Train fault diagnosis method and system based on deep learning

Abstract

The present invention provides a train fault diagnosis method and system based on deep learning, which relates to the field of deep learning technology. By retrieving a priori fault feature chain data that has an associated pattern relationship with the current train fault feature chain data, and generating low-dimensional and high-dimensional a priori fault feature chain data sequences respectively, the fault diagnosis is significantly improved in terms of pertinence and accuracy. Furthermore, the present invention calculates the correlation between each priori fault feature chain data and the current train fault feature chain data, and extracts the associated pattern fault feature chain data based on the correlation, providing reference information for fault diagnosis. Finally, by combining the current train fault feature chain data and multiple associated pattern fault feature chain data, efficient and accurate diagnosis of the target train fault is achieved, greatly improving the intelligence level and diagnostic efficiency of train fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a train fault diagnosis method and system based on deep learning. Background Art

[0002] During train operation, fault diagnosis and elimination are key links to ensure safe and efficient train operation. Traditional train fault diagnosis methods often rely on manual experience and simple rule judgments. This method has great limitations when dealing with complex and changeable fault conditions, and it is difficult to ensure the accuracy and timeliness of diagnosis. With the increasing complexity and intelligence of train systems, traditional fault diagnosis methods can no longer meet the needs of modern train operation. Summary of the invention

[0003] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a train fault diagnosis method based on deep learning, the method comprising:

[0004] Performing deep feature learning on the train fault feature chain data reported by the target train to generate a low-dimensional embedding vector and a high-dimensional detail vector of the train fault feature chain data;

[0005] Retrieving a plurality of priori fault feature chain data whose corresponding low-dimensional embedding vector has an associated pattern relationship with the low-dimensional embedding vector of the train fault feature chain data, generating a first priori fault feature chain data sequence, and retrieving a plurality of priori fault feature chain data whose corresponding high-dimensional detail vector has an associated pattern relationship with the high-dimensional detail vector of the train fault feature chain data, generating a second priori fault feature chain data sequence;

[0006] Calculating the correlation between each priori fault feature chain data in the first priori fault feature chain data sequence and the second priori fault feature chain data sequence and the train fault feature chain data;

[0007] Based on the correlation degree, extracting a plurality of associated pattern fault feature chain data having an associated pattern relationship with the train fault feature chain data from the first a priori fault feature chain data sequence and the second a priori fault feature chain data sequence;

[0008] Based on the train fault feature chain data and the multiple associated mode fault feature chain data, fault diagnosis is performed on the target train.

[0009] On the other hand, an embodiment of the present invention also provides a deep learning-based train fault diagnosis system, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0010] Based on the above aspects, the embodiment of the present application effectively extracts low-dimensional embedded vectors and high-dimensional detail vectors of train fault feature chain data through deep feature learning technology, thereby achieving a comprehensive and refined characterization of fault features. By retrieving prior fault feature chain data that has an associated pattern relationship with the current train fault feature chain data, and generating low-dimensional and high-dimensional prior fault feature chain data sequences respectively, the pertinence and accuracy of fault diagnosis are significantly improved. Furthermore, the present invention calculates the correlation between each prior fault feature chain data and the current train fault feature chain data, and extracts the associated pattern fault feature chain data based on the correlation, providing reference information for fault diagnosis. Finally, by combining the current train fault feature chain data and multiple associated pattern fault feature chain data, efficient and accurate diagnosis of target train faults is achieved, greatly improving the intelligence level and diagnostic efficiency of train fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a schematic diagram of the execution flow of the train fault diagnosis method based on deep learning provided in an embodiment of the present invention.

[0012] Figure 2 It is a schematic diagram of the hardware architecture of a train fault diagnosis system based on deep learning provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0013] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 It is a flow chart of a train fault diagnosis method based on deep learning provided by an embodiment of the present invention. The train fault diagnosis method based on deep learning is introduced in detail below.

[0014] Step S110, performing deep feature learning on the train fault feature chain data reported by the target train to generate a low-dimensional embedding vector and a high-dimensional detail vector of the train fault feature chain data.

[0015] In this embodiment, it is assumed that the target train is a high-speed EMU train, and the train fault feature chain data reported by it contains a lot of complex information. This information covers the operating status data of each subsystem of the train, such as the motor current and voltage fluctuations in the traction system, the brake pressure changes in the brake system, the brake pad wear sensor data, and various control signal states and fault codes in the vehicle control system. And these data form a complex fault feature chain in chronological order or logical causal relationship.

[0016] The server first needs to obtain the initial fault analysis network. This initial fault analysis network has a certain network structure, including information such as the connection method of neurons and the initial weight setting. Then, the server generates a deep feature learning network based on this initial fault analysis network. This process may involve adjustments to the network structure, such as adding hidden layers, adjusting the number of neurons, etc., to meet the needs of deep feature learning of train fault feature chain data.

[0017] Next, the server needs to use the first-instance fault feature chain data corresponding to the deep feature learning network to optimize the parameters of the deep feature learning network. These first-instance fault feature chain data are carefully selected and sorted from a large number of historical train failure cases. For example, the fault feature chain data of a historical failure case is about the failure of the train traction system, which contains detailed information such as abnormal increase in current and unstable voltage fluctuations of the motor under specific working conditions. This information constitutes a complete fault feature chain data sample. The server inputs many such first-instance fault feature chain data into the deep feature learning network and optimizes the parameters of the network through technical means such as the back propagation algorithm.

[0018] During the optimization process, the server will also perform a window masking operation on the local node features in the first-instance fault feature chain data. For example, for the first-instance fault feature chain data of a certain brake system fault, the server may choose to mask the data of the brake pressure sensor within a certain period of time, thereby generating the masked fault feature chain data. At the same time, the target node features corresponding to the masked window are obtained from the original first-instance fault feature chain data, that is, the real data that should have been at this position. Then, the node feature restoration network is used to try to restore the masked node features. Assuming that the node feature restoration network is a neural network-based model, it performs the restoration operation based on the surrounding unshielded node features and the contextual relationship of the entire fault feature chain data. By comparing the restored node features with the target node features, the confidence that the node features corresponding to the masked window obtained by the node feature restoration network are the target node features is generated. For example, if the restored brake pressure data is very close to the original real data, the confidence will be very high. Based on this confidence, the server optimizes the neuron parameter information of the node feature restoration network, and finally generates the target node feature restoration network, and uses it as part of the initialization fault analysis network.

[0019] After completing the above preparations, the server starts deep feature learning of the train fault feature chain data reported by the target train. The server first obtains the fault feature path diagram corresponding to the train fault feature chain data. For a complex mechanical and electrical system such as a train, the fault feature path diagram is very complex. For example, taking the power supply system failure of the train as an example, the fault feature path diagram may start from the grid input voltage, through transformers, rectifiers and other equipment, to the distribution boxes and electrical equipment in each carriage, etc. The changes in voltage, current, power and other data in each link and the logical relationship between them constitute this fault feature path diagram. The server performs path decomposition on the train fault feature chain data, generates an initial fault feature path diagram, and configures the initial node features in the initial node part of the initial fault feature path diagram. For example, in the power supply system fault feature path diagram, the initial node may be the grid input point, and the initial node features may include the rated value of the input voltage, frequency and other information, thereby generating a complete fault feature path diagram.

[0020] The target deep feature learning network of the server includes multiple cascaded deep feature extraction units. For example, the first deep feature extraction unit may focus on extracting the basic feature relationship of each node in the fault feature path graph, and the second deep feature extraction unit further explores the deeper feature relationship based on the first unit. The server uses the first deep feature extraction unit to extract the graph convolution results corresponding to multiple fault feature paths from the fault feature path graph. For the fault feature path of "grid input-transformer-rectifier" in the fault feature path graph of the power supply system, the first deep feature extraction unit will calculate a graph convolution result based on the voltage, current and other characteristics of each node on this fault feature path and the connection relationship between them. This result contains the logical relationship characteristics of the path in the entire fault feature path graph, such as the transmission ratio of voltage, the transformation relationship of current, etc. Then, each other deep feature extraction unit is used to extract the graph convolution results corresponding to multiple fault feature paths from the graph convolution result set extracted by its forward deep feature extraction unit. For example, the second deep feature extraction unit further analyzes the interactive influence relationship between each fault feature path based on the graph convolution result set obtained by the first deep feature extraction unit to generate a new graph convolution result set.

