Wheel polygon detection method and device for high-speed railway and medium

Abstract

The invention discloses a wheel polygon detection method and device for a high-speed railway and a medium, and relates to the technical field of data detection, and the method comprises the steps: carrying out the preprocessing of a collected vertical vibration acceleration signal of an axle box, and obtaining a low-frequency signal; obtaining a target short-time domain signal based on the low-frequency signal, and resampling the target short-time domain signal to obtain an angular domain signal; extracting multi-scale kurtosis features of the angular domain signals based on empirical mode decomposition; and constructing a neural network model, training the neural network model based on the polygonal axle box vibration signals of different orders and amplitudes, generating a wheel polygonal detection model, and obtaining a wheel damage identification result. Through combination of multi-scale kurtosis feature extraction and a neural network model, fine vibration changes caused by wheel damage can be captured from different frequencies and time scales, high-precision data acquisition and detection of polygon damage of wheels can be realized, damage of different orders and amplitudes can be accurately identified, and train operation safety is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of data detection, and particularly to a method, device and medium for detecting wheel polygons of high-speed railways. Background Art

[0002] Wheel polygon wear is a common wear form in high-speed trains. During the high-speed operation of the train, wheel polygon wear will exacerbate the interaction force and high-frequency impact vibration between the wheel and the rail, resulting in a sharp increase in the force on the wheel-rail and vehicle vibration damping components, damaging the track and vehicle components, increasing the noise between the wheel-rail and inside the vehicle, and seriously affecting the safety and stability of train operation. Therefore, timely detection of wheel polygons is crucial for preventing potential safety accidents.

[0003] However, the existing technologies have the following problems in detecting wheel polygons: Some static measurement methods have high requirements for the rotating platform, low monitoring frequency, and are difficult to meet the needs of real-time detection. On-vehicle detection and trackside detection usually indirectly measure the vibration or wheel-rail force changes caused by wheel polygons, with low accuracy and being easily affected by the environment. Some methods fail to fully utilize the multi-scale characteristics of signals during feature extraction, resulting in low detection accuracy.

[0004] In a complex operating environment, such as the variable speed condition of a heavy-haul locomotive, the monitoring signal has strong noise interference and fast frequency modulation characteristics. Especially, the sensors installed trackside are easily affected by electromagnetic interference, resulting in a large amount of noise in the collected signal and affecting the accuracy of the detection result.

[0005] Therefore, the existing technologies have problems such as inaccurate collected data, low detection accuracy, and poor environmental adaptability in wheel polygon detection. Summary of the Invention

[0006] The technical problems to be solved by the present invention are inaccurate collected data, low detection accuracy, and poor environmental adaptability. The purpose is to provide a method, device and medium for detecting wheel polygons of high-speed railways. Through the combination of multi-scale kurtosis feature extraction and a neural network model, it can achieve high-precision data collection and detection of wheel polygon damage, accurately identify damages of different orders and amplitudes, reduce false alarms and missed alarms, effectively solve multiple key problems in wheel polygon damage detection, improve the accuracy, reliability and real-time performance of detection, and provide strong support for the safe operation and maintenance of trains.

[0007] The present invention is achieved through the following technical solutions: The first aspect of the present invention provides a method for detecting wheel polygons of high-speed railways, including the following specific steps: Real-time collect the vertical vibration acceleration signal of the axle box of the running train; Preprocess the collected vertical vibration acceleration signal of the axle box to obtain a low-frequency signal after removing high-frequency noise; Extract the low-frequency signal based on a window function to obtain a target short-time domain signal; Resample the target short-time domain signal to obtain an angular domain signal; Extract the multi-scale kurtosis features of the angular domain signal based on empirical mode decomposition, including: performing empirical mode decomposition on the angular domain signal to obtain N intrinsic mode function components decomposed layer by layer; processing the N intrinsic mode function components to obtain multi-scale kurtosis features; Construct a neural network model, and train the neural network model based on the vibration signals of the polygon wheel axle box with different orders and amplitudes to generate a wheel polygon detection model; Obtain the wheel damage recognition result based on the wheel polygon detection model.

