Traction motor fault diagnosis method based on RPSO-VMD-CatBoost

Abstract

The invention belongs to the field of rail transit traction motor electric variable measurement and fault diagnosis, and particularly relates to a traction motor fault diagnosis method based on RPSO-VMD-CatBoost. Comprising the steps of collecting a traction motor current signal, adaptively optimizing key parameters # imgabs0 # and alpha in a VMD algorithm by using an improved particle swarm optimization algorithm, optimizing an IMF component, extracting a fault feature frequency, extracting time domain features for an original electric signal, optimizing hyper-parameters of a CatBoost classifier, performing model training and the like. According to the method, the RPSO is adopted to optimize the VMD hyper-parameter, the influence of the parameter on feature extraction is reduced, the effective frequency domain feature of the signal is extracted at the same time, the accuracy of fault diagnosis is improved, and a theoretical support with higher credibility can be provided for fault diagnosis of the railway vehicle traction motor in the actual operation process.

Description

Technical Field

[0001] The present invention belongs to the field of electric variable measurement and fault diagnosis of rail transit traction motors, and particularly relates to a traction motor fault diagnosis method based on RPSO-VMD-CatBoost. Background Art

[0002] With the continuous advancements in rail vehicle design and related technologies, their structure and operational complexity are becoming increasingly complex, placing increasing demands on safety, reliability, and ride comfort during operation. As the core power source of rail vehicles, the traction drive system, and the traction motor itself, is the core assembly of the system. Ensuring its stable and reliable operation is crucial. However, due to the structural characteristics and functional requirements of traction motors, their operating environments often involve the complex coupling of multiple physical fields. Long-term operation can lead to fatigue damage and health degradation in the traction motors, resulting in increased failure rates and, in severe cases, even threats to rail vehicle operational safety. Furthermore, traction motor electrical signals exhibit nonstationary, strong noise, multi-component coupling, and nonlinear characteristics, posing significant challenges to the effective extraction of fault signatures. Existing traction motor fault diagnosis methods also suffer from the following technical challenges: Directly performing a Fourier transform on the raw measured electrical signals to generate a spectrum makes it difficult to extract fault signature frequencies; the hyperparameter decomposition scale and penalty factor in the VMD algorithm are difficult to determine manually; and the particle swarm optimization algorithm is prone to falling into local optima and lacks local search capabilities. Therefore, research on traction motor fault diagnosis methods holds significant theoretical and engineering significance. Summary of the Invention

[0003] In order to solve the above technical problems, the present invention provides a traction motor fault diagnosis method based on RPSO-VMD-CatBoost, comprising the following steps:

[0004] Step S1: Install and calibrate high-precision current sensors at the input and output terminals of the traction motor to collect the three-phase stator winding current (I a , I b , I c ) signal is used to reflect the electromagnetic operating state of the motor; the collected original current signals are stored in the database respectively, and their corresponding operating states are marked to provide a basis for feature extraction;

[0005] Furthermore, the sampling rate of the traction motor current signal is set to 100 kHz, and the sampling time is not less than 10 s to obtain complete steady-state and transient signals.

[0006] Step S2: Adaptively optimize the key parameters in the variational mode decomposition (VMD) algorithm using the improved particle swarm optimization (RPSO) algorithm that integrates the sparrow search mechanism S and Cauchy mutation. and α, is the decomposition scale, α is the penalty factor, and includes the following steps:

[0007] Step S201: setting parameters of the improved particle swarm optimization algorithm, including population size, number of iterations, inertia weight, individual learning factor, social learning factor, and setting upper and lower bounds of optimization for key parameters in the variational mode decomposition algorithm;

[0008] Step S202: Initialize the population and calculate the fitness values ​​of key parameters in the variational mode decomposition algorithm;

[0009] Furthermore, the average information entropy is selected as the fitness function of the key parameters in the variational mode decomposition algorithm. The specific expression is as follows:

[0010]

[0011] Where: H(x) ave is the average information entropy of all IMF components, p(x i ) is the The normalized energy probability of the IMF components, n is the total number of IMF components obtained by decomposition;

[0012] Step S203: introducing the sparrow search mechanism and Cauchy mutation to improve the particle swarm optimization algorithm with a certain probability, and updating the speed and position of the particles according to the basic principles of the improved particle swarm optimization algorithm;

[0013] Furthermore, the specific expression of the sparrow search mechanism is as follows:

[0014]

[0015] Where: v ij (t) is the The particle in dimensional speed, c1 is the individual cognitive coefficient, rand(-1,1,size=d) represents a random vector of dimension D, ⊙ represents the element multiplication of the vector, p ij (t) is the The particle in The historical best position in dimension, x ij (t) is the The particle in Current position in dimension;

[0016] The specific expression of the Cauchy variation is as follows:

[0017]

[0018] Where: c is a dimension and v ij(t) The same vector, where each element is independently drawn from a standard Cauchy distribution.