[0021] Based on the graph convolution result set extracted by the last deep feature extraction unit in multiple cascaded deep feature extraction units, the server extracts a high-dimensional detail vector of the train fault feature chain data. Taking the power supply system fault as an example, this high-dimensional detail vector may contain information such as the detailed electrical parameter relationship between various devices and the potential path of fault propagation. At the same time, based on the graph convolution result of each fault feature path, the server calculates the influence weight of each fault feature path in the fault feature path graph. For example, in the power supply system fault feature path graph, if the fault feature path of "transformer-rectifier" has a greater impact on the fault of the entire power supply system, then its influence weight will be higher. Then, based on the preset fault feature path library and the influence weights corresponding to the multiple fault feature paths contained in the fault feature path graph, the server calculates the weights corresponding to the multiple fault feature paths contained in the preset fault feature path library. If a fault feature path in the preset fault feature path library exists in the current train fault feature path map, such as the fault feature path of "rectifier-distribution box" exists in the current fault feature path map, then the server calculates the weight of the fault feature path based on the influence weight of the fault feature path in the fault feature path map and the participation frequency in the map (for example, the number of times the fault feature path appears within a certain period of time). If a fault feature path in the preset fault feature path library does not exist in the current train fault feature path map, such as "special backup power supply line fault" (assuming that the target train does not have such a special backup power supply line), the preset weight is used as the weight of the fault feature path. Finally, based on these weights, the server generates a high-dimensional detail vector of the train fault feature chain data. In addition to the high-dimensional detail vector, the server also calculates the influence weight of each fault feature path in the fault feature path map based on the graph convolution results corresponding to the above multiple fault feature paths, and finally generates a low-dimensional embedding vector of the train fault fault feature chain data. This low-dimensional embedding vector can be regarded as a low-dimensional representation of the train fault feature chain data, which contains the main feature information of the data, while the high-dimensional detail vector focuses more on the detailed relationship between the fault features.

[0022] Step S120, retrieve multiple prior fault feature chain data whose corresponding low-dimensional embedding vectors have an associated pattern relationship with the low-dimensional embedding vectors of the train fault feature chain data, generate a first prior fault feature chain data sequence, and retrieve multiple prior fault feature chain data whose corresponding high-dimensional detail vectors have an associated pattern relationship with the high-dimensional detail vectors of the train fault feature chain data, generate a second prior fault feature chain data sequence.

[0023] In this embodiment, after generating the low-dimensional embedding vector and high-dimensional detail vector of the train fault feature chain data of the target train, the prior fault feature chain data is retrieved. The server has a huge prior fault feature chain data repository, which is accumulated from a large number of train fault cases in the past and contains detailed data of different fault conditions of various types of trains (such as high-speed trains, ordinary trains, freight trains, etc.).

[0024] For the retrieval of low-dimensional embedding vectors, the server compares the low-dimensional embedding vector of the train fault feature chain data of the target train with the low-dimensional embedding vector of each prior fault feature chain data in the repository. For example, assuming that the target train is a high-speed EMU train, the low-dimensional embedding vector of its train fault feature chain data reflects the low-dimensional representation of the comprehensive fault features of the train in terms of the traction system, braking system, and vehicle control system. The server traverses the low-dimensional embedding vectors of all prior fault feature chain data in the repository. If a prior fault feature chain data is a fault record of another high-speed EMU train, and its low-dimensional embedding vector has a similar numerical distribution to the low-dimensional embedding vector of the target train in certain key dimensions (such as dimensions related to the traction system, such as power and speed control), then the server believes that they have an associated pattern relationship. In this way, the server finds multiple prior fault feature chain data that have an associated pattern relationship with the low-dimensional embedding vector of the target train, and generates the first prior fault feature chain data sequence in the retrieval order.

[0025] For the retrieval of high-dimensional detail vectors, the server adopts a similar method. It compares the high-dimensional detail vector of the train fault feature chain data of the target train with the high-dimensional detail vector of the prior fault feature chain data in the repository. The high-dimensional detail vector contains more detailed fault feature relationship information. For example, the high-dimensional detail vector of the target train contains detailed relationships between multiple sub-features such as brake pad wear, brake pressure regulation, and anti-skid system working status in terms of brake system failure. The server searches for the high-dimensional detail vector of the prior fault feature chain data in the repository. If there is a prior fault feature chain data that is a record of a brake system failure of a certain train, and its high-dimensional detail vector has a similar pattern to the high-dimensional detail vector of the target train in terms of the relationship between brake pad wear and brake pressure regulation, for example, both show that the brake pad wear to a certain extent will cause abnormal brake pressure regulation, then the server believes that they have an associated pattern relationship. In this way, the server finds multiple such prior fault feature chain data and generates a second prior fault feature chain data sequence.

[0026] Step S130, calculating the correlation between each priori fault feature chain data in the first priori fault feature chain data sequence and the second priori fault feature chain data sequence and the train fault feature chain data.

[0027] In this embodiment, after obtaining the first priori fault feature chain data sequence and the second priori fault feature chain data sequence, the correlation between each priori fault feature chain data and the train fault feature chain data is calculated.

[0028] For each prior fault feature chain data in the first prior fault feature chain data sequence, the server first calculates the correlation between its low-dimensional embedding vector and the low-dimensional embedding vector of the train fault feature chain data. Assume that a distance-based measurement method, such as Euclidean distance, is used to measure this correlation. For example, for a prior fault feature chain data, its low-dimensional embedding vector represents a low-dimensional representation of some key features of a historical train failure. If this prior fault feature chain data is about a train traction system failure, the values ​​of its low-dimensional embedding vector in dimensions related to traction motor power, speed control, etc. are calculated by Euclidean distance with the values ​​of the low-dimensional embedding vector of the target train fault feature chain data in the corresponding dimensions. If the Euclidean distance is small, it means that they have a higher similarity in the low-dimensional feature space, and the correlation is higher. The server calculates the low-dimensional correlation corresponding to this prior fault feature chain data.

[0029] Then, for the same prior fault feature chain data, the server calculates the correlation between its high-dimensional detail vector and the high-dimensional detail vector of the train fault feature chain data. Since the high-dimensional detail vector contains more complex fault feature relationship information, the server may use a calculation method based on vector similarity, such as cosine similarity. For example, if the high-dimensional detail vector of the prior fault feature chain data describes in detail the relationship between multiple sub-features such as brake pad wear, brake pressure adjustment, and anti-skid system working status in terms of train brake system failure, and is very similar to the high-dimensional detail vector of the target train in the direction of these sub-feature relationships, then their cosine similarity will be close to 1, indicating that the high-dimensional correlation is very high.

[0030] After obtaining the low-dimensional correlation and high-dimensional correlation corresponding to each prior fault feature chain data, the server performs a weighted calculation on the two correlations. Assume that the server sets the weight of the low-dimensional correlation to 0.4 and the weight of the high-dimensional correlation to 0.6 based on the analysis and experience of historical data. For a prior fault feature chain data, its low-dimensional correlation is 0.8 and its high-dimensional correlation is 0.9, then the weighted correlation = 0.4 * 0.8 + 0.6 * 0.9 = 0.86. The server uses this weighted correlation as the correlation between the prior fault feature chain data and the train fault feature chain data.

[0031] For each prior fault feature chain data in the second prior fault feature chain data sequence, the server also performs the above-mentioned operations of calculating the low-dimensional correlation, high-dimensional correlation, and weighted correlation to obtain the correlation between each prior fault feature chain data and the train fault feature chain data.

[0032] Step S140: Based on the correlation degree, extract multiple associated pattern fault feature chain data having an associated pattern relationship with the train fault feature chain data from the first a priori fault feature chain data sequence and the second a priori fault feature chain data sequence.

[0033] In this embodiment, after calculating the correlation between each priori fault feature chain data and the train fault feature chain data in the first priori fault feature chain data sequence and the second priori fault feature chain data sequence, data extraction begins.

[0034] Assume that the server sets a correlation threshold of 0.8. For the prior fault feature chain data in the first prior fault feature chain data sequence, if its correlation is greater than or equal to 0.8, the server considers that this prior fault feature chain data has a strong correlation pattern relationship with the train fault feature chain data. For example, in the scenario of a train traction system failure, there is a prior fault feature chain data, and its correlation calculation result is 0.85, then this prior fault feature chain data will be extracted.

