Wheel polygon fault diagnosis method, device and equipment based on EEMD (ensemble empirical mode decomposition) and LSTM (long short term memory), and medium

Abstract

The invention discloses a wheel polygon fault diagnosis method, device and equipment based on EEMD and LSTM and a medium, and relates to the technical field of wheel detection, and the method comprises the steps: collecting a current vibration signal during the operation of a wheel, carrying out the mode decomposition of the current vibration signal based on an EEMD decomposition method, obtaining a plurality of IMF components, and carrying out the detection of the IMF components; screening out a target IMF component containing most tread impact characteristics from the IMF components based on a maximum similarity principle; performing signal reconstruction based on the target IMF component, and performing feature extraction based on the reconstructed signal to obtain a current target feature set; inputting the current target feature set into a target diagnosis LSTM model to output a fault diagnosis result; the training data of the model comprises a historical target feature set corresponding to the historical vibration signal and a label which corresponds to the historical target feature set and is used for representing whether the polygon fault of the wheel exists or not. And the fault essential information is fully captured and the accuracy of the wheel polygon fault diagnosis is improved.

Description

Technical Field

[0001] This invention relates to the field of wheel inspection technology, and in particular to a method, apparatus, equipment and medium for diagnosing polygonal wheel faults based on EEMD and LSTM. Background Technology

[0002] With the increase in locomotive / train capacity and speed, wheel wear has significantly increased, leading to a significant rise in the probability of polygonal wheel failures. When a polygonal failure occurs in a locomotive, it intensifies the interaction between the wheel and rail, affecting the locomotive's operational performance. The most direct hazards are vibration and noise. On the one hand, noise affects the crew's work, making it difficult to hear the train control commands in the driver's cab, endangering operational safety. On the other hand, vibration can cause malfunctions in related components on the locomotive bogie frame, such as broken primary and secondary damping springs, exploded electrical converter modules, and loose or faulty electrical wiring connections, seriously impacting maintenance, production, and locomotive quality. Currently, polygonal failures cannot be completely cured; they can only be mitigated through tread resurfacing. Therefore, it is necessary to monitor the development of wheel polygonal failures to provide timely guidance and reference for wheel resurfacing.

[0003] Currently, locomotive wheel polygon fault detection systems mainly consist of TPDS (Truck Performance Detecting System, i.e., freight car operation status ground safety monitoring system) and the 6A running gear fault monitoring subsystem. TPDS is a trackside detection device that calculates the wheel polygon order and assesses the severity of wheel polygon faults by collecting wheel-rail forces. However, this method cannot achieve real-time monitoring of the train, nor can it directly determine the severity of polygon faults. The 6A running gear fault monitoring subsystem diagnoses faults through impact signals from axle box measuring points. This device can identify wheel polygon faults and analyze their order, but it still faces difficulties in quickly and accurately extracting key information from massive amounts of detection data. Furthermore, its ability to process nonlinear and non-stationary signals is limited, making it difficult to fully capture the essential information of the fault, thus affecting the accuracy and reliability of fault diagnosis.

[0004] In summary, how to fully capture the essential information of faults and improve the accuracy of wheel polygon fault diagnosis is a problem that needs to be solved in this field. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a method, apparatus, device, and medium for wheel polygon fault diagnosis based on EEMD and LSTM, which fully captures the essential information of the fault and improves the accuracy of wheel polygon fault diagnosis. The specific solution is as follows: In a first aspect, this application discloses a wheel polygon fault diagnosis method based on EEMD and LSTM, including: The current vibration signal of the wheel during operation is collected, and the current vibration signal is decomposed into multiple IMF components based on the EEMD decomposition method. The target IMF component containing the most tread impact features is selected from the IMF components based on the maximum similarity principle. The signal is reconstructed based on the target IMF components, and the current target feature set is obtained by feature extraction based on the reconstructed signal. The current target feature set is input into the target diagnosis LSTM model to output the fault diagnosis result corresponding to the current vibration signal; wherein, the training data of the target diagnosis LSTM model is the historical target feature set corresponding to the historical vibration signal and the label corresponding to the historical target feature set used to characterize whether there is a wheel polygon fault.

[0006] Optionally, the modal decomposition of the current vibration signal based on the EEMD decomposition method to obtain multiple IMF components includes: White noise is added to the current vibration signal in multiple batches to obtain the noise-added signal for each batch; The noisy signal is decomposed using EMD to obtain multiple initial IMF components and decomposition residuals with different serial numbers in each batch; wherein the serial number indicates the position of the initial IMF component in each batch. The initial IMF components of all batches with the same sequence number are averaged to obtain the final IMF component corresponding to each sequence number.

[0007] Optionally, adding white noise to the current vibration signal in multiple batches includes: Determine the standard deviation of the current vibration signal, and determine the target amplitude of the white noise based on the standard deviation; Multiple batches of white noise, each with the target amplitude, are added to the current vibration signal.

[0008] Optionally, the step of selecting the target IMF component containing the most tread impact features from the IMF components based on the maximum similarity principle includes: The similarity between the final IMF component corresponding to each sequence number and the current vibration signal is determined respectively; IMF components that meet preset conditions are selected from the final IMF components and summed up to obtain the target IMF component.