[0019] Step S204: Determine whether the key parameters in the variational mode decomposition algorithm exceed the set boundaries; if so, return to step S203 and repeat the operation; if not, proceed to step S205;

[0020] Step S205: updating key parameters and fitness values ​​in the variational mode decomposition algorithm;

[0021] Step S206, repeat steps S202-S205 until the termination criterion is met and the global optimal solution is obtained. ,α].

[0022] Step S3: Feeding back the key parameters obtained after the optimization into the variational mode decomposition algorithm, and using the algorithm to perform adaptive signal decomposition on the electrical signal collected in step S1, wherein the variational mode decomposition mainly adopts a non-recursive method to construct and solve a finite variation problem, decomposing the complex initial signal into several IMF (Intrinsic Mode Function) component sequences from high to low frequency, and obtaining the corresponding IMF components;

[0023] Furthermore, step S3 specifically includes the following steps:

[0024] Define a finite-bandwidth modal function with strict constraints:

[0025]

[0026] Where: u k (t) is the decomposed modal components; A k (t) is the instantaneous amplitude; Ф k (t) represents the phase function;

[0027] Perform Hilbert transformation on the modal function to obtain the corresponding analytical signal:

[0028]

[0029] Where: δ(t) is the Dirac shock function; is an imaginary unit;

[0030] Utilization Index Adjust the spectrum of the modal function to focus on the corresponding baseband:

[0031]

[0032] Use Gaussian smoothing, that is, calculate L2 The square root of the norm gradient is used to demodulate the signal, so we can get:

[0033]

[0034] in, are the modal functions, is the center frequency of each mode, ∂ t For time The differential operator, f is the original input signal.

[0035] Step S4: determine the optimal IMF component according to the minimum envelope entropy principle, perform envelope spectrum analysis on the optimal IMF component, obtain the envelope spectrum of the fault current signal, and extract the actual fault characteristic frequency on the optimal IMF component.

[0036] Furthermore, the specific expression of the optimal IMF component is determined according to the minimum envelope entropy principle as follows:

[0037]

[0038] Where, is the envelope entropy, p j is the normalized form of a(j), and a(j) is the envelope signal obtained after Hilbert demodulation of signal x(j).

[0039] Step S5: extracting time domain features from the original electrical signal, and extracting frequency domain features and envelope features from the optimal IMF component obtained in step S4 to form a fault feature vector set; wherein the time domain feature values ​​include one or more feature quantities selected from the group consisting of maximum value, maximum absolute value, minimum value, mean value, peak-to-peak value, absolute mean value, root mean square value, root square amplitude, standard deviation, kurtosis, skewness, margin index, waveform index, pulse index, and peak index; the frequency domain feature values ​​include one or more feature quantities selected from the group consisting of center of gravity frequency, root mean square frequency, average frequency, and frequency variance; and the envelope feature values ​​include envelope entropy;

[0040] Step S6: Optimize the hyperparameters of the CatBoost classifier using a grid optimization method to obtain an optimized CatBoost fault diagnosis classification model. The specific steps are as follows:

[0041] Step S601: Set the value range of the hyperparameter iterations, which represents the number of iterations during model training.

[0042] Step S602: Set the value range of the hyperparameter learning_rate, which represents the contribution of each tree to the final prediction;

[0043] Step S603: Set the value range of the hyperparameter depth, which represents the maximum depth of the decision tree;

[0044] Step S604: Setting the value range of the hyperparameter l2_leaf_reg to control the complexity of the leaf nodes;

[0045] Step S605: Obtain the best hyperparameter combination through gridding hyperparameter optimization.

