Traction motor overcurrent fault identification method and system

Abstract

The invention provides a traction motor overcurrent fault identification method and system, and relates to the technical field of overcurrent fault diagnosis, and the method comprises the steps: obtaining a historical signal, carrying out the offline preprocessing, and extracting a first feature identifier; segmenting the first feature identifier in the sliding window, generating a first feature index, carrying out first event sequence conversion, constructing a traction motor overcurrent fault positioning and recognition template, dividing the template into a training set and a test set, training a traction motor overcurrent fault recognition model by using the training set, and obtaining a traction motor overcurrent fault recognition result; the effectiveness of the traction motor overcurrent fault identification model is tested by using the test set; performing online fault detection on the real-time signal, and extracting a second feature identifier; and segmenting the second feature identifier in the sliding window, generating a second feature index, performing second event sequence conversion, obtaining an online template vector, inputting the online template vector into the trained traction motor overcurrent fault recognition model, and outputting a fault recognition result. The algorithm complexity can be effectively reduced, and the fault type can be quickly and accurately identified in real time.

Description

Technical Field

[0001] The present invention relates to the technical field of overcurrent fault diagnosis, and particularly to a method and system for identifying overcurrent faults of traction motors. Background Art

[0002] The high-speed rail traction system is a core component to ensure the normal operation of the train and is known as the heart of the train. The high-speed rail traction system mainly consists of three major parts: a traction transformer, a traction converter, and a traction motor. Its working principle is as follows: 25 kV alternating current is received through the pantograph, stepped down by the traction transformer, converted into direct current by the four-quadrant rectifier, filtered through the intermediate DC link, and finally converted into three-phase alternating current with adjustable frequency and amplitude by the inverter to drive the traction motor, thereby realizing the speed control of the train; the electric locomotive traction drive system is the core power system of the train operation. When working under conditions such as high speed, high power, and complex working conditions, the traction motor is prone to overcurrent faults. Once an overcurrent fault occurs, if it cannot be diagnosed in time and effective measures are not taken, it may cause the paralysis of the traction system, resulting in train delays and even major safety accidents. Therefore, real-time and accurate diagnosis of overcurrent faults of traction motors is of great significance for ensuring the safe operation of trains.

[0003] Currently, the following methods are mainly used for the diagnosis of overcurrent faults of traction motors: One is a simple alarm method based on thresholds, which issues an alarm when the collected motor current signal exceeds the set threshold. This method can only detect the fault phenomenon and cannot accurately locate and judge the cause of the fault; the other is to use artificial intelligence methods such as neural networks and decision trees for fault diagnosis and classification. Although such methods can identify faults to a certain extent, due to the high algorithm complexity and poor real-time performance, it is difficult to be widely applied in actual engineering. The causes of overcurrent faults of traction motors are complex and may be caused by various reasons such as speed sensor faults, inverter IGBT module faults, and motor body faults. After a fault occurs, the system often undergoes a change process of multiple working conditions. Traditional diagnosis methods based on single-moment characteristics or static patterns are difficult to accurately judge the cause of the fault. Summary of the Invention

[0004] To solve the problems of high algorithm complexity, poor real-time performance, and difficulty in quickly and accurately identifying fault types in the above-mentioned existing technologies, the present invention proposes a method and system for identifying overcurrent faults of traction motors, which can effectively reduce the algorithm complexity and quickly and accurately identify fault types in real time.

[0005] To achieve the above technical effects, the technical solution of the present invention is as follows: A method for identifying overcurrent faults of traction motors, comprising the following steps: S1. Obtain the historical signals of the traction drive system, perform offline preprocessing on the historical signals, and obtain the first feature identifier with fault location discrimination; S2. Perform segmentation processing on the first feature identifier within a sliding window to obtain the first feature index, generate a corresponding first event using the first feature index, perform sequence transformation on the first event, and obtain the traction motor overcurrent fault location recognition template; S3. Divide the traction motor overcurrent fault location recognition template into a training set and a test set, use the training set to train the constructed traction motor overcurrent fault recognition model, and use the test set to test the effectiveness of the traction motor overcurrent fault recognition model to obtain the trained traction motor overcurrent fault recognition model; S4. Obtain the real-time signals of the traction drive system, perform online fault detection on the real-time signals, and obtain the second feature identifier with fault location discrimination; S5. Perform segmentation processing on the second feature identifier within a sliding window to obtain the second feature index, generate a corresponding second event using the second feature index, perform sequence transformation on the second event, and obtain the online template vector; S6. Input the online template vector into the trained traction motor overcurrent fault recognition model and output the fault recognition result.

[0006] Preferably, the historical signals include the first phase A current signal , the first phase B current signal , the first intermediate voltage sensor signal and the first speed signal , and the first feature identifier includes the first mean , the first variance of the first intermediate voltage sensor historical signal , the minimum value of the first intermediate voltage sensor signal , the second variance of the first speed signal , the first current minimum value , the first current maximum value , the first current effective value , the first current normalized value and the first current normalized value ; ; ; The performing offline preprocessing on the historical signals to obtain the first feature identifier with fault location discrimination includes: According to the first intermediate voltage sensor historical signal and the first speed signal , the calculation expression for obtaining the first mean is:

[0007] Among them, k represents the running time, represents the k first mean value at the running time , represents an intermediate variable, = , represents the sliding window size, represents taking values of and , ; According to the first intermediate voltage sensor historical signal , the calculation expression for obtaining the first variance is:

[0008] Among them, represents the first variance k at the running time, represents the first mean value k at the running time when taking values of ; According to the first intermediate voltage sensor historical signal , the calculation expression for obtaining the minimum value of the first intermediate voltage sensor signal is:

[0009] Among them, represents the minimum value k of the first intermediate voltage sensor signal at the running time, represents taking the minimum value, the k first intermediate voltage sensor historical signal at the running time, represents the first intermediate voltage sensor historical signal at the running time; According to the first speed signal , the calculation expression for obtaining the second variance is:

[0010] Among them, represents the k second variance at the runtime, represents the first mean value k at the runtime with a value of ; ; According to the first A-phase current signal and the first B-phase current signal , the calculation expression for obtaining the first current minimum value is:

[0011] Among them, represents the first current minimum value k at the runtime, y represents the subscripts with values of a and b , represents the first A-phase current signal k at the runtime and the first B-phase current signal , represents the first A-phase current signal at the runtime and the first B-phase current signal , at the runtime, the first A-phase current signal and the first B-phase current signal ; According to the first A-phase current signal and the first B-phase current signal , the calculation expression for obtaining the current maximum value is:

[0012] Among them, represents taking the maximum value; According to the first A-phase current signal and the first B-phase current signal , the calculation expression for obtaining the first current effective value is:

[0013] Among them, Indicating the k first effective current value at the running moment ; According to the first phase-A current signal and the first phase-B current signal , the calculation expression for obtaining the first current normalization value is:

[0014] wherein, Indicating the k first current normalization value at the running moment .

[0015] Preferably, sliding window feature data is obtained within a sliding window, and the sliding window feature data includes the current frequency of the traction motor , the size of the sliding window , the step size , the number of sliding windows , the number distributed on the left side of the action point of the traction control unit , the number distributed on the right side of the action point of the traction control unit and the data length ; The calculation expression for the size of the sliding window is:

[0016] wherein, represents the sampling period; The calculation expression for the step size is:

[0017] The calculation expression for the number distributed on the left side of the action point of the traction control unit is:

[0018] The calculation expression for the number distributed on the right side of the action point of the traction control unit is:

[0019] The calculation expression for the data length is:

[0020] According to the sliding window feature data, the first feature identifier is segmented to obtain the following calculation expressions for the first feature indicators respectively:

[0021]

[0022]

[0023]

[0024]

[0025] Among them, and and and and represent the first characteristic index, and and and and represent the k first characteristic index at the running time.

