A method for early fault diagnosis of train axle box bearings

Abstract

The present invention provides a method for early fault diagnosis of train axle box bearings, including obtaining train axle box bearing operation data to obtain an original signal data set of the early state of the train axle box bearings; balancing the early original signal data set and obtaining a preliminary time-frequency distribution, and obtaining a preliminary optimized time-frequency graph by reconstruction and denoising; fitting the main energy area of ​​the normalized preliminary optimized time-frequency graph using Gaussian distribution to obtain an enhanced time-frequency graph; obtaining a modal interval based on the denoised and enhanced data; dividing the two-dimensional time-frequency image after encoding processing by a preset ratio according to the state type to obtain the divided samples; and identifying the modal interval based on the trained MIA‑Mamba2 model. The present invention can solve the problem of data imbalance, retain weak signals and remove noise, and effectively improve the problems of incomplete feature extraction and excessive complexity in the early fault diagnosis process of train axle box bearings.

Description

Technical Field

[0001] The invention relates to the technical field of train component fault diagnosis, and in particular to an early fault diagnosis method for a train axle box bearing. Background Art

[0002] With the rapid development of trains, trains have become the preferred means of transportation, so the safety of trains is particularly important. Train axle box bearings refer to a type of axle box bearing design used in railway vehicle bogies. In rail transit, bogies are important components that connect wheels and car bodies. Due to their long-term operation in closed, high-temperature, heavy-load, random alternation, multiple degrees of freedom and nonlinear harsh environments, axle box bearings are one of the components in rail vehicles that are prone to failure. Carrying out early fault diagnosis, identifying potential fault signs in advance, and avoiding the occurrence of faults are of great significance to the safe operation of trains.

[0003] In the existing technology, (1) the energy of the early fault signal of the train bearing is weak and difficult to be detected and identified; it exists in the strong environmental noise interference, the signal-to-noise ratio is low, and the recognition is poor; the weak characteristic signal is often overlapped with other source signals, and the feature recognition is difficult; it is difficult to obtain the characteristics of the early fault in the time domain diagram and spectrum diagram of the original signal;

[0004] (2) Uncertainty caused by data imbalance: In actual working conditions, the data collected by sensors contains much more data on bearings in normal working condition than on faulty conditions, and the amount of data on different early fault types is also unequal. Currently, most early fault diagnosis methods are based on balanced data sets, which results in poor diagnostic effects on unbalanced data sets.

[0005] (3) Accidental uncertainty caused by the harsh working environment of the bearing, sensor accuracy, data acquisition, transmission and storage, and cognitive uncertainty caused by insufficient human understanding of the structural characteristics of the bearing, especially the bogie axle box, and model errors. The deviation caused by the above single type of uncertainty may not be large, but the errors caused by the mixed uncertainty may cause a large deviation in the system response, thereby reducing the reliability of the bearing diagnosis results. Summary of the invention

[0006] Based on this, the purpose of the present invention is to provide a method for early fault diagnosis of train axle box bearings to solve the deficiencies in the above-mentioned prior art.

[0007] The present invention provides a method for diagnosing early faults of train axle box bearings, the method comprising:

[0008] Acquiring train axle box bearing operation data to obtain an early original signal data set of the early state of the train axle box bearing;

[0009] Using synthetic minority class oversampling technology and undersampling technology to balance the early original signal data set, and performing continuous wavelet transform on the early original signal data set to obtain a preliminary time-frequency distribution, and obtaining a two-dimensional time-frequency image based on the preliminary time-frequency distribution;

[0010] The balanced early original signal data set is denoised using a dual-parameter three-segment wavelet threshold denoising method, the denoised wavelet coefficients are amplified or retained at a predetermined scale to highlight the frequency band characteristics of the early original signal data set, and the denoised early original signal data set is reconstructed by an inverse wavelet transform to obtain a preliminary optimized time-frequency diagram;

[0011] Normalizing the preliminary optimized time-frequency diagram, simplifying the complex energy distribution of the normalized preliminary optimized time-frequency diagram through two-dimensional Gaussian fitting to obtain a concentrated representation, and using Gaussian distribution parameters to characterize the main energy area of ​​the normalized preliminary optimized time-frequency diagram to enhance the representation of early weak fault signals in the time-frequency diagram, thereby obtaining an enhanced time-frequency diagram;

[0012] transforming the enhanced time-frequency diagram to obtain a modal interval based on the enhanced time-frequency diagram;

[0013] Performing encoding processing on the two-dimensional time-frequency image, and dividing the two-dimensional time-frequency image after encoding processing according to the state type by a preset ratio to obtain divided samples;