[0046] Step S7: input the fault feature vector set into the optimized CatBoost fault diagnosis classification model for fault diagnosis to obtain a diagnosis result; wherein the fault feature vector set is divided into a training set and a test set according to an 8:2 ratio, and the training set is input into the optimized CatBoost fault diagnosis classification model for training, and then the test set is input into the trained model to obtain a fault classification diagnosis result, and the result is output as a confusion matrix diagram.

[0047] Beneficial effects of the present invention:

[0048] This invention achieves remote, non-contact fault diagnosis of traction motors through signal processing of electrical signals, addressing the drawback of requiring a large number of sensors and providing an effective new approach for remote, non-contact diagnosis of traction motors. This invention improves the particle swarm optimization algorithm by integrating the sparrow search mechanism and Cauchy mutation. This improved particle swarm optimization algorithm retains the original social term of the particle swarm optimization algorithm and introduces uniformly distributed random perturbations by simulating the behavior of followers in the sparrow search algorithm approaching the global optimal individual. This enhances global search capabilities and increases the diversity of particles as they move toward the global optimal direction. Furthermore, the introduction of the Cauchy mutation component replaces the particle swarm algorithm's reliance on uniform random number parameters, supplementing the particle swarm optimization algorithm's lack of an explicit mutation mechanism and improving the algorithm's global search capabilities. Using the improved particle swarm optimization algorithm to globally optimize the key parameters of the variational mode decomposition algorithm can better escape local optimal solutions and effectively determine the optimal parameter combination. This invention extracts frequency domain and envelope features from the optimal IMF component, more intuitively and effectively obtaining the frequency domain and envelope features most strongly correlated with the fault, reducing the impact of noise on feature extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0050] Figure 2 These are the time domain waveform diagrams of experimental signals in four states according to the embodiment of the present invention;

[0051] Figure 3 Spectra of three fault experimental signals and normal signals according to the embodiment of the present invention;

[0052] Figure 4 This is a schematic diagram of the improved particle swarm optimization algorithm flow of the present invention;

[0053] Figure 5 This is a time domain diagram of the normal signal and failure signal of the broken rotor bar after VMD decomposition in an embodiment of the present invention;

[0054] Figure 6 This is a spectrum diagram of the normal signal and failure signal of the broken rotor bar after VMD decomposition in an embodiment of the present invention;

[0055] Figure 7 This is a time domain diagram of the normal signal and failure signal VMD decomposition of the stator inter-turn short circuit in an embodiment of the present invention;

[0056] Figure 8 This is a spectrum diagram after VMD decomposition of the normal signal and failure signal of the stator inter-turn short circuit in an embodiment of the present invention;

[0057] Figure 9 This is a time domain diagram of the normal signal and failure signal of the air gap eccentricity after VMD decomposition in an embodiment of the present invention;

[0058] Figure 10 Spectrum diagram of normal signal and failure signal of air gap eccentricity after VMD decomposition in an embodiment of the present invention;

[0059] Figure 11 The optimal component envelope spectrum of the rotor broken bar failure according to the embodiment of the present invention;

[0060] Figure 12 Optimal component envelope spectrum of stator inter-turn short circuit failure in an embodiment of the present invention;

[0061] Figure 13 The optimal component envelope spectrum of the air gap eccentricity failure in the embodiment of the present invention;

[0062] Figure 14 A confusion matrix for fault diagnosis according to an embodiment of the present invention;

[0063] Figure 15 This is a classification and identification diagram for fault diagnosis according to an embodiment of the present invention. DETAILED DESCRIPTION

[0064] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions and beneficial effects of the present invention will be further described below in conjunction with the drawings and specific examples in the embodiments of the present invention. It is obvious that the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without making creative efforts should fall within the scope of protection of the present invention.

[0065] This embodiment provides a traction motor fault diagnosis method based on RPSO-VMD-CatBoost, including the following steps:

[0066] Step S1: Install and calibrate high-precision current sensors (error ≤ ±0.5%) at the input and output terminals of the traction motor to collect the three-phase stator winding current (I a , I b , I c ) signal is used to reflect the electromagnetic operating state of the motor; the collected original current signals are stored in the database respectively, and their corresponding operating states are marked to provide a basis for feature extraction;