[0026] Preferably, generating a corresponding first event by using the first characteristic index, and performing sequence conversion on the first event to obtain a traction motor overcurrent fault location and identification template, including: S21. Input the first characteristic index and and and and into the first hysteresis comparator respectively, and output the corresponding first event by the first hysteresis comparator. The calculation expression of

[0027] is as follows: represents the subscript taking values of respectively, represents the opening point of the first hysteresis comparator, represents the closing point of the first hysteresis comparator, , ; if is not less than the opening point of the first hysteresis comparator, then the first event outputs 1; if is less than the opening point of the first hysteresis comparator and greater than the closing point of the first hysteresis comparator, then the first event maintains the output state consistent with the previous moment's event; if is not greater than the closing point representing the first hysteresis comparator, then the first event The output is 0; S22. Form a binary time series vector in the order of the first event at the subscript X to form a binary time series vector , and convert the binary time series vector into a decimal number as the following mathematical expression:

[0028] where represents the sliding window position of the currently calculated decimal number , (1, n ), represents the length of the first vector, ; S23. Based on the decimal number , obtain the calculation expression of the traction motor overcurrent fault location identification vector as follows:

[0029] where represents the element of the vector with the number of sliding windows as the length of the first vector , represents taking the minimum value of the decimal number , represents taking the maximum value of the decimal number ; The calculation expression of

[0030] The calculation expression of

[0031] S24. Combine the traction motor overcurrent fault location identification vectors obtained under different fault types to form a set, and obtain the traction motor overcurrent fault location identification template.

[0032] Preferably, the traction motor overcurrent fault identification model includes a sequence input layer, a bidirectional long short-term memory network layer, a self-attention layer, a first fully connected layer, a first activation function layer, a dropout layer, a second fully connected layer, a second activation function layer, and a classification layer connected in sequence; Normalize the overcurrent fault location and identification template of the traction motor into a 16-dimensional template vector and input it into the sequence input layer. The sequence input layer outputs time series data. The bidirectional long short-term memory network layer receives the time series data and outputs the bidirectional dependence features of the time series. The self-attention layer receives the bidirectional dependence features and outputs the time step feature data weighted by attention. The first fully connected layer receives the time step feature data and outputs the first linear feature combination result. The first activation function layer receives the first linear feature combination result and outputs the first activation feature result. The dropout layer receives the first activation feature result and outputs the feature result after dropout. The second fully connected layer receives the feature result after dropout and outputs the second linear feature combination result. The second activation function layer receives the second linear feature combination result and outputs the fault type prediction probability. The classification layer receives the fault type prediction probability and outputs the fault type identification result.

[0033] Preferably, training the overcurrent fault identification model of the traction motor includes: S31. Set the training parameters of the overcurrent fault identification model of the traction motor. The training parameters include: the maximum number of epochs is 150, the batch size is 32, the initial learning rate is 0.005, the learning rate schedule is to decrease by 0.2 times every 50 epochs, the L2 regularization coefficient is 0.0001, and the gradient clipping threshold is 1; S32. Use the Adam optimizer to perform preliminary optimization on the training parameters to obtain the hyperparameters after preliminary optimization; S33. Use the genetic algorithm to perform global optimization on the hyperparameters after preliminary optimization until the hyperparameter combination corresponding to the individual with the highest fitness value is output, and complete the training of the overcurrent fault identification model of the traction motor.

[0034] Preferably, using the genetic algorithm to perform global optimization on the hyperparameters after preliminary optimization includes: S331. Set the hyperparameter configuration of each individual in the population size; S332. Use the hyperparameter configuration of each individual to build an overcurrent fault identification model of the traction motor, use the training set to train the overcurrent fault identification model of the traction motor, and calculate the classification accuracy α and Macro-F1 score β on the test set. Calculate the fitness value X based on the classification accuracy α and Macro-F1 score β as follows: X 0.7α + 0.3β S333. According to the fitness value X, use tournament selection or roulette wheel selection to select excellent individuals from the current population size as parents; S334. Perform crossover operation on the parental individuals. Use single-point crossover or uniform crossover to swap the hyperparameter values of two parental individuals to generate new crossover individuals. S335. With a mutation probability of 5% to 10%, randomly perturb and fine-tune the positions of the hyperparameter values of the new crossover individuals to obtain new mutated individuals. S336. Add the new crossover individuals and the new mutated individuals to the population size, form the next generation, and then perform update iteration. S337. When the number of update iterations reaches the iteration threshold or the change in the fitness value X is less than 0.001 within 5 consecutive generations, end the global optimization and output the hyperparameter combination corresponding to the individual with the highest fitness value.

[0035] Preferably, the real-time signal includes the second phase A current signal , the second phase B current signal , the second intermediate voltage sensor signal and the second speed signal , and the second feature identifier includes the minimum value of the second intermediate voltage sensor signal , the second mean value , the third variance of the second intermediate voltage sensor signal , the fourth variance of the second speed signal , the second current minimum value , the second current maximum value , the second current normalization value , the second current effective value ; Performing online fault detection on the real-time signal to obtain a second feature identifier with fault location discrimination includes: S41. Determine whether the collected second phase A current signal and the second phase B current signal are greater than the current protection threshold within a number of consecutive sampling periods. If so, enter the fault diagnosis link and execute S42; if not, continue to perform online fault detection on the collected second phase A current signal and the second phase B current signal . S42. According to the second intermediate voltage sensor signal and the second speed signal , the calculation expression for obtaining the second mean value is:

[0036] Wherein, kIndicates the running time, Indicates the k Second mean value at the running time , Indicates an intermediate variable, = , Indicates the sliding window size, Indicates that the value is and , ; According to the second intermediate voltage sensor signal , the calculation expression for obtaining the third variance is:

[0037] where Indicates the k Third variance at the running time , Indicates at the k Running time The value is Second mean value ; According to the second intermediate voltage sensor signal , the minimum value of the second intermediate voltage sensor signal is calculated as:

[0038] where Indicates the k Minimum value of the second intermediate voltage sensor signal at the running time , Indicates taking the minimum value, The k Second intermediate voltage sensor signal at the running time , Indicates The second intermediate voltage sensor signal at the running time , Indicates the Second intermediate voltage sensor signal at the running time ; According to the second speed signal , the calculation expression for obtaining the fourth variance is:

[0039] where Indicates thek Fourth variance at runtime , indicating the second mean value k at runtime with a value of ; ; Based on the second A-phase current signal and the second B-phase current signal , the calculation expression for obtaining the second minimum current is:

[0040] wherein represents the second minimum current k at runtime , y represents the subscripts with values of a and b ; indicating the second A-phase current signal k at runtime and the second B-phase current signal ; indicating the second A-phase current signal at runtime and the second B-phase current signal ; at runtime, the second A-phase current signal and the second B-phase current signal ; Based on the second A-phase current signal and the second B-phase current signal , the calculation expression for obtaining the second maximum current is:

[0041] wherein represents taking the maximum value; Based on the second A-phase current signal and the second B-phase current signal , the calculation expression for obtaining the second root-mean-square current is:

[0042] wherein represents the second root-mean-square current k at runtime ; Based on the second A-phase current signal and the second B-phase current signal , the second current normalization value is calculated The calculation expression is:

[0043] wherein, represents the second current normalization value at the k running time .

[0044] Preferably, the second feature identifier is segmented within the sliding window to obtain the following calculation expressions for the second feature indicators respectively:

[0045]

[0046]

[0047]

[0048]

[0049] wherein, , , , , represent the second feature indicators, , , , , represent the k second feature indicators at the running time; Using the second feature indicators to generate corresponding second events, and performing sequence transformation on the second events to obtain an online template vector, including: S51. Input the second feature indicators , , , , into the second hysteresis comparator respectively, and the second hysteresis comparator outputs the corresponding second events ; S52. Arrange the second events in the order of subscript X to form a binary timing event , and convert the binary timing event into a decimal number as the following mathematical expression:

[0050] Among them, represents the sliding window position of the current calculated decimal number of, (1, m ), m represents the second vector length, ; S53. Based on the decimal number , the calculation expression of the template vector is as follows:

[0051] Among them, represents the element of the vector with the number of sliding windows as the second vector length m , represents taking the minimum value of the decimal number , represents taking the maximum value of the decimal number ; The calculation expression of is:

[0052] The calculation expression of is: .