[0009] Optionally, determining the similarity between the final IMF component corresponding to each sequence number and the current vibration signal includes: The correlation coefficient between the final IMF component corresponding to each sequence number and the current vibration signal is determined, and the correlation coefficient is determined as the similarity between the final IMF component corresponding to each sequence number and the current vibration signal.

[0010] Optionally, the step of selecting IMF components that meet preset conditions from each of the final IMF components and accumulating them as the target IMF component includes: The final IMF components corresponding to each sequence number are sorted in descending order of similarity to obtain the IMF component sequence. IMF components that meet the preset similarity criteria are selected from the IMF component sequence, and the IMF components that meet the preset criteria are accumulated. The accumulated result is the target IMF component that contains the most tread impact features.

[0011] Optionally, the signal reconstruction based on the target IMF component includes: The target IMF components are summed at the corresponding time points to obtain the reconstructed signal.

[0012] Optionally, the target features in the current target feature set can be any one of the following features: the time-domain kurtosis of the current vibration signal, the dominant frequency of the wheel tread, the amplitude of the dominant frequency, the dB value of the tread, and the polygon order of the wheel.

[0013] Optionally, the step of extracting features from the reconstructed signal to obtain the current target feature set includes: Determine the signal mean of the reconstructed signal, and use the signal mean and the reconstructed signal to obtain the time-domain kurtosis of the current vibration signal; The reconstructed signal after removing the mean is subjected to Fourier transform to obtain a frequency domain signal, and the amplitude spectrum of the frequency domain signal is generated. The target frequency point with the largest amplitude in the amplitude spectrum is found, and the frequency and amplitude of the target frequency point are determined as the main frequency and main frequency amplitude of the wheel tread, respectively. Determine the fault characteristic spectrum of the wheel, and determine the theoretical high-order spectrum at each high-order number based on the axle curing characteristic spectrum of the wheel tread and each high-order fault. If it is determined that there is a fault impact on the wheel tread based on the theoretical high-order spectrum, then the fault impact is quantified to obtain the tread dB value. Multiple spectral lines to be verified are generated at each order using the fault feature spectral number, and target spectral lines that meet preset prominence conditions are selected from the multiple spectral lines to be verified. The polygon order of the wheel is determined based on the spectral amplitude of the target spectral lines.

[0014] Optionally, determining the fault characteristic spectrum of the wheel includes: If the current vibration signal sampling scenario is speed tracking sampling, then the number of wheel sampling revolutions is determined as the fault characteristic spectrum of the wheel; If the current vibration signal is sampled using an astronomical clock, the fault characteristic spectrum of the wheel is determined based on the wheel speed, the number of signal sampling points, and the sampling frequency.

[0015] Optionally, if the wheel tread is determined to have a fault impact based on the theoretical higher-order spectral lines, then the fault impact is quantified, including: If there is a prominent spectral line in the amplitude spectrum within a first preset range to the left and right of the theoretical higher-order spectral line at the current higher order, then the average value of all spectral lines within a second preset range to the left and right of the prominent spectral line is determined, and a check value is determined based on the average value. If the amplitude of the prominent spectral line is greater than the check value, then it is determined that the current tread higher-order spectral line is prominent and there is a wheel tread fault. The number of high-order numbers for all wheel tread faults is counted. If the number of high-order numbers is greater than a preset threshold, it is determined that there is a fault impact on the wheel tread, and the fault impact is quantified.

[0016] Optionally, quantifying the fault impact to obtain the tread dB value includes: Obtain the product between the rotational speed of the axle containing the wheel and the diameter of the axle containing the wheel; The ratio between the impact value of the fault impact and the product is obtained, and the logarithm of the ratio is taken to obtain the tread dB value.

[0017] Optionally, the step of selecting target spectral lines that meet preset prominence conditions from multiple spectral lines to be verified, and determining the polygon order of the wheel based on the spectral amplitude of the target spectral lines, includes: If the amplitude of the current spectral line to be verified at the current order to be verified is greater than the preset prominence threshold, then the current spectral line to be verified is determined to be the target spectral line that meets the preset prominence condition, so as to obtain the spectral line amplitude of the target spectral line at the current order to be verified. The order with the largest spectral amplitude of the target spectral line among all orders to be verified is determined as the polygon order of the wheel.

[0018] Optionally, inputting the current target feature set into the target diagnosis LSTM model includes: The current target feature set is cleaned and / or its length is processed to obtain a processed current target feature set, which is then input into the target diagnostic LSTM model.

[0019] Optionally, the step of performing data cleaning and / or data length processing on the current target feature set to obtain a processed current target feature set includes: Replace the outliers in the current target feature set with preset feature values ​​to obtain the cleaned current target feature set; Determine the baseline target features from the cleaned current target feature set or from each target feature in the current target feature set; Data padding and / or data truncation are performed on the cleaned current target feature set or the target features in the current target feature set other than the benchmark target feature according to the length of the benchmark target feature, so as to obtain the current target feature set after length processing; The cleaned current target feature set or the length-processed current target feature set is determined as the processed current target feature set.