[0067] This example uses the traction control system fault injection simulation platform TDCS-FIB V2.0 developed by Central South University as an example. This platform primarily consists of a fault injection unit and a real-time data acquisition and monitoring unit. The fault injection unit consists of a fault injection command interface, a fault injection controller, a signal conditioning module, an input / output interface, a level conversion module, a workload, and a fault library. Commands are transmitted to the fault injection controller via the fault injection command interface. Based on these commands, the fault injection controller controls the signal conditioning module and the level conversion module to inject / simulate faults in various components (physical or virtual) of the traction drive control system. Three fault modes with a severity of 0.2, including air gap eccentricity, broken rotor bars, and stator interturn short circuit, were injected into the traction motor, along with the normal state, to generate data for four fault modes. The sampling frequency was 100 kHz, the sampling time was 10 seconds, and there were 1 million sampling points. The fault injection time was set to 1 second after the start of each experiment. Given its greater sensitivity to localized faults and asymmetric characteristics, the B-phase current was selected for subsequent analysis. To simulate actual operating conditions, this embodiment adds an appropriate amount of noise to the current signal to simulate field interference, thereby verifying the diagnostic effectiveness of the present invention in actual application scenarios. It should be noted that the addition of noise is only used to demonstrate the performance of the present invention and is not considered a necessary step in its implementation. Figure 2 、 Figure 3 is the time domain waveform and spectrum of the experimental signal, Figure 2 The time domain waveforms of various faults are basically similar and cannot be identified by the naked eye. Figure 3 The spectra under different fault and normal conditions are also very close, and the fault frequency components cannot be effectively observed.

[0068] Step S2, using the improved particle swarm optimization (RPSO) algorithm that integrates the sparrow search mechanism S and Cauchy mutation to adaptively optimize the key parameters K and α in the variational mode decomposition (VMD) algorithm, where K is the decomposition scale and α is the penalty factor;

[0069] The particle swarm optimization (PSO) is essentially a random search algorithm. It first randomly generates a swarm of particles within a solution space, the dimensionality of which is determined by the number of variables in the problem being optimized. Each particle is given an initial position and velocity, which are then optimized through an iterative process. During each iteration, the particles adjust their positions and velocities within the solution space by tracking two key "extremes." The algorithm uses these behaviors to find optimal parameters, abstracted into mathematical terms.

[0070] Assume that there is a group of N particles in the D-dimensional target search space, where A particle can be represented as a D-dimensional vector X i :

[0071]

[0072] The "flying" speed V of the i-th particle i It is also a D-dimensional vector, recorded as:

[0073]

[0074] The optimal position reached by the i-th particle during its search process is called the individual extreme value P best , recorded as:

[0075]

[0076] The optimal position found during the entire particle swarm iterative search process is called the global optimal value g best , recorded as:

[0077]

[0078] When these two optimal values ​​are found, the particle updates its speed and position according to the following two formulas:

[0079]

[0080]

[0081] Where: v ij (t+1) is the The particle in The updated speed in the dimension, x ij (t+1) is the The particle in The updated position in the dimension; c1 and c2 are learning factors, also known as acceleration constants; r1 and r2 are uniformly distributed random numbers in the interval [0,1]; j = 1, 2…, D; v ij represents the velocity of the particle, and v ij ∈[-vmax ,v max ], where v max is a user-defined constant used to limit the maximum "flying" speed of the particle; r1 and r2 are random variables between 0 and 1, which introduce random characteristics to the "flying" process of the particle;

[0082] The present invention makes the following improvements to the particle swarm algorithm:

[0083] The sparrow search mechanism is introduced with a certain probability. The specific formula is:

[0084]

[0085] Where c1 is the individual cognitive coefficient; rand(-1,1,size=d) represents a random vector of dimension D (the dimension of the particle), whose element values ​​are between -1 and 1; ⊙ represents the element-wise multiplication of the vector.

[0086] Cauchy mutation is introduced with a certain probability. The specific formula is:

[0087]

[0088] Where, is a dimension with v ij (t) are identical vectors, where each element is independently drawn from a standard Cauchy distribution with location parameter 0 and scale parameter 1.

