Fault detection method for train transmission system

Through the graph attention network and node feature extractor, a fault diagnosis model of the train transmission system is constructed, and the interaction between components is considered, which improves the accuracy of fault diagnosis and anti-interference ability, and solves the problem of untimely fault detection in the existing methods. question.

CN118936869APending Publication Date: 2024-11-12QIQIHAR UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410941458.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-15
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

Existing train transmission system fault detection methods ignore the interaction between components, resulting in untimely failure detection in high noise states and low fault prediction accuracy.

Method used

A graph attention network is used to combine node feature extractor and classifier to build a train transmission system fault diagnosis model, and through the topological relationship diagram, new node features are generated and fault classification is performed.

Benefits of technology

It improves the accuracy of fault diagnosis of train transmission system, especially the accuracy of motor fault diagnosis, enhances anti-interference ability, and solves the problem that existing methods cannot effectively diagnose.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118936869A_ABST
    Figure CN118936869A_ABST
Patent Text Reader

Abstract

The invention discloses a train transmission system fault detection method, and relates to a mechanical part test method. The invention relates to a train transmission system fault detection method, and aims to solve the problems that the existing train transmission system fault detection method neglects the relevance of the adjacent component fault to the self fault, so that the transmission system fault is not found in time and the fault prediction accuracy is low in the high noise state of the train in the running process. The method comprises the following steps: simultaneously acquiring an operation signal of each component in the train transmission system; the components are divided into a target component and at least one other component in direct physical connection with the target component; inputting the operation signal into a transmission system fault diagnosis model, and judging whether a component in the transmission system has a fault or not; and when the target component is diagnosed as a fault, outputting a corresponding fault type.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a testing method for mechanical components. Background Art

[0002] During the operation and service phase, which has the longest window period in the entire life cycle of high-speed trains, any minor fault in the train transmission system, including the traction motor, transmission gearbox, and axlebox, may lead to serious consequences if it cannot be diagnosed and handled effectively in time. Therefore, real-time fault diagnosis of the train transmission system is crucial to ensure the safety and reliability of train operation.

[0003] The existing train transmission system fault detection method is to establish a separate fault diagnosis network for each component in the transmission system, but the separate network does not consider the interaction between adjacent components. Therefore, the correlation between the adjacent component failure of the component and the component's own failure is ignored, resulting in the transmission system fault not being discovered in time and the low fault prediction accuracy under the high noise state during the train operation. Summary of the invention

[0004] The purpose of the present invention is to overcome the problem that the existing train transmission system fault detection method ignores the correlation between the fault of adjacent components of the component and its own fault, resulting in the transmission system fault is not discovered in time and the fault prediction accuracy is low under the high noise state during the operation of the train, and provides a train transmission system fault detection method.

[0005] The train transmission system fault detection method of the present invention comprises the following steps:

[0006] Simultaneously acquiring operating signals of each component in the train drive system; the components are divided into a target component and at least one other component that has a direct physical connection with the target component;

[0007] The operation signal is input into the transmission system fault diagnosis model to determine whether a component in the transmission system is faulty; and when the target component is diagnosed as faulty, the corresponding fault type is output.

[0008] Furthermore, the train transmission system fault diagnosis model includes a sequentially connected node feature extractor, a graph attention network, and a classifier.

[0009] Furthermore, the process of training the train transmission system fault diagnosis model is as follows:

[0010] Step 1: Collect data sets;

[0011] The data set is the operating signal of each component collected simultaneously in advance;

[0012] The operation signal includes vibration signal data during normal operation and vibration signal data corresponding to different types of faults;

[0013] Step 2: Generate a topological relationship diagram of the train transmission system;

[0014] All components are nodes in the topology diagram;

[0015] The physical connections between components are all used as edges in the topology graph;

[0016] Step 3: Use the node feature extractor to extract the node features and classification of the nodes corresponding to each component from the data set;

[0017] Step 4: Take the topological relationship graph, node features and classification of the node as the input of the graph attention network, and generate new node features for each node under the premise of considering the node features of the node and the node features of the node's neighbor nodes;

[0018] Step 5: The new node features of the nodes corresponding to the target components are classified into fault categories through a classifier, and after verification and testing, the training of the train transmission system fault diagnosis model is completed.

[0019] Furthermore, the step of acquiring vibration signal data also includes: a process of converting the vibration signal data from a time domain signal to a frequency domain signal.

