A high-speed train bearing fault diagnosis method based on intelligent hyperparameter optimization

Abstract

The application discloses a high-speed train bearing fault diagnosis method based on intelligent super parameter optimization. The method collects multiple vibration signals, carries out wavelet adaptive threshold denoising and speed normalization preprocessing, and extracts 30-dimensional multi-modal features in time domain, frequency domain and time-frequency domain; in view of sample imbalance, a random mask hybrid synthesis method is used for data enhancement, and the feature dimension is reduced to 14 dimensions through correlation analysis two-stage screening; a particle swarm optimization random forest main diagnosis model is constructed to realize global automatic optimization of key super parameters of the random forest; meanwhile, a genetic algorithm optimized long short-term memory network auxiliary diagnosis model is constructed to assist in verifying the main model result. The experimental results show that the main model diagnosis accuracy reaches 90%, all kinds of AUC are more than 0.97, which is better than 88% of the default parameter random forest; the macro average F1-score of the auxiliary model reaches 0.72. The application realizes automatic fusion of intelligent super parameter optimization and high-precision fault classification.

Description

Technical Field

[0001] This invention relates to the fields of intelligent operation and maintenance and machine learning technology in rail transit, specifically to a method and system for intelligent fault diagnosis of high-speed train bearings based on particle swarm optimization random forest (PSO-RF) and genetic algorithm optimization long short-term memory network (GA-LSTM). Background Technology

[0002] High-speed train bearings are the core rotating load-bearing components of the train running gear, playing a decisive role in the safety and stable operation of trains. Bearings are subjected to high speed, alternating heavy load, and complex operating conditions with varying temperature and humidity for a long time, making them prone to early failures such as wear and pitting. If these failures are not identified in time, they can lead to wheelset seizure, line shutdown, and in severe cases, derailment accidents. Therefore, accurate diagnosis of bearing failures is a core task of rail transit operation and maintenance.

[0003] Traditional monitoring methods rely heavily on maintenance personnel manually interpreting vibration signals using fixed time- and frequency-domain indicators. This approach is significantly affected by noise and variable-speed operating conditions, and lacks accuracy in identifying subtle early-stage faults, exhibiting poor generalization performance. With the continuous maturation of machine learning and deep learning technologies, various data-driven diagnostic solutions are increasingly being applied to bearing condition identification. These solutions can automatically uncover hidden fault features in vibration signals, overcoming the limitations of manual interpretation. For example, patent application CN122364782A discloses a high-speed train bearing fault diagnosis method based on transfer learning. This method collects labeled source domain vibration data from the laboratory and unlabeled target vibration data from actual vehicles. After basic signal preprocessing, it extracts time, frequency, and time-frequency features. Feature selection is performed using t-SNE, and a Stacking ensemble model composed of random forest, SVM, and AdaBoost is built as the basic diagnostic model. Single-scale subspace alignment is used to reduce the distribution difference between the source and target domains. Fault determination is completed based on single-scale output results. The transfer learning approach addresses the problem of scarce fault labels on actual vehicles, significantly improving cross-condition identification performance compared to traditional manual detection. However, the solution relies on manual tuning of all model hyperparameters, lacks a global automatic optimization mechanism, and does not construct a complete multimodal feature system. It also lacks a targeted sample balancing algorithm and a two-level redundant feature removal process, relying solely on a single diagnostic model, resulting in significant instability in fault identification under complex working conditions. Summary of the Invention

[0004] To address the shortcomings of existing high-speed train bearing diagnosis methods, such as manual parameter tuning, limited features, and imbalanced samples, this invention provides an intelligent hyperparameter optimization bearing fault diagnosis method and system. The method first collects and preprocesses multiple vibration signals, extracts and balances multimodal features, and then filters them. It then uses a PSO-optimized random forest as the primary model and a GA-optimized LSTM as the auxiliary model to automatically identify faults. This invention relies on intelligent algorithms to achieve global hyperparameter optimization, improving the accuracy of bearing fault diagnosis and providing reliable support for intelligent train operation and maintenance.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A high-speed train bearing fault diagnosis method based on intelligent hyperparameter optimization is implemented according to the following steps:

[0007] S1. Vibration signal acquisition and preprocessing:

[0008] Vibration acceleration signals from the bearing drive end, fan end, and base of a high-speed train were collected. The original signals were subjected to a uniform sampling rate setting and fixed duration truncation, and noise reduction was performed using an adaptive threshold denoising method based on improved wavelet transform.

[0009] Then the normalized frequency formula was used. The frequency axis is normalized to eliminate frequency deviation under different speed conditions.

[0010] S2. Multimodal feature extraction: Extract 14-dimensional time-domain feature variables from the time domain, including mean, standard deviation, variance, root mean square, peak value, peak-to-peak value, mean absolute value, waveform factor, kurtosis factor, impulse factor, margin factor, gap factor, kurtosis and skewness.

[0011] We used the Welch power spectral density estimation method to extract nine-dimensional frequency domain feature variables from the frequency domain, including spectral centroid, spectral bandwidth, spectral skewness, spectral kurtosis, spectral entropy, and bandpass energy of the 1, 2, 3, and 5 times the frequency shift neighborhood.

[0012] Short-time Fourier Transform (STFT) was performed on the vibration signal and combined with Daubechies-4 wavelet 4-level decomposition to extract 7 time-frequency domain feature variables, including time-frequency entropy, time-frequency mean frequency and 5-dimensional wavelet energy ratio, to form a 30-dimensional multimodal feature set.

[0013] S3. Sample Balancing and Feature Selection: To address the issue of uneven sample numbers across different categories in the source domain data, a random masking hybrid synthesis method is used to augment the data of each fault category to achieve sample balance.

[0014] Specifically, for two original samples x a and x bGenerate a random binary mask M∈{0,1}, and the value of the i-th position in the synthesized sample satisfies: when M i When x = 1, take x a，i When M i When x = 0, take x b，i .