[0020] Optionally, determining the baseline target feature from each target feature in the cleaned current target feature set includes: Determine the number of measurement points for each target feature in the current target feature set after cleaning, and determine the target feature with the largest number of measurement points as the baseline target feature; Accordingly, the step of filling the target features in the cleaned current target feature set, excluding the benchmark target feature, with data according to the length of the benchmark target feature includes: The target features in the current cleaned target feature set, excluding the baseline target features, are determined as target features to be filled. The preset filling feature value is filled into the target feature to be filled according to the length of the reference target feature.

[0021] Optionally, before inputting the current target feature set into the target diagnosis LSTM model, the method further includes: Construct an initial diagnostic LSTM model that includes an input layer, an LSTM layer, a dynamic dropout layer, and a fully connected layer; Collect historical vibration signals of the wheel during operation, obtain the historical target feature set corresponding to the historical vibration signals, and obtain the label corresponding to the historical target feature set to characterize whether there is a polygonal fault in the wheel. Construct training data and verification set including the historical target feature set and the label corresponding to the historical target feature set. The initial diagnostic LSTM model is trained using the training data to obtain the trained diagnostic LSTM model. The trained diagnostic LSTM model is validated using the validation set. If the validation result indicates that the trained diagnostic LSTM model meets the preset fine-tuning stopping condition, then the trained diagnostic LSTM model is determined as the target diagnostic LSTM model. If the validation result indicates that the trained diagnostic LSTM model does not meet the preset fine-tuning stopping condition, then the trained diagnostic LSTM model is fine-tuned, and the fine-tuned diagnostic LSTM model is determined as the target diagnostic LSTM model.

[0022] Optionally, the input layer is used to convert the training data into a three-dimensional tensor, the LSTM layer is used to extract temporal features from the three-dimensional tensor, the dynamic discard layer is used to discard some of the temporal features, and the fully connected layer is used to obtain the fault probability value based on the discarded temporal features; wherein, the three-dimensional tensor includes the number of training samples in a single training iteration, the temporal length of a single training sample, and the feature dimension of a single time step.

[0023] Secondly, this application discloses a wheel polygon fault diagnosis device based on EEMD and LSTM, comprising: The component acquisition module is used to collect the current vibration signal when the wheel is running, perform modal decomposition on the current vibration signal based on the EEMD decomposition method to obtain multiple IMF components, and select the target IMF component containing the most tread impact features from the IMF components based on the maximum similarity principle. The feature extraction module is used to reconstruct the signal based on the target IMF components, and to extract features from the reconstructed signal to obtain the current target feature set. The result output module is used to input the current target feature set into the target diagnosis LSTM model to output the fault diagnosis result corresponding to the current vibration signal; wherein, the training data of the target diagnosis LSTM model is the historical target feature set corresponding to the historical vibration signal and the label corresponding to the historical target feature set used to characterize whether there is a wheel polygon fault.

[0024] Thirdly, this application discloses an electronic device, including: Memory, used to store computer programs; A processor is used to execute the computer program to implement the steps of the aforementioned disclosed method for diagnosing wheel polygon faults based on EEMD and LSTM.

[0025] Fourthly, this application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the aforementioned disclosed wheel polygon fault diagnosis method based on EEMD and LSTM.

[0026] The beneficial effects of this application are as follows: This application collects the current vibration signal of the wheel during operation, performs modal decomposition on the current vibration signal based on the EEMD decomposition method to obtain multiple IMF components, and selects the target IMF component containing the most tread impact features from the IMF components based on the maximum similarity principle; performs signal reconstruction based on the target IMF component, and performs feature extraction based on the reconstructed signal to obtain the current target feature set; inputs the current target feature set into the target diagnosis LSTM model to output the fault diagnosis result corresponding to the current vibration signal; wherein, the training data of the target diagnosis LSTM model is the historical target feature set corresponding to the historical vibration signal and the label corresponding to the historical target feature set used to characterize whether there is a polygonal fault in the wheel. Therefore, this application utilizes the EEMD decomposition method to perform modal decomposition on the current vibration signal, which effectively suppresses modal aliasing and improves the accuracy of signal analysis. Based on the principle of maximum similarity, it selects the target IMF component containing the most tread impact features and reconstructs the signal, concentrating the vibration components related to tread impact in the original signal and reducing vibration interference caused by non-tread factors. This makes the reconstructed signal closer to the actual operating state of the wheel, providing high-quality data support for subsequent feature extraction. Extracting the current target feature set related to the wheel polygon fault from the reconstructed signal accurately quantifies fault-related characteristics, providing a representative and discriminative basis for fault diagnosis. Inputting the current target feature set into the target diagnosis LSTM model leverages the LSTM model's advantages in processing time-series data and the learned correlation between historical target feature sets and fault labels, enabling effective diagnosis of wheel polygon faults. Furthermore, the target diagnosis LSTM model trained based on historical target feature sets and related labels corresponding to historical vibration signals can fully explore long-term dependencies and fault characteristic patterns in the data, improving the accuracy and reliability of fault diagnosis. Simultaneously, the entire process requires minimal manual intervention, achieving automated signal processing, feature extraction, and fault diagnosis, thus improving the efficiency of fault diagnosis. Attached Figure Description