[0014] A MIA-Mamba2 model is established, and the divided samples are input into the MIA-Mamba2 model so that the MIA-Mamba2 model is trained based on the divided samples to obtain a trained MIA-Mamba2 model, and the modal interval is identified based on the trained MIA-Mamba2 model.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: by adopting the synthetic minority class oversampling technology and the undersampling technology to balance the original signal data set, the problem of data imbalance is solved; by adopting the dual-parameter three-segment wavelet threshold denoising method to denoise the balanced original signal data set, the characteristics of the weak signal are retained while the noise is removed; the preliminary time-frequency distribution is obtained by continuous wavelet transform, which can retain the high-quality vibration signal characteristics; and the early fault diagnosis characteristics are further enhanced by two-dimensional Gaussian fitting, which accurately reflects the operating status of the train axle box bearing under different early fault conditions; at the same time, with the help of the modal interval method, the uncertainty problem of the train axle box bearing is further dealt with; finally, based on the MIA-Mamba2 model, the feature extraction and pattern recognition classification tasks are completed, and the MIA-Mamba2 model can assist in determining the specific type and location of the early fault.

[0016] Furthermore, the step of acquiring the train axle box bearing operation data to obtain an early original signal data set of the early state of the train axle box bearing includes:

[0017] Acquiring the operating data of the train axle box bearing based on the sensor device;

[0018] The operation data is converted into digital signals and stored in a PC to obtain the early original signal data set.

[0019] Furthermore, the step of balancing the early original signal data set by using the synthetic minority class oversampling technology and the undersampling technology includes:

[0020] Selecting nearest neighbor samples of several fault types in the early original signal data set based on the Euclidean distance;

[0021] Based on the synthetic minority class oversampling method and in a linear interpolation manner, new synthetic samples are generated, and based on the undersampling technology, blurred areas in the synthetic samples are eliminated.

[0022] Furthermore, the step of obtaining a two-dimensional time-frequency image based on the preliminary time-frequency distribution includes:

[0023] Calculating the instantaneous frequency of each time-frequency point of the preliminary time-frequency distribution by using the phase derivative of the wavelet transform result;

[0024] The energy of the wavelet transform coefficients is distributed to the corresponding instantaneous frequencies.

[0025] Furthermore, the step of denoising the balanced early original signal data set using a dual-parameter three-segment wavelet threshold denoising method includes:

[0026] Performing wavelet transformation on the balanced early original signal data set, and converting the early original signal data set into time-frequency domain to decompose into wavelet coefficients of different scales and frequency bands;

[0027] A dual parameter is introduced to control the threshold processing of the wavelet coefficients to obtain the processed wavelet coefficients, wherein the threshold processing is a three-segment threshold function, and the expression of the three-segment threshold function is:

[0028] ;

[0029] In the formula, represents the three-segment threshold function, Represents the first parameter, Represents the second parameter, represents the wavelet coefficients;

[0030] The processed wavelet coefficients are subjected to inverse wavelet transform and normalization processing.

[0031] Furthermore, the expression of the continuous wavelet transform is:

[0032] ;

[0033] In the formula, represents the time-frequency coefficient, represents the scale parameter, represents the time displacement parameter, represents the time domain signal, represents the mother wavelet function, Indicates time, Represents the amplitude of the time domain signal.

[0034] Furthermore, the normalized preliminary optimized time-frequency diagram is simplified by two-dimensional Gaussian fitting to obtain a concentrated representation of the complex energy distribution of the normalized preliminary optimized time-frequency diagram, and Gaussian distribution parameters are used to characterize the main energy area of ​​the normalized preliminary optimized time-frequency diagram to enhance the representation of the early weak fault signal in the time-frequency diagram. The step of obtaining the enhanced time-frequency diagram includes:

[0035] Normalizing the preliminary optimized time-frequency graph and extracting initial parameters of the two-dimensional Gaussian distribution to obtain the energy centroid, standard deviation in time, standard deviation in frequency direction and amplitude of the preliminary optimized time-frequency graph;

[0036] Constructing a fitting model and extracting Gaussian fitting parameters to generate a Gaussian distribution diagram, and superimposing the Gaussian distribution diagram on the preliminary optimized time-frequency diagram to obtain an enhanced time-frequency diagram to enhance the representation of early weak fault signals in the time-frequency diagram;

[0037] Among them, the expression of the energy center of mass is:

[0038]

[0039] In the formula, Indicates time, Indicates frequency, represents the energy value in the time-frequency graph, It represents the weighted average value of energy in the time-frequency diagram in the time direction, and represents the central position of energy in time. It represents the weighted average value of energy in the frequency direction in the time-frequency diagram, indicating the center position of energy in frequency;

[0040] The calculation expressions of the standard deviation in time and the standard deviation in frequency direction are:

[0041] ;

[0042] In the formula, Indicates the standard deviation of time, representing the discrete degree of energy distribution of the time-frequency graph on the time axis. Indicates the standard deviation in the frequency direction, representing the degree of discreteness of the energy distribution of the time-frequency graph on the frequency axis;

[0043] The calculation expression of the amplitude is:

[0044] ;

[0045] In the formula, It represents the maximum value of energy in the time-frequency diagram and the peak value of energy distribution.