[0053] The present invention also proposes a traction motor overcurrent fault identification system, including: An offline preprocessing module, configured to obtain historical signals of the traction drive system, perform offline preprocessing on the historical signals, and obtain a first feature identifier with fault location discrimination; An offline segmentation and transformation module, configured to perform segmentation processing on the first feature identifier within a sliding window to obtain a first feature index, generate a corresponding first event using the first feature index, and perform sequence transformation on the first event to obtain a traction motor overcurrent fault location identification template; A traction motor overcurrent fault identification model construction module, configured to construct a traction motor overcurrent fault identification model, divide the traction motor overcurrent fault location identification template into a training set and a test set, train the traction motor overcurrent fault identification model using the training set, and test the effectiveness of the traction motor overcurrent fault identification model using the test set to obtain a trained traction motor overcurrent fault identification model; An online preprocessing module, configured to obtain real-time signals of the traction drive system, perform online fault detection on the real-time signals, and obtain a second feature identifier with fault location discrimination; An online segmentation and transformation module is used to perform segmentation processing on the second feature identifier within a sliding window to obtain a second feature index, generate a corresponding second event using the second feature index, and perform sequence transformation on the second event to obtain an online template vector; An output module is used to input the online template vector into a trained traction motor overcurrent fault recognition model and output a fault recognition result.

[0054] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: The present invention proposes a method and system for identifying traction motor overcurrent faults. First, in the offline stage, the historical signal is preprocessed offline to obtain a first feature identifier. The first feature identifier is segmented according to the sliding window feature data to obtain a first feature index. A corresponding first event is generated using the first feature index, and sequence transformation is performed on the first event. The purpose is to reduce the offline data dimension to one dimension, thereby effectively reducing the computational complexity. Then, a traction motor overcurrent fault recognition model is constructed and trained to obtain a trained traction motor overcurrent fault recognition model; then, online fault detection is performed on the real-time signal to obtain a second feature identifier with fault location discrimination. The second feature identifier is segmented within a sliding window to obtain a second feature index. A corresponding second event is generated using the second feature index, and sequence transformation is performed on the second event. The purpose is to reduce the online data dimension to one dimension, thereby further effectively reducing the computational complexity; finally, the online template vector is input into the trained traction motor overcurrent fault recognition model to achieve rapid and accurate discrimination of the fault type, significantly improving the efficiency and accuracy of real-time fault recognition. The present invention can improve the speed and real-time performance of data matching through the traction motor overcurrent fault recognition model on the premise of ensuring the accuracy of fault recognition, so as to better meet the requirements of real-time diagnosis of motor overcurrent faults. Description of the Drawings

[0055] Figure 1 It represents a flowchart showing a method for identifying traction motor overcurrent faults proposed in an embodiment of the present invention; Figure 2 It represents another flowchart showing a method for identifying traction motor overcurrent faults proposed in an embodiment of the present invention; Figure 3 It represents a power generation circuit topology structure model diagram of a traction drive system proposed in an embodiment of the present invention; Figure 4 It represents a structural block diagram of a system for identifying traction motor overcurrent faults proposed in an embodiment of the present invention. Detailed Embodiments

[0056] The drawings are only for illustrative purposes and should not be construed as limiting the present invention; In order to better illustrate the present embodiment, some parts of the drawings may be omitted, enlarged or reduced, and do not represent actual sizes. The description of the directions of parts such as "upper" and "lower" does not limit the present invention. It is understandable to those skilled in the art that some well-known contents may be omitted in the drawings; In order to facilitate understanding of this embodiment, first, the prior art information of this embodiment is introduced as follows: The existing patent literature discloses a traction converter fault diagnosis method, equipment, medium and product. In this application, the intermediate DC voltage data is obtained as a data set, and the data set is decomposed based on variational mode decomposition to obtain a multi-channel standard fault feature set. Variational mode decomposition has high frequency domain resolution and can separate different frequency characteristics of the signal in the presence of noise. According to its characteristics, it is determined to use the PE-Spearman rank correlation coefficient as the supporting channel weighting layer. Based on this, a supporting channel weighting layer-one-dimensional deep separable convolutional neural network model is established. The supporting channel weighting layer serves as a "bridge" between variational mode decomposition and the neural network model. Finally, the multi-channel standard fault feature set is input into the model to obtain the fault classification. This method is relatively complex to calculate, and it is difficult to output results quickly. It has high requirements on hardware performance, otherwise it cannot meet the fault diagnosis requirements that require rapid response. In addition, the parameter structure of the model established by this method is relatively fixed, and the adaptability of the model is poor.

[0057] The existing patent literature also discloses a method for real-time identification and diagnosis of inverter overcurrent based on a time-series operating condition event set. In this application, it consists of two parts: offline design and online diagnosis. In the offline design stage, the historical fault data is first preprocessed to obtain the window size, the relevant characteristic variables are calculated in the window to obtain the characteristic index, the event is generated in the hysteresis comparator, and a template library is established for the time-series operating condition event set of different fault types. The online stage contains fault detection and fault decision modules. The fault detection module extracts the inverter overcurrent related sensor signal in real time, compares it with the detection threshold, and generates a diagnostic enable flag that lasts for a certain period to the fault decision module. After receiving the flag and the adaptive window size, the fault decision module generates a working condition event set for the analog quantity collected by the detection module, matches it in real time according to the fault diagnosis template library of the time-series operating condition event set, and outputs the fault type. This method reads a large number of data sets, and it takes a long time to calculate the characteristic index and event set. In addition, this method has high requirements on the quality of data acquisition, and a complete match is required to output the result.

[0058] An existing paper discloses an analysis of motor overcurrent faults during braking of a mining electric wheel dump truck. This method uses the measured voltage output by the inverter to infer the real-time values ​​of the amplitude and phase of the spatial voltage reference vector in the control program for overcurrent fault detection, but fails to trace the source of the overcurrent fault.

[0059] The existing papers also disclose the research on the fault diagnosis method of high-speed train traction motors based on the T-S fuzzy model. This method establishes the T-S fuzzy model of the traction motor. Based on the T-S fuzzy model, the robust fault diagnosis problem of the traction motor affected by uncertain factors such as parameter changes, interference, and noise is studied. It mainly studies the current sensor fault and the inter-turn short circuit fault of the traction motor stator, but cannot accurately diagnose some other types of faults.

[0060] The existing papers also disclose the real-time diagnosis of overcurrent of traction motors based on time-series feature pattern recognition. This method constructs an event set through time-series feature indicators, establishes a diagnostic template library in the offline state, and realizes the online diagnosis of overcurrent faults. However, in the paper, only the overcurrent faults of the speed signal type can be distinguished, and other fault causes are not studied in depth.

[0061] To sum up, at present, the overcurrent diagnosis of the traction drive system motor mostly adopts the method based on model construction, and mostly can only diagnose the inverter overcurrent of a single fault source type. And the model construction of the traction drive system is complex and it is difficult to represent it with a single mathematical model. A small interference will have a great impact on the diagnosis result. In the current diagnostic methods, the method of calculating the characteristic indicators mostly adopts the fixed sliding window size mode, and the time spent on the calculation for a large amount of data will be relatively long, which requires high hardware performance and affects the maintenance efficiency. The current matching method in the diagnostic process has high requirements for the reading data accuracy and the setting of the event trigger threshold, and has certain requirements for the data time series length, which makes the application range small. Therefore, to solve the problems of high algorithm complexity, poor real-time performance, and difficulty in quickly and accurately identifying the fault type existing in the prior art, the present invention proposes a method and system for identifying overcurrent faults of traction motors.

[0062] The terms used to describe the positional relationship in the drawings are only for illustrative purposes and should not be construed as a limitation to the present invention; The technical solutions of the present invention will be further described below with reference to the drawings and embodiments.