[0015] Then, the recursive feature elimination algorithm based on random forest (RF-RFE) is used to compress the 30-dimensional features to 20-dimensional features, and then the Pearson correlation coefficient heatmap is used to remove highly redundant features, finally obtaining 14-dimensional effective discriminative features;

[0016] S4. Automatic Hyperparameter Optimization and Diagnosis of Random Forest Based on Particle Swarm Optimization Algorithm:

[0017] With the core principle of automatically searching for the optimal hyperparameter combination without human intervention, a random forest fault diagnosis framework driven by particle swarm optimization (PSO) algorithm is constructed.

[0018] The four key hyperparameters of random forest, n_estimators, max_depth, min_samples_split, and min_samples_leaf, are encoded as position vectors of particles in four-dimensional space. The macro-average F1-score is used as the fitness function. A global search is performed in the hyperparameter space through the iterative update mechanism of particle velocity and position to automatically determine the hyperparameter combination that optimizes classification performance.

[0019] The random forest classifier is retrained with the obtained optimal hyperparameters to achieve high-precision automatic identification of four types of bearing states: normal state (N), inner ring fault (IR), outer ring fault (OR), and rolling element fault (B).

[0020] Compared to the Stacking ensemble model used in CN122364782A cited in the background section, which relies on manual parameter tuning, this step achieves global automatic search of hyperparameters through PSO, eliminating manual intervention.

[0021] S5. Genetic Algorithm Optimization of Long Short-Term Memory Network for Auxiliary Diagnosis: The number of hidden layer units, dropout rate, learning rate, number of fully connected layer units, batch size, and number of training rounds of the LSTM network are encoded as real chromosomes. Through population initialization, roulette wheel selection, crossover and mutation operations, the model is iteratively evolved and the near-optimal hyperparameter combination is searched with the macro-average F1-score as the fitness function to train the LSTM-assisted diagnostic model.

[0022] S6. Model Evaluation and Output: The model comprehensively evaluates the diagnostic results using multiple indicators, including accuracy, precision, recall, F1-score, ROC curve, and AUC value, and outputs bearing fault type labels to provide decision-making support for maintenance personnel.

[0023] Furthermore, the threshold calculation formula for the adaptive threshold denoising method described in step S1 is as follows:

[0024]

[0025] Where, λ i σ is the threshold for the detail coefficients of the i-th layer. i N represents the standard deviation of the detail coefficients. i For the length of the coefficient, This represents the energy of this layer. A soft thresholding function is used to process the detail coefficients, preserving the fault impact characteristics in the large-amplitude coefficients and suppressing the small-amplitude noise coefficients.

[0026] The denoised vibration signal is obtained by wavelet reconstruction.

[0027] Furthermore, the particle velocity update formula for the particle swarm optimization algorithm described in step S4 is:

[0028] v i，d (t+1)=ω·v i，d (t)+c1·r1(pbest i,d -x i,d (t))+c2·r2(gbest d -x i.d (t))

[0029] The particle position update formula is: x i，d (t+1)=x i，d (t)+v i，d (t+1)

[0030] Where ω is the inertia weight, with a value of 0.72; c1 and c2 are both learning factors, with a value of 1.49; r1 and r2 are random numbers in the interval [0, 1].

[0031] pbest_{i, d} represents the individual's historical best position; gbest_d represents the global best position.

[0032] The hyperparameter search range is: n_estimators∈[50,1000], max_depth∈[3,50], min_samples_split∈[2,20], min_samples_leaf∈[1,20]; the particle swarm size is set to 20, and the maximum number of iterations is 25.

[0033] Furthermore, in step S5, the fitness function of the genetic algorithm adopts the macro-average F1-score, and the LSTM model is evaluated through hierarchical K-fold cross-validation.

[0034] The encoding scheme normalizes the hyperparameters to be optimized and maps them to real numbers in the interval [0, 1] to encode chromosomes; the decoding scheme decodes the chromosomes into actual hyperparameter values ​​through an inverse mapping function.

[0035] Furthermore, the random forest classifier uses the Gini index (GI) and information gain (IG) as node splitting criteria, constructs multiple decision trees through bootstrap sampling with replacement, randomly selects some features as candidate splitting attributes each time a node splits, and finally determines the fault category by combining the outputs of all trees through a voting mechanism.

[0036] The present invention also provides a high-speed train bearing fault diagnosis system for implementing the above method, characterized in that it includes:

[0037] Signal acquisition module: used to acquire vibration acceleration signals from the drive end, fan end, and base.

[0038] Preprocessing module: used to perform sampling rate unification, adaptive wavelet denoising, and speed normalization on the original signal;

[0039] Feature extraction module: used to extract 30-dimensional multimodal features from the time domain, frequency domain, and time-frequency domain;

[0040] Feature filtering module: used to perform recursive feature elimination and correlation analysis based on random forest to obtain 14 effective features;

[0041] PSO-RF diagnostic module: used to run the particle swarm optimization algorithm to search for the optimal combination of hyperparameters in a random forest and output fault category labels;

[0042] GA-LSTM Auxiliary Module: Used to run genetic algorithms to optimize LSTM hyperparameters and provide auxiliary diagnostic results;

[0043] Evaluation output module: Used to calculate evaluation metrics such as accuracy, F1-score, and ROC-AUC, and output diagnostic reports.

[0044] The beneficial effects of this invention are:

[0045] 1. A multimodal feature system with 30 dimensions, including 14 dimensions in the time domain, 9 dimensions in the frequency domain, and 7 dimensions in the time-frequency domain, was constructed. Combined with the speed normalization strategy, it can comprehensively reflect the impact characteristics, energy distribution, and non-stationary laws of bearing vibration signals, and significantly improve the feature discrimination ability.