[0046] Furthermore, the expression for constructing the fitting model and extracting the Gaussian fitting parameters is:

[0047] ;

[0048] In the formula, Represents the fitted two-dimensional Gaussian distribution, representing the energy concentration area of ​​the time-frequency diagram;

[0049] The Gaussian distribution diagram is superimposed on the preliminary optimized time-frequency diagram to obtain an expression of the enhanced time-frequency diagram:

[0050] ;

[0051] In the formula, represents the enhanced time-frequency diagram, represents the enhancement weight coefficient, Represents the fitted two-dimensional Gaussian distribution, representing the energy concentration area of ​​the time-frequency graph, represents the time-frequency diagram after denoising, ,in, represents the time-frequency coefficient, represents the scale parameter, Represents the time displacement parameter.

[0052] Furthermore, the mathematical definition expression of the modal interval is:

[0053] ;

[0054] In the formula, represents the modal interval, represents the lower bound, represents the upper bound, represents an interval in the modal interval.

[0055] Further, the step of inputting the divided samples into the MIA-Mamba2 model so that the MIA-Mamba2 model is trained based on the divided samples to obtain the trained MIA-Mamba2 model, and identifying the modal interval based on the trained MIA-Mamba2 model includes:

[0056] Inputting the training set in the divided samples into the MIA-Mamba2 model, so that the MIA-Mamba2 model is trained based on the training set and through the loss function to obtain a trained MIA-Mamba2 model;

[0057] Validating the trained MIA-Mamba2 model based on the validation set in the divided samples;

[0058] The test set in the divided samples is input into the trained MIA-Mamba2 model, and the modal interval is identified by the trained MIA-Mamba2 model to output a diagnosis result. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 The present invention is a flow chart of a method for diagnosing early faults of train axle box bearings in an embodiment of the present invention.

[0060] The following specific implementation manner will further illustrate the present invention in conjunction with the above-mentioned drawings. DETAILED DESCRIPTION

[0061] In order to facilitate the understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. Several embodiments of the present invention are given in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive.

[0062] It should be noted that when an element is referred to as being "fixed to" another element, it may be directly on the other element or there may be a central element. When an element is considered to be "connected to" another element, it may be directly connected to the other element or there may be a central element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are for illustrative purposes only.

[0063] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which the present invention belongs. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.

[0064] See also Figure 1 , which shows a method for diagnosing early faults of train axle box bearings in an embodiment of the present invention, the method comprises steps S1 to S7:

[0065] S1, acquiring the train axle box bearing operation data to obtain the original signal data set of the early state of the train axle box bearing;

[0066] Specifically, the step S1 includes steps S11 to S12:

[0067] S11, acquiring the operating data of the train axle box bearing based on a sensor device;

[0068] S12, converting the operation data into digital signals and storing them in a PC to obtain the original signal data set;

[0069] It can be understood that the train axle box bearing is taken as the research object, and an experimental platform is built. The experimental equipment includes: an acceleration sensor, a data acquisition card, a bogie model, and a PC, etc. The acceleration sensor is installed in the vertical direction of the bearing end cover of the train axle box bearing on the experimental platform to obtain the operating data of the train axle box bearing. The data acquisition card receives the data collected by the acceleration sensor, converts the analog signal into a digital signal, and stores it in the PC to form the original signal data set of the early state of the train axle box bearing. In this embodiment, the original signal data set is recorded as .

[0070] S2, using a synthetic minority class oversampling technique and an undersampling technique to balance the early original signal data set, and performing a continuous wavelet transform on the early original signal data set to obtain a preliminary time-frequency distribution, and obtaining a two-dimensional time-frequency image based on the preliminary time-frequency distribution;

[0071] Specifically, the step S2 includes steps S21 to S24:

[0072] S21, selecting nearest neighbor samples of several fault types in the original signal data set based on the Euclidean distance;

[0073] S22, generating new synthetic samples based on a synthetic minority class oversampling method and by linear interpolation, and eliminating fuzzy areas in the synthetic samples based on an undersampling technique;

[0074] S23, calculating the instantaneous frequency of each time-frequency point of the preliminary time-frequency distribution by using the phase derivative of the wavelet transform result;

[0075] S24, distributing the energy of the wavelet transform coefficient to the corresponding instantaneous frequency;