[0035] Similarly, for the priori fault feature chain data in the second priori fault feature chain data sequence, the server also judges and extracts according to the correlation threshold. If the correlation of a priori fault feature chain data reaches or exceeds 0.8, the server determines it as the correlation mode fault feature chain data that has a correlation mode relationship with the train fault feature chain data.

[0036] In this way, multiple correlation pattern fault feature chain data are extracted from the first priori fault feature chain data sequence and the second priori fault feature chain data sequence. These data are all data that have a strong correlation pattern relationship with the train fault feature chain data of the target train in terms of low-dimensional embedding vectors and high-dimensional detail vectors, and have important reference value for fault diagnosis of the target train.

[0037] Step S150: performing fault diagnosis on the target train based on the train fault feature chain data and the multiple associated mode fault feature chain data.

[0038] Assume that the train fault feature chain data shows that the target train has a fault in the braking system, which is manifested by features such as longer braking distance, abnormal noise during braking, and fluctuations in brake pressure. The server finds multiple prior fault feature chain data related to the brake system fault from the associated pattern fault feature chain data.

[0039] One of the related mode fault feature chain data shows that the previous train had similar longer braking distance, abnormal noise and brake pressure fluctuations due to severe wear of the brake pads, and the wear of the brake pads caused the brake disc surface to be uneven, which in turn caused abnormal noise and brake pressure fluctuations during braking. At the same time, the wear of the brake pads reduced the braking efficiency, resulting in a longer braking distance. Another related mode fault feature chain data shows that this fault may also be caused by the presence of air in the hydraulic pipeline of the brake system. The compressibility of the air causes brake pressure fluctuations, which in turn affects the braking effect, causing the braking distance to increase and abnormal noise to be generated.

[0040] The server diagnoses the fault of the target train based on the specific parameters in the train fault feature chain data (such as the specific value of the braking distance exceeding the normal range, the frequency range of the noise, the amplitude of the brake pressure fluctuation, etc.) and the fault causes and solutions in the associated mode fault feature chain data. For example, the server finds that the braking distance of the target train exceeds the normal range by a large margin, the brake pressure fluctuation amplitude is large, and the brake pads are found to be severely worn through inspection. At the same time, the prior data on the fault caused by brake pad wear in the associated mode fault feature chain data has a high degree of correlation. Then the server can preliminarily diagnose that the brake fault of the target train is caused by severe brake pad wear.

[0041] The server can also provide maintenance personnel with corresponding maintenance guidance based on the maintenance suggestions in the associated mode fault feature chain data, such as replacing brake pads, venting the brake system (if there is air in the hydraulic pipeline), and other operation suggestions, thereby realizing fault diagnosis and maintenance guidance for the target train.

[0042] Based on the above steps, the embodiment of the present application effectively extracts low-dimensional embedded vectors and high-dimensional detail vectors of train fault feature chain data through deep feature learning technology, thereby achieving a comprehensive and refined characterization of fault features. By retrieving prior fault feature chain data that has an associated pattern relationship with the current train fault feature chain data, and generating low-dimensional and high-dimensional prior fault feature chain data sequences respectively, the pertinence and accuracy of fault diagnosis are significantly improved. Furthermore, the present invention calculates the correlation between each prior fault feature chain data and the current train fault feature chain data, and extracts the associated pattern fault feature chain data based on the correlation, providing reference information for fault diagnosis. Finally, by combining the current train fault feature chain data and multiple associated pattern fault feature chain data, efficient and accurate diagnosis of target train faults is achieved, greatly improving the intelligence level and diagnostic efficiency of train fault diagnosis.

[0043] In a possible implementation manner, the low-dimensional embedding vector and the high-dimensional detail vector of the train fault feature chain data are generated by performing deep feature learning on the train fault feature chain data using a target deep feature learning network. The method further includes:

[0044] Step A110, obtaining an initialized fault analysis network, and generating a deep feature learning network based on the initialized fault analysis network.

[0045] Step A120: Based on the first sample fault feature chain data corresponding to the deep feature learning network, the parameters of the deep feature learning network are optimized to generate the target deep feature learning network.

[0046] In a possible implementation, the method further includes:

[0047] Step A101, performing window masking on local node features in the first sample fault feature chain data to generate masked fault feature chain data, and obtaining target node features corresponding to the masking window of the masked fault feature chain data from the first sample fault feature chain data.

[0048] Step A102, using a node feature restoration network to restore the node features corresponding to the shielding window of the shielding fault feature chain data, and generating a confidence that the node features corresponding to the shielding window obtained by the node feature restoration network are the target node features.

[0049] Step A103, optimizing the neuron parameter information of the node feature restoration network based on the confidence, generating a target node feature restoration network, and using the target node feature restoration network as the initialization fault analysis network.

[0050] In this embodiment, it is assumed that there are various types of trains in the railway system managed by the server, including high-speed EMU trains, ordinary passenger trains, and freight trains. The initialization fault analysis network is a pre-built basic network structure with certain neuron connection methods and initial weight settings. This network structure may be built based on the previous general understanding of the train system and some preliminary fault analysis models. For example, for the traction system of the train, the initialization network may preliminarily set the neuron connections related to the motor, as well as some weight settings related to speed control, current and voltage monitoring. These settings are based on the understanding of the basic operating principles of the traction system.

[0051] Based on this initialized fault analysis network, a deep feature learning network can be generated. This process involves adjusting the network structure to adapt to the deep feature learning needs of train fault feature chain data. Taking high-speed EMU trains as an example, due to the complexity of its system, the server may need to add hidden layers to mine deeper fault feature relationships. For example, the braking system of a high-speed EMU train contains multiple subsystems, such as brake discs, brake pads, hydraulic systems, etc., and the fault relationships between these subsystems are very complex. When generating a deep feature learning network, the server may add a hidden layer specifically for analyzing the interactive fault characteristics between the subsystems of the braking system. At the same time, the number of neurons may also be adjusted, such as increasing the number of neurons in the part related to the vehicle control system, because the vehicle control system involves numerous sensors and control signals, and more neurons are needed to process this complex information.

[0052] In this embodiment, a huge database of historical train fault cases is pre-built, from which the first example fault feature chain data is selected. These first example fault feature chain data are the key basis for the deep feature learning network to optimize parameters. For example, there is a historical fault case about a traction system failure of a high-speed EMU train. In this case, the traction motor was abnormal when the train was running at high speed, and its fault feature chain data contained detailed information such as abnormal increase in current and unstable voltage fluctuations of the motor under specific working conditions. This complete fault feature chain data sample becomes part of the first example fault feature chain data.

[0053] The server inputs many such first-instance fault feature chain data into the deep feature learning network, and optimizes the parameters of the network through technical means such as the back-propagation algorithm. Suppose another first-instance fault feature chain data is about the fault of the brake system of an ordinary passenger train, with severe wear of the brake pads and abnormal brake pressure regulation. When this first-instance fault feature chain data is input into the deep feature learning network, the parameters such as weights and biases in the network will be adjusted according to the difference between the input data and the expected output (in this case, accurate identification of the fault type and related features). For example, if the network makes an inaccurate judgment on the relationship between brake pad wear and brake pressure regulation in the initial state, through the input of this sample data and the optimization of the back-propagation algorithm, the network will adjust the neuron connection weights related to the brake system, so that this fault relationship can be more accurately identified when similar fault feature chain data is processed later.

[0054] In the process of optimizing the deep feature learning network, the local node features in the first-instance fault feature chain data can be window-shielded. Taking the vehicle control system failure of a freight train as an example, it is assumed that the first-instance fault feature chain data contains data from multiple sensors in the vehicle control system, such as temperature sensors, speed sensors, etc. The server may choose to shield the data of the temperature sensor within a certain period of time, thereby generating shielded fault feature chain data. In this shielding operation, the shielded temperature sensor data corresponds to the shielding window.

[0055] At the same time, the target node feature corresponding to the shielded window is obtained from the original first sample fault feature chain data, that is, the real data that should have been at this position. For example, the actual temperature value of the original temperature sensor data during the shielded time period is the target node feature. This target node feature is very important for the subsequent verification of the accuracy of the node feature restoration network.