[0008] Further, the preprocessing of the collected vertical vibration acceleration signal of the axle box to obtain a low-frequency signal after removing high-frequency noise specifically includes: Decompose the vertical vibration acceleration signal of the axle box by using wavelet transform, perform threshold processing on the wavelet coefficients, and then reconstruct the signal to obtain a low-frequency signal; Perform Gaussian weighted moving average filtering on the low-frequency signal to obtain a smoothed signal; Eliminate abnormal signals from the smoothed signal to obtain the preprocessed signal.

[0009] Further, the extraction of the stable short-time domain signal based on the window function for the low-frequency signal specifically includes: Construct a sliding truncation function, set the length and step size of the sliding truncation function, and extract the initial short-time domain signal from the low-frequency signal according to the sliding truncation function; Divide the initial short-time domain signal into multiple segments; Calculate the consistency correlation coefficient between each initial short-time domain signal segment; Select the initial short-time domain signal segments with the consistency correlation coefficient within the set threshold as the target short-time domain signal.

[0010] Further, the calculation of the consistency correlation coefficient between each initial short-time domain signal segment specifically includes: Obtain the time series of each segment, perform spectral analysis and probability density analysis on the time series of each segment to obtain a spectral curve and a probability density curve; Calculate the consistent correlation coefficient of the spectral curve and the probability density curve; Obtain the average value of the consistent correlation coefficient; Store the short-time domain signals corresponding to the average coherence correlation coefficient greater than the set threshold in a temporary group, and select the short initial short-time domain signals with the coherence correlation coefficient within the set threshold from the temporary group as the target short-time domain signals.

[0011] Further, resampling the target short-time domain signal to obtain an angular domain signal specifically includes: Calculate the wheel speed curve according to the target short-time domain signal; Determine the angular change curve of the wheel according to the speed curve; Determine the time points corresponding to each equal angular interval according to the angular change curve; Calculate the signal values corresponding to the time points of each equal angular interval through interpolation to obtain the angular domain signal.

[0012] Further, processing the N intrinsic mode function components to obtain multi-scale kurtosis features specifically includes: Perform empirical mode decomposition on the angular domain signal to obtain N intrinsic mode function components decomposed layer by layer; Extract the k intrinsic mode function components with the highest correlation among the N mode function components as the main components based on the correlation coefficient principle; Based on the WVD time-frequency analysis method, obtain the signal time-frequency distributions of the k main components respectively; Based on the signal time-frequency distributions of the k main components, obtain the kurtosis values of multiple mode function components; Combine the kurtosis values of multiple mode function components to obtain a multi-scale kurtosis feature vector.

[0013] Further, constructing the wheel polygon detection model specifically includes: Construct a neural network model based on multi-scale kurtosis features and initialize the neural network parameters; Use the polygon wheel axle box vibration signals of different orders and amplitudes as the input of the neural network to construct the mapping relationship between multi-scale kurtosis features and wheel damage; Train the neural network model based on the mapping relationship to obtain the wheel polygon detection model.

[0014] Further, constructing the mapping relationship between multi-scale kurtosis features and wheel damage specifically includes: Perform time-frequency analysis on the polygon wheel axle box vibration signals of different orders and amplitudes using wavelet packet decomposition to obtain the wavelet packet node energy levels corresponding to the wheel polygon under different orders and amplitudes; Based on the wavelet packet node energy levels as the index for quantifying the polygon amplitude, use Kriging interpolation to establish the mapping relationship between wheel polygon features and axle box vibration features.

[0015] In a second aspect of the present invention, there is provided an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, a method for detecting wheel polygons of high-speed railways is implemented.

[0016] In a third aspect of the present invention, there is provided a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, a method for detecting wheel polygons of high-speed railways is implemented.