[0089] The improved particle swarm optimization algorithm (RPSO) is used to adaptively optimize the key parameters in the variational mode decomposition algorithm for four different state experimental signals, and the global optimal solution is obtained; the key parameters include the decomposition scale K and the penalty factor α; Figure 4 As shown, the following steps are included:

[0090] Step S201: setting parameters of the improved particle swarm optimization algorithm, including population size, number of iterations, inertia weight, individual learning factor, social learning factor, and setting upper and lower bounds of optimization for key parameters in the variational mode decomposition algorithm;

[0091] Step S202: Initialize the population and calculate the fitness values ​​of key parameters in the variational mode decomposition algorithm;

[0092] Furthermore, the average information entropy is selected as the fitness function of the key parameters in the variational mode decomposition algorithm. The specific expression is as follows:

[0093]

[0094] Where: H(x) ave is the average information entropy of all IMF components, p(xi ) is the The normalized energy probability of the IMF components, n is the total number of IMF components obtained by decomposition;

[0095] Step S203: updating the speed and position of the particles according to the basic principle of the improved particle swarm optimization algorithm;

[0096] Step S204: Determine whether the key parameters in the variational mode decomposition algorithm exceed the set boundaries; if so, return to step S203 and repeat the operation; if not, proceed to step S205;

[0097] Step S205: updating key parameters and fitness values ​​in the variational mode decomposition algorithm;

[0098] Step S206, repeat steps S202-S205 until the termination criterion is met and the global optimal solution is obtained. ,α].

[0099] In this embodiment, the decomposition scale K is set in the range of (3, 9), the penalty factor α is set in the range of (100, 3000), the particle population size is set to 30, the number of iterations is set to 50, the inertia weight is set to 0.5, the individual learning factor is set to 0.2, and the social learning factor is set to 0.5.

[0100] Step S3: Feeding back the key parameters obtained after the optimization into the variational mode decomposition algorithm, and using the algorithm to perform adaptive signal decomposition on the electrical signal collected in step S1, wherein the variational mode decomposition mainly adopts non-recursive technology to construct and solve a finite variation problem, decomposing the complex initial signal into several IMF (Intrinsic Mode Function) component sequences from high to low frequency, and obtaining the corresponding IMF components; the specific steps include:

[0101] Define a finite-bandwidth modal function with strict constraints:

[0102]

[0103] Where: u k (t) is the kth modal component obtained by decomposition; A k (t) is the instantaneous amplitude; Ф k (t) represents the phase function;

[0104] Perform Hilbert transformation on the modal function to obtain the corresponding analytical signal:

[0105]

[0106] Where: δ(t) is the Dirac shock function; is an imaginary unit;

[0107] Utilization Index Adjust the spectrum of the modal function to focus on the corresponding baseband:

[0108]

[0109] Use Gaussian smoothing, that is, calculate L 2 The square root of the norm gradient is used to demodulate the signal, so we can get:

[0110]

[0111] in, are the modal functions, is the center frequency of each mode, ∂ t For time The differential operator is f, and f is the original input signal. The three fault experimental signals are decomposed by VMD to obtain the time-frequency domain diagram as shown in the figure below: Figure 5-10 shown.

[0112] Step S4: determine the optimal IMF component according to the minimum envelope entropy principle, perform envelope spectrum analysis on the optimal IMF component, obtain the envelope spectrum of the fault current signal, and extract the actual fault characteristic frequency on the optimal IMF component.

[0113] Furthermore, the specific expression of the optimal IMF component is determined according to the minimum envelope entropy principle as follows:

[0114]

[0115] Where, is the envelope entropy, p j is the normalized form of a(j), and a(j) is the envelope signal obtained after Hilbert demodulation of signal x(j).

[0116] The envelope spectrum of the three fault experimental signals is obtained by performing envelope spectrum analysis on the optimal component. Figure 11-13 As shown in Figure 3, the fault characteristic frequency can be effectively extracted on the optimal IMF component.

[0117] Step S5: extracting time domain features from the original electrical signal, and extracting frequency domain features and envelope features from the optimal IMF component obtained in step S4 to form a fault feature vector set; wherein the time domain feature values ​​include maximum value, maximum absolute value, minimum value, mean value, peak-to-peak value, absolute mean value, root mean square value, root square amplitude, standard deviation, kurtosis, skewness, margin index, waveform index, pulse index, peak index and other feature quantities; the frequency domain feature values ​​include feature quantities such as center of gravity frequency, frequency root mean square, average frequency, frequency variance and other feature quantities; the envelope feature values ​​include feature quantities such as envelope entropy; the specific expressions are shown in Table 1 and Table 2;

[0118] Table 1 Time domain feature calculation expression

[0119]

[0120] Note: X = [X(1), X(2), …, X(N)] is a time domain signal sample with a length of N.