[0076] It is understandable that in the early failure of train axle box bearings, due to the sporadic nature of bearing failures, there is an imbalance in the samples of various types of failures, which greatly affects the accuracy of pattern recognition. It is proposed to use a method combining synthetic minority class oversampling technology and undersampling technology to balance the early failure data. For each sample of the minority fault type , first select its The selection of the nearest neighbor samples of the minority class is usually based on the Euclidean distance, and the expression of the Euclidean distance is:

[0077]

[0078] In the formula, , Respectively represent samples No. feature values, samples No. eigenvalues, express and The distance between represents the feature dimension, The index representing the feature;

[0079] After finding the minority class samples of After the nearest neighbors, the synthetic minority class oversampling technique uses linear interpolation to Generate new synthetic samples between its nearest neighbor samples Synthetic Sample The generation formula is:

[0080] ;

[0081] In the formula, represents a synthetic sample, Represents a sample, Indicates from of A sample randomly selected from the nearest neighbors, Represents a random number used to control the location of synthetic sample generation;

[0082] The new samples generated by the synthetic minority oversampling technique may appear near the boundary between the minority and majority class samples, causing the decision boundary to become unclear and increasing confusion. This situation can easily lead to the classifier being unable to distinguish between faulty and normal class samples well. Therefore, based on the synthetic minority oversampling technique, undersampling technology is used to identify and delete majority class samples that are too close to minority class samples, and clean up the fuzzy areas on the decision boundary. This allows the classifier to better learn clear category boundaries, which helps reduce misclassification. Specifically, undersampling technology is defined as: for sample pairs , if the following two conditions are met, then constitutes an undersampling, and Belonging to different categories, samples It is a sample The most recent sample, and the sample It is a sample The most recent sample is:

[0083] ;

[0084] In the formula, express and The distance between Represents a collection A sample point in Indicates the sample we are looking for Cannot be a sample itself), Representing a collection A sample point in Indicates the sample we are looking for Cannot be a sample itself), represents a sample set, Represents searching for a set All not equal to Among the sample points, The distance to the sample point with the smallest distance, Represents searching for a set All not equal to Among the sample points, The distance of the sample point with the smallest distance is usually calculated using the Euclidean distance.

[0085] Once an undersampling is detected , to check their category labels, if Belongs to the majority class, then delete ,if Belongs to the majority class, then delete In this way, by combining the synthetic minority class oversampling technology and undersampling technology, the data set can be balanced and the classification performance of the model can be enhanced.

[0086] S3, using a dual-parameter three-segment wavelet threshold denoising method to denoise the balanced early original signal data set, amplifying or retaining the denoised wavelet coefficients at a predetermined scale to highlight the frequency band characteristics of the early original signal data set, and reconstructing the denoised early original signal data set through an inverse wavelet transform to obtain a preliminary optimized time-frequency diagram;

[0087] Specifically, the step S3 includes steps S31 to S33:

[0088] S31, performing wavelet transform on the balanced early original signal data set, and converting the early original signal data set into time-frequency domain to decompose into wavelet coefficients of different scales and frequency bands;

[0089] S32, introducing dual parameters to control the threshold processing of the wavelet coefficients to obtain processed wavelet coefficients, wherein the threshold processing is a three-segment threshold function, and the expression of the three-segment threshold function is:

[0090] S33, performing inverse wavelet transform on the processed wavelet coefficients and performing normalization processing;

[0091] It should be noted that the expression of the continuous wavelet transform is:

[0092] ;

[0093] In the formula, represents the time-frequency coefficient, represents the scale parameter, represents the time displacement parameter, represents the time domain signal, represents the mother wavelet function, Indicates time, Represents the amplitude of the time domain signal.

[0094] It is understandable that a suitable function such as Morlet wavelet is usually selected; is the scale parameter, which is related to frequency. The smaller it is, the higher the frequency; is the time displacement parameter; It is the time-frequency coefficient of wavelet transform, which is used to describe the signal at different times and scale The following features.

[0095] The time-frequency distribution is obtained by wavelet transform After that, the instantaneous frequency of each time-frequency point needs to be calculated , which can be estimated by the phase derivative of the wavelet transform result. The formula for instantaneous frequency is:

[0096] ;

[0097] in, express The phase of Represents the symbol of partial derivative.

[0098] By calculating the instantaneous frequency After that, the next step is to transform the wavelet transform coefficients The energy is redistributed to the corresponding frequency Position. By compressing the energy, a higher frequency resolution is obtained. The specific formula for synchronous compression is:

[0099] ;

[0100] in: is the time-frequency distribution after synchronous compression, indicating that the signal In time and frequency The energy density at is the Dirac function, which transforms the wavelet transform result The energy in the The corresponding frequency position Department; is the scale normalization term.

