High-speed train traction motor fault diagnosis method based on migration random forest

Abstract

The invention discloses a high-speed train traction motor fault diagnosis method based on a migration random forest, and belongs to the technical field of high-speed train fault diagnosis, and the method specifically comprises the following steps: collecting original signals of a high-speed train traction motor in multiple states under multiple working conditions, extracting time domain features and frequency domain features from the original signals and integrating the time domain features and the frequency domain features to form a feature data set; carrying out distribution alignment on the source working condition data and the target working condition data; constructing an initial random forest fault diagnosis model; updating model parameters by using the target working condition data set after distribution alignment, and constructing a migration random forest fault diagnosis model; and a sample of a current working condition is collected, and online fault diagnosis is realized based on the migration random forest fault diagnosis model. According to the method, the problem of data distribution difference caused by multiple working conditions can be overcome, fault information under related working conditions is fully utilized, and the fault diagnosis accuracy of the traction motor is improved.

Description

Technical Field

[0001] The present invention belongs to the field of high-speed train fault diagnosis, and in particular relates to a high-speed train traction motor fault diagnosis method based on migration random forest. Background Art

[0002] The traction motor of a high-speed train is the core component that drives the train forward and is the power source of the high-speed train. The traction motor operates in a long-term harsh and complex environment, and is affected by various factors such as high temperature, vibration and electromagnetic interference, so failures are inevitable. Traction motors usually have mechanical failures such as bearing wear, gear failure, rotor imbalance, electrical failures such as winding short circuit, insulation aging, overload, thermal failures such as performance degradation or damage caused by overheating, and control system failures such as inverter failure and sensor failure. Once a traction motor failure occurs, it may cause the train to lose power or even completely lose power, which in turn causes a major safety accident. Therefore, it is crucial to perform effective fault diagnosis on traction motors.

[0003] At present, the fault diagnosis methods for high-speed train traction motors include: analytical model methods, knowledge-based methods, and data-driven methods. The first two methods rely too much on accurate analytical models and expert knowledge, but the actual high-speed train traction motor has a complex structure, and accurate models and sufficient expert knowledge are difficult to obtain. Data-driven fault diagnosis methods require sufficient high-speed train operation history data, that is, it is necessary to be able to obtain operation data covering multiple working conditions of the traction motor, and then establish an accurate mapping relationship between output data and faults. However, most of the data collected during the operation of high-speed trains are healthy operating status data, and semi-physical tests can only obtain data under specific working conditions, which is difficult to meet the needs of training, testing and verification of traction drive system operation status monitoring methods for complex service conditions. At the same time, it is very difficult to obtain and label new working condition data. In this case, a small sample problem will occur under new working conditions, which will cause the machine learning-based fault diagnosis model to overfit and thus reduce the fault diagnosis performance. Due to the differences in data distribution caused by multiple working conditions, data-driven fault diagnosis models are often difficult to achieve good results directly in similar new working conditions. Therefore, a method is needed to effectively solve the problem of performance degradation of fault diagnosis algorithms caused by multiple working conditions, while overcoming data distribution differences and improving the performance of fault diagnosis algorithms. Summary of the invention

[0004] In order to solve the above problems, the present invention proposes a high-speed train traction motor fault diagnosis method based on migration random forest, which is used to diagnose the fault of the high-speed train traction motor. By constructing a migration random forest fault diagnosis model, the present invention can overcome the distribution differences of training data caused by multiple working conditions, thereby improving the fault diagnosis accuracy of the fault diagnosis algorithm for target working conditions with scarce samples.

[0005] The technical solution adopted by the present invention is as follows: A high-speed train traction motor fault diagnosis method based on migration random forest is used to diagnose the fault of the traction motor, which specifically includes the following steps: Step 1: collect original signals of the traction motor of the high-speed train under multiple working conditions in multiple states, extract time domain features and frequency domain features from the original signals and integrate them to form a feature data set; Step 2: Design a data nonlinear transformation method based on balanced distribution adaptation to align the distribution of source working condition data and target working condition data, and eliminate the distribution difference caused by working condition changes; Step 3: Use the source operating condition data set after distribution alignment to build an initial random forest fault diagnosis model; Step 4: Maintain the initial random forest fault diagnosis model structure, use the target working condition data set after distribution alignment to update the model parameters, and build a migration random forest fault diagnosis model; Step 5: Collect samples of the current working condition, use the conversion matrix to convert the samples of the current working condition, input the converted samples into the migration random forest fault diagnosis model, obtain the fault category of the samples, and realize online fault diagnosis.