[0063] Embodiment 1 As Figure 1 and Figure 2 shown, this embodiment proposes a method for identifying overcurrent faults of traction motors, including the following steps: S1. Obtain the historical signals of the traction drive system, and perform offline preprocessing on the historical signals to obtain a first feature identifier with fault location discrimination; See Figure 3, which is a power generation circuit topology model of the traction drive system, consists of three major parts: a traction transformer, a traction converter, and a traction motor. Single-phase AC 25 kV alternating current flows into the car body through the pantograph, the main circuit breaker (VCB), and the primary winding of the traction transformer. After being stepped down by the traction transformer, single-phase AC power is provided to the train traction converter through the secondary winding. The electrical energy undergoes AC-DC-AC conversion in the converter in sequence to supply power to the train traction motor. The historical signals include the first phase A current signal , the first phase B current signal , the first intermediate voltage sensor signal , and the first speed signal . Through the analysis of the motor overcurrent fault mechanism, the historical signals are preprocessed offline to obtain the first characteristic identifier, including the first mean value , the first variance of the first intermediate voltage sensor historical signal , the minimum value of the first intermediate voltage sensor signal , the second variance of the first speed signal , the first current minimum value , the first current maximum value , the first current effective value , and the first current normalized value , the first current maximum value , the first current effective value , and the first current normalized value . Among them, the analysis of the motor overcurrent fault mechanism includes: S11: When the train is running normally, the output current waveform of the motor overcurrent is sinusoidal in a cycle, and the absolute value of each sampling period value is less than the protection threshold, and the enable protection will not be triggered. The first intermediate voltage sensor signal and the first speed signal are disturbed within a reasonable range and there will be no large deviation. At this time, it can be judged that the train is in a normal running state.

[0064] S12: When a motor overcurrent fault of the speed signal type occurs in the traction drive system, before the traction control unit TCU works, the waveforms of the first phase A current signal and the first phase B current signal will have obvious distortion phenomena. After the traction control unit TCU works, both drop to 0, and the first speed signal has positive and negative value jumps. The first intermediate voltage sensor signal does not change significantly during the whole process.

[0065] S13: When a motor overcurrent fault of the traction motor type occurs in the traction drive system, before the traction control unit TCU works, the waveform does not distort, but shows a divergent trend. There are obvious burr disturbances. After the traction control unit TCU works both drop to zero, and the signal of the first intermediate voltage sensor The burr disturbance phenomenon disappears.

[0066] S14: When a motor overcurrent fault of the intermediate voltage sensor signal type occurs in the traction drive system, before the traction control unit TCU works, the signal of the first intermediate voltage sensor during normal operation suddenly drops, and at the same time rapidly diverges and exceeds the protection threshold. After the traction control unit TCU works, the current returns to zero, and the signal of the first intermediate voltage sensor returns to the state during normal operation. The first speed signal has no obvious change during the whole process.

[0067] S15: When a motor overcurrent fault of the converter type occurs in the traction drive system, before the traction control unit TCU works, the converter suddenly fails during normal operation, and the signal of the first intermediate voltage sensor instantly drops to near 0 value, and at the same time rapidly diverges and exceeds the protection threshold, and the first speed signal also fluctuates. After the traction control unit TCU works, the current drops to 0, and the signal of the first intermediate voltage sensor returns to the normal level and then rapidly drops.

[0068] S16: From the summary of the motor overcurrent fault mechanism of the traction drive system from S11 to S15, it can be clearly obtained that collecting 、 and has good discrimination, and thus the first characteristic identifier is constructed to include 、 、 、 、 、 、 .

[0069] S17: Through MATLAB-Simulink, a simulation of the motor overcurrent fault of the train traction drive system is carried out, and the data of the system in normal condition and in the case of motor overcurrent fault are obtained as the original data, including the normal state and 4 types of motor overcurrent fault states, a total of 5 groups of data, with 17,500 sample data in each group.

[0070] The offline preprocessing of the historical signal to obtain a first feature identifier with fault location discrimination includes: Based on the historical signal of the first intermediate voltage sensor and the first speed signal a first mean value is obtained, and the calculation expression is:

[0071] where k represents the running time, represents the k first mean value at the running time , represents an intermediate variable, = , represents the sliding window size, represents taking values of and , ; Based on the historical signal of the first intermediate voltage sensor a first variance is obtained, and the calculation expression is:

[0072] where represents the first variance k at the running time , represents the first mean value k at the running time when taking values of ; Based on the historical signal of the first intermediate voltage sensor the minimum value of the first intermediate voltage sensor signal is obtained, and the calculation expression is:

[0073] where represents the minimum value k of the first intermediate voltage sensor signal at the running time , represents taking the minimum value, the k first intermediate voltage sensor historical signal at the running time , represents the first intermediate voltage sensor historical signal at the running time​ , represents the historical signal of the first intermediate voltage sensor at the running moment ; Based on the said first speed signal , the calculation expression for obtaining the second variance is:

[0074] wherein, represents the k second variance at the running moment , represents the k first mean value with a value of at the running moment ; ; Based on the said first phase-A current signal and the first phase-B current signal , the calculation expression for obtaining the first minimum current is:

[0075] wherein, represents the k first minimum current at the running moment , y represents the subscript with values of a and b , represents the k first phase-A current signal at the running moment and the first phase-B current signal , represents the first phase-A current signal at the running moment and the first phase-B current signal , at the running moment, the first phase-A current signal and the first phase-B current signal ; Based on the said first phase-A current signal and the first phase-B current signal , the calculation expression for obtaining the maximum current is:

[0076] wherein, represents taking the maximum value; According to the first phase-A current signal and the first phase-B current signal , the calculation expression for obtaining the first effective current value is:

[0077] wherein, represents the first effective current value k at the th running moment; According to the first phase-A current signal and the first phase-B current signal , the calculation expression for obtaining the first current normalization value is:

[0078] wherein, represents the first current normalization value k at the th running moment.

[0079] S2. Perform segmentation processing on the first feature identifier within a sliding window to obtain a first feature index, generate a corresponding first event using the first feature index, and perform sequence transformation on the first event to obtain a traction motor overcurrent fault location recognition template; In S2, obtain sliding window feature data within the sliding window. The sliding window feature data includes the current frequency of the traction motor obtained through fast Fourier transform (FFT) . Calculate and extract the data length based on the current frequency, and calculate the sliding window size , step size , number of sliding windows , number of those distributed on the left side of the traction control unit action point , number of those distributed on the right side of the traction control unit action point and data length ; The sliding window size is dynamically selected according to the magnitude of the current frequency . The processing and analysis of the S2 sliding window are as follows: S21: When analyzing and processing the identifier, adopt the sliding window idea to process the data. The sliding window size and step size are determined based on the collected current data through fast Fourier transform to obtain the current frequency . Determine the sliding window size and step size according to the current frequency . The current frequency used for fast Fourier transform To enable the first 5,000 data points out of the 17,500 data points extracted after diagnosis, this can make the calculated current frequency not affected by the post-fault jump and ensure the calculation accuracy. In addition, it can also ensure that there are at least two complete current cycles in the calculation.

[0080] S22: Obtain the number of sliding windows required for different frequencies through training with offline data The number of sliding windows can be obtained according to the current frequency in S21 Taking the action time of the traction control unit TCU as the base point, distribute the sliding window on both sides of the base point.

[0081] The size of the sliding window The calculation expression is:

[0082] Among them, represents the sampling period; The step size The calculation expression is:

[0083] Obtain the number of sliding windows required for different current frequencies through training with offline data ; The number distributed on the left side of the traction control unit action point The calculation expression is:

[0084] The number distributed on the right side of the traction control unit action point The calculation expression is:

[0085] The data length The calculation expression is: .

[0086] The first feature identifier is segmented according to the sliding window feature data to obtain the following calculation expressions for the first feature indicators respectively:

[0087]

[0088]

[0089]

[0090]

[0091] Among them, , , , , represent the first characteristic index, , , , , represent the k first characteristic index at the running time.