[0027] 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0028] Figure 1 This is a flowchart of a wheel polygon fault diagnosis method based on EEMD and LSTM disclosed in this application; Figure 2This is a schematic diagram of a polygonal wheel fault diagnosis device based on EEMD and LSTM disclosed in this application. Figure 3 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0029] The technical solutions of 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 the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0030] With the increase in locomotive / train capacity and speed, wheel wear has significantly increased, leading to a significant rise in the probability of polygonal wheel failures. When a polygonal failure occurs in a locomotive, it intensifies the interaction between the wheel and rail, affecting the locomotive's operational performance. The most direct hazards are vibration and noise. On the one hand, noise affects the crew's work, making it difficult to hear the train control commands in the driver's cab, endangering operational safety. On the other hand, vibration can cause malfunctions in related components on the locomotive bogie frame, such as broken primary and secondary damping springs, exploded electrical converter modules, and loose or faulty electrical wiring connections, seriously impacting maintenance, production, and locomotive quality. Currently, polygonal failures cannot be completely cured; they can only be mitigated through tread resurfacing. Therefore, it is necessary to monitor the development of wheel polygonal failures to provide timely guidance and reference for wheel resurfacing.

[0031] Currently, locomotive wheel polygon fault detection systems mainly consist of TPDS (Truck Performance Detecting System, i.e., freight car operation status ground safety monitoring system) and the 6A running gear fault monitoring subsystem. TPDS is a trackside detection device that calculates the wheel polygon order and assesses the severity of wheel polygon faults by collecting wheel-rail forces. However, this method cannot achieve real-time monitoring of the train, nor can it directly determine the severity of polygon faults. The 6A running gear fault monitoring subsystem diagnoses faults through impact signals from axle box measuring points. This device can identify wheel polygon faults and analyze their order, but it still faces difficulties in quickly and accurately extracting key information from massive amounts of detection data. Furthermore, its ability to process nonlinear and non-stationary signals is limited, making it difficult to fully capture the essential information of the fault, thus affecting the accuracy and reliability of fault diagnosis.

[0032] Therefore, this application provides a wheel polygon fault diagnosis scheme based on EEMD and LSTM, which fully captures the essential information of the fault and improves the accuracy of wheel polygon fault diagnosis.

[0033] See Figure 1 As shown in the figure, this application discloses a method for diagnosing polygonal wheel faults based on EEMD and LSTM, including: Step S11: Collect the current vibration signal of the wheel during operation, perform modal decomposition on the current vibration signal based on the EEMD decomposition method to obtain multiple IMF components, and select the target IMF component containing the most tread impact features from the IMF components based on the maximum similarity principle.

[0034] The current vibration signal during wheel operation is collected. The vibration signal can be any one or both of acceleration and impact signals. Ensemble Empirical Mode Decomposition (EEMD) is applied to this current vibration signal. EEMD is an improvement on Empirical Mode Decomposition (EMD). While EMD can decompose complex signals into a series of Intrinsic Mode Functions (IMFs), it suffers from mode aliasing. EEMD, by adding white noise to the original signal multiple times before EMD decomposition and averaging the results, effectively suppresses mode aliasing and more accurately extracts the signal's different scale features. The final result is the multiple IMF components of the current vibration signal after EEMD decomposition.

[0035] In this embodiment, the modal decomposition of the current vibration signal based on the EEMD decomposition method to obtain multiple IMF components includes: adding white noise to the current vibration signal in multiple batches to obtain the noisy signal corresponding to each batch; performing EMD decomposition on the noisy signal to obtain multiple initial IMF components with different indices and decomposition residuals in each batch; wherein, the indices represent the position of the initial IMF components in each batch; and averaging the initial IMF components with the same indices in all batches to obtain the final IMF components corresponding to each indices.

[0036] White noise with an amplitude of 0.1 to 0.2 times the standard deviation of the original signal is added to multiple batches of the current vibration signal to obtain the noisy signal corresponding to each batch. EEMD improves the signal discontinuity caused by intermittent noise or impact pulses in the original signal by adding noise to the original signal. Finally, the interference caused by white noise is suppressed by averaging multiple sets of IMFs. Therefore, EEMD can suppress mode aliasing and improve the accuracy of signal analysis. Empirical mode decomposition is performed on the noisy signals of each batch to obtain the initial IMF components and decomposition residuals of multiple different sequences under each batch; that is, the initial IMF components and decomposition residuals of different output positions under each batch are obtained. The initial IMF components of the same sequence number of all batches are integrated and averaged to obtain the final IMF component corresponding to each sequence number, thus completing the EEMD mode decomposition of the current vibration signal.

[0037] In this embodiment, adding white noise to the current vibration signal in multiple batches includes: determining the standard deviation of the current vibration signal, and determining the target amplitude of the white noise based on the standard deviation; adding white noise with an amplitude of the target amplitude to the current vibration signal in multiple batches.

[0038] First, calculate the standard deviation of the current vibration signal. Then, determine 0.1 to 0.2 times this standard deviation as the target amplitude of the white noise. Add white noise to the current vibration signal in multiple batches according to this target amplitude, with each batch of white noise having the same target amplitude. That is, add white noise of the same amplitude to the current vibration signal x multiple times. The formula is as follows: ; In the formula, a is the amplitude of white noise, which is generally taken as 0.1 to 0.2 times the standard deviation of the original signal; The noise is white noise (i = 1, 2, 3, ..., N), where N represents the number of times noise is added to the original signal; Let be the noisy signal corresponding to the i-th iteration.