[0046] 2. A method is proposed to introduce Particle Swarm Optimization (PSO) algorithm into the hyperparameter optimization of bearing fault diagnosis in Random Forest. This method enables global automatic search of four-dimensional hyperparameters such as the number of decision trees and maximum depth, fundamentally replacing the existing hyperparameter setting methods that rely on manual trial and error, grid search, or fixed default values. Experimental results show that PSO-RF improves the diagnostic accuracy by about 2 percentage points to 90% compared with Random Forest with default parameters, and all types of AUC exceed 0.97. Moreover, PSO converges in only the first 5 generations, and its computational efficiency is significantly better than exhaustive grid search.

[0047] 3. A genetic algorithm (GA) is proposed to automatically optimize the hyperparameters of the LSTM network. Through roulette wheel selection, crossover and mutation operations, the automatic search of LSTM hyperparameters is realized. The macro average F1-score of the obtained model reaches 0.72, which can be used as an auxiliary verification method for the PSO-RF master diagnostic model and provide reference and verification for the diagnostic results of the master model.

[0048] 4. A data augmentation method based on random mask hybrid synthesis is proposed. Unlike the linear interpolation strategy of the existing SMOTE method, this method synthesizes original samples of the same type by selecting each element through random binary masks, which preserves the nonlinear distribution characteristics of real samples and avoids the defect of SMOTE in interpolating noisy samples in high-dimensional space. At the same time, a two-level screening method of random forest recursive feature elimination and Pearson correlation coefficient is adopted to reduce the feature dimension from 30 dimensions to 14 dimensions, effectively suppressing overfitting.

[0049] 5. The entire diagnostic process is highly automated, requiring no manual intervention in hyperparameter setting, and has good engineering usability and scalability.

[0050] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 A schematic diagram of the overall process of the high-speed train bearing fault diagnosis method based on intelligent hyperparameter optimization provided in the embodiments of this application;

[0053] Figure 2 A comparison diagram of the t-SNE distribution of samples in the feature space provided for embodiments of this application;

[0054] Figure 3A comparison chart of the accuracy of different diagnostic models provided in the embodiments of this application across various fault categories;

[0055] Figure 4 The genetic algorithm parameter optimization curve provided in the embodiments of this application;

[0056] Figure 5 A heatmap of the confusion matrix of the PSO-RF diagnostic model provided in this application embodiment on the test set;

[0057] Figure 6 ROC curves of the PSO-RF diagnostic model provided in the embodiments of this application for each fault category;

[0058] Figure 7 This is a domain adaptation training curve provided for an embodiment of this application. Detailed Implementation

[0059] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0060] Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application.

[0061] refer to Figure 1 The diagram shows a flowchart of a high-speed train bearing fault diagnosis method based on intelligent hyperparameter optimization provided in this application embodiment.

[0062] This application provides a high-speed train bearing fault diagnosis method based on intelligent hyperparameter optimization, the method comprising:

[0063] S1. Collect vibration acceleration signals from the drive end, fan end, and base of the high-speed train bearing. Set a uniform sampling rate and extract the original signals for a fixed duration. Then, use an adaptive threshold denoising method based on improved wavelet transform to denoise the signals. Finally, use the normalized frequency formula to normalize the frequency axis to eliminate frequency shifts under different speed conditions.

[0064] S2. Extract 14-dimensional time-domain feature variables from the time domain; extract 9-dimensional frequency-domain feature variables from the frequency domain using the Welch power spectral density estimation method; perform short-time Fourier transform on the vibration signal and combine it with Daubechies-4 wavelet 4-level decomposition to extract 7-dimensional time-frequency domain feature variables, forming a 30-dimensional multimodal feature set.

[0065] S3. A random mask hybrid synthesis method is used to augment the data of each fault category to achieve sample balance; a recursive feature elimination algorithm based on random forest is used to compress the 30-dimensional features to 20-dimensional features, and then the Pearson correlation coefficient heatmap is used to remove highly redundant features to obtain 14-dimensional effective discriminative features.

[0066] S4. Construct a random forest fault diagnosis framework driven by particle swarm optimization algorithm. Encode the four key hyperparameters of random forest as position vectors of particles in four-dimensional space. Use the macro-average F1-score as the fitness function. Perform a global search in the hyperparameter space through the iterative update mechanism of particle velocity and position to automatically determine the hyperparameter combination that optimizes classification performance. Retrain the random forest classifier with the obtained optimal hyperparameters to achieve high-precision automatic identification of four types of bearing states: normal bearing state, inner ring fault, outer ring fault, and rolling element fault.

[0067] S5. Encode the number of hidden layer units, dropout rate, learning rate, number of fully connected layer units, batch size, and number of training rounds of the Long Short-Term Memory Network as real chromosomes. Iterate through population initialization, roulette wheel selection, crossover and mutation operations, and search for the near-optimal hyperparameter combination with the macro-average F1-score as the fitness function to train the LSTM-assisted diagnostic model.

[0068] S6. The system comprehensively evaluates the diagnostic results using multiple indicators, including accuracy, precision, recall, F1-score, ROC curve, and AUC value, and outputs bearing fault type labels to provide decision-making support for maintenance personnel.

[0069] In the embodiments of this application, during the high-speed train bearing fault diagnosis task, the vibration acceleration signals from the drive end, fan end, and base are first collected and preprocessed. Background noise is effectively suppressed by wavelet adaptive threshold denoising, and frequency offset under different speed conditions is eliminated by speed normalization. Then, the system extracts 30-dimensional multimodal features in the time domain, frequency domain, and time-frequency domain. Sample balance is achieved by random mask mixing and synthesis. 14-dimensional effective discriminative features are obtained through two-level screening of recursive feature elimination and correlation analysis. Finally, the optimal hyperparameter combination of random forest is automatically searched using particle swarm optimization algorithm to train a high-precision fault classifier, realizing the automatic identification of four health states of the bearing.

[0070] The vibration signal acquisition and preprocessing steps provided in this application embodiment include:

[0071] S11. Collect vibration acceleration signals from three sources: the bearing drive end, the fan end, and the base of the high-speed train. The sampling frequency is 32kHz, the duration of each recorded signal is 8 seconds, and each measurement cycle contains 256,000 sampling points.