[0017] Compared with the prior art, the present invention has the following advantages and beneficial effects: Through multi-scale kurtosis feature extraction, it is possible to capture the subtle vibration changes caused by wheel damage from different frequencies and time scales. The multi-scale feature extraction method can effectively process vibration signals under complex working conditions. Based on the mapping relationship between multi-scale kurtosis features and wheel damage, a wheel polygon detection model is generated. The training and application processes of the neural network model are efficient, and it can quickly detect and classify wheel damage. It can be applied to real-time monitoring during train operation, timely detect wheel damage, effectively distinguish normal signals from damage signals, reduce false alarms and missed alarms, and ensure the safety of train operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following will briefly introduce the drawings required for the embodiments. It should be understood that the following drawings only show some embodiments of the present invention, and therefore should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts. In the drawings: Figure 1 is the wheel polygon detection process in the embodiment of the present invention; Figure 2 is the schematic diagram of a polygon wheel in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] To make the objectives, technical solutions, and advantages of the present invention more clearly understood, the following will further elaborate on the present invention in conjunction with the embodiments and the drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0020] As a possible implementation, such as Figure 1 and Figure 2As shown in the figure, this embodiment provides a method for detecting wheel polygons of high-speed railways, including the following specific steps: Collecting the vertical vibration acceleration signals of the axle boxes of the running train in real time; Preprocessing the collected vertical vibration acceleration signals of the axle boxes to obtain low-frequency signals after removing high-frequency noise; Extracting the low-frequency signals based on a window function to obtain target short-time domain signals; Resampling the target short-time domain signals to obtain angular domain signals; Extracting multi-scale kurtosis features of the angular domain signals based on empirical mode decomposition, including: Performing empirical mode decomposition on the angular domain signals to obtain N intrinsic mode function components decomposed layer by layer; Processing the N intrinsic mode function components to obtain multi-scale kurtosis features; Constructing a neural network model and training the neural network model based on the vibration signals of the polygon wheels of different orders and amplitudes to generate a wheel polygon detection model; Obtaining the wheel damage recognition result based on the wheel polygon detection model. Through the extraction of multi-scale kurtosis features, this embodiment can capture the subtle vibration changes caused by wheel damage from different frequencies and time scales. The multi-scale feature extraction method can effectively process vibration signals under complex working conditions. Based on the mapping relationship between multi-scale kurtosis features and wheel damage, a wheel polygon detection model is generated. The training and application process of the neural network model is efficient, which can quickly detect and classify wheel damage, be applicable to real-time monitoring during train operation, timely detect wheel damage, effectively distinguish normal signals from damage signals, reduce false alarms and missed alarms, and ensure the safety of train operation.

[0021] In some possible implementation manners, preprocessing the collected vertical vibration acceleration signals of the axle boxes to obtain low-frequency signals after removing high-frequency noise specifically includes: Selecting an appropriate wavelet basis and decomposing the vertical vibration acceleration signals of the axle boxes by wavelet transform to obtain wavelet coefficients; Performing threshold processing on the wavelet coefficients; Then reconstructing the signals to obtain low-frequency signals; Performing Gaussian weighted moving average filtering on the low-frequency signals to obtain smoothed signals, which is to further smooth the small fluctuations in the low-frequency signals, highlight the trend components, and based on Gaussian weighting, higher weights can be given to the central points, reduce phase delay, and retain the signal time series characteristics; Removing abnormal signals from the smoothed signals to obtain preprocessed signals; Removing abnormal signals from the smoothed signals includes eliminating abnormal values caused by sensor interference, instantaneous impact, etc.

[0022] In some possible embodiments, the wavelet coefficients include high-frequency detail coefficients and low-frequency approximation coefficients. Threshold processing of the wavelet coefficients includes: using the soft threshold or hard threshold method for the high-frequency wavelet coefficients to suppress the noise components. Only retaining the low-frequency approximation coefficients to reconstruct the signal, filtering out the high-frequency noise and irrelevant vibration components. By retaining the low-frequency components (such as bearing wear, structural resonance) that reflect the overall movement trend of the axle box and removing the high-frequency random noise (such as track impact, sensor noise), more accurate wheel data can be extracted.

[0023] In some possible embodiments, based on the window function, the low-frequency signal is extracted to obtain a stable short-time domain signal, specifically including: Construct a sliding truncation function. The sliding truncation divides the non-stationary signal into short-time quasi-stationary segments, effectively retaining the local time-frequency characteristics of the signal. Set the length and step size of the sliding truncation function, and extract the initial short-time domain signal from the low-frequency signal according to the sliding truncation function; wherein, the length and step size of the sliding truncation function are set according to the dominant frequency of the signal to ensure that the window contains at least 2 - 3 complete cycles Divide the initial short-time domain signal into multiple segments; then through the screening of the consistency correlation coefficient, eliminate the segments contaminated by noise or with abnormal interference to improve the signal quality; The screening of the consistency correlation coefficient includes: calculating the consistency correlation coefficient between each initial short-time domain signal segment; selecting the initial short-time domain signal segments with the consistency correlation coefficient within the set threshold as the target short-time domain signal.