[0006] Furthermore, the specific process of step 1 is as follows: Step 1.1, based on the existing high-speed rail traction control system simulation platform, a multi-sensor fusion mode is adopted to collect sensor signals of the traction motor under four types of states, namely, normal state, rotor bar broken fault, turn-to-turn short circuit fault and air gap eccentricity fault, at different speeds and working conditions of the high-speed train as original signals; Step 1.2, the original signal is composed of several segments, and the time domain features are extracted from each original signal segment according to the empirical formula; the time domain features include maximum value, minimum value, peak value, peak-to-peak value, average value, absolute average value, root mean square value, root square amplitude, standard deviation, pulse factor, peak factor, and margin factor; Step 1.3, perform fast Fourier transform on each original signal segment to obtain a frequency domain signal, and extract frequency domain features from the frequency domain signal; the frequency domain features include spectrum energy mean, spectrum energy variance, center of gravity frequency, frequency standard deviation, mean square frequency, and spectrum kurtosis; Step 1.4, integrating the time domain features and frequency domain features extracted from the same original signal segment to form feature samples, determining sample labels according to the state type, and synthesizing all feature samples to obtain a feature data set; Step 1.5: Determine the source condition data set and the target condition data set according to the migration conditions between each condition; define The source condition data set of samples is ,Include The target operating condition data set of samples is ; Each source condition data , , is the number of source condition samples; For the source condition samples; the number of features of the source condition samples is the same as the number of features of the target condition samples, both defined as ; For the The first sample of source conditions Features Each target condition data , , is the number of target working condition samples; For the target working condition samples; For the The first target working condition sample Features.

[0007] Furthermore, the specific process of step 2 is as follows: Step 2.1: Calculate the marginal distribution difference between the source working condition data and the target working condition data Difference from conditional distribution , the formula is as follows: ; ; in, and are the labels of the source and target condition samples respectively; For the Source condition samples; For the target working condition samples; represents a Hilbert space; is a mapping function that maps the original sample to the regenerated Hilbert space; is the category of the fault sample; is the number of fault sample categories; Indicates that it contains indivual Source operating condition data set of class fault samples; Indicates that it contains indivual Target operating condition dataset of fault-like samples; Step 2.2: According to and Determine the balance factor , calculate the distribution distance between the source working condition data and the target working condition data: ; in, for and The distribution distance of Step 2.3: Use the kernel function method to transform the problem of minimizing the distribution distance into an optimization problem of minimizing the trace of the transformation matrix. The specific formula is as follows: ; in, Represents the transformation matrix traces; yes The transpose of Indicated by and The input matrix composed of express Transpose of a matrix; is the regularization parameter; express The square of the norm; and is a matrix of different parameters determined by the number of samples; is the identity matrix; is the center matrix, is a 1 matrix, that is, a matrix with all elements equal to 1; Solution Need to construct Lagrangian function ,as follows: ; in, is the Lagrange multiplier; make , you can find the transformation matrix ; Step 2.4: Based on the conversion matrix, the source working condition data and the target working condition data after distribution alignment are obtained: ; ; in, is the source working condition data after distribution alignment; It is the target working condition data after distribution alignment; is the number of features after distribution alignment; All the source working condition data after distribution alignment constitute the source working condition data set after distribution alignment, and all the target working condition data after distribution alignment constitute the target working condition data set after distribution alignment.