[0072] S12. Perform wavelet transform adaptive threshold denoising on the original signal. Use the Daubechies-4 wavelet basis to decompose the signal into multiple layers to obtain approximation coefficients and detail coefficients. Perform adaptive threshold processing on the detail coefficients of each layer to retain the fault impact characteristics in the large amplitude coefficients and suppress the small amplitude noise coefficients. The denoised vibration signal is obtained by wavelet reconstruction.

[0073] S13. Divide the frequency f by the rotational frequency f r The dimensionless normalized frequency is obtained, thus eliminating rotational speed deviation.

[0074] In the embodiments of this application, the bearing vibration signal data is collected under different operating conditions on the test bench, covering four bearing states: normal state (N), inner ring fault (IR), outer ring fault (OR), and rolling element fault (B). The raw signal contains a large amount of background noise and outliers, and directly using the raw data is not conducive to the stability and accuracy of fault identification.

[0075] Wavelet transform denoising utilizes multi-resolution characteristics to decompose a signal into approximation coefficients and detail coefficients at different scales. The original signal x(t) after wavelet decomposition is represented as:

[0076]

[0077] in, Let ψ be the scaling function. j，k (t) is the wavelet basis function. The approximation coefficients mainly contain the low-frequency part (the signal when the bearing is running stably), while the detail coefficients reflect the high-frequency part (impact and noise components).

[0078] The adaptive threshold calculation formula described in claim 2 is used to calculate the threshold for the detail coefficients of each layer. This threshold is adaptively adjusted according to the energy distribution of each layer, which is superior to the traditional fixed threshold method. A soft threshold function is used to process the detail coefficients, retaining the large amplitude coefficients (fault impact characteristics) and suppressing the small amplitude coefficients (noise interference). The final denoised signal is obtained by wavelet reconstruction, in which high-frequency noise is significantly suppressed and fault impact characteristics (such as periodic pulses) are clearly preserved.

[0079] Subsequently, speed normalization is introduced: The frequency axis was normalized to a dimensionless scale to eliminate frequency offsets under different speed conditions. After the above preprocessing, the distribution differences of vibration signals under different bearing conditions became more significant: the signal was most stable under normal conditions and concentrated near zero, while the signal fluctuations increased significantly and the dispersion was higher under fault conditions; the signal distributions of different fault types (inner ring fault, outer ring fault, rolling element fault) under the same sensor had their own characteristics, laying a high-quality signal foundation for subsequent feature extraction and fault classification.

[0080] The multimodal feature extraction steps provided in this application embodiment include:

[0081] S21. Extract the 14-dimensional time-domain feature variables as described in claim 3 from the time domain;

[0082] S22. The Welch power spectral density estimation method is used to extract 9-dimensional frequency domain feature variables from the frequency domain, including spectral centroid, spectral bandwidth, spectral skewness, spectral kurtosis, spectral entropy, and bandpass energy in the normalized frequency axis at approximately 1, 2, 3, and 5 times the frequency shift neighborhood.

[0083] S23. Perform a short-time Fourier transform with a window length of nperseg=1024 to obtain the time-frequency power matrix, and extract the time-frequency entropy and the time-frequency mean frequency as 2D features. At the same time, use the Daubechies-4 wavelet basis to decompose the signal into 4 levels, obtain 5 sets of decomposition coefficients, calculate the energy of each level and normalize it into proportional features, extract 5-dimensional wavelet energy proportional features, and a total of 7-dimensional time-frequency domain features.

[0084] In the embodiments of this application, when a rolling bearing fails, the damaged part continuously contacts the rolling element during rotation, forming periodic impact characteristics in the vibration signal. Time-domain characteristics can directly characterize the amplitude distribution, impact intensity, and non-Gaussianity of the signal over time, and are the most basic and intuitive feature type in fault diagnosis.

[0085] Let the signal be a vector of length N. The extracted time-domain features are as follows:

[0086] mean Reflects the DC bias of the signal;

[0087] Standard deviation Reflects the overall fluctuation intensity;

[0088] variance Reflects the distribution of signal energy;

[0089] Root Mean Square Reflects the effective value and overall energy / load level;

[0090] Peak = max n |x n | reflects the maximum instantaneous impact amplitude;

[0091] Peak-to-peak value P2P = max n (x n )-min n (x n This reflects the maximum dynamic range;

[0092] Average absolute value It reflects the average energy level.

[0093] Simultaneously extract the dimensionless index: waveform factor. kurtosis factor Pulse factor margin factor gap factor

[0094] Where SQM is the square root mean. ∈ represents a very small number used for numerical stability;

[0095] Kurtosis reflects the degree of signal spikes. The kurtosis of a normal bearing is close to a Gaussian distribution (threshold is about 3). When a bearing is subjected to periodic impact, the kurtosis value can exceed 10.

[0096] Skewness reflects the symmetry of signal distribution.

[0097] Frequency domain features can extract deeper information related to faults.

[0098] The Welch method is used to estimate the power spectrum of the signal, with a sampling rate of f. s The bearing frequency is Obtain the frequency axis {f k} and power spectral density {(P Xx (f k ))}.

[0099] The normalized energy of the spectrum is defined as And introduce a rotational speed normalization frequency

[0100] Based on this, the spectral centroid (reflecting the location of the spectral energy center), spectral bandwidth (reflecting the degree of energy dispersion), spectral skewness (reflecting spectral symmetry), spectral kurtosis (reflecting spectral peak state), spectral entropy (reflecting spectral complexity), and bandpass energy in the approximately 1, 2, 3, and 5 times the frequency shift neighborhood on the normalized frequency axis are extracted to capture the harmonic response characteristics of periodic impacts.

[0101] Time-frequency domain characteristics are used to reveal the dynamic changes of a signal in the time-frequency plane. By performing a STFT on the vibration signal (window length nperseg = 1024), the time-frequency matrix Z(f, t) is obtained, and the time-frequency power power = |Z| is calculated. 2 And perform global normalization.