[0101] It is worth mentioning that the dual-parameter three-segment wavelet threshold denoising method is an innovative signal denoising method, which mainly achieves a more efficient denoising effect by optimizing the traditional wavelet threshold denoising algorithm. The core principle of this method is to introduce dual parameters and three-segment threshold functions, so as to achieve more flexible threshold control in different signal processing stages, avoid signal distortion and effectively remove noise. In the specific implementation, the signal is first transformed by wavelet transform, and the time domain signal is converted into time-frequency domain so as to decompose it into wavelet coefficients of different scales and frequency bands. Then the dual-parameter adjustment mechanism and the three-segment threshold function are introduced to perform threshold processing on each layer of wavelet coefficients. Dual parameter adjustment: Use two adjustable parameters to control the strength and smoothness of threshold processing. The first parameter Controls the coefficient threshold of the low-value area. The coefficients in this area are usually regarded as noise and are directly set to zero. The second parameter The coefficients of the high-value area are controlled. Most of these coefficients are useful signals and remain unchanged. The expression of the three-segment threshold function is:

[0102] ;

[0103] In the formula, represents the three-segment threshold function, Represents the first parameter, Represents the second parameter, represents the wavelet coefficients;

[0104] It needs to be explained that the low value range , directly transform the wavelet coefficients Set to 0; middle range : Smooth adjustment of these coefficients; high value range : Leave these coefficients unchanged.

[0105] It is understandable that through the dual parameters and Control the threshold processing of wavelet coefficients. Perform inverse wavelet transform on the processed wavelet coefficients to reconstruct the denoised signal. Then normalize the denoised signal to obtain the data set. , which is also a time domain signal.

[0106] S4, normalizing the preliminary optimized time-frequency diagram, simplifying the complex energy distribution of the normalized preliminary optimized time-frequency diagram by two-dimensional Gaussian fitting, so as to obtain a concentrated representation, and using Gaussian distribution parameters to characterize the main energy area of ​​the normalized preliminary optimized time-frequency diagram, so as to enhance the representation of the early weak fault signal in the time-frequency diagram, and obtain an enhanced time-frequency diagram;

[0107] Specifically, the step S4 includes step S41 to step S42:

[0108] S41, normalizing the preliminary optimized time-frequency graph and extracting initial parameters of a two-dimensional Gaussian distribution to obtain the energy centroid, time standard deviation, frequency standard deviation and amplitude of the preliminary optimized time-frequency graph;

[0109] S42, constructing a fitting model and extracting Gaussian fitting parameters to generate a Gaussian distribution diagram, and superimposing the Gaussian distribution diagram on the preliminary optimized time-frequency diagram to obtain an enhanced time-frequency diagram to enhance the representation of the early weak fault signal in the time-frequency diagram;

[0110] Among them, the expression of the energy center of mass is:

[0111]

[0112] In the formula, Indicates time, Indicates frequency, represents the energy value in the time-frequency graph, It represents the weighted average value of energy in the time-frequency diagram in the time direction, and represents the central position of energy in time. It represents the weighted average value of energy in the frequency direction in the time-frequency diagram, indicating the center position of energy in frequency;

[0113] The calculation expressions of the standard deviation in time and the standard deviation in frequency direction are:

[0114] ;

[0115] In the formula, Indicates the standard deviation of time, representing the discrete degree of energy distribution of the time-frequency graph on the time axis. Indicates the standard deviation in the frequency direction, representing the degree of discreteness of the energy distribution of the time-frequency graph on the frequency axis;

[0116] The calculation expression of the amplitude is:

[0117] ;

[0118] In the formula, Indicates the maximum value of energy in the time-frequency diagram, representing the peak value of energy distribution;

[0119] The expression for constructing the fitting model and extracting the Gaussian fitting parameters is:

[0120] ;

[0121] In the formula, Represents the fitted two-dimensional Gaussian distribution, representing the energy concentration area of ​​the time-frequency diagram;

[0122] The Gaussian distribution diagram is superimposed on the preliminary optimized time-frequency diagram to obtain the expression of the enhanced time-frequency diagram:

[0123] ;

[0124] In the formula, represents the enhanced time-frequency diagram, represents the enhancement weight coefficient, Represents the fitted two-dimensional Gaussian distribution, representing the energy concentration area of ​​the time-frequency graph, represents the time-frequency diagram after denoising, ,in, represents the time-frequency coefficient, represents the scale parameter, Represents the time displacement parameter.