[0056] Assuming that the node feature restoration network is a neural network-based model, the restoration operation is performed based on the surrounding unshielded node features and the contextual relationship of the entire fault feature chain data. Taking the above-mentioned freight train vehicle control system failure as an example, the node feature restoration network will try to restore the shielded temperature sensor data based on the speed sensor data related to the temperature sensor data, other control signal data, and the overall operating status of the vehicle control system in the entire fault feature chain data.

[0057] If the restored temperature data is very close to the original real data, the confidence level will be high. For example, the original temperature data shows that the temperature continues to rise to 50 degrees Celsius during the blocked time period, while the temperature data restored by the node feature restoration network is between 48 and 52 degrees Celsius, which indicates that the restoration result is relatively accurate, so the server will generate a higher confidence level for this restoration result. Conversely, if the restoration result is significantly different from the original data, the confidence level will be lower.

[0058] The server optimizes the neuron parameter information of the node feature restoration network based on the generated confidence. If in a first example fault feature chain data, the node feature restoration network has a high restoration confidence for the node features corresponding to the shielded window (such as the freight train temperature sensor data mentioned above), it means that the current neuron parameter setting is more effective in this case. However, if the confidence is low, the server needs to adjust the neuron parameters in the node feature restoration network. For example, it may adjust the weights related to the surrounding node features, or adjust parameters such as the neuron activation function in the hidden layer.

[0059] By performing such operations on multiple first-instance fault feature chain data, the neuron parameter information of the node feature restoration network is continuously optimized, and finally a target node feature restoration network is generated. This target node feature restoration network can more accurately restore the shielded node features and use them as part of the initialization fault analysis network. This helps to improve the processing capability of the entire fault analysis network for train fault feature chain data, because it can better handle local missing or abnormal situations in the data, thereby more accurately analyzing the train fault features.

[0060] In this way, a more optimized fault analysis network can be constructed, laying the foundation for the subsequent accurate analysis and fault diagnosis of the fault feature chain data reported by the target train.

[0061] In a possible implementation manner, the low-dimensional embedding vector and the high-dimensional detail vector of the train fault feature chain data are generated by performing deep feature learning on the train fault feature chain data using a target deep feature learning network. The method further includes:

[0062] Step B110, obtaining second sample fault feature chain data and sample association mode fault feature chain data corresponding to the second sample fault feature chain data.

[0063] Step B120, using a deep feature learning network to perform deep feature learning on the second sample fault feature chain data and the sample association mode fault feature chain data respectively, to generate a low-dimensional embedding vector and a high-dimensional detail vector of the second sample fault feature chain data, as well as a low-dimensional embedding vector and a high-dimensional detail vector of the sample association mode fault feature chain data.

[0064] Step B130, calculate the correlation between the low-dimensional embedding vector of the sample association mode fault feature chain data and the low-dimensional embedding vector of the second sample fault feature chain data, generate the low-dimensional correlation corresponding to the sample association mode fault feature chain data, and calculate the correlation between the high-dimensional detail vector of the sample association mode fault feature chain data and the high-dimensional detail vector of the second sample fault feature chain data, generate the high-dimensional correlation corresponding to the sample association mode fault feature chain data.

[0065] Step B140, calculating the network learning error based on the low-dimensional correlation corresponding to the sample association mode fault feature chain data, generating low-dimensional network learning error parameters, and calculating the network learning error based on the high-dimensional correlation corresponding to the sample association mode fault feature chain data, generating high-dimensional network learning error parameters.

[0066] Step B150, perform weighted calculation on the low-dimensional correlation and high-dimensional correlation corresponding to the sample association mode fault feature chain data, generate the weighted correlation corresponding to the sample association mode fault feature chain data, and calculate the network learning error based on the weighted correlation corresponding to the sample association mode fault feature chain data, and generate the weighted network learning error parameter.

[0067] Step B160, based on the low-dimensional network learning error parameter, the high-dimensional network learning error parameter and the weighted network learning error parameter, calculate the target network learning error parameter, and optimize the deep feature learning network based on the target network learning error parameter to generate the target deep feature learning network.

[0068] In a possible implementation manner, the sample association pattern fault feature chain data includes positive sample fault feature chain data having an association pattern relationship with the second sample fault feature chain data, and negative sample fault feature chain data that does not match the second sample fault feature chain data.

[0069] Step B140 includes: calculating the network learning error based on the low-dimensional correlation corresponding to the positive sample fault feature chain data and the low-dimensional correlation corresponding to the negative sample fault feature chain data, and generating the low-dimensional network learning error parameter, wherein the low-dimensional network learning error parameter is inversely correlated with the low-dimensional correlation corresponding to the positive sample fault feature chain data, and is positively correlated with the low-dimensional correlation corresponding to the negative sample fault feature chain data.

[0070] In this embodiment, first, the server obtains the second sample fault feature chain data and the corresponding sample associated mode fault feature chain data, which are also derived from the huge historical train fault case database stored by the server. Taking the high-speed EMU train as an example, the second sample fault feature chain data may be a related data chain about the train power supply system failure. This data chain contains detailed information about each link in the process from the power grid input to the electrical equipment in each carriage, such as the voltage conversion of the transformer, the current distribution of the distribution box, etc. The corresponding sample associated mode fault feature chain data, part of which is the positive sample fault feature chain data with an associated mode relationship with the power supply system failure, for example, there is a positive sample fault feature chain data recording the power supply system failure caused by the short circuit of the internal line of the distribution box, and the data of each link in its fault feature chain has a similar mode relationship with the power supply system failure in the current second sample fault feature chain data. At the same time, there are also negative sample fault feature chain data that do not match the second sample fault feature chain data, for example, a negative sample fault feature chain data is about the train brake system failure, which has no associated mode relationship with the power supply system failure.

[0071] Next, the server uses the deep feature learning network to perform deep feature learning on the second sample fault feature chain data and the sample association mode fault feature chain data. For the power supply system fault data in the second sample fault feature chain data, the deep feature learning network will mine the deep feature relationship between its various links. For example, the deep feature learning network will analyze the potential relationship between the grid input voltage fluctuation and the transformer output voltage stability, and convert these feature relationships into low-dimensional embedding vectors and high-dimensional detail vectors through complex calculations of multiple neurons. For the positive sample fault feature chain data, such as the data of the short circuit in the distribution box, the deep feature learning network will also analyze the relationship between each node in its fault feature chain, such as the current change before and after the short circuit point, the situation of other components in the distribution box being affected by the short circuit, etc., and then generate its low-dimensional embedding vector and high-dimensional detail vector. For the negative sample fault feature chain data, although it is not related to the second sample fault feature chain data, the deep feature learning network still processes it according to the established algorithm to generate the corresponding low-dimensional embedding vector and high-dimensional detail vector.

[0072] Then, the server calculates the correlation between the low-dimensional embedding vector of the sample association mode fault feature chain data and the low-dimensional embedding vector of the second sample fault feature chain data to generate the low-dimensional correlation corresponding to the sample association mode fault feature chain data. In the example of power supply system failure, for the positive sample fault feature chain data (distribution box short circuit), this embodiment can compare its low-dimensional embedding vector with the low-dimensional embedding vector of the second sample fault feature chain data (power supply system failure). Assuming that the low-dimensional embedding vector has a numerical representation in the dimensions related to voltage stability and current distribution, if the numerical value of the positive sample fault feature chain data is close to the numerical value of the second sample fault feature chain data in these dimensions, then according to the pre-set calculation method (such as the Euclidean distance calculation method), the low-dimensional correlation between the two will be high. For the negative sample fault feature chain data (brake system failure), since it has no correlation with the power supply system failure, the numerical difference in the relevant dimensions of the low-dimensional embedding vector is large, so the low-dimensional correlation will be very low. At the same time, the server also calculates the correlation between the high-dimensional detail vector of the sample association mode fault feature chain data and the high-dimensional detail vector of the second sample fault feature chain data, and generates the high-dimensional correlation corresponding to the sample association mode fault feature chain data. For example, for the positive sample fault feature chain data (distribution box short circuit), its high-dimensional detail vector contains the detailed fault relationship of each component in the distribution box, such as the relationship between the local overheating caused by the short circuit and the performance degradation of other components. In this embodiment, this high-dimensional detail vector can be compared with the high-dimensional detail vector of the second sample fault feature chain data (power supply system fault), and a calculation method such as cosine similarity may be used. If the vector representations of the two in terms of component relationship, fault propagation path, etc. are similar, then the high-dimensional correlation will be high; while the high-dimensional detail vectors of the negative sample fault feature chain data (brake system fault) and the power supply system fault are completely different in content and structure, and the high-dimensional correlation is extremely low.