[0008] Furthermore, the specific process of step 3 is as follows: Step 3.1: Source condition data after distribution alignment Based on this, the source condition sample subset is constructed by using the Bootstrap resampling method; the specific process of the Bootstrap resampling method is as follows: go through Sampling with replacement, we get a The source condition sample subset consists of samples ; The samples that have never been selected in the sampling process constitute the source condition out-of-bag data set; Step 3.2: After distribution alignment From the feature set consisting of features, we extract Features form feature subsets; Step 3.3, complete the growth of the decision tree on the subset pairs consisting of the source working condition sample subset and the feature subset, and gradually complete the node splitting from the root node until all child nodes contain only one type of sample; the split feature of each non-leaf node during the node splitting process and split point The choice of is determined by the minimum Gini index, as follows: ; in, is the Gini index; , is the set of source operating condition data in the left child node and the right child node; , They are , The number of samples; is the total number of samples at the parent leaf node; On the data subset pair consisting of the source condition sample subset and the feature subset, the decision tree is grown in the same manner as in step 3.3; Step 3.4: Input the out-of-bag data of the source working condition into the constructed initial random forest fault diagnosis model, obtain the output result of the random forest, and compare it with the true label of the out-of-bag data of the source working condition to obtain the out-of-bag estimation result of the source working condition; the out-of-bag sample of the source working condition enters from the root node, and its characteristic value is compared with the node splitting threshold at the node and divided into child nodes. In this way, the sample finally falls into the leaf node, and the fault type of the sample is determined according to the fault type corresponding to the leaf node; each decision tree will classify the sample, and by voting on the fault classification results of all decision trees, the fault category with the most votes is selected as the final fault classification result of the random forest; specifically: ; in, This is the fault classification result of random forest; is the input source condition out-of-bag data; For the The classification results of a decision tree; The number of decision trees that satisfy the expression in brackets; is the total number of decision trees; Step 3.5: Determine the optimal number of decision trees of the random forest based on the out-of-bag estimation results of the source operating condition. The process of finding the optimal number of decision trees is essentially equivalent to the optimization process of the parameters of the initial random forest fault diagnosis model. The model is tested by using the out-of-bag data of the source operating condition. The model classification accuracy will gradually increase and converge with the increase in the number of decision trees. When the accuracy converges, the minimum number of decision trees is the optimal number of decision trees, and the initial random forest fault diagnosis model is obtained.

[0009] Furthermore, the specific process of step 4 is as follows: Step 4.1: Target operating condition data after distribution alignment The Bootstrap method is used to generate a target working condition sample subset with the same number of decision trees as the initial random forest model. , The samples that have never been selected in the sampling process constitute the target working condition out-of-bag data set; Step 4.2: Maintain the initial random forest fault diagnosis model structure, and use the target working condition sample subset after distribution alignment to complete the update of the decision tree node splitting threshold. The new node splitting threshold is determined by the divergence gain and the Gini index. The specific formula is as follows: ; ; in, is the splitting threshold of the decision tree node in the initial random forest fault diagnosis model; is the divergence gain; The splitting threshold of the decision tree node in the random forest fault diagnosis model after the node threshold is updated; is the distribution difference function, i.e., the square of the maximum mean difference; and are the sets of target operating condition data in the left child node and the right child node respectively; and are the sets of source operating condition data in the left child node and the right child node respectively; for Possible values ​​of the node splitting threshold within the neighborhood; is an arbitrarily small positive number; The distribution difference between the source condition data and the target condition data in the child node is measured by the following formula: ; ; in, and are the number of target working condition samples in the left child node and the right child node respectively; Step 4.3: Starting from the root node, gradually update the node splitting thresholds of all non-leaf nodes according to the rules of step 4.2; Step 4.4: using the target operating condition sample subset formed in step 4.1, update the node splitting thresholds of all decision trees according to the updating method of the node splitting thresholds from step 4.2 to step 4.3; Step 4.5: For the parent nodes with empty data in the child nodes, the original splitting characteristics and node splitting threshold are maintained, and the decision tree is not pruned; if the target working condition sample categories in the leaf nodes are inconsistent, the node splitting is continued according to the principle of minimizing the Gini index until the decision tree stops growing. After the update is completed, the migration random forest fault diagnosis model is obtained; Step 4.6: Use the target operating condition out-of-bag data in step 4.1 to complete the test of building a migration random forest fault diagnosis model.

[0010] Furthermore, in step 5, the conversion formula is: ; in, is a sample of the current working condition; is the converted sample.

[0011] The beneficial technical effects brought by the present invention are as follows: the present invention designs a nonlinear feature transformation method to complete the distribution alignment of source working condition data and target working condition data, thereby alleviating the distribution difference between the two; constructs an initial random forest fault diagnosis model on the transformed source working condition data; proposes an update method for the node splitting threshold of a comprehensive divergence gain and a Gini index to obtain a migration random forest fault diagnosis model; fully utilizes the prior knowledge in the source working condition data, overcomes the distribution difference between the source working condition and the target working condition, and improves the fault diagnosis accuracy of the migration random forest fault diagnosis model in the target working condition. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is a flow chart of the high-speed train traction motor fault diagnosis method based on migration random forest of the present invention.