[0102] Rotational speed normalization is used on the frequency axis to eliminate the influence of different rotational speeds, and the time-frequency entropy (reflecting the dispersion of energy in the time-frequency plane, the smaller the value, the more concentrated the energy) and the time-frequency mean frequency (representing the energy centroid under the normalized frequency) are extracted.

[0103] Simultaneously, the signal is decomposed into four levels using the Daubechies-4 wavelet basis, yielding five sets of coefficients (approximate components and four levels of detail components), and the energy E of each level is calculated. i =∑ci 2 And normalized to proportional characteristics Low-scale components correspond to low-frequency trend features, while high-scale components characterize medium- to high-frequency impact features.

[0104] The 7-dimensional time-frequency domain feature set constructed in this way can reasonably reflect the signal's stationary state, transient impact characteristics, and cross-frequency energy distribution.

[0105] The sample balancing and feature selection steps provided in this application embodiment include:

[0106] S31. Use a random masking hybrid synthesis method to augment data for each fault category to achieve sample balance;

[0107] S32. A recursive feature elimination algorithm based on random forest is used to compress 30-dimensional features to 20-dimensional features;

[0108] S33. Using the Pearson correlation coefficient heatmap to remove highly redundant features, a total of 14 effective discriminative features are finally obtained.

[0109] refer to Figure 2 As shown in the figure, the feature space t-SNE dimensionality reduction distribution comparison diagram provided in the embodiments of this application is presented. The figure distinguishes various bearing samples in the source domain and target domain. After random masking enhancement and two-level redundancy removal, the similar samples of the features are closely clustered, and the boundaries of different fault categories are clear, verifying that feature screening effectively improves separability.

[0110] In the embodiments of this application, the source domain data exhibits a clear class imbalance: the outer ring fault (OR) state has the most data, reaching 77; rolling element faults (B) and inner ring faults (IR) both have 44; and the normal state (N) has the fewest, with only 4. If a classification model is directly trained based on this data, the small number of samples will result in insufficient learning of the minority class and inadequate generalization ability. Therefore, data augmentation and class balancing are necessary.

[0111] Based on the bearing failure mechanism, a method for synthesizing new samples by random mask mixing is designed.

[0112] For minority class samples, repeat the following steps until the number of samples in that class reaches the target number: randomly select two original samples x from that class. a and x b Generate a random binary mask M∈{0,1}, and P(Mi=1)=mix_ratio. The value of the i-th position in the synthesized sample satisfies: when M i When x = 1, take x a，iWhen M i When x = 0, take x b，i .

[0113] This method addresses class imbalance through data augmentation, enabling the model to learn the features of each class more fairly during training. It is particularly suitable for scenarios in industrial fault diagnosis where "minority class samples are scarce but accurate identification is required."

[0114] Feature selection is then performed. Recursive Feature Elimination (RFE) works like a feature "promotion competition": all features are initially included, and in each round, features ranked lower are eliminated based on model performance. Winners advance to the next round, and this process is iterated until a specified number of superior features remain. This paper uses a random forest as the base learner for RFE, leveraging its ability to effectively evaluate feature importance to determine whether a feature should be retained, ultimately selecting the 20 most discriminative features.

[0115] Based on this, the Pearson correlation coefficient was used for correlation analysis. The Pearson correlation coefficient is a statistic that measures the degree of linear correlation between two variables, with a value range of [-1, 1]. A coefficient close to 1 indicates a strong positive linear correlation, close to -1 indicates a strong negative linear correlation, and close to 0 indicates a weak linear correlation.

[0116] By visually observing the significant linear correlation between multiple sets of features through heatmaps, for feature pairs with extremely strong linear correlation (such as a correlation coefficient of 0.94 between variance and standard deviation, and a correlation coefficient of 1.0 between root mean square and standard deviation), basic features are retained and redundant features are eliminated based on the fundamental and interpretive nature of statistical definitions.

[0117] For peak values ​​and peak-to-peak values ​​(correlation coefficient 0.92), the peak value, which more directly reflects the maximum amplitude characteristic of the signal, is retained. Finally, variance, peak value, mean absolute value, skewness, spectral kurtosis, gap factor, spectral entropy, time-frequency entropy, time-frequency mean frequency, and 5-dimensional wavelet energy features (wavelet energy 0 to wavelet energy 4) are retained, for a total of 14 effective discriminant features.

[0118] This two-stage screening strategy eliminates feature redundancy and suppresses the risk of overfitting, while also taking into account the feature diversity in the time domain, frequency domain, and various dimensions of the frequency domain.

[0119] The steps for finding hyperparameters in a random forest using particle swarm optimization provided in this application embodiment include:

[0120] S41. Encode the four key hyperparameters of random forest, n_estimators, max_depth, min_samples_split, and min_samples_leaf, into position vectors of particles in four-dimensional space.

[0121] S42. Using the macro-average F1-score as the fitness function, a global search is performed in the hyperparameter space through an iterative update mechanism of particle velocity and position.

[0122] S43. Automatically determine the hyperparameter combination that optimizes classification performance, retrain the random forest classifier with the obtained optimal hyperparameters, and output the recognition results of the four types of bearing states.

[0123] In the embodiments of this application, random forest, as an efficient ensemble learning algorithm, has model performance that largely depends on the values ​​of its hyperparameters.

[0124] The relationship between the hyperparameters of random forests and model performance is non-linear and non-explicit, and cannot be directly expressed by mathematical formulas; it belongs to the category of black-box optimization problems.

[0125] Particle swarm optimization (PSO) does not rely on the derivative information or mathematical expression of the objective function. It can search for the optimal solution simply by the "flight" and "information sharing" of particles in the solution space, making it very suitable for hyperparameter optimization scenarios where it is impossible to establish a clear mapping relationship.