[0073] The network learning error is calculated based on the low-dimensional correlation corresponding to the fault feature chain data of the sample association pattern, and the low-dimensional network learning error parameter is generated. For the low-dimensional correlation corresponding to the positive sample fault feature chain data and the low-dimensional correlation corresponding to the negative sample fault feature chain data, since the low-dimensional network learning error parameter is inversely correlated with the low-dimensional correlation corresponding to the positive sample fault feature chain data and is positively correlated with the low-dimensional correlation corresponding to the negative sample fault feature chain data, when calculating, if the low-dimensional correlation of the positive sample fault feature chain data is high, it means that the network recognizes this association pattern better, then the low-dimensional network learning error parameter should be small; while it is normal for the low-dimensional correlation of the negative sample fault feature chain data to be low, if this low-dimensional correlation is abnormally increased (indicating that the network may mistakenly judge irrelevant data as relevant), then the low-dimensional network learning error parameter will be large. For example, when the low-dimensional correlation of the positive sample fault feature chain data (distribution box short circuit) is calculated to be 0.8 (close to 1 means high correlation), according to the preset calculation rules, the low-dimensional network learning error parameter may be 0.2 (small); while if the low-dimensional correlation of the negative sample fault feature chain data (brake system fault) abnormally rises to 0.5 (should be very low), the low-dimensional network learning error parameter may become 0.5 (large). Similarly, the network learning error is calculated based on the high-dimensional correlation corresponding to the sample association pattern fault feature chain data to generate a high-dimensional network learning error parameter. If the high-dimensional correlation of the positive sample fault feature chain data is high, it means that the network is accurate in identifying the high-dimensional feature relationship, and the high-dimensional network learning error parameter should be small; conversely, if the high-dimensional correlation of the negative sample fault feature chain data is abnormal, the high-dimensional network learning error parameter will be large.

[0074] Afterwards, the server performs weighted calculation on the low-dimensional correlation and high-dimensional correlation corresponding to the sample correlation mode fault feature chain data, generates the weighted correlation corresponding to the sample correlation mode fault feature chain data, and calculates the network learning error based on this, and generates the weighted network learning error parameter. Assume that the server sets the weight of the low-dimensional correlation to 0.4 and the weight of the high-dimensional correlation to 0.6 based on historical data and experience. For the positive sample fault feature chain data (distribution box short circuit), its low-dimensional correlation is 0.8 and its high-dimensional correlation is 0.9, then the weighted correlation = 0.4 * 0.8 + 0.6 * 0.9 = 0.86. Then the weighted network learning error parameter is calculated based on this weighted correlation. If the weighted correlation is close to 1 (indicating a high correlation), the weighted network learning error parameter will be smaller; otherwise, it will be larger.

[0075] Finally, the server calculates the target network learning error parameter based on the low-dimensional network learning error parameter, the high-dimensional network learning error parameter and the weighted network learning error parameter, and optimizes the deep feature learning network based on the target network learning error parameter to generate a target deep feature learning network. For example, if the low-dimensional network learning error parameter is 0.2, the high-dimensional network learning error parameter is 0.15, and the weighted network learning error parameter is 0.18, this embodiment can calculate the target network learning error parameter according to a pre-set comprehensive calculation method (which may be a weighted average method). Assuming that the target network learning error parameter is calculated to be 0.17, this embodiment can adjust and optimize the neuron weights, biases and other parameters in the deep feature learning network according to this target network learning error parameter. For the neuron connection part related to the power supply system fault, if it is found that a certain neuron has a large error when processing the features related to voltage stability (reflected by the target network learning error parameter), this embodiment can adjust the connection weight between the neuron and other neurons to improve the network's learning and recognition capabilities for the power supply system fault features, and finally generate a target deep feature learning network, which will be more accurate and efficient in the subsequent processing of train fault feature chain data.

[0076] In a possible implementation, step S110 includes:

[0077] Step S111, obtaining a fault feature path graph corresponding to the train fault feature chain data, and extracting graph convolution results corresponding to a plurality of fault feature paths contained in the fault feature path graph, wherein the graph convolution result of each fault feature path includes the logical relationship features of each fault feature path in the fault feature path graph.

[0078] Step S112: extracting a high-dimensional detail vector of the train fault feature chain data based on the graph convolution results corresponding to the multiple fault feature paths respectively.

[0079] In this embodiment, when deep feature learning is performed on the train fault feature chain data reported by the target train to generate a low-dimensional embedding vector and a high-dimensional detail vector, the fault feature path map corresponding to the train fault feature chain data must first be obtained. Taking a high-speed EMU train as an example, the fault feature chain data of the train covers many subsystems, such as traction system, braking system, power supply system, etc. For the power supply system, the construction of the fault feature path map needs to consider the entire process from the grid input, through a series of equipment such as transformers, rectifiers, etc., to the distribution boxes and electrical equipment in each carriage. In this process, the changes in voltage, current, power and other data in each link and the logical relationship between them must be taken into consideration. For example, the fluctuation of the grid input voltage will affect the output voltage of the transformer, and the change of the transformer output voltage will further affect the working state of the rectifier. The connection relationship and data transmission relationship between these devices together constitute the fault feature path map.

[0080] After obtaining the fault characteristic path diagram, the server begins to extract the graph convolution results corresponding to the multiple fault characteristic paths contained in the fault characteristic path diagram. Still taking the power supply system as an example, one of the fault characteristic paths may be "grid input-transformer-rectifier". This embodiment can calculate the graph convolution result based on the voltage, current and other characteristics of each node on this fault characteristic path and the connection relationship between them. This graph convolution result contains the logical relationship characteristics of the path in the entire fault characteristic path diagram, such as the transmission ratio of voltage from the grid input to the transformer, the transformation relationship of current after passing through the transformer, etc. For another fault characteristic path "rectifier-distribution box", the server will also analyze the voltage and current relationship between the rectifier output and the distribution box input on this fault characteristic path, including possible logical relationships such as power loss, and obtain the corresponding graph convolution result.

[0081] Based on the graph convolution results corresponding to multiple fault feature paths, the server extracts the high-dimensional detail vector of the train fault feature chain data. In the power supply system failure scenario, the high-dimensional detail vector may contain the detailed electrical parameter relationship between each device, such as the precise relationship between the transformer ratio and the actual output voltage, the relationship between the rectifier efficiency and the input and output current, etc. At the same time, it will also include potential paths for fault propagation. For example, when a transformer fails, the fault may affect the rectifier through voltage fluctuations, and then affect the distribution box and electrical equipment. These detailed fault propagation path relationships will be reflected in the high-dimensional detail vector. For the train's braking system, the fault feature paths in its fault feature path graph are such as "brake pedal-brake booster-brake disc-brake pad". The high-dimensional detail vector obtained by analyzing the graph convolution results corresponding to these fault feature paths may contain detailed fault-related detail features such as the relationship between the brake pedal stroke and the brake booster pressure, the relationship between the brake disc temperature and the brake pad wear, etc. In this way, by analyzing the graph convolution results of each fault feature path in the fault feature path graph, the server can comprehensively extract the high-dimensional detail vector of the train fault feature chain data and provide detailed fault feature information for subsequent fault diagnosis.

[0082] In a possible implementation, step S111 includes:

[0083] The train fault feature chain data is path-decomposed to generate an initial fault feature path graph, and initial node features are configured in the initial node part of the initial fault feature path graph to generate the fault feature path graph.

[0084] Step S112 includes:

[0085] The graph convolution result corresponding to the initial node feature is used as a high-dimensional detail vector of the train fault feature chain data.

[0086] In a possible implementation, step S111 may further include:

[0087] Obtain a target deep feature learning network, wherein the target deep feature learning network includes a plurality of cascaded deep feature extraction units, wherein the plurality of cascaded deep feature extraction units include a first deep feature extraction unit and other deep feature extraction units except the first deep feature extraction unit.