[0121] Table 2 Frequency domain feature calculation expression

[0122]

[0123] Note: f k For the The frequency value corresponding to the frequency point; S(k) is the power spectrum density value of the signal at this frequency point; is the total number of frequency points involved in the calculation.

[0124] Step S6: Optimize the hyperparameters of the CatBoost classifier using a grid optimization method to obtain an optimized CatBoost fault diagnosis classification model. The specific steps are as follows:

[0125] Step S601: Set the value range of the hyperparameter iterations. In this example, the hyperparameter iterations in the CatBoost classifier is set to the range [100, 200, 300, 400], which represents the number of iterations during model training, that is, the number of gradient boosts to be constructed.

[0126] Step S602: Set the value range of the hyperparameter learning_rate, which represents the contribution of each tree to the final prediction. In this example, the hyperparameter learning_rate is set to the range [0.05, 0.1, 0.2, 0.3, 0.4].

[0127] Step S603: Set the value range of the hyperparameter depth, which represents the maximum depth of the decision tree. In this example, the hyperparameter depth is set to the range [3, 5, 7].

[0128] Step S604: Set the value range of the hyperparameter l2_leaf_reg. In this example, the hyperparameter l2_leaf_reg is set to the range [3, 5, 7]. This is a regularization parameter used to control the complexity of leaf nodes.

[0129] Step S605: Obtain the best hyperparameter combination through gridding hyperparameter optimization.

[0130] Step S7: input the fault feature vector set into the optimized CatBoost fault diagnosis classification model for fault diagnosis to obtain a diagnosis result; wherein the fault feature vector set is divided into a training set and a test set according to an 8:2 ratio, and the training set is input into the optimized CatBoost fault diagnosis classification model for training, and then the test set is input into the trained model to obtain a fault classification diagnosis result, and the result is output as a confusion matrix diagram.

[0131] In this embodiment, every 2048 sampling points constitutes a sample. To enhance the data, 1000 sampling points are overlapped, so that each pattern corresponds to 848 samples, and the four types of data have a total of 3392 samples. 80% of the failure data samples in each type are randomly selected as the training set, and the remaining 20% ​​are used as the test set.

[0132] Figure 14 The confusion matrix for fault diagnosis is shown; Figure 15 A classification recognition diagram for fault diagnosis is shown.

Claims

1. A traction motor fault diagnosis method based on RPSO-VMD-CatBoost, characterized by: The steps include: Step S1: Install and calibrate high-precision current sensors at the input and output terminals of the traction motor to collect three-phase stator winding current signals under normal and fault conditions during the operation of the traction motor to reflect the electromagnetic operating state of the motor; store the collected raw current signals separately and mark their corresponding operating states; Step S2: Adaptively optimize the key parameters in the variational mode decomposition algorithm using the improved particle swarm optimization algorithm that integrates the sparrow search mechanism and Cauchy mutation and α, is the decomposition scale, α is the penalty factor, and includes the following steps: Step S201: setting parameters of the improved particle swarm optimization algorithm, including population size, number of iterations, inertia weight, individual learning factor, social learning factor, and setting upper and lower bounds of optimization for key parameters in the variational mode decomposition algorithm; Step S202: Initialize the population and calculate the fitness values ​​of key parameters in the variational mode decomposition algorithm; Step S203: introducing the sparrow search mechanism and Cauchy mutation to improve the particle swarm optimization algorithm with a certain probability, and updating the speed and position of the particles according to the basic principles of the improved particle swarm optimization algorithm; Step S204: Determine whether the key parameters in the variational mode decomposition algorithm exceed the set boundaries; if so, return to step S203 and repeat the operation; if not, proceed to step S205; Step S205: updating key parameters and fitness values ​​in the variational mode decomposition algorithm; Step S206, repeat steps S202-S205 until the termination criterion is met and the global optimal solution is obtained. ,α]; Step S3: Feeding back the key parameters obtained after the optimization to the variational mode decomposition algorithm, using the algorithm to perform adaptive signal decomposition on the electrical signal collected in step S1, decomposing the complex initial signal into several IMF component sequences according to the frequency from high to low, and obtaining the corresponding IMF components; Step S4: determining the optimal IMF component according to the minimum envelope entropy principle, performing envelope spectrum analysis on the optimal IMF component to obtain the envelope spectrum of the fault current signal, and extracting the actual fault characteristic frequency on the optimal IMF component; Step S5: extracting time domain features from the original electrical signal, and extracting frequency domain features and envelope features from the optimal IMF component obtained in step S4 to form a fault feature vector set; Step S6: Optimize the hyperparameters of the CatBoost classifier using a grid optimization method to obtain an optimized CatBoost fault diagnosis classification model; Step S7: divide the fault feature vector set into a training set and a test set, and input them into the optimized CatBoost fault diagnosis classification model for training and fault diagnosis.