[0092] The method of generating a corresponding first event by using the first characteristic index and performing sequence transformation on the first event to obtain a traction motor overcurrent fault location and identification template includes: S21. Input the first characteristic indexes , , , , into the first hysteresis comparator respectively, and output the corresponding first event by the first hysteresis comparator. The calculation expression is as follows:

[0093] Among them, represents the subscript taking values respectively. There are 5 first hysteresis comparators. represents the opening point of the first hysteresis comparator, represents the closing point of the first hysteresis comparator, , ; if is not less than the opening point of the first hysteresis comparator, then the first event outputs 1; if is less than the opening point of the first hysteresis comparator and greater than the closing point of the first hysteresis comparator, then the first event maintains the output state consistent with the previous moment event; if is not greater than the closing point representing the first hysteresis comparator, then the first event outputs 0; S22. Form a binary time series vector from the first events X in the order of the subscript, and convert the binary time series vector into a decimal number as the following mathematical expression: ​

[0094] Among them, represents the sliding window position of the current calculated decimal number of, (1, n ), represents the first vector length, ; S23. Based on the decimal number , obtain the calculation expression of the traction motor overcurrent fault location and identification vector as follows:

[0095] Among them, represents the element of the vector with the number of sliding windows as the first vector length of, represents taking the minimum value of the decimal number of, represents taking the maximum value of the decimal number of; The calculation expression of

[0096] The calculation expression of

[0097] S24. Combine the traction motor overcurrent fault location and identification vectors obtained under different fault types to form a set, and obtain the traction motor overcurrent fault location and identification template.

[0098] Each first event has two states, 0 and 1. Write the first event as a binary time series vector, and convert the binary time series vector into a normalized decimal number , count the sequence characteristics, and generate a traction motor overcurrent fault location and identification vector with the number of sliding windows as the length.

[0099] S3. Construct a traction motor overcurrent fault identification model, divide the traction motor overcurrent fault location and identification template into a training set and a test set, use the training set to train the traction motor overcurrent fault identification model, and use the test set to test the effectiveness of the traction motor overcurrent fault identification model to obtain a trained traction motor overcurrent fault identification model; S4. Obtain the real-time signal of the traction drive system, perform online fault detection on the real-time signal, and obtain a second feature identifier with fault location discrimination; S5. Perform segmentation processing on the second feature identifier within the sliding window to obtain a second feature index, generate a corresponding second event using the second feature index, perform sequence transformation on the second event, and obtain an online template vector; S6. Input the online template vector into the trained traction motor overcurrent fault identification model and output a fault identification result.

[0100] In this embodiment, first in the offline stage, perform offline preprocessing on the historical signal to obtain a first feature identifier, perform segmentation processing on the first feature identifier according to the sliding window feature data to obtain a first feature index, generate a corresponding first event using the first feature index, and perform sequence transformation on the first event. The purpose is to reduce the offline data dimension to one dimension, thereby effectively reducing the computational complexity. Then, construct and train a traction motor overcurrent fault identification model to obtain a trained traction motor overcurrent fault identification model; then perform online fault detection on the real-time signal to obtain a second feature identifier with fault location discrimination; perform segmentation processing on the second feature identifier within the sliding window to obtain a second feature index, generate a corresponding second event using the second feature index, and perform sequence transformation on the second event. The purpose is to reduce the online data dimension to one dimension, thereby further effectively reducing the computational complexity; finally, input the online template vector into the trained traction motor overcurrent fault identification model to achieve fast and accurate discrimination of the fault type, significantly improving the efficiency and accuracy of real-time fault identification. The present invention can improve the speed and real-time performance of data matching through the traction motor overcurrent fault identification model on the premise of ensuring the accuracy of fault identification, so as to better meet the requirements of real-time diagnosis of motor overcurrent faults.

[0101] Embodiment 2 This embodiment further explains S3 in the above-mentioned traction motor overcurrent fault identification method.

[0102] The traction motor overcurrent fault identification model in S3 includes a sequentially connected sequence input layer, a bidirectional long short-term memory network layer with 50 hidden units, a self-attention layer with 2 attention heads and 8 hidden units, a first fully connected layer, a first activation function layer, a dropout layer, a second fully connected layer, a second activation function layer, and a classification layer; Normalize the traction motor overcurrent fault location and identification template into a 16-dimensional template vector and input it into the sequence input layer. The sequence input layer outputs time series data. The bidirectional long short-term memory network layer receives the time series data and outputs the bidirectional dependence features of the time series. The self-attention layer receives the bidirectional dependence features and outputs the time step feature data weighted by attention. The first fully connected layer receives the time step feature data and outputs a 30-dimensional first linear feature combination result. The first activation function layer receives the first linear feature combination result and outputs the first activation feature result. The dropout layer receives the first activation feature result, sets the dropout rate to 0.2, and outputs the feature result after dropout. The second fully connected layer receives the feature result after dropout and outputs a second linear feature combination result. The second activation function layer receives the second linear feature combination result and outputs the fault type prediction probability. The classification layer receives the fault type prediction probability and outputs the fault type identification result.

[0103] Extract features from and normalize (0–1) the traction motor overcurrent fault location and identification template, and divide it into a training set and a test set (ratio 7:3). Each sample is reconstructed into an LSTM sequence format, and the label data is converted into categorical variables. Training the traction motor overcurrent fault identification model includes: S31. Set the training parameters of the traction motor overcurrent fault identification model. The training parameters include: the maximum number of epochs is 150, the batch size is 32, the initial learning rate is 0.005, the learning rate schedule is to decrease by 0.2 times every 50 epochs, the L2 regularization coefficient is 0.0001, and the gradient clipping threshold is 1; S32. Use the Adam optimizer to perform preliminary optimization on the training parameters to obtain the preliminarily optimized hyperparameters; S33. Use the genetic algorithm to perform global optimization on the preliminarily optimized hyperparameters until the hyperparameter combination corresponding to the individual with the highest fitness value is output, completing the training of the traction motor overcurrent fault identification model.

[0104] Using the genetic algorithm to perform global optimization on the preliminarily optimized hyperparameters includes: S331. Set the hyperparameter configuration of each individual in the population size; preferably 20 individuals, ranging from 10 to 50 individuals. Randomly generate the initial population within the defined hyperparameter value range. Each individual corresponds to a combination of a set of hyperparameter configurations, including the number of bidirectional LSTM hidden units (30–100), the number of self-attention heads (2–8), the initial learning rate (0.001–0.01), the batch size (16–64), and the L2 regularization coefficient (0.00001–0.001); S332. Use the hyperparameter configuration of each individual to build a traction motor overcurrent fault identification model, train the traction motor overcurrent fault identification model using the training set, and calculate the classification accuracy α and Macro-F1 score β on the test set. Calculate the fitness value X based on the classification accuracy α and Macro-F1 score β as follows:

[0105] S333. According to the fitness value X, use tournament selection or roulette wheel selection to select excellent individuals from the current population size as parents to retain hyperparameter configurations with high fitness; S334. Perform crossover operations on the parent individuals. Use single-point crossover or uniform crossover to swap the hyperparameter values of two parent individuals to generate new crossed individuals, so as to enhance population diversity; S335. With a mutation probability of 5% to 10%, randomly perturb and fine-tune the positions of the hyperparameter values of the new crossed individuals to obtain new mutated individuals, so as to avoid falling into local optima; S336. Add the new crossed individuals and the new mutated individuals to the population size, form the next generation and perform update iterations; S337. When the number of update iterations reaches the iteration threshold or the change in the fitness value X is less than 0.001 within 5 consecutive generations, end the global optimization, output the hyperparameter combination corresponding to the individual with the highest fitness value, and the hyperparameter combination corresponding to the individual with the highest fitness value finally output is used for the construction of the final traction motor overcurrent fault identification model.

[0106] The traction motor overcurrent fault identification model is used to extract deep features of complex time series and optimize the model hyperparameters through the genetic algorithm (GA) to achieve high accuracy, high robustness, and high real-time performance. The final performance of the traction motor overcurrent fault identification model is evaluated by the classification accuracy, confusion matrix, Macro-F1 score, etc. of the training set and the test set. Combining the self-attention mechanism and masking analysis, the importance of each time step for the prediction result can be determined, providing effective support for the mechanism analysis and location of traction motor overcurrent faults.