[0126] In the algorithm, each "particle" represents a potential solution (i.e., the hyperparameter combination of the random forest). By tracking the individual optimal solution (pbest) and the global optimal solution (gbest), it is iteratively updated according to the velocity update formula and position update formula described in claim 5, where the inertia weight ω = 0.72, the learning factor c1 = c2 = 1.49, and r1 and r2 are random numbers in the interval [0, 1].

[0127] The hyperparameter search range is set as follows: n_estimators∈[50, 1000], max_depth∈[3, 50], min_samples_split∈[2, 20], min_samples_leaf∈[1, 20]; particle dimension NUM_DIM=4 (consistent with the number of hyperparameters to be optimized);

[0128] Particle swarm size NUM_PARTICLES = 20; maximum number of iterations NUM_ITERS = 25.

[0129] The dataset was divided into training and testing sets in a 7:3 ratio, and a fixed random seed was set to ensure reproducibility.

[0130] The macro-average F1-score is used as the fitness function to evaluate the quality of particles. Through iterative optimization, the PSO algorithm converges to the optimal solution in the first 5 generations.

[0131] The optimal combination of hyperparameters is: n_estimators = 605, max_depth = 21, min_samples_split = 2, min_samples_leaf = 1.

[0132] The random forest classifier was retrained with the optimal combination of hyperparameters. Multiple decision trees were constructed using Bootstrap sampling with replacement, and the Gini index (GI) and information gain (IG) were used as node splitting criteria.

[0133] The formula for the Gini index is:

[0134] Where pi is the probability of the i-th class sample in node t;

[0135] The information gain formula is:

[0136] Where H(D) is the entropy. Finally, a voting mechanism is used to combine the outputs of all trees to determine the fault category.

[0137] Table 1 below shows a performance comparison of the PSO-RF diagnostic model with other diagnostic methods, with the best results in each column highlighted in bold.

[0138] Table 1. Quantitative comparison of the performance of the PSO-RF diagnostic model with other methods in bearing fault diagnosis.

[0139]

[0140] As can be seen from Table 1, the PSO-RF diagnostic model provided in this application outperforms the default parameter random forest (88%, 0.88), 1D-CNN (81%, 0.80) and GA-LSTM (70%, 0.69) in key metrics such as accuracy (90%) and macro-average F1-score (0.90).

[0141] refer to Figure 3 As shown in the schematic diagram comparing the accuracy of different diagnostic models for various fault categories provided in the embodiments of this application, the identification accuracy of each model in the four fault categories of B, OR, IR and N is further refined. The accuracy of PSO-RF in each fault category is comprehensively better than the comparison model, which verifies the effectiveness of the method of this application from the perspective of classification granularity.

[0142] The AUC values ​​for each category are as follows: N for normal condition is 1.00, IR for inner ring fault is 0.98, OR for outer ring fault is 0.97, and B for rolling element fault is 0.96. All of them exceed 0.95, indicating that the model has excellent discrimination ability for each type of fault.

[0143] Meanwhile, the PSO algorithm converges to the optimal solution in just 5 generations, and its computational efficiency is significantly better than that of exhaustive grid search.

[0144] The steps for optimizing LSTM hyperparameters using a genetic algorithm provided in this application embodiment include:

[0145] S51. Encode the number of hidden layer units, dropout rate, learning rate, number of fully connected layer units, batch size, and number of training rounds of the LSTM network into real number chromosomes;

[0146] S52. Through iterative evolution of population initialization, roulette wheel selection, crossover and mutation operations, the approximate optimal hyperparameter combination is searched using the macro-average F1-score as the fitness function.

[0147] S53. Decode the optimal chromosome into actual hyperparameter values ​​and train the LSTM-assisted diagnostic model.

[0148] In the embodiments of this application, Long Short-Term Memory (LSTM) network is a model based on an improvement of recurrent neural network. It is designed to solve the gradient vanishing and gradient explosion problems encountered by traditional RNNs during training, so as to effectively learn and remember long-term dependencies in sequence data.

[0149] The core of LSTM lies in cell state and gating mechanisms (including forgetting gate, input gate, and output gate), and the main calculation formulas are as follows:

[0150] Forgotten Gate: f t =σ(W f h t-1 +W f X t +b f )

[0151] Input gate: i t =σ(W f h t-1 +W f X t +b i ),

[0152] Output gate: O t =σ(W o h t-1 +W o X t +b o ), h t =O t ×tanh(C t )

[0153] To address the problem of low efficiency in manual hyperparameter tuning due to the large number and complex dimensions of hyperparameters in LSTM models, this application employs a genetic algorithm for automatic hyperparameter search and optimization.

[0154] This method transforms the hyperparameter optimization problem into a population optimization problem in the process of biological evolution, and efficiently obtains near-optimal solutions by simulating natural selection and genetic mechanisms.

[0155] Encoding scheme: The LSTM hyperparameters to be optimized (number of hidden layer units, dropout rate, learning rate, number of fully connected layer units, batch size, number of training rounds) are normalized and mapped to a real number encoding array in the interval [0, 1]. This array is regarded as a "chromosome" individual, representing a complete hyperparameter configuration scheme.

[0156] Decoding scheme: Design an inverse mapping function decode_chromosome to decode the chromosome into the actual hyperparameter values ​​required for model training. For integer parameters (such as the number of hidden layer units, the number of fully connected layer units, batch size, and the number of training epochs), a rounding strategy is adopted;

[0157] For continuous parameters (such as dropout rate and learning rate), linear or nonlinear scaling is used to restore their original range of values.

[0158] Fitness function: The fitness function is the core indicator for evaluating the quality of chromosomes. The hyperparameters of the chromosome are used to configure the LSTM model, and the hierarchical K-fold cross-validation method is used for training and evaluation. The macro-average F1-score is selected as the performance indicator for cross-validation, which comprehensively reflects the overall performance of the model across all categories.