[0020] Furthermore, the step of acquiring the status data also includes: a process of adding a noise signal to the vibration signal data.

[0021] Furthermore, the signal-to-noise ratio of the noise signal is 30 dB to 6 dB.

[0022] Further, the assembly includes a motor bearing, a gearbox drive gear, a gearbox driven gear, a gearbox drive gear bearing and an axlebox bearing.

[0023] Furthermore, the specific process of step 4 is as follows:

[0024] Step 41: Calculate the original attention coefficient of node j to node i:

[0025]

[0026] Among them, node j is the neighbor node of node i; and are the node features of node i and node j respectively; W is the weight matrix; a(·) is the self-attention-shared attention mechanism function;

[0027] Step 42: Normalize the original attention coefficient to obtain the normalized attention coefficient:

[0028]

[0029] Among them, N i is a set of node i and its neighbor nodes;

[0030] Step 4.3: The original attention coefficient in the above formula is converted into a weight vector Parameterized and expanded after applying the LeakyReLU nonlinear function:

[0031]

[0032] in, for The transposed matrix of ; ∥ is the concatenation operation;

[0033] Step 4. Calculate the new node feature of node i using the above formula:

[0034]

[0035] Here, σ represents the sigma function.

[0036] Furthermore, the node feature extractor is a convolutional neural network.

[0037] Furthermore, the classifier is a Softmax classifier or a prototype-based classifier.

[0038] The beneficial effects of the present invention are:

[0039] The train transmission system fault detection method of the present invention constructs a system-level diagnostic network based on a component spatial relationship graph, effectively models the influence between adjacent components, and integrates the fault diagnosis of each component of the transmission system into a unified framework. It is used to detect the faults of each component in the train transmission system, and constructs a graph attention neural network to complete the integrated fault diagnosis of the train transmission system. The component spatial relationship graph enables the diagnostic network to effectively learn and capture the propagation and interaction of faults between different components, thereby improving the fault diagnosis accuracy of the train transmission system, solving the problem that the existing fault detection method does not consider the interaction between adjacent components and cannot effectively and stably diagnose the train transmission system under high noise conditions, especially improving the accuracy of fault diagnosis of the important component of the train motor. Compared with the individual construction of the network for monitoring each component in the transmission system, it has higher accuracy and stronger anti-interference ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A topological relationship diagram of a train transmission system in a train transmission system fault detection method of the present invention;

[0041] Among them, a is the train motor, b is the gearbox, c and d are the axle boxes respectively;

[0042] Figure 2 A schematic diagram of a flow chart of a train transmission system fault diagnosis model for training a train transmission system fault detection method;

[0043] Figure 3 Confusion matrix diagram obtained by testing the existing train transmission system fault detection method;

[0044] Figure 4 The confusion matrix diagram is obtained by testing the train transmission system fault detection method of the present invention. DETAILED DESCRIPTION

[0045] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0047] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention. Specific implementation method 1

[0049] The train transmission system fault detection method of this embodiment comprises the following steps:

[0050] The operating signals of each component in the train transmission system are obtained respectively; the components are divided into a target component and at least one other component that is directly physically connected to the target component;

[0051] The operation signal is input into the transmission system fault diagnosis model to determine whether a component in the transmission system is faulty; and when the target component is diagnosed as faulty, the corresponding fault type is output. Specific implementation method 2

[0053] This implementation is a further explanation of implementation one. In this embodiment, the train transmission system fault diagnosis model includes a sequentially connected node feature extractor, a graph attention network, and a classifier. Specific implementation method three

[0055] This embodiment is a further explanation of the second embodiment. In this embodiment, the process of training the train transmission system fault diagnosis model is as follows:

[0056] Step 1: Collect data sets;

[0057] The data set is the operating signal of each component collected simultaneously in advance;

[0058] The operation signal includes vibration signal data during normal operation and vibration signal data corresponding to different types of faults;

[0059] Step 2: Generate a topological relationship diagram of the train transmission system;

[0060] All components are nodes in the topology diagram;

[0061] The physical connections between components are all used as edges in the topology graph;

[0062] Step 3: Use the node feature extractor to extract the node features and classification of the nodes corresponding to each component from the data set;

[0063] Step 4: Take the topological relationship graph, node features and classification of the node as the input of the graph attention network, and generate new node features for each node under the premise of considering the node features of the node and the node features of the node's neighbor nodes;

[0064] Step 5: The new node features of the nodes corresponding to the target components are classified into fault categories through a classifier, and after verification and testing, the training of the train transmission system fault diagnosis model is completed.