[0107] Example 3 See Figure 2 , the real-time signal in S4 includes the second phase A current signal , the second phase B current signal , the second intermediate voltage sensor signal and the second speed signal , the second feature identifier includes the minimum value of the second intermediate voltage sensor signal , the second mean value , the second mean , the second intermediate voltage sensor signal third variance , the second speed signal fourth variance , the second minimum current , the second maximum current , the second current normalization value , the second root mean square current ; Performing online fault detection on the real-time signal to obtain a second feature identifier with fault location discrimination includes: S41. Judging whether the collected second phase A current signal and the second phase B current signal are greater than the current protection threshold within 5 consecutive sampling periods. If so, enter the fault diagnosis process and execute S42; if not, continue to perform online fault detection on the collected second phase A current signal and the second phase B current signal ; For step S31, the further condition for judging whether the system has a motor overcurrent fault is: the current sampling period is . If the absolute value of the current sampling data is greater than the current protection threshold in 5 consecutive sampling periods, it means that the system has a motor overcurrent fault, and the system sends an enable diagnosis flag bit; if the absolute value of the current sampling data cannot be greater than the current protection threshold in 5 consecutive sampling periods, it means that the system is in a normal state.

[0108] S42. According to the second intermediate voltage sensor signal and the second speed signal , the calculation expression for obtaining the second mean is:

[0109] where k represents the running time, represents the k second mean at the running time , represents an intermediate variable, = , represents the sliding window size, represents taking values of and , ; According to the second intermediate voltage sensor signal , the calculation expression for obtaining the third variance is:

[0110] Among them, represents the k third-party variance at the th running time, k and represents the second mean value with a value of at the th running time according to the second intermediate voltage sensor signal; The calculation expression for obtaining the minimum value

[0111] of the second intermediate voltage sensor signal k at the th running time is: where represents the k minimum value of the second intermediate voltage sensor signal at the th running time, represents taking the minimum value, and the second intermediate voltage sensor signal at the

[0112] th running time; k The calculation expression for obtaining the fourth variance according to the second speed signal k at the th running time is: ; According to the second phase A current signal and the second phase B current signal, the calculation expression for obtaining the

[0113] Among them, represents the k second minimum current value at the running time , y represents the subscript with values of a and b , represents the second phase-A current signal k at the running time and the second phase-B current signal , represents the second phase-A current signal at the running time and the second phase-B current signal , At the running time, the second phase-A current signal and the second phase-B current signal ; According to the second phase-A current signal and the second phase-B current signal , the calculation expression for obtaining the second maximum current value is:

[0114] Among them, represents taking the maximum value; According to the second phase-A current signal and the second phase-B current signal , the calculation expression for obtaining the second current effective value is:

[0115] Among them, represents the k second current effective value at the running time ; According to the second phase-A current signal and the second phase-B current signal , the calculation expression for obtaining the second current normalization value is:

[0116] Among them, represents the k second current normalization value at the running time .

[0117] Extract a fault data segment with a data length of 17500. First, perform a fast Fourier transform on the extracted current data to obtain the current frequency , determine the sliding window size according to the current frequency and the step size , finally obtain the number of sliding windows and the number of them distributed on the right side of the action point of the traction control unit and the data length ; based on the obtained sliding window size and the step size , perform the calculation of the identifier, and calculate the second feature identifier in each window to obtain the second feature index , , , , .

[0118] In step S5, the second feature identifier is segmented in the sliding window to obtain the following calculation expressions for the second feature index respectively:

[0119]

[0120]

[0121]

[0122]

[0123] where , , , , represent the second feature index, , , , , represent the k second feature index at the running time.

[0124] In step S5, generate the corresponding second event by using the second feature index, and perform sequence conversion on the second event to obtain a template vector, including: S51. Input the second feature index , , , , into the second hysteresis comparator respectively, and the second hysteresis comparator outputs the corresponding second event ; S52. Form a binary timing event with the second event in the order of the subscript X ​ , convert the binary timing event into a decimal number using the following mathematical expression:

[0125] where represents the sliding window position for calculating the current decimal number , (1, m ), m represents the length of the second vector, ; S53. Based on the decimal number , obtain the calculation expression for the template vector as follows:

[0126] where represents the element of the vector with the number of sliding windows as the length of the second vector m , represents taking the minimum value of the decimal number , represents taking the maximum value of the decimal number ; The calculation expression for

[0127] The calculation expression for .

[0128] See Figure 2 , S6 normalize the online template vector to a 16-dimensional template vector and input it into the trained traction motor overcurrent fault identification model, and output the fault identification result by the trained traction motor overcurrent fault identification model.

[0129] The present invention aims to protect the binary encoding of the feature identifiers in each sliding window and the conversion of the obtained binary sequence into a one-dimensional normalized decimal time series. This process effectively reduces the dimension of the data, significantly reduces the amount of computation, improves the computational efficiency and real-time performance, ensures the consistency of the data in the numerical range, and thus greatly improves the identification efficiency of the traction motor overcurrent fault identification model.

[0130] The present invention includes two parts: offline modeling and online diagnosis. In the offline stage, first, fault mechanism analysis is carried out on the current signal, voltage signal, and speed signal in the historical data, and characteristic indexes such as the minimum value, mean value, and variance are extracted; then, the current frequency is calculated through fast Fourier transform to determine the sliding window; the characteristic indexes are calculated within the sliding window, and a hysteresis comparator is used to generate characteristic events, and finally, a template library containing different fault types is constructed. In the online stage, the current signal is collected in real time and compared with the protection threshold. If a continuous overcurrent phenomenon is detected, the fault diagnosis process is triggered. Subsequently, the cached data segment is extracted, the size and step length of the sliding window are adjusted according to the current frequency, the real-time characteristic indexes are calculated within the window, and the corresponding event sequence is generated. The sequence is converted from binary to a normalized decimal sequence. Compared with the prior art, the advantages of the present invention are 1. It is difficult to construct a mathematical model for the traction drive system. The existing diagnostic methods based on model construction are difficult to accurately represent the system and are easily affected by interference, which affects the diagnostic results. At the same time, the traditional fault diagnosis method based on the dynamic time warping (DTW) algorithm has a large amount of calculation and a long time when dealing with multi-dimensional data, and it is difficult to meet the real-time requirements. However, the present invention does not rely on a single mathematical model. By converting the multi-dimensional binary time series into a normalized one-dimensional decimal time series, it can diagnose the motor overcurrent fault more stably and accurately, and reduce the influence of interference on the diagnostic results.

[0131] 2. In the existing diagnostic methods, the fixed sliding window size mode has low efficiency when dealing with a large amount of data and requires high hardware performance. The present invention dynamically adjusts the size and step length of the sliding window according to the current frequency, which can effectively reduce the amount of calculation, improve the data processing efficiency, reduce the dependence on hardware performance, and thus improve the maintenance efficiency.

[0132] 3. The method of dynamically adjusting the size and step length of the sliding window in the present invention effectively reduces the dependence on hardware performance and optimizes the processing efficiency. Compared with the traditional method with a fixed window size, it can more flexibly adapt to the data characteristics under different working conditions, further improving the calculation efficiency and fault detection accuracy.

[0133] 4. The method proposed in this paper can accurately locate 4 types of motor overcurrent faults, including speed signal faults, traction motor faults, intermediate DC voltage signal faults, and converter module faults. Compared with some existing methods that can only diagnose a single fault source type or have inaccurate diagnosis for some faults, the comprehensiveness and accuracy of fault diagnosis are significantly improved.

[0134] It should also be specifically stated that in terms of feature extraction, in addition to the currently constructed feature identifiers and index calculation methods, the Hidden Markov Model (HMM) can be tried. HMM can describe the transition process of system states and the observation probabilities in different states, and can be used to analyze the state changes during the overcurrent fault process of the motor. However, since HMM requires accurate estimation of model parameters and has relatively high requirements for the assumptions of system states, reasonable adjustment and optimization may be needed according to specific situations in practical applications.