[0125] S5, transforming the enhanced time-frequency diagram to obtain a modal interval based on the enhanced time-frequency diagram;

[0126] It is understandable that there may be uncertainty and other issues in the processing of train axle box bearings. The noise-reduced data is transformed into the form of modal intervals, which are recorded as , , in which, is the lower limit of the train traction motor bearing data set, is the lower limit of the train traction motor bearing data set. The mathematical definition expression of the modal interval is:

[0127] ;

[0128] In the formula, represents the modal interval, represents the lower bound, represents the upper bound, represents an interval in the modal interval. In this embodiment, the algorithm basis of the modal interval comes from the Kaucher algorithm.

[0129] S6, encoding the two-dimensional time-frequency image, and dividing the two-dimensional time-frequency image after encoding according to the state type by a preset ratio to obtain divided samples;

[0130] It can be understood that the vibration time-frequency diagrams of various state types after the train axle box bearing conversion are encoded and processed, and the various state types of the train axle box bearing are proportionally divided into training samples, verification samples and test samples according to 6:2:2. In this example, four working conditions are set: normal, air gap eccentricity, bearing failure and rotor bar breakage, with 1000 sets of data for each working condition, for a total of 4000 sets of data.

[0131] S7, establishing a MIA-Mamba2 model, inputting the divided samples into the MIA-Mamba2 model, so that the MIA-Mamba2 model is trained based on the divided samples to obtain a trained MIA-Mamba2 model, and identifying the modal interval based on the trained MIA-Mamba2 model;

[0132] It should be noted that the MIA-Mamba2 model includes: state space model (SSM), selection scanning mechanism (S6) and cross scanning module (CSM), which together constitute the MIA-Mamba2 model. The MIA-Mamba2 model is a further extension of the Mamba2 model, which performs modal interval processing on the parameters of Mamba2. In the output part of the MIA-Mamba2 model, combined with the analysis results of MIA, a deeper understanding and analysis of abnormal phenomena is carried out. At the same time, image features related to modal parameter abnormalities are extracted to assist in determining the type and location of early faults.

[0133] Specifically, the step S7 includes steps S71 to S73:

[0134] S71, inputting the training set in the divided samples into the MIA-Mamba2 model, so that the MIA-Mamba2 model is trained based on the training set and through a loss function, to obtain a trained MIA-Mamba2 model;

[0135] S72, verifying the trained MIA-Mamba2 model based on the verification set in the divided samples;

[0136] It should be explained that the MIA-Mamba2 model originated from the need for time series data modeling. It combines the characteristics of the state space model (SSM) and the Transformer. Mamba2 integrates the state transfer matrix of the SSM with the attention mechanism, making up for the high computational complexity of the traditional attention mechanism.

[0137] Input layer: The modal area of ​​the time-frequency graph is converted into a serialized feature representation through a convolutional layer (such as CNN) for processing by subsequent modules.

[0138] State Space Model (SSM): Local feature extraction,The state space model (SSM) extracts local dependencies of time series through a state transfer matrix (A matrix) and input weight matrix (B matrix).

[0139] A dynamic evolution model of the state vector is established to simulate the changes of local features in the time series through the state transfer matrix A and the input control matrix B. The convolution form of SSM can effectively capture the change rules of time series signals in a short period of time.

[0140] The state space formula is as follows:

[0141]

[0142] in, represents the state space, represents the state vector, Represents input data.

[0143] Selective scanning mechanism (S6): Based on the attention mechanism, S6 uses a mask matrix to limit the attention range, so that the model can focus on important areas, while only selecting the time-frequency areas most relevant to the current time point for modeling, avoiding the interference of irrelevant areas on feature extraction.

[0144] Cross Scanning Module (CSM): It is a feature integration module that enables the model to understand global context information. Through the cross scanning mechanism, the features of the local area in the time-frequency graph are gradually integrated into the global features.

[0145] Classification prediction stage: The test set in the divided samples is input into the trained MIA-Mamba2 model, the time-frequency graph of the input test set is modally binned and converted into a serialized feature representation through a convolutional layer, the SSM module extracts local features of the time-frequency graph, the S6 module captures global patterns and long-distance dependencies, and the CSM module fuses local and global features to generate the final feature representation. The final feature representation is input into the fully connected classification layer, and the early fault category of the train axle box bearing is output.

[0146] S73, inputting the test set in the divided samples into the trained MIA-Mamba2 model, and identifying the modal interval by the trained MIA-Mamba2 model to output a diagnosis result;

[0147] It should be explained that the two-dimensional time-frequency images are input into the MIA-Mamba2 model as training sets and validation sets according to the divided ratio. Select a suitable loss function, such as the cross entropy loss function, to measure the difference between the model prediction output and the true label during the parameter optimization process, and guide the model to continuously adjust the parameters through optimization methods such as gradient descent; introduce dynamic learning rate scheduling such as cosine annealing learning rate, and dynamically adjust the learning rate to ensure that the model can converge to the best performance; introduce regularization technology (such as Dropout) to improve the robustness of the model and avoid overfitting when the CSM module extracts too many redundant features. The convergence of the model is judged by indicators such as accuracy, recall, and F1 value on the validation set. When the performance on the validation set reaches stability, it indicates that the model has been trained.