[0013] Figure 2 This is a model parameter update diagram of the present invention.

[0014] Figure 3 This is a performance comparison chart of different algorithms in the experiment of the present invention. DETAILED DESCRIPTION

[0015] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments: like Figure 1 As shown, the present invention proposes a high-speed train traction motor fault diagnosis method based on migration random forest, which performs fault diagnosis on the traction motor, specifically comprising the following steps: Step 1: Collect the original signals of the high-speed train traction motor under multiple working conditions, extract the time domain features and frequency domain features from the original signals and integrate them to form a feature data set. The specific process is as follows: Step 1.1, based on the high-speed rail traction control system simulation platform jointly developed by Central South University and Zhuzhou Electric Locomotive Research Institute, a multi-sensor fusion mode is adopted to collect sensor signals of the traction motor in four states of normal, broken rotor bar fault, inter-turn short circuit fault and air gap eccentricity fault under three working conditions of high-speed train speeds of 20 km / h, 169 km / h and 280 km / h as original signals. In the embodiment of the present invention, the sensor signals include four types: stator three-phase current, rectifier AC side current, DC voltage and speed, so the original signal is a 7-dimensional signal; Step 1.2, the original signal is composed of several segments, and the time domain features are extracted from each original signal segment according to the empirical formula, including maximum value, minimum value, peak value, peak-to-peak value, average value, absolute average value, root mean square value, root square amplitude, standard deviation, pulse factor, peak factor, and margin factor; Step 1.3, perform fast Fourier transform on each original signal segment to obtain frequency domain signal, and extract frequency domain features from the frequency domain signal; frequency domain features include spectrum energy mean, spectrum energy variance, center of gravity frequency, frequency standard deviation, mean square frequency, spectrum kurtosis; the characteristic parameters and expressions of all time domain features and frequency domain features are shown in Table 1: Table 1 List of time domain features and frequency domain features .

[0016] in, is the maximum value; is the minimum value; is the peak value; is the peak-to-peak value; is the average value; is the absolute average value; is the root mean square; is the square root amplitude; is the standard deviation; is the pulse factor; is the peak factor; is the margin factor; is the mean spectral energy; is the spectrum energy variance; is the center of gravity frequency; is the frequency standard deviation; is the root mean square frequency; is the spectrum kurtosis; is a time domain signal of a certain current; is the number of the sampling point, ; is the total number of sampling points; is the frequency domain signal; is the number of the spectral line, ; is the total number of spectral lines; It is The frequency of a spectral line.

[0017] Step 1.4: Integrate the time domain features and frequency domain features extracted from the same original signal segment to form feature samples, determine the sample labels according to the state type, and obtain the feature data set by integrating all feature samples.

[0018] Step 1.5: Determine the source working condition data set and the target working condition data set according to the migration conditions between the three working conditions; in the embodiment of the present invention, the six migration tasks determined are shown in Table 2, and the definition includes Source condition data set of samples and contains Target operating condition dataset of samples . Each source condition data , , is the number of source condition samples; For the source condition samples; the number of features of the source condition samples is the same as the number of features of the target condition samples, both defined as ; For the The first sample of source conditions Features Each target condition data , , is the number of target working condition samples; For the target working condition samples; For the The first target working condition sample Features.

[0019] Table 2 Migration task list .

[0020] in, , , , , , for six different migration tasks; e.g. Indicates the migration task from the source condition V-20 km / h to the target condition B-160 km / h.