2. The traction motor fault diagnosis method based on RPSO-VMD-CatBoost according to claim 1 is characterized by: In step S1, the sampling rate of the traction motor current signal is set to 100 kHz, and the sampling time is not less than 10 s to obtain complete steady-state and transient signals.

3. The traction motor fault diagnosis method based on RPSO-VMD-CatBoost according to claim 1 is characterized in that: In step S202, the average information entropy is selected as the fitness function of the key parameters in the variational mode decomposition algorithm, and the expression is as follows: ， Where: H(x) ave is the average information entropy of all IMF components, p(x i ) is the is the normalized energy probability of the IMF components, and n is the total number of IMF components obtained by decomposition.

4. The traction motor fault diagnosis method based on RPSO-VMD-CatBoost according to claim 1 is characterized in that: In step S203, the expression of the sparrow search mechanism is as follows: ， Where: v ij (t) is the The particle in dimensional speed, c1 is the individual cognitive coefficient, rand(-1,1,size=d) represents a random vector of dimension D, ⊙ represents the element multiplication of the vector, p ij (t) is the The particle in The historical best position in dimension, x ij (t) is the The particle in Current position in dimension; The expression of the Cauchy variation is as follows: ， Where: c is a dimension and v ij (t) The same vector, where each element is independently drawn from a standard Cauchy distribution.

5. The traction motor fault diagnosis method based on RPSO-VMD-CatBoost according to claim 1 is characterized in that: Step S3 includes: Define a finite-bandwidth modal function with strict constraints: ， Where: u k (t) is the decomposed modal components; A k (t) is the instantaneous amplitude; Ф k (t) represents the phase function; Perform Hilbert transformation on the modal function to obtain the corresponding analytical signal: ， Where: δ(t) is the Dirac shock function; is an imaginary unit; Utilization Index Adjust the spectrum of the modal function to focus on the corresponding baseband: ； Use Gaussian smoothing, that is, calculate L 2 The square root of the norm gradient is used to demodulate the signal, so we can get: ， in, are the modal functions, is the center frequency of each mode, ∂ t For time The differential operator, f is the original input signal.

6. The traction motor fault diagnosis method based on RPSO-VMD-CatBoost according to claim 1, characterized in that: In step S4, the optimal IMF component expression is determined according to the minimum envelope entropy principle as follows: ， Where, is the envelope entropy, p j is the normalized form of a(j), and a(j) is the envelope signal obtained after Hilbert demodulation of signal x(j).

7. The traction motor fault diagnosis method based on RPSO-VMD-CatBoost according to claim 1, characterized in that: In step S5, the time domain eigenvalues ​​include one or more of the following: maximum value, maximum absolute value, minimum value, mean value, peak-to-peak value, absolute mean value, root mean square value, root amplitude, standard deviation, kurtosis, skewness, margin index, waveform index, pulse index, and peak index; the frequency domain eigenvalues ​​include one or more of the following: center of gravity frequency, frequency root mean square, average frequency, and frequency variance; and the envelope eigenvalues ​​include envelope entropy.

8. The traction motor fault diagnosis method based on RPSO-VMD-CatBoost according to claim 1, characterized in that: Step S6 includes: Set the value range of the hyperparameter iterations, which represents the number of iterations during model training; Set the value range of the hyperparameter learning_rate, which represents the contribution of each tree to the final prediction; Set the value range of the hyperparameter depth, which represents the maximum depth of the decision tree; Set the value range of the hyperparameter l2_leaf_reg to control the complexity of leaf nodes; The best hyperparameter combination is obtained through grid hyperparameter optimization.

Images