[0039] The noisy signal is decomposed using EMD to obtain the initial IMF components and decomposition residuals for multiple different serial numbers in each batch. The specific formula is shown below: ; In the formula: Let j be the j-th IMF component of the i-th noise-added signal; n is the number of IMF components; Let be the decomposition residual of the i-th noisy signal.

[0040] The final IMF component for each batch with the same serial number is obtained by averaging the initial IMF components. The specific formula is as follows: ; In the above process, EEMD adds noise to the original signal to improve the signal discontinuity caused by intermittent noise or impulse pulses in the original signal. Finally, it suppresses the interference caused by white noise by averaging multiple IMFs. Therefore, EEMD can both suppress mode aliasing and improve the accuracy of signal analysis.

[0041] In this embodiment, the step of selecting the target IMF component containing the most tread impact features from the IMF components based on the maximum similarity principle includes: determining the similarity between the final IMF component corresponding to each sequence number and the current vibration signal; selecting IMF components that meet preset conditions from each of the final IMF components, and accumulating them as the target IMF component.

[0042] The correlation coefficient method is used to determine the similarity between the final IMF component corresponding to each sequence number and the current vibration signal. Components that meet the preset similarity conditions are selected from each final IMF component. These selected IMF components are superimposed and accumulated to obtain the target IMF component.

[0043] In this embodiment, determining the similarity between the final IMF component corresponding to each sequence number and the current vibration signal includes: determining the correlation coefficient between the final IMF component corresponding to each sequence number and the current vibration signal, and determining the correlation coefficient as the similarity between the final IMF component corresponding to each sequence number and the current vibration signal.

[0044] For each sequence number corresponding to the final IMF component, the correlation coefficient method is used to calculate its correlation coefficient with the current vibration signal. The formula for calculating the correlation coefficient is as follows: ; The correlation coefficient reflects the degree of linear correlation between the two, and its value ranges between [-1, 1]. The closer the absolute value is to 1, the higher the similarity between the two. The calculation process is based on the number of signal sampling points. After deducting the average value of the current vibration signal and the average value of the final IMF component of the corresponding sequence number, the solution is completed. The correlation coefficient calculated in this way is directly determined as the similarity between the final IMF component corresponding to each sequence number and the current vibration signal.

[0045] In this embodiment, the step of selecting IMF components that meet preset conditions from each of the final IMF components and accumulating them as the target IMF component includes: sorting the final IMF components corresponding to each sequence number in descending order of similarity to obtain an IMF component sequence; selecting IMF components that meet preset conditions of similarity from the IMF component sequence, and accumulating each IMF component that meets the preset conditions, with the accumulation result serving as the target IMF component containing the most tread impact features.

[0046] Based on the order of similarity between the final IMF component corresponding to each sequence number and the current vibration signal from high to low, all final IMF components are sorted to form an ordered IMF component sequence. Then, IMF components with similarity meeting preset conditions are selected from this IMF component sequence. These components are considered to contain more wheel tread impact information. Specifically, a preset number of IMF components can be selected from the IMF component sequence. The number of IMF components meeting the preset conditions can be determined according to actual needs and signal characteristics. Generally, an appropriate number is selected through experiments or experience to ensure that interference from non-tread factors can be effectively removed while retaining sufficient tread impact characteristics. Then, the IMF components with the same sequence number meeting the preset conditions are added point by point. The accumulated result is taken as the target IMF component that concentrates the tread impact characteristics in the original signal and contains the most tread impact characteristics.

[0047] Step S12: Reconstruct the signal based on the target IMF component, and extract features based on the reconstructed signal to obtain the current target feature set.

[0048] In this embodiment, the signal reconstruction based on the target IMF component includes: adding the target IMF components at corresponding time points to obtain the reconstructed signal.

[0049] The target IMF components containing a large amount of wheel tread impact information are selected as the core reconstruction basis. The signals of each target IMF component are summed point by point at the corresponding time point. In this way, vibration interference components caused by non-tread factors are removed from the original vibration signal, and the core vibration characteristics related to wheel tread impact are retained. Finally, the reconstructed signal focusing on the tread impact signal is obtained. The specific formula is as follows: ; Where S is the index set of the target IMF components, and y(t) is the reconstructed signal. The reconstructed signal concentrates the vibration components most related to tread impact in the original signal, reduces vibration interference caused by non-tread factors, and thus is closer to the actual state.

[0050] In this embodiment, the target features in the current target feature set are any one of the following features: the time-domain kurtosis of the current vibration signal, the dominant frequency of the wheel tread, the amplitude of the dominant frequency, the dB value of the tread, and the polygon order of the wheel.

[0051] The target features in the current target feature set are the time-domain kurtosis of the current vibration signal extracted from the reconstructed signal, the dominant frequency of the wheel tread, the amplitude of the dominant frequency, the dB value of the tread, and the polygon order of the wheel. Among them, temporal kurtosis is a time-domain statistical measure that reflects the sharpness or flatness of the signal distribution pattern. It is highly sensitive to impact signals and can serve as an important basis for detecting the presence and degree of impact in a signal. The dominant frequency of the wheel tread refers to the frequency component in which the signal energy is most concentrated in the frequency domain. It represents the main vibration frequency characteristics of the tread and can reflect the vibration characteristics of the tread and the frequency information of possible faults. The amplitude of the dominant frequency is the amplitude value of the signal at the dominant frequency, which reflects the energy of the signal at the main vibration frequency component. Its change can intuitively reflect the severity of the fault. The tread dB value is a characteristic parameter that feeds back the presence and magnitude of tread impact information and is an important basis for analyzing the health status of the tread. The polygon order of the wheel is the core characteristic parameter for determining whether the current sample is caused by polygonal impact of the wheel. When no tread dB value is detected or the fault is not classified as polygonal, the polygon order is 0. When a dB value is detected and the fault is determined to be polygonal, the corresponding polygon order is extracted.