[0159] The core operations of genetic algorithms include:

[0160] Population initialization—a certain number of chromosomes are randomly generated in the solution space to form the initial population;

[0161] Selection operation – using roulette wheel selection, the probability of an individual being selected is proportional to its fitness value;

[0162] Crossover operation – exchanging gene fragments between paired individuals with a preset probability to produce new offspring individuals;

[0163] Mutation operation – sets a low probability to randomly change the values ​​of certain gene loci in an individual, enhancing local search capabilities and helping to escape local optima.

[0164] After optimization using a genetic algorithm, the optimal combination of hyperparameters is: 61 hidden layer units, 0.15 dropout rate, 125 fully connected layer units, 0.0039 learning rate, 32 batch size, and 29 training epochs.

[0165] refer to Figure 4As shown in the figure, the genetic algorithm provided in this application optimizes the LSTM hyperparameter iterative optimization curve. The horizontal axis represents the number of evolutionary iterations, and the vertical axis represents the macro-average F1 score of the model. It can be seen that the performance of the algorithm steadily increases and converges quickly during the iterative process, verifying that GA can efficiently complete the automatic optimization of LSTM hyperparameters.

[0166] The macro-average F1-score of the GA-LSTM model reached 0.72, verifying the effectiveness of the genetic algorithm in LSTM hyperparameter optimization.

[0167] The model evaluation and result output steps provided in this application embodiment include:

[0168] S61. Comprehensively evaluate diagnostic results using multiple indicators, including accuracy, precision, recall, F1-score, ROC curve, and AUC value.

[0169] S62 outputs bearing fault type labels (N / IR / OR / B) to provide decision-making basis for maintenance personnel.

[0170] In the embodiments of this application, the following indicators are used to comprehensively evaluate the diagnostic model:

[0171] Precision: It represents the proportion of samples predicted as class c that actually belong to class c. High precision means low false alarm rate.

[0172] Recall: The ability of a model to identify all true positive samples is measured by its recall rate; a high recall rate means a low false negative rate.

[0173] F1 score: The harmonic mean that combines precision and recall;

[0174] Accuracy: Measure the proportion of all samples that are correctly classified;

[0175] ROC curve and AUC value: The ROC curve depicts how the true positive rate changes with the false positive rate when the classification threshold changes. The closer the curve is to the upper left corner, the better the performance. The AUC value is the area under the ROC curve. The closer it is to 1, the stronger the model's ability to distinguish between positive and negative samples.

[0176] refer to Figure 5 As shown in the figure, the confusion matrix heatmap of the PSO-RF model test set provided in this application embodiment is shown. The horizontal axis represents the predicted fault label and the vertical axis represents the actual bearing state label. It intuitively shows the distribution of correct and incorrect identification of the four types of faults (N normal, IR inner race, OR outer race, B rolling element). It can be clearly seen that the model has only a small number of OR and B misclassified samples, and the overall identification accuracy is high.

[0177] Table 2 below provides a detailed evaluation report of the PSO-RF diagnostic model in each category.

[0178] Table 2. Detailed evaluation report of the PSO-RF diagnostic model for each fault category.

[0179]

[0180] As shown in Table 2, the PSO-RF model demonstrates excellent diagnostic precision and recall for the four bearing conditions. The normal condition N is classified completely correctly (precision, recall, and F1-score are all 1.00), and the F1-score for the inner race fault IR reaches 0.93, indicating that the model has extremely high diagnostic reliability.

[0181] ROC curve analysis showed that the AUC values ​​for all categories exceeded 0.96, further validating the model's excellent ability to distinguish between various types of faults.

[0182] Confusion matrix analysis shows that the main errors in the model are concentrated in the misclassification between OR and B classes. This is because the vibration characteristics of these two types of faults are somewhat similar. However, the overall classification accuracy remains at a high level of 90%.

[0183] refer to Figure 6 As shown in the figure, the ROC curves of the PSO-RF model for four types of faults provided in this application embodiment are as follows. The closer the curve is to the upper left corner, the stronger the classification and discrimination ability. In this scheme, the AUC in the normal state is 1.00, and the AUC for inner ring, outer ring, and rolling element faults all reach above 0.97, proving that the model has extremely strong discrimination performance for various bearing faults.

[0184] To verify the effectiveness of the PSO-RF method provided in this application, a comparative experiment was conducted, and the experimental results are shown in Table 1. Under the same dataset and partition ratio, the diagnostic accuracy of the random forest with default parameters was 88%, while the diagnostic accuracy of the random forest after PSO hyperparameter optimization increased to 90%, an improvement of about 2 percentage points, and the AUC of all types exceeded 0.97.

[0185] Meanwhile, PSO converges in just the first 5 generations, with computational efficiency significantly superior to exhaustive grid search. GA-LSTM and 1D-CNN achieve accuracies of 70% and 81%, respectively, both lower than the method described in this application.

[0186] The experimental results above demonstrate that the high-speed train bearing fault diagnosis method based on intelligent hyperparameter optimization provided in this application has higher diagnostic accuracy and F1-score compared to other comparative methods, thus verifying the effectiveness of this method.

[0187] To verify the effectiveness of the diagnostic method in actual engineering scenarios, the trained PSO-RF diagnostic model was applied to the target domain bearing vibration data (16 samples) under actual operating conditions.

[0188] The results show that the prediction confidence of the model for the target domain samples is higher than 0.85, and the prediction confidence of the normal state samples reaches above 0.98.

[0189] refer to Figure 7 As shown in the figure, this application provides various loss curves and Wasserstein distance change graphs for domain adaptation training. These include reconstruction loss, CORAL, and MMD loss curves. After feature alignment, the difference in sample distribution between the source and target domains continuously decreases, proving that this solution is adaptable to complex real-world vehicle operating conditions and has strong generalization ability.