[0021] Step 2: Design a data nonlinear transformation method based on balanced distribution adaptation to align the distribution of source working condition data and target working condition data, and eliminate the distribution difference caused by working condition changes; the specific process of the data nonlinear transformation method based on balanced distribution adaptation is as follows: Step 2.1: Use the maximum mean discrepancy (MMD) as the metric for distribution difference to calculate the marginal distribution difference between the source and target operating data. Difference from conditional distribution , the formula is as follows: ; ; in, and are the labels of the source and target condition samples respectively; For the Source condition samples; For the target working condition samples; represents a Hilbert space; is a mapping function that maps the original sample to the regenerated Hilbert space; is the category of the fault sample; is the number of fault sample categories; Indicates that it contains indivual Source operating condition data set of class fault samples; Indicates that it contains indivual Target operating condition dataset of fault-like samples; Step 2.2: According to and Determine the balance factor , calculate the distribution distance between the source working condition data and the target working condition data: ; in, for and The distribution distance of Step 2.3: The prerequisite for using Hilbert space is that the inner product of the function is complete, and the regenerative Hilbert space is regenerative. and satisfy ,in is the kernel function. After the square of the maximum mean difference (MMD) is expanded, the inner product in the regenerated Hilbert space can be converted into a kernel function, so the final maximum mean difference can be directly calculated by the kernel function. The kernel function method is used to transform the problem of minimizing the distribution distance into the optimization problem of minimizing the trace of the transformation matrix; the formula for the optimization problem of minimizing the trace of the transformation matrix is ​​as follows: ; in, Represents the transformation matrix traces; yes The transpose of Indicated by and The input matrix composed of express Transpose of a matrix; is the balance factor; is the regularization parameter; express The square of the norm; and is a matrix of different parameters determined by the number of samples; is the identity matrix; is the center matrix, is a 1 matrix, that is, a matrix with all elements equal to 1; Solution Need to construct Lagrangian function ,as follows: ; in, is the Lagrange multiplier; make , you can find the transformation matrix ; Step 2.4: Based on the conversion matrix, the source working condition data and the target working condition data after distribution alignment are obtained: ; ; in, is the source working condition data after distribution alignment; It is the target working condition data after distribution alignment; is the number of features after distribution alignment; At this time, the source working condition data after distribution alignment Target operating condition data after alignment with distribution Minimal distribution differences; All the source working condition data after distribution alignment constitute the source working condition data set after distribution alignment, and all the target working condition data after distribution alignment constitute the target working condition data set after distribution alignment.

[0022] Step 3: Use the source operating condition data set after distribution alignment to build an initial random forest fault diagnosis model. The specific process is as follows: Step 3.1: Source condition data after distribution alignment Based on this, the source condition sample subset is constructed by using the Bootstrap resampling method; the specific process of the Bootstrap resampling method is as follows: go through Sampling with replacement, we get a The source condition sample subset consists of samples ; The samples that have never been drawn in the sampling process are called out-of-bag data, and the set composed of these out-of-bag data is called the out-of-bag data set; therefore, The samples that have never been selected in the sampling process constitute the source condition out-of-bag data set; Step 3.2: After distribution alignment From the feature set consisting of features, we extract Features form feature subsets; Step 3.3, complete the growth of the decision tree on the subset pairs consisting of the source working condition sample subset and the feature subset, and gradually complete the node splitting from the root node until all child nodes contain only one type of sample; the split feature of each non-leaf node during the node splitting process and split point The choice of is determined by the minimum Gini index, as follows: ; in, is the Gini index; , is the set of source operating condition data in the left child node and the right child node; , They are , The number of samples; is the total number of samples at the parent leaf node.

[0023] On the data subset pair consisting of the source condition sample subset and the feature subset, the decision tree is grown according to the above decision tree growth method; Step 3.4: Input the out-of-bag data of the source working condition into the constructed initial random forest fault diagnosis model, obtain the output result of the random forest, and compare it with the true label of the out-of-bag data of the source working condition to obtain the out-of-bag estimation result of the source working condition; the out-of-bag sample of the source working condition enters from the root node, and its characteristic value is compared with the node splitting threshold at the node and divided into child nodes. In this way, the sample finally falls into the leaf node, and the fault type of the sample is determined according to the fault type corresponding to the leaf node; each decision tree will classify the sample, and by voting on the fault classification results of all decision trees, the fault category with the most votes is selected as the final fault classification result of the random forest; specifically: ; in, This is the fault classification result of random forest; is the input source condition out-of-bag data; For the The classification results of a decision tree; Indicates the random forest The classification result of the decision tree is ; The number of decision trees that satisfy the expression in brackets; is the total number of decision trees.

[0024] Step 3.5: Determine the optimal number of decision trees of the random forest based on the out-of-bag estimation results of the source operating condition. The process of finding the optimal number of decision trees is essentially equivalent to the optimization process of the parameters of the initial random forest fault diagnosis model. The model is tested by using the out-of-bag data of the source operating condition. The model classification accuracy will gradually increase and converge with the increase in the number of decision trees. When the accuracy converges, the minimum number of decision trees is the optimal number of decision trees, and the initial random forest fault diagnosis model is obtained.