[0088] The first deep feature extraction unit is used to extract graph convolution results corresponding to the multiple fault feature paths from the fault feature path graph, and a graph convolution result set extracted by the first deep feature extraction unit is generated.

[0089] Utilizing each other deep feature extraction unit, the graph convolution results corresponding to the multiple fault feature paths are extracted from the graph convolution result set extracted by the forward deep feature extraction unit of each other deep feature extraction unit, to generate the graph convolution result set extracted by each other deep feature extraction unit.

[0090] Step S112 may further include:

[0091] Based on the graph convolution result set extracted by the last deep feature extraction unit in the multiple cascaded deep feature extraction units, a high-dimensional detail vector of the train fault feature chain data is extracted.

[0092] Furthermore, the step S110 may further include:

[0093] Based on the graph convolution result of each fault feature path, an influence weight of each fault feature path in the fault feature path graph is calculated.

[0094] Based on a preset fault feature path library and the influence weights corresponding to the multiple fault feature paths contained in the fault feature path diagram, the weights corresponding to the multiple fault feature paths contained in the preset fault feature path library are calculated, and based on the weights, a high-dimensional detail vector of the train fault feature chain data is generated.

[0095] Among them, in the process of calculating the weight of any fault characteristic path included in the preset fault characteristic path library, if the fault characteristic path diagram contains any fault characteristic path, the weight of any fault characteristic path is calculated based on the influence weight of any fault characteristic path in the fault characteristic path diagram and the participation frequency of any fault characteristic path in the fault characteristic path diagram.

[0096] If the fault characteristic path graph does not include any of the fault characteristic paths, the preset weight is used as the weight of the any of the fault characteristic paths.

[0097] In this embodiment, first, in terms of obtaining the fault feature path map corresponding to the train fault feature chain data. Taking the high-speed EMU train as an example, the train fault feature chain data contains many complex subsystem information, such as traction system, braking system, power supply system and vehicle control system. The server performs path decomposition on the train fault feature chain data to generate an initial fault feature path map. For the power supply system, this embodiment can decompose the entire power supply process, starting from the power grid input, and gradually analyze the transformer, rectifier, distribution box and various power-consuming equipment, which preliminarily forms an initial fault feature path map. Then, the initial node feature is configured in the initial node part of the initial fault feature path map to generate a complete fault feature path map. In the example of the power supply system, the initial node may be the power grid input point, and the initial node feature may include the rated value of the input voltage, frequency and other information. For the braking system, the initial node may be the brake pedal, and the initial node feature includes the initial travel range of the pedal, the normal pressure value, etc. In this way, by configuring the corresponding initial node features at the initial nodes of each subsystem, the server constructs a comprehensive fault feature path map.

[0098] Next, the graph convolution results corresponding to the multiple fault feature paths contained in the fault feature path diagram are extracted. The server must first obtain the target deep feature learning network, which contains multiple cascaded deep feature extraction units, including the first deep feature extraction unit and other deep feature extraction units except the first deep feature extraction unit. For the fault feature path diagram of a high-speed EMU train, such as the fault feature path diagram of the power supply system, the first deep feature extraction unit is used to extract the graph convolution results corresponding to the multiple fault feature paths from the fault feature path diagram, and generate a set of graph convolution results extracted by the first deep feature extraction unit. For example, for the fault feature path of "grid input-transformer", the first deep feature extraction unit will obtain the graph convolution result corresponding to this fault feature path through complex calculations based on the voltage and current of the grid input and the rated parameters of the transformer. This result includes logical relationship features in this fault feature path, such as the voltage transfer ratio and the current transformation relationship. Similarly, for fault feature paths such as "transformer-rectifier" and "rectifier-distribution box", the first deep feature extraction unit will calculate the corresponding graph convolution results, which together constitute the graph convolution result set extracted by the first deep feature extraction unit.

[0099] Then, using each other deep feature extraction unit, the graph convolution results corresponding to multiple fault feature paths are extracted from the graph convolution result set extracted by the forward deep feature extraction unit of each other deep feature extraction unit, and the graph convolution result set extracted by each other deep feature extraction unit is generated. Taking the second deep feature extraction unit as an example, it will further analyze the interactive influence relationship between each fault feature path based on the graph convolution result set obtained by the first deep feature extraction unit, and generate a new graph convolution result set. For example, in the power supply system, the first deep feature extraction unit obtains the graph convolution results of each single fault feature path, and the second deep feature extraction unit will consider the influence of the graph convolution result of the fault feature path of "grid input-transformer" on the fault feature path of "transformer-rectifier", and the influence of the graph convolution result of the fault feature path of "rectifier-distribution box" on the fault feature path of "distribution box-electrical equipment", etc., and in this way, a new graph convolution result set after considering the interactive influence relationship is calculated. Subsequent deep feature extraction units are also calculated in a similar manner.

[0100] In terms of extracting high-dimensional detail vectors of train fault feature chain data based on graph convolution results corresponding to multiple fault feature paths. The server extracts high-dimensional detail vectors of train fault feature chain data based on the graph convolution result set extracted by the last deep feature extraction unit in multiple cascaded deep feature extraction units. Taking the power supply system as an example, the graph convolution result set extracted by the last deep feature extraction unit contains the final logical relationship features of each fault feature path after multi-layer analysis. This high-dimensional detail vector may contain very detailed electrical parameter relationships, such as the relationship between the precise transformation ratio of the transformer and the actual output voltage under different working conditions, the precise relationship between the efficiency of the rectifier and the input and output current under different loads, and potential paths for fault propagation. For example, when a fuse in the distribution box blows, a series of chain reactions may be caused, including the impact on other electrical equipment and the impact on the voltage and current distribution of the entire power supply system. Detailed information will be reflected in this high-dimensional detail vector. For the braking system, the high-dimensional detail vector may include the precise relationship between the degree of brake pad wear and the change in brake disc temperature, the relationship between the change in brake booster pressure and the change in brake pedal travel, and the fault propagation path between different sub-components when the braking system fails.

[0101] At the same time, in the step of performing deep feature learning on the train fault feature chain data reported by the target train and generating a low-dimensional embedding vector and a high-dimensional detail vector of the train fault feature chain data, it also includes calculating the influence weight of each fault feature path in the fault feature path diagram based on the graph convolution result of each fault feature path. In the fault feature path diagram of the power supply system, for the fault feature path of "transformer-rectifier", this embodiment can calculate its influence weight in the entire fault feature path diagram based on factors such as the electrical parameter relationship and the possibility of fault propagation in the graph convolution result of this fault feature path. If the instability of the transformer output voltage in this fault feature path has a greater impact on the working state of the rectifier, and this instability may further affect the subsequent distribution box and electrical equipment, then the influence weight of this fault feature path will be higher.

[0102] Based on the influence weights corresponding to the multiple fault characteristic paths contained in the preset fault characteristic path library and the fault characteristic path diagram, the weights corresponding to the multiple fault characteristic paths contained in the preset fault characteristic path library are calculated, and based on the weights, a high-dimensional detail vector of the train fault characteristic chain data is generated. The server pre-configures a preset fault characteristic path library, which contains various possible fault characteristic paths. These fault characteristic paths are constructed based on historical train fault data and a comprehensive understanding of the train system. When calculating the weight of any fault characteristic path contained in the preset fault characteristic path library, if the fault characteristic path diagram contains this any fault characteristic path, such as the fault characteristic path of "rectifier-distribution box" is included in the fault characteristic path diagram of the power supply system, then based on the influence weight of this fault characteristic path in the fault characteristic path diagram and the participation frequency of this fault characteristic path in the diagram (for example, the number of times this fault characteristic path appears within a certain period of time), the weight of this fault characteristic path is calculated. Assuming that this fault characteristic path appears frequently in multiple fault monitorings and has a greater impact on the power supply system each time it appears, its weight will be higher. If the fault feature path diagram does not contain any fault feature path, for example, there is a path about the special backup power supply line fault in the preset fault feature path library, but the target train does not have this special backup power supply line, then the preset weight is used as the weight of this fault feature path. In this way, considering all the weight information comprehensively, the server generates a high-dimensional detail vector of the train fault feature chain data. This high-dimensional detail vector comprehensively reflects the various relationships and characteristics in the train fault feature chain data, and provides a detailed basis for subsequent fault diagnosis and other operations.