[0190] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0191] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A high-speed train bearing fault diagnosis method based on intelligent hyperparameter optimization, characterized in that, Includes the following steps: S1. Collect bearing vibration signals, preprocess the vibration signals, and eliminate frequency shifts under different speed conditions; S2. Extract time-domain features, frequency-domain features, and time-frequency-domain features from the preprocessed vibration signal to form a multimodal feature set; S3. Perform data augmentation on the multimodal feature set to achieve sample balance, and then filter the augmented features to obtain effective discriminative features; S4. Construct a random forest fault diagnosis model driven by particle swarm optimization algorithm. Encode the key hyperparameters of random forest into particle position vectors. Use the classification performance index as the fitness function to perform a global search in the hyperparameter space to automatically determine the optimal hyperparameter combination. Train the random forest classifier with the obtained optimal hyperparameters and output the bearing fault type. S5. Construct a genetic algorithm-driven long short-term memory network-assisted diagnostic model. Encode the key hyperparameters of the LSTM network into chromosomes. Iterate through genetic operations and search for the optimal hyperparameter combination using the classification performance index as the fitness function to train the LSTM-assisted diagnostic model. S6. Model Evaluation and Output: The model comprehensively evaluates the diagnostic results using multiple indicators, including accuracy, precision, recall, F1-score, ROC curve, and AUC value, and outputs bearing fault type labels to provide decision-making support for maintenance personnel.

2. The method according to claim 1, characterized in that, The preprocessing described in step S1 specifically includes: Three vibration acceleration signals were collected from the drive end, fan end and base, with a sampling frequency of 32kHz and a single recording duration of 8 seconds. An adaptive threshold denoising method based on improved wavelet transform is used for noise reduction. The threshold calculation formula is as follows: Where, λ i σ is the threshold for the detail coefficients of the i-th layer. i N represents the standard deviation of the detail coefficients. i For the length of the coefficient, This refers to the energy of that layer; A soft thresholding function is used to process the detail coefficients, preserving the fault impact characteristics in the large amplitude coefficients and suppressing the small amplitude noise coefficients; the denoised vibration signal is obtained by wavelet reconstruction. Then the normalized frequency formula was used. The frequency axis is normalized by rotational speed to eliminate frequency offset under different rotational speed conditions.

3. The method according to claim 1, characterized in that, The time-domain features mentioned in step S2 include 14 dimensions: mean, standard deviation, variance, root mean square, peak value, peak-to-peak value, mean absolute value, waveform factor, kurtosis factor, impulse factor, margin factor, gap factor, kurtosis, and skewness. The frequency domain features include 9 dimensions: spectral centroid, spectral bandwidth, spectral skewness, spectral kurtosis, spectral entropy, and bandpass energy of the 1x, 2x, 3x, and 5x frequency shift neighborhoods. The time-frequency domain features include 7 dimensions: time-frequency entropy, time-frequency mean frequency, and 5-dimensional wavelet energy ratio. The multimodal feature set has a total of 30 dimensions.

4. The method according to claim 1, characterized in that, In step S3, a random masking method is used to synthesize new samples to address the problem of imbalanced sample distribution; For two original samples x a and x b Generate a random binary mask M∈{0,1} and synthesize samples. The number of samples in each category is balanced to match that of the category with the most samples, and then recursive feature elimination and correlation screening are performed.

5. The method according to claim 1, characterized in that, The particle velocity update formula for the particle swarm optimization algorithm described in step S4 is: v i,d (t+1)=ω·v i，d (t)+c1·r1(pbest i,d -x i,d (t))+c2·r2(gbest d -x i，d (t)) The particle position update formula is: x i，d (t+1)=x i，d (t)+v i，d (t+1) Where ω is the inertia weight, with a value of 0.72; c1 and c2 are learning factors, with a value of 1.49; r1 and r2 are random numbers in the interval [0, 1]; pbest_{i, d} is the individual's historical best position; gbest_d is the global best position; the hyperparameter search range is: n_estimators∈[50, 1000], max_depth∈[3, 50], min_samples_split∈[2, 20], min_samples_leaf∈[1, 20]; the particle swarm size is set to 20, and the maximum number of iterations is 25.

6. The method according to claim 1, characterized in that, The key hyperparameters of the LSTM network mentioned in step S5 include the number of hidden layer units, dropout rate, learning rate, number of fully connected layer units, batch size, and number of training rounds; The genetic operations include: encoding hyperparameters into real chromosomes, iteratively evolving through roulette wheel selection, crossover, and mutation operations, searching for the optimal combination of hyperparameters using the macro-average F1-score as the fitness function, and then training an LSTM-assisted diagnostic model after decoding.

7. The method according to claim 1, characterized in that, The random forest classifier uses the Gini index (GI) and information gain (IG) as node splitting criteria, and constructs multiple decision trees through bootstrap sampling with replacement; Each time a node splits, some features are randomly selected as candidate splitting attributes, and the failure category is finally determined by combining the outputs of all trees using a voting mechanism. The goal is to minimize the expected classification error on the training set using an O-1 loss function.

8. A high-speed train bearing fault diagnosis system implementing the method of any one of claims 1 to 7, characterized in that, include: Signal acquisition module: used to acquire vibration acceleration signals from the drive end, fan end, and base. Preprocessing module: used to perform sampling rate unification, noise reduction, and speed normalization on the original signal; Feature extraction module: used to extract 30-dimensional multimodal features from the time domain, frequency domain, and time-frequency domain; Feature filtering module: used to perform recursive feature elimination and correlation analysis based on random forest to obtain 14 effective features; PSO-RF diagnostic module: used to run the particle swarm optimization algorithm to search for the optimal combination of hyperparameters in a random forest and output fault category labels; GA-LSTM Auxiliary Module: Used to run genetic algorithms to optimize LSTM hyperparameters and provide auxiliary diagnostic results; Evaluation output module: Used to calculate evaluation metrics such as accuracy, F1-score, and ROC-AUC, and output diagnostic reports.

9. A terminal device, characterized in that, The terminal device includes a memory and a processor; The memory is used to store a high-speed train bearing fault diagnosis program based on intelligent hyperparameter optimization. When the program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program; When the computer program is executed, it implements the steps of the method as described in any one of claims 1 to 7.

Images