[0025] Step 4: Maintain the initial random forest fault diagnosis model structure, use the target working condition data set after distribution alignment to update the model parameters, and build a migration random forest fault diagnosis model, such as Figure 2 As shown, the specific process is as follows: Step 4.1: Target operating condition data after distribution alignment The Bootstrap method is used to generate a target working condition sample subset with the same number of decision trees as the initial random forest model. , The samples that have never been selected in the sampling process constitute the target working condition out-of-bag data set; Step 4.2: Maintain the initial random forest fault diagnosis model structure, and use the target working condition sample subset after distribution alignment to complete the update of the decision tree node splitting threshold. The new node splitting threshold is determined by the divergence gain and the Gini index. The specific formula is as follows: ; ; in, is the splitting threshold of the decision tree node in the initial random forest fault diagnosis model; is the divergence gain; The splitting threshold of the decision tree node in the random forest fault diagnosis model after the node threshold is updated; is the distribution difference function, i.e., the square of the maximum mean difference; and are the sets of target operating condition data in the left child node and the right child node respectively; and are the sets of source operating condition data in the left child node and the right child node respectively; for Possible values ​​of the node splitting threshold within the neighborhood; is an arbitrarily small positive number; The distribution difference between the source condition data and the target condition data in the child node is measured by the following formula: ; ; in, is a mapping function that maps the original sample to the regenerated Hilbert space. and are the number of target working condition samples in the left child node and the right child node respectively; Step 4.3: Starting from the root node, gradually update the node splitting thresholds of all non-leaf nodes according to the rules of step 4.2, from Updated to , and Respectively represent the node splitting threshold before and after updating Decision Tree Layer The node splitting threshold of non-leaf nodes; Figure 2 As shown, from Updated to , and other similarities are achieved, thereby achieving the migration random forest from random forest.

[0026] Step 4.4: using the target working condition sample subset formed in step 4.1, update the node splitting thresholds of all decision trees according to the node splitting threshold updating method described above; Step 4.5: For the parent nodes with empty data in the child nodes, the original splitting characteristics and node splitting threshold are maintained, and the decision tree is not pruned; if the target working condition sample categories in the leaf nodes are inconsistent, the node splitting is continued according to the principle of minimizing the Gini index until the decision tree stops growing. After the update is completed, the migration random forest fault diagnosis model is obtained; Step 4.6: Use the target operating condition out-of-bag data in step 4.1 to complete the test of building a migration random forest fault diagnosis model.

[0027] Step 5: Collect samples of the current working condition, use the conversion matrix to convert the samples of the current working condition, input the converted samples into the migration random forest fault diagnosis model, obtain the fault category of the samples, and realize online fault diagnosis. The conversion formula is: ; in, is a sample of the current working condition; is the converted sample.

[0028] In order to prove the feasibility and superiority of the method of the present invention, the following experiment was carried out.

[0029] First, the migration random forest algorithm (i.e., the method of fault diagnosis using the migration random forest fault diagnosis model of the present invention) is compared with the traditional random algorithm in terms of fault diagnosis accuracy. The results are as follows: Figure 3 As shown, Indicated in The traditional random forest model trained under working conditions is used The results of each condition, Indicated in The transfer random forest model trained under working conditions is used for The results of each condition, , , For different source conditions. For example, Indicated in The traditional random forest model trained under working conditions is used The results of each working condition; Indicated in The transfer random forest model trained under working conditions is used for The results of the working condition. Figure 3 The comparison results show that the fault diagnosis accuracy of the migration random forest algorithm under multiple migration tasks is significantly improved compared with the traditional random forest algorithm. Based on the above comparison results, it is proved that the migration random forest algorithm can effectively solve the fault diagnosis problem of high-speed train traction motor under multiple working conditions.

[0030] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A high-speed train traction motor fault diagnosis method based on migration random forest, characterized in that: The fault diagnosis of the traction motor includes the following steps: Step 1: collect original signals of the traction motor of the high-speed train under multiple working conditions in multiple states, extract time domain features and frequency domain features from the original signals and integrate them to form a feature data set; Step 2: Design a data nonlinear transformation method based on balanced distribution adaptation to align the distribution of source working condition data and target working condition data, and eliminate the distribution difference caused by working condition changes; Step 3: Use the source operating condition data set after distribution alignment to build an initial random forest fault diagnosis model; Step 4: Maintain the initial random forest fault diagnosis model structure, use the target working condition data set after distribution alignment to update the model parameters, and build a migration random forest fault diagnosis model; Step 5: Collect samples of the current working condition, use the conversion matrix to convert the samples of the current working condition, input the converted samples into the migration random forest fault diagnosis model, obtain the fault category of the samples, and realize online fault diagnosis.