[0052] In this embodiment, the step of extracting features based on the reconstructed signal to obtain the current target feature set includes: determining the signal mean of the reconstructed signal; using the signal mean and the reconstructed signal to obtain the temporal kurtosis of the current vibration signal; performing a Fourier transform on the reconstructed signal after removing the mean to obtain a frequency domain signal, and generating the amplitude spectrum of the frequency domain signal; finding the target frequency point with the largest amplitude in the amplitude spectrum, and determining the frequency and amplitude of the target frequency point as the dominant frequency and amplitude of the wheel tread, respectively; determining the fault feature spectrum of the wheel; determining the theoretical high-order spectral lines at each high order based on the axle solidification feature spectrum of the wheel tread and each high-order fault; if it is determined that there is a fault impact on the wheel tread based on the theoretical high-order spectral lines, then quantifying the fault impact to obtain the tread dB value; generating multiple unverified spectral lines at each unverified order using the fault feature spectrum, and selecting the target spectral line that meets the preset prominence conditions from the multiple unverified spectral lines; and determining the polygon order of the wheel based on the spectral amplitude of the target spectral line.

[0053] First, the mean value of the reconstructed signal is calculated. Then, according to the formula for calculating time-domain kurtosis, the reconstructed signal value and the mean value are substituted into the calculation to obtain the time-domain kurtosis of the current vibration signal. This time-domain kurtosis can reflect the sharpness or flatness of the signal distribution pattern and has high sensitivity to wheel tread impact signals. It can serve as an important basis for detecting whether there is an impact component and the degree of impact in the signal. The specific calculation formula is as follows: ; In the formula, N is the number of signal sampling points, and y(t) is the reconstructed signal. This is the signal mean of the reconstructed signal.

[0054] The reconstructed signal after mean removal undergoes a Fast Fourier Transform (FFT) to convert the time-domain reconstructed signal into a frequency-domain signal FFTData(t). The amplitude spectrum |FFTData(t)| of this frequency-domain signal is then calculated. The target frequency point fmax with the largest amplitude is found within the amplitude spectrum. The frequency value corresponding to this target frequency point is determined as the dominant frequency of the wheel tread. This dominant frequency represents the main vibration frequency characteristics of the tread and reflects the vibration characteristics and potential fault frequency information of the tread. Simultaneously, the amplitude value corresponding to this target frequency point is determined as the dominant frequency amplitude. This amplitude reflects the energy of the signal at the main vibration frequency components, and its changes directly reflect the severity of the wheel tread fault. It is important to note that in actual calculations, since the frequency resolution after the FFT transformation is related to the number of sampling points and the sampling frequency, it may be necessary to perform appropriate interpolation on the frequency value corresponding to the target frequency point and determine the dominant frequency of the wheel tread from the interpolated frequency value to improve the accuracy of the dominant frequency extraction.

[0055] In this embodiment, determining the fault characteristic spectrum of the wheel includes: if the sampling scenario of the current vibration signal is speed tracking sampling, then the number of wheel sampling revolutions is determined as the fault characteristic spectrum of the wheel; if the sampling scenario of the current vibration signal is astronomical clock sampling, then the fault characteristic spectrum of the wheel is determined according to the wheel speed, the number of signal sampling points, and the sampling frequency.

[0056] If the current vibration signal sampling scenario is speed tracking sampling, then the number of wheel sampling revolutions is directly determined as the wheel's fault characteristic spectrum fn, which serves as the basis for analyzing wheel tread faults. If the current vibration signal sampling scenario is astronomical clock sampling, then the wheel's fault characteristic spectrum fn is calculated using the formula combined with the wheel speed V (in r / min), the number of signal sampling points N, and the sampling frequency fs. The calculation formula is as follows: ; This calculation method is adapted to the signal characteristics of astronomical clock sampling and can accurately obtain the core frequency parameters reflecting wheel tread faults in the corresponding sampling scenarios, providing an accurate frequency reference for subsequent searching of tread fault spectrum lines and higher-order spectrum lines, and judging tread faults.

[0057] Next, based on the axle curing characteristic spectrum of the wheel tread and the preset high-order faults, the theoretical high-order spectrum at each high-order fault is calculated by formula. If the wheel tread is determined to have a fault impact based on the theoretical high-order spectrum, the fault impact is quantified.

[0058] In this embodiment, if the wheel tread is determined to have a fault impact based on the theoretical higher-order spectral line, then the fault impact is quantified, which includes: if there is a prominent spectral line in the amplitude spectrum within a first preset range to the left and right of the theoretical higher-order spectral line at the current higher order, then the average value of all spectral lines within a second preset range to the left and right of the prominent spectral line is determined, and a check value is determined based on the average value. If the amplitude of the prominent spectral line is greater than the check value, then it is determined that the current tread higher-order spectral line is prominent and there is a wheel tread fault. The number of all higher-order spectral lines with the wheel tread fault is counted. If the number of higher-order spectral lines is greater than a preset threshold, then it is determined that there is a fault impact on the wheel tread, and the fault impact is quantified.