[0135] Embodiment 4 Refer to Figure 4 , this embodiment also proposes a traction motor overcurrent fault identification system, including: An offline preprocessing module, used to obtain the historical signals of the traction drive system, perform offline preprocessing on the historical signals, and obtain a first feature identifier with fault location discrimination; An offline segmentation and transformation module, used to perform segmentation processing on the first feature identifier within a sliding window to obtain a first feature index, generate a corresponding first event using the first feature index, and perform sequence transformation on the first event to obtain a traction motor overcurrent fault location identification template; A traction motor overcurrent fault identification model construction module, used to construct a traction motor overcurrent fault identification model, divide the traction motor overcurrent fault location identification template into a training set and a test set, train the traction motor overcurrent fault identification model using the training set, and test the effectiveness of the traction motor overcurrent fault identification model using the test set to obtain a trained traction motor overcurrent fault identification model; An online preprocessing module, used to obtain the real-time signals of the traction drive system, perform online fault detection on the real-time signals, and obtain a second feature identifier with fault location discrimination; An online segmentation and transformation module, used to perform segmentation processing on the second feature identifier within a sliding window to obtain a second feature index, generate a corresponding second event using the second feature index, and perform sequence transformation on the second event to obtain an online template vector; An output module, used to input the online template vector into the trained traction motor overcurrent fault identification model and output a fault identification result.

[0136] In this embodiment, first in the offline stage, the historical signal is preprocessed offline to obtain a first feature identifier. The first feature identifier is segmented according to the sliding window feature data to obtain a first feature index. A corresponding first event is generated using the first feature index, and the first event is subjected to sequence transformation. The purpose is to reduce the offline data dimension to one dimension, thereby effectively reducing the computational complexity. Then, a traction motor overcurrent fault identification model is constructed and trained to obtain a trained traction motor overcurrent fault identification model. Next, online fault detection is performed on the real-time signal to obtain a second feature identifier with fault location discrimination. The second feature identifier is segmented within the sliding window to obtain a second feature index. A corresponding second event is generated using the second feature index, and the second event is subjected to sequence transformation. The purpose is to reduce the online data dimension to one dimension, thereby further effectively reducing the computational complexity. Finally, by inputting the online template vector into the trained traction motor overcurrent fault identification model, rapid and accurate discrimination of the fault type is achieved, significantly improving the efficiency and accuracy of real-time fault identification. The present invention can, on the premise of ensuring the accuracy of fault identification, improve the speed and real-time performance of data matching through the traction motor overcurrent fault identification model, thereby better meeting the requirements of real-time diagnosis of motor overcurrent faults.

[0137] Obviously, the above embodiments of the present invention are only examples for clearly illustrating the present invention, and are not limitations on the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all the implementation manners here. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the claims of the present invention.

Claims

1. A method for identifying overcurrent faults in a traction motor, characterized in that: The following steps are involved: S1. Obtaining historical signals of the traction drive system, preprocessing the historical signals offline, and obtaining a first characteristic identifier with fault location discrimination; S2. Segmenting the first feature identifier in the sliding window to obtain a first feature index, using the first feature index to generate a corresponding first event, performing sequence conversion on the first event, and obtaining a traction motor overcurrent fault location identification template; S3. construct a traction motor overcurrent fault recognition model, divide the traction motor overcurrent fault location recognition template into a training set and a test set, use the training set to train the traction motor overcurrent fault recognition model, and use the test set to test the effectiveness of the traction motor overcurrent fault recognition model to obtain a trained traction motor overcurrent fault recognition model; S4. Acquire a real-time signal of the traction drive system, perform online fault detection on the real-time signal, and obtain a second characteristic identifier having a fault location distinction; S5. Segmenting the second feature identifier in the sliding window to obtain a second feature index, using the second feature index to generate a corresponding second event, performing sequence conversion on the second event to obtain an online template vector; S6. Input the online template vector into the trained traction motor overcurrent fault recognition model and output the fault recognition result.

2. The traction motor overcurrent fault identification method according to claim 1, characterized in that: The historical signal includes a first A-phase current signal , the first B phase current signal , the first intermediate voltage sensor signal and the first speed signal , the first feature identifier includes a first mean , the first intermediate voltage sensor historical signal The first variance , the first intermediate voltage sensor signal Minimum value of , first speed signal The second variance , the first current minimum , the first current maximum value , the first current effective value and the first current normalized value ; The offline preprocessing of the historical signal to obtain a first feature identifier with fault location discrimination includes: According to the historical signal of the first intermediate voltage sensor and the first speed signal , get the first mean The calculation expression is: in, k Indicates the running time, Indicates k The first mean of the running time , represents the intermediate variable, = , represents the sliding window size, Indicates that the value is and , ; According to the historical signal of the first intermediate voltage sensor , get the first variance The calculation expression is: in, Indicates k First variance of running time , Indicated in k Running time The value is The first mean ; According to the historical signal of the first intermediate voltage sensor , get the first intermediate voltage sensor signal Minimum value of The calculation expression is: in, Indicates k The first intermediate voltage sensor signal at the time of operation Minimum value of , Indicates taking the minimum value, No. k The historical signal of the first intermediate voltage sensor at the time of operation , express The historical signal of the first intermediate voltage sensor at the time of operation , Indicates The historical signal of the first intermediate voltage sensor at the time of operation ; According to the first speed signal , and get the second variance The calculation expression is: in, Indicates k Run-time second variance , Indicated in k Running time The value is The first mean ; According to the first A phase current signal and the first B phase current signal , calculate the first current minimum value The calculation expression is: in, Indicates k The first minimum current value during operation , y Indicates that the value is a and b The subscript of Indicated in k The first A-phase current signal at the time of operation and the first B phase current signal , Indicated in The first A-phase current signal at the time of operation and the first B phase current signal , exist The first A-phase current signal at the time of operation and the first B phase current signal ; According to the first A phase current signal and the first B phase current signal , calculate the maximum current The calculation expression is: in, Indicates taking the maximum value; According to the first A phase current signal and the first B phase current signal , calculate the first current effective value The calculation expression is: in, Indicates k The first current effective value at the time of operation ; According to the first A phase current signal and the first B phase current signal , calculate the first current normalized value The calculation expression is: in, Indicates k The first normalized current value at the time of operation .

3. The traction motor overcurrent fault identification method according to claim 2 is characterized in that: Acquire sliding window characteristic data in the sliding window, wherein the sliding window characteristic data includes the current frequency of the traction motor , sliding window size , Step Length , Number of sliding windows , the number of distribution on the left side of the traction control unit action point , the number of nodes distributed on the right side of the traction control unit action point and data length ; The sliding window size The calculation expression is: in, Indicates the sampling period; The step length The calculation expression is: The number of the distribution on the left side of the traction control unit action point The calculation expression is: The number of the distribution on the right side of the traction control unit action point The calculation expression is: The data length The calculation expression is: The first feature identifier is segmented according to the sliding window feature data to obtain the first feature indexes, which are respectively calculated as follows: in, , , , , represents the first characteristic index, , , , , Indicates k The first characteristic indicator of running time.

4. The traction motor overcurrent fault identification method according to claim 3 is characterized in that: The method of using the first characteristic indicator to generate a corresponding first event and performing sequence conversion on the first event to obtain a traction motor overcurrent fault location identification template includes: S21. The first characteristic index , , , , are respectively input into the first hysteresis comparator, and the first hysteresis comparator outputs the corresponding first event The calculation expression is as follows: in, Indicates that the values ​​are The subscript of Indicates the first hysteresis comparator start point, represents the first hysteresis comparator shutdown point, , ;like Not less than the first hysteresis comparator opening point , then the first event The output is 1; if Less than the first hysteresis comparator turn-on point and is greater than the first hysteresis comparator shutdown point , then the first event Keep the output state consistent with the previous moment event; if Not greater than the first hysteresis comparator shutdown point , then the first event The output is 0; S22. The first event Press the X The order forms a binary time series vector , the binary time series vector Convert to decimal The mathematical expression is as follows: in, Indicates the current calculated decimal number The sliding window position, (1, n ), represents the length of the first vector, ; S23. Based on the decimal number , get the traction motor overcurrent fault location identification vector The calculation expression is as follows: in, Indicates that the number of sliding windows is the length of the first vector The elements of the vector, Indicates a decimal number The minimum value of Indicates a decimal number The maximum value of The calculation expression is: The calculation expression is: S24. The traction motor overcurrent fault location identification vector obtained under different fault types A collection is formed to obtain the traction motor overcurrent fault location and identification template.