2. The high-speed train traction motor fault diagnosis method based on migration random forest according to claim 1 is characterized in that: The specific process of step 1 is as follows: Step 1.1, based on the existing high-speed rail traction control system simulation platform, a multi-sensor fusion mode is adopted to collect sensor signals of the traction motor under four types of states, namely, normal state, rotor bar broken fault, turn-to-turn short circuit fault and air gap eccentricity fault, at different speeds and working conditions of the high-speed train as original signals; Step 1.2, the original signal is composed of several segments, and the time domain features are extracted from each original signal segment according to the empirical formula; Time domain characteristics include maximum value, minimum value, peak value, peak-to-peak value, average value, absolute average value, RMS value, RMS amplitude, standard deviation, pulse factor, peak factor, and margin factor; Step 1.3, perform fast Fourier transform on each original signal segment to obtain a frequency domain signal, and extract frequency domain features from the frequency domain signal; the frequency domain features include spectrum energy mean, spectrum energy variance, center of gravity frequency, frequency standard deviation, mean square frequency, and spectrum kurtosis; Step 1.4, integrating the time domain features and frequency domain features extracted from the same original signal segment to form feature samples, determining sample labels according to the state type, and synthesizing all feature samples to obtain a feature data set; Step 1.5: Determine the source condition data set and the target condition data set according to the migration conditions between each condition; define The source condition data set of samples is ,Include The target operating condition data set of samples is ; Each source condition data , , is the number of source condition samples; For the source condition samples; the number of features of the source condition samples is the same as the number of features of the target condition samples, both defined as ; For the The first sample of source conditions Features Each target condition data , , is the number of target working condition samples; For the target working condition samples; For the The first target working condition sample Features.

3. The high-speed train traction motor fault diagnosis method based on migration random forest according to claim 2 is characterized in that: The specific process of step 2 is: Step 2.1: Calculate the marginal distribution difference between the source working condition data and the target working condition data Difference from conditional distribution , the formula is as follows: ； ； in, and are the labels of the source and target condition samples respectively; For the Source condition samples; For the target working condition samples; represents a Hilbert space; is a mapping function that maps the original sample to the regenerated Hilbert space; is the category of the fault sample; is the number of fault sample categories; Indicates that it contains indivual Source operating condition data set of class fault samples; Indicates that it contains indivual Target operating condition dataset of fault-like samples; Step 2.2: According to and Determine the balance factor , calculate the distribution distance between the source working condition data and the target working condition data: ； in, for and The distribution distance of Step 2.3: Use the kernel function method to transform the problem of minimizing the distribution distance into an optimization problem of minimizing the trace of the transformation matrix. The specific formula is as follows: ； in, Represents the transformation matrix traces; yes The transpose of Indicated by and The input matrix composed of express Transpose of a matrix; is the regularization parameter; express The square of the norm; and is a matrix of different parameters determined by the number of samples; is the identity matrix; is the center matrix, is a 1 matrix, that is, a matrix with all elements equal to 1; Solution Need to construct Lagrangian function ,as follows: ； in, is the Lagrange multiplier; make , you can find the transformation matrix ; Step 2.4: Based on the conversion matrix, the source working condition data and the target working condition data after distribution alignment are obtained: ； ； in, is the source working condition data after distribution alignment; It is the target working condition data after distribution alignment; is the number of features after distribution alignment; All the source working condition data after distribution alignment constitute the source working condition data set after distribution alignment, and all the target working condition data after distribution alignment constitute the target working condition data set after distribution alignment.