[0059] First, check the theoretical higher-order spectral lines in the amplitude spectrum of the current higher-order signal in the frequency domain, PH=roud(n f), where n is the order and f is the shaft curing characteristic spectrum of the bearing or tread. For example, the theoretical spectrum of the first-order fault is 40.2, then the second-order is 80, and the third-order is 121. If there is a prominent spectrum line that is significantly higher than the surrounding spectrum lines within the first preset range to the left and right of the theoretical high-order spectrum line, the average amplitude of all spectrum lines within the second preset range to the left and right of the prominent spectrum line is calculated, and a corresponding verification value is set based on the average value. If the amplitude of the prominent spectrum line is greater than the verification value, it is determined that the current tread high-order spectrum line is prominent and there is a wheel tread fault. Then, all the high-order numbers of the wheel tread to be detected are checked one by one and the number of high-order numbers with the above wheel tread fault is counted. If the proportion of the number of high-order numbers is greater than the preset number threshold, it is determined that there is a fault impact on the wheel tread. The magnitude of the fault impact is quantified by the special calculation formula of the tread dB value to obtain the tread dB value that can reflect the degree of tread impact.

[0060] In this embodiment, quantifying the fault impact to obtain the tread dB value includes: obtaining the product between the rotational speed of the axle where the wheel is located and the axle diameter where the wheel is located; obtaining the ratio between the impact value of the fault impact and the product, and taking the logarithm of the ratio to obtain the tread dB value.

[0061] Obtain the product between the rotational speed N (in r / min) of the axle and the axle diameter D (in mm); obtain the ratio between the impact value SV of the fault impact and the product, and take the logarithm of the ratio to obtain the tread dB value. The specific calculation formula is as follows: ; Where dB is the fault level difference; SV is the impact value of the fault impact (1SV represents the corresponding acceleration measure when the half-wave sinusoidal impact pulse width is 20us and the peak value is 0.01g (g is the gravitational acceleration)). SV is obtained by superimposing the three-order spectrum lines in front of the tread; N is the rotational speed of the shaft where the monitored component is located; D is the shaft diameter of the shaft where the monitored object is located.

[0062] Using the established wheel fault characteristic spectrum, for each order to be verified where x ranges from 8 to 40, [(x-1)] is generated. pn, x pn, (x+1) pn, (2 x-1) pn,2 x pn, (2 x+1) The wheel has multiple spectral lines to be verified, where pn is the tread feature spectrum. Target spectral lines that meet the preset prominence conditions are selected from the multiple spectral lines to be verified corresponding to each order. The polygon order of the wheel is determined based on the spectral amplitude of the selected target spectral lines that meet the preset prominence conditions.

[0063] In this embodiment, the step of selecting target spectral lines that meet preset prominence conditions from multiple spectral lines to be verified, and determining the polygon order of the wheel based on the spectral amplitude of the target spectral lines, includes: if the amplitude of the current spectral line to be verified at the current order is greater than a preset prominence threshold, then the current spectral line to be verified is determined to be a target spectral line that meets the preset prominence conditions, so as to obtain the spectral amplitude of the target spectral line at the current order to be verified; and the order with the largest spectral amplitude of the target spectral lines among all orders to be verified is determined as the polygon order of the wheel.

[0064] If the amplitude of the current spectral line to be verified at the current order to be verified is greater than the preset prominence threshold, then the current spectral line to be verified is determined to be a target spectral line that meets the preset prominence condition. The spectral amplitudes of all target spectral lines that meet the preset prominence condition at the current order to be verified are counted one by one to obtain the spectral amplitude of the target spectral line at the current order to be verified. The order to be verified is a value from 8 to 40. For each order to be verified, the spectral line to be verified is generated, the target spectral line is determined, and the corresponding number of targets is counted according to this rule. After the statistics of all orders to be verified are completed, the order with the largest spectral amplitude of the target spectral line in each order to be verified is determined as the polygon order of the wheel. If the tread dB value is 0, the polygon order of the wheel is directly determined to be 0. If the wheel polygon fault is not determined according to the target spectral line, the polygon order of the wheel is also determined to be 0.

[0065] Step S13: Input the current target feature set into the target diagnosis LSTM model to output the fault diagnosis result corresponding to the current vibration signal; wherein, the training data of the target diagnosis LSTM model is the historical target feature set corresponding to the historical vibration signal and the label corresponding to the historical target feature set used to characterize whether there is a wheel polygon fault.

[0066] In this embodiment, inputting the current target feature set into the target diagnosis LSTM model includes: performing data cleaning and data length processing on the current target feature set to obtain a processed current target feature set, and inputting the processed current target feature set into the target diagnosis LSTM model.

[0067] Data cleaning and data length standardization are performed sequentially on the current target feature set to obtain the processed current target feature set. The processed current target feature set is then input into the trained target diagnosis LSTM model. The data to be cleaned can be data with abnormally large or small values. Furthermore, data length standardization refers to data padding and / or data truncation to ensure that the lengths of each feature input into the target diagnosis LSTM model are consistent, thereby ensuring the quality and consistency of the feature set and enabling it to better adapt to subsequent modeling tasks.