5. The traction motor overcurrent fault identification method according to claim 4 is characterized in that: The traction motor overcurrent fault recognition model comprises a sequentially connected sequence input layer, a bidirectional long short-term memory network layer, a self-attention layer, a first fully connected layer, a first activation function layer, a discard layer, a second fully connected layer, a second activation function layer and a classification layer; The traction motor overcurrent fault location and identification template is normalized into a 16-dimensional template vector and input into the sequence input layer, and the sequence input layer outputs time series data. The bidirectional long short-term memory network layer receives the time series data and outputs the bidirectional dependency features of the time series. The self-attention layer receives the bidirectional dependency features and outputs the time step feature data weighted by attention. The first fully connected layer receives the time step feature data and outputs the first linear feature combination result. The first activation function layer receives the first linear feature combination result and outputs the first activation feature result. The discard layer receives the first activation feature result and outputs the feature result after discarding. The second fully connected layer receives the feature result after discarding and outputs the second linear feature combination result. The second activation function layer receives the second linear feature combination result and outputs the fault type prediction probability. The classification layer receives the fault type prediction probability and outputs the fault type identification result.

6. The traction motor overcurrent fault identification method according to claim 5, characterized in that: The training of the traction motor overcurrent fault identification model includes: S31. Set the training parameters of the traction motor overcurrent fault identification model, the training parameters include: maximum rounds of 150, batch size of 32, initial learning rate of 0.005, learning rate scheduling of 0.2 times every 50 rounds, L2 regularization coefficient of 0.0001, gradient clipping threshold of 1; S32. Using the Adam optimizer to preliminarily optimize the training parameters to obtain preliminarily optimized hyperparameters; S33. Use a genetic algorithm to globally optimize the hyperparameters after preliminary optimization until the hyperparameter combination corresponding to the individual with the highest fitness value is output, thereby completing the training of the traction motor overcurrent fault identification model.

7. The method for identifying overcurrent fault of a traction motor according to claim 6, characterized in that: The method of using a genetic algorithm to globally optimize the initially optimized hyperparameters includes: S331. Set the hyperparameter configuration of each individual in the population size; S332. Use the hyperparameter configuration of each individual to build a traction motor overcurrent fault recognition model, use the training set to train the traction motor overcurrent fault recognition model, and calculate the classification accuracy α and Macro-F1 score β on the test set. Calculate the fitness value X based on the classification accuracy α and Macro-F1 score β as follows: X 0.7α+0.3β S333. According to the fitness value X, a tournament selection or a roulette wheel selection is adopted to select a good individual from the current population size as a parent generation; S334. Perform a crossover operation on the parent individuals, using single-point crossover or uniform crossover, and swap the hyperparameter values ​​of the two parent individuals to generate a new crossover individual; S335. Randomly perturb and fine-tune the hyperparameter value sites of the new individuals of the crossover with a mutation probability of 5% to 10% to obtain the mutated new individuals; S336. Add the new individuals of crossover and mutation to the population scale, and then update and iterate after forming the next generation; S337. When the number of update iterations reaches the iteration threshold or the change of the fitness value X in 5 consecutive generations is less than 0.001, the global optimization is terminated and the hyperparameter combination corresponding to the individual with the highest fitness value is output.

8. The method for identifying overcurrent fault of a traction motor according to claim 7, characterized in that: The real-time signal includes a second A-phase current signal , the second B phase current signal , the second intermediate voltage sensor signal and the second speed signal The second characteristic identifier includes a second intermediate voltage sensor signal Minimum value of , the second mean , the second intermediate voltage sensor signal The third party difference , Second speed signal The fourth variance , the second current minimum , the second current maximum value , the second current normalized value , the second current effective value ; The performing online fault detection on the real-time signal to obtain a second feature identifier with fault location discrimination includes: S41. Determine the collected second A phase current signal and the second B phase current signal Is it greater than the current protection threshold in several consecutive sampling cycles? If yes, then the fault diagnosis phase is entered and S42 is executed; if no, the collected second A phase current signal is continued. and the second B phase current signal Conduct online fault detection; S42. According to the second intermediate voltage sensor signal and the second speed signal , and obtain the second mean The calculation expression is: in, k Indicates the running time, Indicates k The second mean of running time , represents the intermediate variable, = , represents the sliding window size, Indicates that the value is and , ; According to the second intermediate voltage sensor signal , get the third party difference The calculation expression is: in, Indicates k Third-party differences at runtime , Indicated in k Running time The value is The second mean ; According to the second intermediate voltage sensor signal , get the second intermediate voltage sensor signal Minimum value of The calculation expression is: in, Indicates k The second intermediate voltage sensor signal at the operating time Minimum value of , Indicates taking the minimum value, No. k The second intermediate voltage sensor signal at the operating time , express The second intermediate voltage sensor signal at the operating time , Indicates The second intermediate voltage sensor signal at the operating time ; According to the second speed signal , and get the fourth variance The calculation expression is: in, Indicates k Run-time fourth variance , Indicated in k Running time The value is The second mean ; According to the second A phase current signal and the second B phase current signal , calculate the second current minimum value The calculation expression is: in, Indicates k The second minimum current during operation , y Indicates that the value is a and b The subscript of Indicated in k The second A phase current signal at the running time and the second B phase current signal , Indicated in The second A phase current signal at the running time and the second B phase current signal , exist The second A phase current signal at the running time and the second B phase current signal ; According to the second A phase current signal and the second B phase current signal , calculate the second current maximum value The calculation expression is: in, Indicates taking the maximum value; According to the second A phase current signal and the second B phase current signal , calculate the second current effective value The calculation expression is: in, Indicates k The second current effective value at the time of operation ; According to the second A phase current signal and the second B phase current signal , calculate the second current normalized value The calculation expression is: in, Indicates k The second current normalized value at the time of operation .

9. The method for identifying overcurrent fault of a traction motor according to claim 8, characterized in that: The second feature identifier is segmented in the sliding window to obtain the second feature indexes, which are respectively calculated as follows: in, , , , , represents the second characteristic index, , , , , Indicates k The second characteristic indicator of the running time; Generating a corresponding second event using the second characteristic indicator, performing sequence conversion on the second event, and obtaining an online template vector, including: S51. The second characteristic index , , , , are respectively input into the second hysteresis comparator, and the second hysteresis comparator outputs the corresponding second event ; S52. The second event Press the X The sequence of binary sequential events , the binary timing events Convert to decimal The mathematical expression is as follows: in, Indicates the current calculated decimal number The sliding window position, (1, m ), m represents the length of the second vector, ; S53. Based on the decimal number , get the online template vector The calculation expression is as follows: in, Indicates that the number of sliding windows is the length of the second vector m The elements of the vector, Indicates a decimal number The minimum value of Indicates a decimal number The maximum value of The calculation expression is: The calculation expression is: 。 10. A traction motor overcurrent fault identification system, characterized in that: include: An offline preprocessing module, used for acquiring a historical signal of the traction drive system, performing offline preprocessing on the historical signal, and obtaining a first feature identifier with fault location discrimination; an offline segmentation and conversion module, configured to segment the first feature identifier in a sliding window to obtain a first feature index, generate a corresponding first event using the first feature index, and perform sequence conversion on the first event to obtain a traction motor overcurrent fault location and recognition template; A traction motor overcurrent fault identification model construction module is used to construct a traction motor overcurrent fault identification model, divide the traction motor overcurrent fault location identification template into a training set and a test set, use the training set to train the traction motor overcurrent fault identification model, and use the test set to test the effectiveness of the traction motor overcurrent fault identification model to obtain a trained traction motor overcurrent fault identification model; An online preprocessing module, used for acquiring a real-time signal of a traction drive system, performing online fault detection on the real-time signal, and obtaining a second characteristic identifier with fault location discrimination; An online segmentation conversion module, used to perform segmentation processing on the second feature identifier in a sliding window to obtain a second feature index, generate a corresponding second event using the second feature index, and perform sequence conversion on the second event to obtain an online template vector; The output module is used to input the online template vector into the trained traction motor overcurrent fault recognition model and output the fault recognition result.

Images