[0065] Specifically, the train transmission system fault diagnosis model is trained and obtained in the following manner.

[0066] S1. Collect dataset:

[0067] At the same time, various operation signals of the transmission system and other components in the train transmission system are obtained (expressed as vibration signals. And including vibration signal data during normal operation, and vibration signal data corresponding to different fault types during faults).

[0068] This embodiment uses the public data set of the National Key Laboratory of Advanced Rail Transit Autonomous Operation of Beijing Jiaotong University as the training data set.

[0069] Among them, the motor bearing model is SKF6205-2RSH; the gear of the reduction gearbox is a helical gear, the number of teeth of the driving gear is 16, and the number of teeth of the driven gear is 107; the driving gear support bearing model is Harbin Bearing Factory 32305; the axle box bearing model is Harbin Bearing Factory 352213. The experiment collected 9 channels, and the sampling frequency of each channel was 64kHz; a total of 5 types of fault data including normal state were collected. Considering the scarcity of fault data in real scenarios, 20 samples were used for experiments, of which 70% were used as training sets, 10% as validation sets, and 20% as test sets, and then the training sets and test sets were randomly scattered to increase the randomness of the samples.

[0070] S2, generating a topological relationship diagram of the train transmission system;

[0071] Construct a topological relationship diagram based on the physical structure of the actual train transmission system.

[0072] like Figure 1 As shown, according to the structure of the train transmission system, each component is regarded as a node, and the spatial adjacency relationship between components is regarded as an edge, thereby generating a topological relationship graph G=(V, E), where V represents a set of nodes and E represents a set of edges.

[0073] S3. Extract node features and classifications of nodes corresponding to each component from the training set of the data set;

[0074] A convolutional neural network is used as a node feature extractor to capture low-dimensional node attributes from the vibration signal data collected in S1, and the node features of each component corresponding to the node in the topological relationship diagram are classified and extracted.

[0075] Each node is feature extracted and classified using a convolutional neural network, and node label indexes (for corresponding component names and fault types) are added to the extracted labels and features. These node feature extractors are constructed by alternately stacking convolutional layers, pooling layers, and batch normalization layers. This process can be expressed as:

[0076] h i =E i (x i ), i=1,2,3,...N

[0077] In the formula, i represents the ordinal number of the node, N represents the number of nodes, and h i is the node feature of the i-th node, E i (·) is the nonlinear mapping of the node feature extractor, x i is the vibration signal data of the i-th node input.

[0078] S4. Taking the topological relationship graph, node features and classification of the nodes as the input of the graph attention network, a new node feature is generated for each node under the premise of considering the node features of the node and the node features of the node's neighboring nodes;

[0079] A graph attention network is introduced to realize information propagation between adjacent nodes in the topological graph, thereby modeling and decoupling the interactions between components in the train transmission system.

[0080] The input is a set of node features Where F is the number of node features in each node. The graph attention network produces a new set of node features (potentially different cardinality F 0): As output, F' is the new number of node features for each node.

[0081] The multi-head attention mechanism of the graph attention network is used to splice the node features under the same label. For example, the node feature of the target node label 1 is only spliced ​​with the node feature of the adjacent node label 1.

[0082] S5. The new node features of the nodes corresponding to each component in the transmission system are classified into fault categories through a classifier, and after verification and testing, the training of the train transmission system fault diagnosis model is completed.

[0083] The new vector features of the transmission system nodes are classified into fault categories through the Softmax classifier and output to the corresponding labels to achieve fault diagnosis.

[0084] The test set divided in step S2 is input into the trained and fitted train motor fault diagnosis model to obtain the corresponding fault classification result.

[0085] After testing, the train transmission system fault diagnosis model of this implementation has good robustness and accuracy.

[0086] This implementation transforms the original time domain vibration signal extracted from each component in the train transmission system into a frequency domain signal, and then uses a convolutional neural network to extract features. At the same time, based on the physical structure of the train transmission system, a topological relationship diagram of the physical structure is drawn, and the features of weighted neighbor nodes are aggregated by using a graph attention mechanism. Finally, a Softmax classifier is used to classify the node features of each component in the transmission system. This can effectively enhance the robustness of the model for fault classification and improve the fault diagnosis of the target component, especially when the target component is a motor, the accuracy of fault diagnosis.