[0024] In some possible embodiments, calculating the consistency correlation coefficient between each initial short-time domain signal segment specifically includes: First, convert the non-stationary vibration signal into a quasi-stationary short-time analysis unit for subsequent time-frequency and statistical characteristic analysis: divide the preprocessed signal according to a fixed duration, and obtain the time series of each segment according to the fixed duration to avoid the limitations of single-dimensional analysis; Based on the power spectral density estimation, perform spectral analysis and probability density analysis on the time series of each segment to obtain the spectral curve and the probability density curve; Calculate the consistent correlation coefficient of the spectral curve and the probability density curve; Obtain the average value of the consistent correlation coefficient: Calculate the mean and standard deviation of the consistent correlation coefficients of all segments as the global consistency benchmark; Store the short-time domain signals corresponding to the average consistent correlation coefficient greater than the set threshold (the segments with the consistent correlation coefficient greater than the sum of the mean and the standard deviation) in a temporary group, and select the short initial short-time domain signals with the consistent correlation coefficient within the set threshold from the temporary group as the target short-time domain signals, excluding the segments with extremely high consistency (possibly noise saturation) and low consistency (transient interference), and retaining the typical state signals.

[0025] In some possible embodiments, since the angular domain signal makes the fault characteristics (such as the bearing fault frequency) appear as a constant order, independent of the rotational speed and facilitating diagnosis, in order to avoid the non-stationarity of the time-domain signal caused by rotational speed fluctuations and enhance the extraction of fault characteristics, the target short-time domain signal is resampled to obtain the angular domain signal, which specifically includes: Calculate the wheel rotational speed curve according to the target short-time domain signal; Determine the angular change curve of the wheel according to the rotational speed curve; Traverse the angular change curve to determine the time points corresponding to each equal angular interval, that is, the time points satisfying the set threshold; Calculate the signal values corresponding to the time points corresponding to each equal angular interval by interpolation method to obtain the angular domain signal.

[0026] Let the target short-time domain signal be x(t), and the known time points corresponding to each equal angular interval be tn, and the signal values corresponding to the time points calculated by the interpolation method are x(tn).

[0027] In some possible embodiments, processing N intrinsic mode function components to obtain multi-scale kurtosis features specifically includes: Perform empirical mode decomposition on the angular domain signal to obtain N intrinsic mode function components decomposed layer by layer; Extract the k intrinsic mode function components with the highest correlation among the N mode function components as the main components based on the correlation coefficient principle; Based on the WVD time-frequency analysis method, respectively obtain the time-frequency distributions of the signals of the k main components; Based on the time-frequency distributions of the signals of the k main components, obtain the kurtosis values of multiple mode function components; Combine the kurtosis values of multiple mode function components to obtain a multi-scale kurtosis feature vector. Since the multi-scale time-frequency kurtosis combines the frequency band information of different mode function components, it can enhance the fault feature expression ability.

[0028] In some possible embodiments, the multi-scale kurtosis feature is a feature that can effectively characterize the non-Gaussian characteristics of the signal. By calculating the kurtosis of the signal at different scales, different frequency components and local features in the wheel axle box vibration signal can be captured.

[0029] In some possible embodiments, constructing a wheel polygon detection model specifically includes: Construct a neural network model based on the multi-scale kurtosis feature and initialize the neural network parameters; Collect the vibration signals of the polygonal wheel axle box with different orders and amplitudes. These signals can be measured through experiments or generated by simulation. For each wheel damage condition, record the corresponding vibration signals and calculate their multi-scale kurtosis features. Use the vibration signals of the polygonal wheel axle box with different orders and amplitudes collected as the input of the neural network, that is, use the calculated multi-scale kurtosis features as the input of the neural network, and the wheel damage state (such as damage degree, position, etc.) as the output. Train the neural network to make it learn the non-linear mapping relationship between the input features and the output damage state, and train the neural network model based on the mapping relationship to obtain the wheel polygon detection model.