[0103] In a possible implementation, step S130 includes:

[0104] Step S131, calculating the correlation between the low-dimensional embedding vector of each priori fault feature chain data and the low-dimensional embedding vector of the train fault feature chain data, and generating the low-dimensional correlation corresponding to each priori fault feature chain data.

[0105] Step S132, calculating the correlation between the high-dimensional detail vector of each priori fault feature chain data and the high-dimensional detail vector of the train fault feature chain data, and generating the high-dimensional correlation corresponding to each priori fault feature chain data.

[0106] Step S133, perform weighted calculation on the low-dimensional correlation and high-dimensional correlation corresponding to each priori fault feature chain data, generate a weighted correlation corresponding to each priori fault feature chain data, and use the weighted correlation corresponding to each priori fault feature chain data as the correlation between each priori fault feature chain data and the train fault feature chain data.

[0107] In this embodiment, continuing to take the high-speed EMU train as an example, the server first calculates the correlation between the low-dimensional embedding vector of each prior fault feature chain data and the low-dimensional embedding vector of the train fault feature chain data, thereby generating the low-dimensional correlation corresponding to each prior fault feature chain data. The server stores a large amount of prior fault feature chain data, which comes from numerous previous train failure cases. Assume that the current train fault feature chain data reflects that a certain fault has occurred in the train traction system, such as abnormal current fluctuations in the traction motor, and some control signals in the vehicle control system are also abnormal. For a priori fault feature chain data, if it is also about the train traction system fault, its low-dimensional embedding vector contains low-dimensional representations related to key features of the traction system, such as motor power, speed control, etc. In this embodiment, a distance-based metric method, such as Euclidean distance, can be used to calculate the correlation between the low-dimensional embedding vector of the priori fault feature chain data and the low-dimensional embedding vector of the current train fault feature chain data. In this example, the closer the values ​​of the low-dimensional embedding vector of the prior fault feature chain data in the dimensions related to motor power and speed control are to the values ​​of the low-dimensional embedding vector of the current train fault feature chain data in the corresponding dimensions, the smaller the Euclidean distance between them. For example, if the motor power in the prior fault feature chain data is 1000 kilowatts under specific working conditions, and the motor power in the current train fault feature chain data is 1050 kilowatts under similar working conditions, and there are similar close values ​​in the speed control dimension, then the distance value calculated by the Euclidean distance formula will be smaller, which indicates that they have a higher similarity in the low-dimensional feature space, and the corresponding low-dimensional correlation is higher. If the prior fault feature chain data is about the brake system fault, since the low-dimensional embedding vector of the current train traction system fault is quite different in the key dimensions, the distance value calculated by the Euclidean distance will be larger, and the low-dimensional correlation will be lower.

[0108] Next, the server calculates the correlation between the high-dimensional detail vector of each prior fault feature chain data and the high-dimensional detail vector of the train fault feature chain data to generate the high-dimensional correlation corresponding to each prior fault feature chain data. For high-dimensional detail vectors, since they contain more complex fault feature relationship information, the server may use a calculation method based on vector similarity, such as cosine similarity. Continuing with the train traction system fault as an example, the high-dimensional detail vector of the current train fault feature chain data contains the detailed relationship between the abnormal fluctuation of the traction motor current and the abnormal control signal in the vehicle control system, such as the time series relationship between the current fluctuation amplitude and the abnormal control signal. For a priori fault feature chain data, if it is also about the traction system fault, its high-dimensional detail vector contains similar detailed relationships such as between the motor fault and the control system feedback. The server determines the high-dimensional correlation between the two high-dimensional detail vectors by calculating the cosine similarity between them. If the two high-dimensional detail vectors are very similar in the direction of the relationship between the motor fault and the control system, for example, the logical relationship of the motor current fluctuation leading to the control system feedback adjustment is consistent in the two vectors, then their cosine similarity will be close to 1, indicating that the high-dimensional correlation is very high. On the contrary, if a priori fault feature chain data is about a brake system fault, its high-dimensional detail vector contains the internal relationships of the brake system, such as brake pad wear and brake pressure regulation, which is completely different from the high-dimensional detail vector of the current train traction system fault in content and structure. Then the calculated cosine similarity will be close to 0, indicating that the high-dimensional correlation is very low.

[0109] After obtaining the low-dimensional correlation and high-dimensional correlation corresponding to each prior fault feature chain data, the server needs to perform weighted calculation on the two correlations to generate the weighted correlation corresponding to each prior fault feature chain data, and use this weighted correlation as the correlation between each prior fault feature chain data and the train fault feature chain data. Assume that the server sets the weight of the low-dimensional correlation to 0.4 and the weight of the high-dimensional correlation to 0.6 based on the analysis and experience of historical data. For a prior fault feature chain data, if its low-dimensional correlation is 0.8, it indicates that it has a high similarity with the current train fault feature chain data in the low-dimensional feature space; its high-dimensional correlation is 0.9, indicating that it also has a high similarity in the high-dimensional detail relationship. Then the weighted correlation = 0.4 * 0.8 + 0.6 * 0.9 = 0.86. This weighted correlation takes into account the low-dimensional and high-dimensional correlations, and can more comprehensively reflect the degree of correlation between the prior fault feature chain data and the current train fault feature chain data. For another prior fault feature chain data, if the low-dimensional correlation is 0.3 (indicating that the low-dimensional features are quite different from the current train fault feature chain data), and the high-dimensional correlation is 0.5 (not very similar in high-dimensional detail relationships), then the weighted correlation = 0.4 * 0.3 + 0.6 * 0.5 = 0.42, which means that the correlation between this prior fault feature chain data and the current train fault feature chain data is low. Through such a weighted calculation method, the server can more accurately measure the correlation between each prior fault feature chain data and the train fault feature chain data, providing an important basis for subsequent fault diagnosis and other operations. This correlation calculation plays a key role in the entire train fault diagnosis system. It can help the server filter out the data most relevant to the current train fault from a large number of prior fault feature chain data, thereby improving the accuracy and efficiency of fault diagnosis.

[0110] Figure 2 The hardware structure of the deep learning-based train fault diagnosis system 100 for implementing the deep learning-based train fault diagnosis method according to an embodiment of the present invention is shown. Figure 2 As shown, the deep learning-based train fault diagnosis system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0111] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data acquired from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the deep learning-based train fault diagnosis system 100 uses to execute or use to complete the exemplary method described in the present invention.

[0112] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the train fault diagnosis method based on deep learning in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0113] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned deep learning-based train fault diagnosis system 100. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0114] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer executable instructions are preset. When a processor executes the computer executable instructions, the above-mentioned train fault diagnosis method based on deep learning is implemented.

[0115] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, various features are sometimes combined into one embodiment, drawing or description thereof.

Claims

1. A train fault diagnosis method based on deep learning, characterized in that: The method comprises: Performing deep feature learning on the train fault feature chain data reported by the target train to generate a low-dimensional embedding vector and a high-dimensional detail vector of the train fault feature chain data; Retrieving a plurality of priori fault feature chain data whose corresponding low-dimensional embedding vector has an associated pattern relationship with the low-dimensional embedding vector of the train fault feature chain data, generating a first priori fault feature chain data sequence, and retrieving a plurality of priori fault feature chain data whose corresponding high-dimensional detail vector has an associated pattern relationship with the high-dimensional detail vector of the train fault feature chain data, generating a second priori fault feature chain data sequence; Calculating the correlation between each priori fault feature chain data in the first priori fault feature chain data sequence and the second priori fault feature chain data sequence and the train fault feature chain data; Based on the correlation degree, extracting a plurality of associated pattern fault feature chain data having an associated pattern relationship with the train fault feature chain data from the first a priori fault feature chain data sequence and the second a priori fault feature chain data sequence; Based on the train fault characteristic chain data and the plurality of associated mode fault characteristic chain data, performing fault diagnosis on the target train; The calculating the correlation between each priori fault feature chain data in the first priori fault feature chain data sequence and the second priori fault feature chain data sequence and the train fault feature chain data comprises: Calculating the correlation between the low-dimensional embedding vector of each priori fault feature chain data and the low-dimensional embedding vector of the train fault feature chain data, and generating a low-dimensional correlation corresponding to each priori fault feature chain data; Calculating the correlation between the high-dimensional detail vector of each priori fault feature chain data and the high-dimensional detail vector of the train fault feature chain data, and generating the high-dimensional correlation corresponding to each priori fault feature chain data; A weighted calculation is performed on the low-dimensional correlation and the high-dimensional correlation corresponding to each priori fault feature chain data to generate a weighted correlation corresponding to each priori fault feature chain data, and the weighted correlation corresponding to each priori fault feature chain data is used as the correlation between each priori fault feature chain data and the train fault feature chain data.