[0087] According to the instruction manual Figure 3 and 4 As shown, the train transmission system fault diagnosis model of this embodiment is used to diagnose motor faults, and the accuracy rate reaches 94.96%, while the existing method for motor fault diagnosis is only 88.88%. Specific implementation method four

[0089] This implementation is a further description of implementation one, two or three. In this implementation, the step of acquiring vibration signal data also includes: a process of converting the vibration signal data from a time domain signal to a frequency domain signal.

[0090] The other technical features of this embodiment are exactly the same as those of embodiment one or two. Specific implementation method five

[0092] This embodiment is a further explanation of the fourth embodiment. In this embodiment, the step of acquiring the state data further includes: a process of adding a noise signal to the vibration signal data.

[0093] The other technical features of this embodiment are exactly the same as those of the third embodiment.

[0094] Specifically, considering the existence of noise under actual working conditions, a noise signal with a signal-to-noise ratio of 10 dB is added to all vibration signal data. Specific implementation method 6

[0096] This embodiment is a further description of the fifth embodiment. In this embodiment, the signal-to-noise ratio of the noise signal is 30 dB to 6 dB.

[0097] The other technical features of this embodiment are exactly the same as those of embodiment 4. Specific implementation method seven

[0099] This embodiment is a further description of the first, second, third, fifth or sixth embodiment. In this embodiment, the components include a transmission system bearing, a gearbox driving gear, a gearbox driven gear, a gearbox driving gear bearing and an axle box bearing.

[0100] The other technical features of this embodiment are exactly the same as those of embodiments one, two, four or five.

[0101] Specifically, the target component can be one of a transmission system bearing, a gearbox driving gear, a gearbox driven gear, a gearbox driving gear bearing and an axle box bearing. When one of the components, such as a motor, needs to be tested, the motor is used as the target component, and there is a direct transmission connection between the motor and the gearbox. Specific implementation method eight

[0103] This implementation is a further description of implementation three, five or six. In this implementation, the specific process of step four is as follows:

[0104] Step 41: Calculate the original attention coefficient of node j to node i:

[0105]

[0106] Among them, node j is the neighbor node of node i; and are the node features of node i and node j respectively; W is the weight matrix; a(·) is the self-attention-shared attention mechanism function;

[0107] Step 42: Normalize the original attention coefficient to obtain the normalized attention coefficient:

[0108]

[0109] Among them, N i is a set of node i and its neighbor nodes;

[0110] Step 4.3: The original attention coefficient in the above formula is converted into a weight vector Parameterized and expanded after applying the LeakyReLU nonlinear function:

[0111]

[0112] in, for The transposed matrix of ; ∥ is the concatenation operation;

[0113] Step 4. Calculate the new node feature of node i using the above formula:

[0114]

[0115] Here, σ represents the sigma function.

[0116] The other technical features of this embodiment are exactly the same as those of embodiments two, four or five.

[0117] Specifically, in order to obtain sufficient expressive power and transform the input features into higher-level features, at least one learnable linear transformation is required. To this end, as an initial step, a shared linear transformation is applied to each node, which is represented by the weight matrix Parameterized.

[0118]

[0119] in, and is the node feature (vector) of node i and node j, and a is the attention function used to calculate the association between node i and node j. By performing attention, the graph structure of the topological relationship graph is injected into the mechanism, and e is only calculated for node j. ij ,in N i is the neighbor node (first-order neighbor) of node i. In order to make the coefficients easy to compare between different nodes, all the choices of node j are normalized using the Softmax function:

[0120]

[0121] The attention mechanism a is a single-layer feed-forward neural network composed of the weight vector Parameterized and LeakyReLU nonlinear function is applied. After full expansion, the coefficients calculated by the attention mechanism can be expressed as:

[0122]

[0123] In the formula, ∥ is the splicing operation, is the neighborhood of node i (including node i itself and the set of neighboring nodes of node i), h i is the eigenvector of the central node i, h j is the feature vector of one of the neighbor nodes j, h k is the characteristic vector of any neighbor node k, and exp is an exponential function.