[0030] In some possible implementation manners, after the training is completed, use an independent validation data set to verify the model and evaluate the generalization ability of the wheel polygon detection model for unknown data. According to the verification results, optimize the wheel polygon detection model and adjust the network structure, parameters, learning rate, etc. of the wheel polygon detection model to further improve the robustness of the model. The finally obtained neural network model can accurately detect the wheel polygon damage according to the multi-scale kurtosis features of the input vibration signals of the wheel axle box.

[0031] In some possible implementation manners, construct the mapping relationship between the multi-scale kurtosis features and the wheel damage, specifically including: Perform time-frequency analysis on the vibration signals of the polygonal wheel axle box with different orders and amplitudes by using wavelet packet decomposition to obtain the wavelet packet node energy levels corresponding to the wheel polygon under different orders and amplitudes; Based on the wavelet packet node energy levels as the index for quantifying the polygon amplitude, use Kriging interpolation to establish the mapping relationship between the wheel polygon features and the axle box vibration features.

[0032] In some possible implementation manners, perform time-frequency analysis on the vibration signals of the polygonal wheel axle box with different orders and amplitudes by using wavelet packet decomposition to obtain the wavelet packet node energy levels corresponding to the wheel polygon under different orders and amplitudes, specifically including: Perform time-frequency analysis on the vibration signals of the polygonal wheel axle box with different orders and amplitudes by using wavelet packet decomposition to obtain the wavelet packet node energy levels corresponding to the wheel polygon under different orders and amplitudes. According to the characteristics of the vibration signals of the wheel axle box, select appropriate wavelet bases and decomposition levels. Among them, the decomposition level determines the division accuracy of the signal in different frequency ranges, and the decomposition level is set according to the signal frequency. Input the vibration signals of the polygonal wheel axle box with different orders and amplitudes into the wavelet packet decomposition algorithm to obtain the wavelet packet coefficients of different nodes. The coefficients of each node reflect the energy distribution of the signal in a specific frequency range. Perform the sum of squares operation on the coefficients of each wavelet packet node to obtain the energy level of the corresponding node. These energy levels can characterize the characteristics of the wheel polygon under different orders and amplitudes.

[0033] In some possible embodiments, based on the wavelet packet node energy level as an index for quantifying the polygon amplitude, Kriging interpolation is used to establish the mapping relationship between the wheel polygon characteristics and the axle box vibration characteristics, which specifically includes: Taking the wavelet packet node energy level as the known point attribute value: The node energy level obtained by wavelet packet decomposition is used as the attribute value of the known point, and at the same time, the wheel polygon order and amplitude information corresponding to these known points are recorded; Determining the parameters of Kriging interpolation: According to the distribution of the known points, a suitable Kriging interpolation model (such as ordinary Kriging, simple Kriging, etc.) is selected. Determine the parameters in the model, such as the parameters of the semi-variance model (including nugget effect, sill value, range, etc.). These parameters can be obtained by fitting experimental data; Performing Kriging interpolation: Using the energy level of the known points and the corresponding wheel polygon characteristics, calculate the energy level of the unknown points through the Kriging interpolation formula. According to the interpolated energy level, the wheel polygon characteristics corresponding to the unknown points can be predicted, thereby establishing the mapping relationship between the wheel polygon characteristics and the axle box vibration characteristics.

[0034] As a possible implementation manner, this embodiment provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, a wheel polygon detection method for high-speed railways is implemented.

[0035] As a possible implementation manner, this embodiment provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, a wheel polygon detection method for high-speed railways is implemented.

[0036] The specific implementation manners described above further elaborate on the purpose, technical solutions, and beneficial effects of the present invention. It should be understood that the above description is only the specific implementation manners of the present invention and is not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A wheel polygon detection method for high-speed railway, characterized in that: The specific steps include: Real-time collection of vertical vibration acceleration signals of the axle box of running trains; Preprocess the collected vertical vibration acceleration signal of the axle box to obtain a low-frequency signal after removing high-frequency noise; The low-frequency signal is extracted based on the window function to obtain the target short-time domain signal; Resample the target short-time domain signal to obtain the angular domain signal; Extracting multi-scale kurtosis features of angular domain signals based on empirical mode decomposition, including: performing empirical mode decomposition on angular domain signals to obtain N intrinsic mode function components decomposed layer by layer; processing the N intrinsic mode function components to obtain multi-scale kurtosis features; Construct a neural network model, train the neural network model based on polygonal wheel axle box vibration signals of different orders and amplitudes, and generate a wheel polygon detection model; The wheel damage recognition results are obtained based on the wheel polygon detection model.