[0148] In addition, the divided test set is input into the trained train axle box bearing MIA-Mamba2 model, and the output results are obtained through the model. Combined with the analysis results of MIA, a deeper understanding and analysis of the abnormal situation is carried out. At the same time, the image features related to the abnormal modal parameters are identified to determine the type of early fault.

[0149] In summary, the early fault diagnosis method for train axle box bearings in the above-mentioned embodiments of the present invention balances the original signal data set by adopting synthetic minority class oversampling technology and undersampling technology, thereby solving the problem of data imbalance, and denoises the balanced original signal data set by adopting a dual-parameter three-segment wavelet threshold denoising method, thereby retaining the characteristics of weak signals while removing noise, and obtaining a preliminary time-frequency distribution through continuous wavelet transform, which can retain high-quality vibration signal characteristics, and further enhance the early fault diagnosis characteristics through two-dimensional Gaussian fitting, accurately reflecting the operating status of the train axle box bearings under different early fault conditions, and at the same time, with the help of the modal interval method, further dealing with the uncertainty problem of the train axle box bearings, and finally, completing the feature extraction and pattern recognition classification tasks based on the MIA-Mamba2 model, and the MIA-Mamba2 model can assist in determining the specific type and location of the early fault.

[0150] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0151] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for early fault diagnosis of train axle box bearings, characterized in that: The method comprises: Acquiring train axle box bearing operation data to obtain an early original signal data set of the early state of the train axle box bearing; Using synthetic minority class oversampling technology and undersampling technology to balance the early original signal data set, and performing continuous wavelet transform on the early original signal data set to obtain a preliminary time-frequency distribution, and obtaining a two-dimensional time-frequency image based on the preliminary time-frequency distribution; The balanced early original signal data set is denoised using a dual-parameter three-segment wavelet threshold denoising method, the denoised wavelet coefficients are amplified or retained at a predetermined scale to highlight the frequency band characteristics of the early original signal data set, and the denoised early original signal data set is reconstructed by an inverse wavelet transform to obtain a preliminary optimized time-frequency diagram; Normalizing the preliminary optimized time-frequency diagram, simplifying the complex energy distribution of the normalized preliminary optimized time-frequency diagram through two-dimensional Gaussian fitting to obtain a concentrated representation, and using Gaussian distribution parameters to characterize the main energy area of ​​the normalized preliminary optimized time-frequency diagram to enhance the representation of early weak fault signals in the time-frequency diagram, thereby obtaining an enhanced time-frequency diagram; transforming the enhanced time-frequency diagram to obtain a modal interval based on the enhanced time-frequency diagram; Performing encoding processing on the two-dimensional time-frequency image, and dividing the two-dimensional time-frequency image after encoding processing according to the state type at a preset ratio to obtain divided samples; A MIA-Mamba2 model is established, and the divided samples are input into the MIA-Mamba2 model so that the MIA-Mamba2 model is trained based on the divided samples to obtain a trained MIA-Mamba2 model, and the modal interval is identified based on the trained MIA-Mamba2 model.

2. The method for early fault diagnosis of train axle box bearings according to claim 1, characterized in that: The step of acquiring the train axle box bearing operation data to obtain an early original signal data set of the early state of the train axle box bearing comprises: Acquiring the operating data of the train axle box bearing based on the sensor device; The operation data is converted into digital signals and stored in a PC to obtain the early original signal data set.

3. The method for early fault diagnosis of train axle box bearings according to claim 1, characterized in that: The step of balancing the early original signal data set by using the synthetic minority class oversampling technology and the undersampling technology comprises: Selecting nearest neighbor samples of several fault types in the early original signal data set based on the Euclidean distance; Based on the synthetic minority class oversampling method and in a linear interpolation manner, new synthetic samples are generated, and based on the undersampling technology, blurred areas in the synthetic samples are eliminated.

4. The method for early fault diagnosis of train axle box bearings according to claim 1, characterized in that: The step of obtaining a two-dimensional time-frequency image based on the preliminary time-frequency distribution comprises: Calculating the instantaneous frequency of each time-frequency point of the preliminary time-frequency distribution by using the phase derivative of the wavelet transform result; The energy of the wavelet transform coefficients is distributed to the corresponding instantaneous frequencies.

