample is paired with its true category or pseudo-label category prototype to form a positive relationship pair, and the sample is paired with other category prototypes or other candidate category prototypes to form a negative relationship pair. During diagnosis, the relationship score of the candidate fault cause category corresponding to the sample to be diagnosed is normalized and the probability of the fault cause is output. The multilayer perceptron includes two hidden layers.

10. A safety fault diagnosis system for locomotive signal on-board equipment, characterized in that, include: The data acquisition module is used to acquire the operating data, fault event records, and equipment maintenance records of the locomotive signal onboard equipment during the fault event period; The sample construction module is used to slice the runtime data into time windows, fill in missing data, and filter validity, and to divide the data into labeled and unlabeled samples based on equipment maintenance records. The feature extraction module is used to input time-series samples into the feature extraction network to obtain sample embedding features; The prototype building module is used to build initial category prototypes based on the sample embedding features of labeled samples; The pseudo-label screening module is used to screen reliable unlabeled samples from the candidate fault cause set of unlabeled samples based on the nearest prototype distance condition and the second nearest prototype interval condition, and to determine the pseudo-label and pseudo-label confidence weight. The prototype update module is used to update the category prototypes in a weighted manner based on labeled samples and pseudo-labeled samples; The relation discrimination module is used to construct sample-prototype relation pair features and output relation scores and fault cause diagnosis results.

Images