2. The wheel polygon detection method for high-speed railway according to claim 1, characterized in that: The preprocessing of the collected vertical vibration acceleration signal of the axle box to obtain a low-frequency signal after removing high-frequency noise specifically includes: The vertical vibration acceleration signal of the axle box is decomposed by wavelet transform, the wavelet coefficients are threshold processed, and then the signal is reconstructed to obtain a low-frequency signal. Perform Gaussian weighted moving average filtering on the low-frequency signal to obtain a smooth signal; Abnormal signals are removed from the smoothed signal to obtain a preprocessed signal.

3. The wheel polygon detection method for high-speed railway according to claim 1, characterized in that: The extracting of the low-frequency signal based on the window function to obtain a stable short-time domain signal specifically includes: Construct a sliding truncation function, set the length and step size of the sliding truncation function, and extract an initial short-time domain signal from the low-frequency signal according to the sliding truncation function; Dividing the initial short-time domain signal into a plurality of segments; Calculate the consistency correlation coefficient between each initial short-time domain signal segment; An initial short-time domain signal segment whose consistency correlation coefficient is within a set threshold is selected as the target short-time domain signal.

4. The wheel polygon detection method for high-speed railway according to claim 1, characterized in that: The calculating of the consistency correlation coefficient between each initial short-time domain signal segment specifically includes: Obtain the time series of each segment, perform spectrum analysis and probability density analysis on the time series of each segment, and obtain a spectrum curve and a probability density curve; Calculate the consistent correlation coefficient between the spectrum curve and the probability density curve; Get the average of the consistent correlation coefficients; The short time domain signals corresponding to the average consistent correlation coefficient greater than the set threshold are stored in a temporary group, and the short initial short time domain signals whose consistent correlation coefficient is within the set threshold are selected from the temporary group as the target short time domain signals.

5. The wheel polygon detection method for high-speed railway according to claim 1, characterized in that: The resampling of the target short-time domain signal to obtain the angle domain signal specifically includes: Calculate the wheel speed curve according to the target short-time domain signal; According to the speed curve, determine the angle change curve of the wheel; According to the angle change curve, determine the time point corresponding to each equal angle interval; The signal value corresponding to each time point corresponding to each equal angle interval is calculated by the interpolation method to obtain the angular domain signal.

6. The wheel polygon detection method for high-speed railway according to claim 5, characterized in that: The processing of the N intrinsic mode function components to obtain a multi-scale kurtosis feature specifically includes: Based on the correlation coefficient principle, the k modal function components with the highest correlation among the N modal function components are extracted as the main components; Based on the WVD time-frequency analysis method, the signal time-frequency distribution of k principal components is obtained respectively; Based on the signal time-frequency distribution of k principal components, the kurtosis values ​​of multiple modal function components are obtained; The kurtosis values ​​of multiple modal function components are combined to obtain a multi-scale kurtosis eigenvector.

7. The wheel polygon detection method for high-speed railway according to claim 1, characterized in that: The construction of the wheel polygon detection model specifically includes: Construct a neural network model based on multi-scale kurtosis features and initialize the neural network parameters; The polygonal wheel axle box vibration signals of different orders and amplitudes are used as the input of the neural network to construct the mapping relationship between multi-scale kurtosis features and wheel damage. The neural network model is trained based on the mapping relationship to obtain the wheel polygon detection model.

8. The wheel polygon detection method for high-speed railway according to claim 1, characterized in that: The constructing of the mapping relationship between the multi-scale kurtosis feature and the wheel damage specifically includes: Wavelet packet decomposition is used to perform time-frequency analysis on the vibration signals of polygonal wheel axle boxes of different orders and amplitudes, and the wavelet packet node energy levels corresponding to the wheel polygons of different orders and amplitudes are obtained. Based on the wavelet packet node energy level as an indicator to quantify the polygon amplitude, Kriging interpolation is used to establish the mapping relationship between the wheel polygon features and the axle box vibration features.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the wheel polygon detection method for high-speed railway is implemented as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the wheel polygon detection method for a high-speed railway as described in any one of claims 1 to 8 is implemented.

Images