5. The method for early fault diagnosis of train axle box bearings according to claim 1, characterized in that: The step of denoising the balanced early original signal data set using a dual-parameter three-segment wavelet threshold denoising method comprises: Performing wavelet transformation on the balanced early original signal data set, and converting the early original signal data set into time-frequency domain to decompose into wavelet coefficients of different scales and frequency bands; A dual parameter is introduced to control the threshold processing of the wavelet coefficients to obtain the processed wavelet coefficients, wherein the threshold processing is a three-segment threshold function, and the expression of the three-segment threshold function is: ； In the formula, represents the three-segment threshold function, Represents the first parameter, Represents the second parameter, represents the wavelet coefficients; The processed wavelet coefficients are subjected to inverse wavelet transform and normalization processing.

6. The method for early fault diagnosis of train axle box bearings according to claim 1, characterized in that: The expression of the continuous wavelet transform is: ； In the formula, represents the time-frequency coefficient, represents the scale parameter, represents the time displacement parameter, represents the time domain signal, represents the mother wavelet function, Indicates time, Represents the amplitude of the time domain signal.

7. The method for early fault diagnosis of train axle box bearings according to claim 1, characterized in that: The step of normalizing the preliminary optimized time-frequency diagram, simplifying the complex energy distribution of the normalized preliminary optimized time-frequency diagram by two-dimensional Gaussian fitting to obtain a concentrated representation, and using Gaussian distribution parameters to characterize the main energy area of ​​the normalized preliminary optimized time-frequency diagram to enhance the representation of the early weak fault signal in the time-frequency diagram, and obtaining the enhanced time-frequency diagram includes: Normalizing the preliminary optimized time-frequency graph and extracting initial parameters of the two-dimensional Gaussian distribution to obtain the energy centroid, standard deviation in time, standard deviation in frequency direction and amplitude of the preliminary optimized time-frequency graph; Constructing a fitting model and extracting Gaussian fitting parameters to generate a Gaussian distribution diagram, and superimposing the Gaussian distribution diagram on the preliminary optimized time-frequency diagram to obtain an enhanced time-frequency diagram to enhance the representation of early weak fault signals in the time-frequency diagram; Among them, the expression of the energy center of mass is: In the formula, Indicates time, Indicates frequency, represents the energy value in the time-frequency graph, It represents the weighted average value of energy in the time-frequency diagram in the time direction, and represents the central position of energy in time. It represents the weighted average value of energy in the frequency direction in the time-frequency diagram, indicating the center position of energy in frequency; The calculation expressions of the standard deviation in time and the standard deviation in frequency direction are: ； In the formula, Indicates the standard deviation of time, representing the discrete degree of energy distribution of the time-frequency graph on the time axis. Indicates the standard deviation in the frequency direction, representing the degree of discreteness of the energy distribution of the time-frequency graph on the frequency axis; The calculation expression of the amplitude is: ； In the formula, It represents the maximum value of energy in the time-frequency diagram and the peak value of energy distribution.

8. The method for early fault diagnosis of train axle box bearings according to claim 7, characterized in that: The expression for constructing the fitting model and extracting the Gaussian fitting parameters is: ； In the formula, Represents the fitted two-dimensional Gaussian distribution, representing the energy concentration area of ​​the time-frequency diagram; The Gaussian distribution diagram is superimposed on the preliminary optimized time-frequency diagram to obtain the expression of the enhanced time-frequency diagram: ； In the formula, represents the enhanced time-frequency diagram, represents the enhancement weight coefficient, Represents the fitted two-dimensional Gaussian distribution, representing the energy concentration area of ​​the time-frequency graph, represents the time-frequency diagram after denoising, ,in, represents the time-frequency coefficient, represents the scale parameter, Represents the time displacement parameter.

9. The method for early fault diagnosis of train axle box bearings according to claim 1, characterized in that: The mathematical definition expression of the modal interval is: ； In the formula, represents the modal interval, represents the lower bound, represents the upper bound, Represents an interval in the modal interval.

10. The method for early fault diagnosis of train axle box bearings according to claim 1, characterized in that: The step of inputting the divided samples into the MIA-Mamba2 model so that the MIA-Mamba2 model is trained based on the divided samples to obtain the trained MIA-Mamba2 model, and identifying the modal interval based on the trained MIA-Mamba2 model includes: Inputting the training set in the divided samples into the MIA-Mamba2 model, so that the MIA-Mamba2 model is trained based on the training set and through the loss function to obtain a trained MIA-Mamba2 model; Validating the trained MIA-Mamba2 model based on the validation set in the divided samples; The test set in the divided samples is input into the trained MIA-Mamba2 model, and the modal interval is identified by the trained MIA-Mamba2 model to output a diagnosis result.

Images