4. The high-speed train traction motor fault diagnosis method based on migration random forest according to claim 3 is characterized in that: The specific process of step 3 is as follows: Step 3.1: Source condition data after distribution alignment Based on this, the source condition sample subset is constructed by using the Bootstrap resampling method; the specific process of the Bootstrap resampling method is as follows: go through Sampling with replacement, we get a The source condition sample subset consists of samples ; The samples that have never been selected in the sampling process constitute the source condition out-of-bag data set; Step 3.2: After distribution alignment From the feature set consisting of features, we extract Features form feature subsets; Step 3.3, complete the growth of the decision tree on the subset pairs consisting of the source working condition sample subset and the feature subset, and gradually complete the node splitting from the root node until all child nodes contain only one type of sample; the split feature of each non-leaf node during the node splitting process and split point The choice of is determined by the minimum Gini index, as follows: ； in, is the Gini index; , is the set of source operating condition data in the left child node and the right child node; , They are , The number of samples; is the total number of samples at the parent leaf node; On the data subset pair consisting of the source condition sample subset and the feature subset, the decision tree is grown in the same manner as in step 3.3; Step 3.4: Input the out-of-bag data of the source working condition into the constructed initial random forest fault diagnosis model, obtain the output result of the random forest, and compare it with the true label of the out-of-bag data of the source working condition to obtain the out-of-bag estimation result of the source working condition; the out-of-bag sample of the source working condition enters from the root node, and its characteristic value is compared with the node splitting threshold at the node and divided into child nodes. In this way, the sample finally falls into the leaf node, and the fault type of the sample is determined according to the fault type corresponding to the leaf node; each decision tree will classify the sample, and by voting on the fault classification results of all decision trees, the fault category with the most votes is selected as the final fault classification result of the random forest; specifically: ； in, This is the fault classification result of random forest; is the input source condition out-of-bag data; For the The classification results of a decision tree; The number of decision trees that satisfy the expression in brackets; is the total number of decision trees; Step 3.5: Determine the optimal number of decision trees of the random forest based on the out-of-bag estimation results of the source operating condition. The process of finding the optimal number of decision trees is essentially equivalent to the optimization process of the parameters of the initial random forest fault diagnosis model. The model is tested by using the out-of-bag data of the source operating condition. The model classification accuracy will gradually increase and converge with the increase in the number of decision trees. When the accuracy converges, the minimum number of decision trees is the optimal number of decision trees, and the initial random forest fault diagnosis model is obtained.

5. The high-speed train traction motor fault diagnosis method based on migration random forest according to claim 4 is characterized in that: The specific process of step 4 is as follows: Step 4.1: Target operating condition data after distribution alignment The Bootstrap method is used to generate a target working condition sample subset with the same number of decision trees as the initial random forest model. , The samples that have never been selected in the sampling process constitute the target working condition out-of-bag data set; Step 4.2: Maintain the initial random forest fault diagnosis model structure, and use the target working condition sample subset after distribution alignment to complete the update of the decision tree node splitting threshold. The new node splitting threshold is determined by the divergence gain and the Gini index. The specific formula is as follows: ； ； in, is the splitting threshold of the decision tree node in the initial random forest fault diagnosis model; is the divergence gain; The splitting threshold of the decision tree node in the random forest fault diagnosis model after the node threshold is updated; is the distribution difference function, i.e., the square of the maximum mean difference; and are the sets of target operating condition data in the left child node and the right child node respectively; and are the sets of source operating condition data in the left child node and the right child node respectively; for Possible values ​​of the node splitting threshold within the neighborhood; is an arbitrarily small positive number; The distribution difference between the source condition data and the target condition data in the child node is measured by the following formula: ； ； in, and are the number of target working condition samples in the left child node and the right child node respectively; Step 4.3: Starting from the root node, gradually update the node splitting thresholds of all non-leaf nodes according to the rules of step 4.2; Step 4.4: using the target operating condition sample subset formed in step 4.1, update the node splitting thresholds of all decision trees according to the updating method of the node splitting thresholds from step 4.2 to step 4.3; Step 4.5: For the parent nodes with empty data in the child nodes, the original splitting characteristics and node splitting threshold are maintained, and the decision tree is not pruned; if the target working condition sample categories in the leaf nodes are inconsistent, the node splitting is continued according to the principle of minimizing the Gini index until the decision tree stops growing. After the update is completed, the migration random forest fault diagnosis model is obtained; Step 4.6: Use the target operating condition out-of-bag data in step 4.1 to complete the test of building a migration random forest fault diagnosis model.

6. The high-speed train traction motor fault diagnosis method based on migration random forest according to claim 5 is characterized in that: In step 5, the conversion formula is: ； in, is a sample of the current working condition; is the converted sample.

Images