[0124] Regarding the process of updating node features, first the feature conversion is performed. The feature vector of each node is converted to a new feature space through the weight matrix W, i.e. h′ i =Wh i Then, for each pair of neighbor nodes i and j, the concatenated linear transformation of their feature vectors is calculated, and then the LeakyReLU function is used to increase the nonlinearity of the network:

[0125] e ij =”LeakyReLu”(a T [h′ i ||h′ j ])

[0126] Then use the Softmax function to calculate all the e ij Normalize to get the final attention coefficient α ij Next, the new features of each node are obtained by concatenation, i.e. Where σ is a nonlinear activation function. These steps together define the information transfer and feature update mechanism in graph attention, enabling the network to effectively learn on graph structured data. Specific implementation method nine

[0128] This implementation is a further description of implementation modes two, three, five or six. In this implementation mode, the node feature extractor is a convolutional neural network. Specific implementation method 10

[0130] This implementation is a further description of the ninth implementation. In this implementation, the classifier is a Softmax classifier or a prototype-based classifier.

[0131] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. It should therefore be understood that many modifications may be made to the exemplary embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the features of the various dependent claims and herein may be combined in a manner different from that described in the original claims. It should also be understood that the features described in conjunction with the individual embodiments may be used in other embodiments.

Claims

1. A train transmission system fault detection method, characterized in that: Here are the steps: Acquire the operation signal of each component in the train transmission system respectively; the components are divided into a target component and at least one other component that is directly physically connected to the target component; The operation signal is input into a transmission system fault diagnosis model to determine whether a component in the transmission system is faulty; and when a target component is diagnosed as faulty, a corresponding fault type is output.

2. The train transmission system fault detection method according to claim 1, characterized in that: The train transmission system fault diagnosis model includes a sequentially connected node feature extractor, a graph attention network and a classifier.

3. The train transmission system fault detection method according to claim 2, characterized in that: The process of training the train transmission system fault diagnosis model is as follows: Step 1: Collect data sets; The data set is the operation signal of each component collected simultaneously in advance; The operation signal includes vibration signal data during normal operation and corresponding vibration signal data when different types of faults occur; Step 2: Generate a topological relationship diagram of the train transmission system; The components are all used as nodes of the topological relationship diagram; The physical connection relationships between the components are all used as edges of the topological relationship graph; Step 3: Using a node feature extractor to extract node features and classifications of nodes corresponding to each component from the data set; Step 4: Take the topological relationship graph, node features and classification of the nodes as the input of the graph attention network, and generate new node features for each node under the premise of considering the node features of a node and the node features of its neighboring nodes; Step 5: The new node features of the nodes corresponding to the target components are classified into fault categories through a classifier, and after verification and testing, the training of the train transmission system fault diagnosis model is completed.

4. The train transmission system fault detection method according to claim 1, 2 or 3, characterized in that: The step of acquiring vibration signal data also includes: a process of converting the vibration signal data from a time domain signal to a frequency domain signal.

5. The train transmission system fault detection method according to claim 4, characterized in that: The step of acquiring the state data also includes: a process of adding a noise signal to the vibration signal data.

6. The train transmission system fault detection method according to claim 5, characterized in that: The signal-to-noise ratio of the noise signal is 30 dB to 6 dB.

7. The train transmission system fault detection method according to claim 1, 2, 3, 5 or 6, characterized in that: The assembly includes the motor bearing, gearbox drive gear, gearbox driven gear, gearbox drive gear bearing and axlebox bearing.

8. The train transmission system fault detection method according to claim 3, 5 or 6, characterized in that: The specific process of step 4 is as follows: Step 41: Calculate the original attention coefficient of node j to node i: Among them, node j is the neighbor node of node i; and are the node features of node i and node j respectively; W is the weight matrix; a(·) is the self-attention-shared attention mechanism function; Step 42: Normalize the original attention coefficient to obtain a normalized attention coefficient: Among them, N i is a set of node i itself and its neighbor nodes; Step 4.3: The original attention coefficient in the above formula is converted into a weight vector Parameterized and expanded after applying the LeakyReLU nonlinear function: in, for The transposed matrix of ; ∥ is the concatenation operation; Step 4. Calculate the new node feature of node i using the above formula: Here, σ represents the sigma function.

9. The train transmission system fault detection method according to claim 2, 3, 5 or 6, characterized in that: The node feature extractor is a convolutional neural network.

10. The train transmission system fault detection method according to claim 9, characterized in that: The classifier is a Softmax classifier or a prototype-based classifier.