2. The train fault diagnosis method based on deep learning according to claim 1 is characterized in that: The low-dimensional embedding vector and high-dimensional detail vector of the train fault feature chain data are generated by performing deep feature learning on the train fault feature chain data using a target deep feature learning network; the method further includes: Obtaining an initialized fault analysis network, and generating a deep feature learning network based on the initialized fault analysis network; Based on the first sample fault feature chain data corresponding to the deep feature learning network, the parameters of the deep feature learning network are optimized to generate the target deep feature learning network.

3. The train fault diagnosis method based on deep learning according to claim 2 is characterized in that: The method further comprises: Performing window masking on local node features in the first sample fault feature chain data to generate masked fault feature chain data, and obtaining target node features corresponding to the masking window of the masked fault feature chain data from the first sample fault feature chain data; Using a node feature restoration network to restore the node features corresponding to the shielding window of the shielding fault feature chain data, and generating a confidence that the node features corresponding to the shielding window obtained by the node feature restoration network are the target node features; The neuron parameter information of the node feature restoration network is optimized based on the confidence level to generate a target node feature restoration network, and the target node feature restoration network is used as the initialization fault analysis network.

4. The train fault diagnosis method based on deep learning according to claim 1 is characterized in that: The low-dimensional embedding vector and high-dimensional detail vector of the train fault feature chain data are generated by performing deep feature learning on the train fault feature chain data using a target deep feature learning network; the method further includes: Acquire second sample fault feature chain data and sample association mode fault feature chain data corresponding to the second sample fault feature chain data; Using a deep feature learning network, respectively perform deep feature learning on the second sample fault feature chain data and the sample association mode fault feature chain data to generate a low-dimensional embedding vector and a high-dimensional detail vector of the second sample fault feature chain data, as well as a low-dimensional embedding vector and a high-dimensional detail vector of the sample association mode fault feature chain data; Calculating the correlation between the low-dimensional embedding vector of the sample association mode fault feature chain data and the low-dimensional embedding vector of the second sample fault feature chain data to generate the low-dimensional correlation corresponding to the sample association mode fault feature chain data, and calculating the correlation between the high-dimensional detail vector of the sample association mode fault feature chain data and the high-dimensional detail vector of the second sample fault feature chain data to generate the high-dimensional correlation corresponding to the sample association mode fault feature chain data; Calculating the network learning error based on the low-dimensional correlation corresponding to the sample association mode fault feature chain data to generate a low-dimensional network learning error parameter, and calculating the network learning error based on the high-dimensional correlation corresponding to the sample association mode fault feature chain data to generate a high-dimensional network learning error parameter; Performing weighted calculation on the low-dimensional correlation degree and the high-dimensional correlation degree corresponding to the sample association mode fault feature chain data to generate the weighted correlation degree corresponding to the sample association mode fault feature chain data, and calculating the network learning error based on the weighted correlation degree corresponding to the sample association mode fault feature chain data to generate the weighted network learning error parameter; Based on the low-dimensional network learning error parameter, the high-dimensional network learning error parameter and the weighted network learning error parameter, a target network learning error parameter is calculated, and based on the target network learning error parameter, the deep feature learning network is optimized to generate the target deep feature learning network.

5. The train fault diagnosis method based on deep learning according to claim 4 is characterized in that: The sample association pattern fault feature chain data includes positive sample fault feature chain data having an association pattern relationship with the second sample fault feature chain data, and negative sample fault feature chain data that does not match the second sample fault feature chain data; The calculating of the network learning error based on the low-dimensional correlation degree corresponding to the sample correlation mode fault feature chain data and generating the low-dimensional network learning error parameter comprises: The network learning error is calculated based on the low-dimensional correlation corresponding to the positive sample fault feature chain data and the low-dimensional correlation corresponding to the negative sample fault feature chain data, and the low-dimensional network learning error parameter is generated, wherein the low-dimensional network learning error parameter is inversely correlated with the low-dimensional correlation corresponding to the positive sample fault feature chain data, and is positively correlated with the low-dimensional correlation corresponding to the negative sample fault feature chain data.

6. The train fault diagnosis method based on deep learning according to claim 1, characterized in that: The deep feature learning is performed on the train fault feature chain data reported by the target train to generate a low-dimensional embedding vector and a high-dimensional detail vector of the train fault feature chain data, including: Obtain a fault feature path graph corresponding to the train fault feature chain data, and extract graph convolution results corresponding to a plurality of fault feature paths contained in the fault feature path graph; wherein the graph convolution result of each fault feature path includes the logical relationship features of each fault feature path in the fault feature path graph; Based on the graph convolution results corresponding to the multiple fault feature paths respectively, a high-dimensional detail vector of the train fault feature chain data is extracted.

7. The train fault diagnosis method based on deep learning according to claim 6 is characterized in that: The obtaining of the fault characteristic path diagram corresponding to the train fault characteristic chain data includes: Performing path decomposition on the train fault feature chain data to generate an initial fault feature path graph, and configuring initial node features in an initial node part of the initial fault feature path graph to generate the fault feature path graph; The step of extracting a high-dimensional detail vector of the train fault feature chain data based on the graph convolution results respectively corresponding to the multiple fault feature paths includes: The graph convolution result corresponding to the initial node feature is used as a high-dimensional detail vector of the train fault feature chain data.

8. The train fault diagnosis method based on deep learning according to claim 6, characterized in that: The extracting graph convolution results corresponding to the multiple fault feature paths included in the fault feature path graph includes: Acquire a target deep feature learning network; wherein the target deep feature learning network comprises a plurality of cascaded deep feature extraction units, wherein the plurality of cascaded deep feature extraction units comprises a first deep feature extraction unit and other deep feature extraction units except the first deep feature extraction unit; Utilizing the first deep feature extraction unit, extracting graph convolution results corresponding to the plurality of fault feature paths from the fault feature path graph, and generating a graph convolution result set extracted by the first deep feature extraction unit; Utilizing each other deep feature extraction unit, extracting the graph convolution results corresponding to the multiple fault feature paths respectively from the graph convolution result set extracted by the forward deep feature extraction unit of each other deep feature extraction unit, and generating the graph convolution result set extracted by each other deep feature extraction unit; The step of extracting a high-dimensional detail vector of the train fault feature chain data based on the graph convolution results respectively corresponding to the multiple fault feature paths includes: Extracting a high-dimensional detail vector of the train fault feature chain data based on a graph convolution result set extracted by a last deep feature extraction unit in the plurality of cascaded deep feature extraction units; And, the step of performing deep feature learning on the train fault feature chain data reported by the target train to generate a low-dimensional embedding vector and a high-dimensional detail vector of the train fault feature chain data also includes: Based on the graph convolution result of each fault feature path, calculating the influence weight of each fault feature path in the fault feature path graph; Based on a preset fault feature path library and the influence weights respectively corresponding to the multiple fault feature paths contained in the fault feature path diagram, the weights respectively corresponding to the multiple fault feature paths contained in the preset fault feature path library are calculated, and based on the weights, a high-dimensional detail vector of the train fault feature chain data is generated; Wherein, in the process of calculating the weight of any one of the fault characteristic paths included in the preset fault characteristic path library, if the fault characteristic path graph includes the any one of the fault characteristic paths, the weight of the any one of the fault characteristic paths is calculated based on the influence weight of the any one of the fault characteristic paths in the fault characteristic path graph and the participation frequency of the any one of the fault characteristic paths in the fault characteristic path graph; If the fault characteristic path graph does not include any of the fault characteristic paths, the preset weight is used as the weight of the any of the fault characteristic paths.

9. A train fault diagnosis system based on deep learning, characterized in that: The deep learning-based train fault diagnosis system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the deep learning-based train fault diagnosis method described in any one of claims 1 to 8.

Images