[0068] In this embodiment, the step of performing data cleaning and data length processing on the current target feature set to obtain a processed current target feature set includes: replacing outliers in the current target feature set with preset feature values ​​to obtain a cleaned current target feature set; determining a baseline target feature from the cleaned current target feature set or each target feature in the current target feature set; performing data padding and / or data truncation on the cleaned current target feature set or target features in the current target feature set other than the baseline target feature according to the length of the baseline target feature to obtain a length-processed current target feature set; and determining the cleaned current target feature set or the length-processed current target feature set as the processed current target feature set.

[0069] Outliers in the current target feature set are replaced with preset feature values ​​of 0. For example, kurtosis is a statistic that measures the sharpness of data distribution. An abnormally large kurtosis may indicate the presence of extreme values ​​or noise in the data. Therefore, abnormally large time-domain kurtosis values ​​are replaced with 0. The main spectrum frequency reflects the main energy distribution of the signal. A main spectrum that is too low or too high may indicate the presence of abnormal patterns or noise in the data. For the analysis of tread polygon faults, the main frequency is considered to be below the first-order main spectrum number of the tread and above 100 times the first-order main spectrum number of the tread. Therefore, abnormal main frequency values ​​and their corresponding main frequency amplitudes that are below the first-order main spectrum number of the tread and above 100 times the first-order main spectrum number of the tread are all replaced with 0. This is how the cleaned current target feature set is obtained. In a specific embodiment, the cleaned current target feature set can be directly used as the processed current target feature set.

[0070] Since the sample data collected or saved at different measurement points may have inconsistent lengths, in order to ensure the quality and consistency of the feature set and make it better suited for subsequent modeling tasks, it is necessary to perform corresponding adjustments to the feature set. This helps the model skip invalid data during training and reduces the interference of invalid data on the model. A baseline target feature is determined from the cleaned current target feature set or the measurement point data corresponding to each target feature in the current target feature set. Data padding and / or truncation are performed on the target features in the cleaned current target feature set according to the length of the baseline target feature. Then, the `pack_padded_sequence` method in the PyTorch package is used to compress and package the padded feature sequence, allowing the model to skip invalid data during subsequent training and reducing the interference of invalid data. This results in a current target feature set with uniform specifications and valid data length processing. In one specific embodiment, the current target feature set with length processing is determined as the processed current target feature set.

[0071] In this embodiment, determining the baseline target feature from each target feature in the cleaned current target feature set includes: determining the number of measurement points for each target feature in the cleaned current target feature set, and determining the target feature with the largest number of measurement points as the baseline target feature.

[0072] The number of sample data points collected and saved for each target feature in the current target feature set after cleaning, including temporal kurtosis, dominant frequency of wheel tread, dominant frequency amplitude, tread dB value, and polygon order of the wheel, is determined. The target feature corresponding to the measurement point with the largest number of sample data points is determined as the baseline target feature. The sample data length of the baseline target feature is used as a unified standard for subsequent data filling and truncation of other target features to ensure that the sample data length of each target feature is consistent to adapt to the training and diagnostic needs of the subsequent LSTM model.

[0073] In this embodiment, the step of filling data into the target features other than the benchmark target feature in the cleaned current target feature set according to the length of the benchmark target feature includes: determining the target features other than the benchmark target feature in the cleaned current target feature set as target features to be filled; and filling the target features to be filled with preset filling feature values ​​according to the length of the benchmark target feature.

[0074] The target features other than the baseline target features in the current target feature set after cleaning are identified as target features to be filled. Specifically, for target features whose length is less than the length of the baseline target features, their historical time temporal kurtosis, dominant frequency, dominant frequency amplitude, tread dB value, and polygon order are all filled with a preset feature value of 0.

[0075] In this embodiment, before inputting the current target feature set into the target diagnostic LSTM model, the method further includes: constructing an initial diagnostic LSTM model comprising an input layer, an LSTM layer, a dynamic dropout layer, and a fully connected layer; acquiring historical vibration signals from wheel operation, obtaining historical target feature sets corresponding to the historical vibration signals, and obtaining labels corresponding to the historical target feature sets to characterize whether a polygonal wheel fault exists; constructing training data and a validation set comprising the historical target feature sets and the labels corresponding to the historical target feature sets; training the initial diagnostic LSTM model using the training data to obtain a trained diagnostic LSTM model; validating the trained diagnostic LSTM model using the validation set; if the validation result indicates that the trained diagnostic LSTM model meets a preset fine-tuning stopping condition, then the trained diagnostic LSTM model is determined as the target diagnostic LSTM model; if the validation result indicates that the trained diagnostic LSTM model does not meet the preset fine-tuning stopping condition, then the trained diagnostic LSTM model is fine-tuned, and the fine-tuned diagnostic LSTM model is determined as the target diagnostic LSTM model.

[0076] First, an initial diagnostic LSTM (Long Short-Term Memory) model is constructed, which includes an input layer, an LSTM layer, a dynamic dropout layer, and a fully connected layer. As a special type of recurrent neural network (RNN), LSTM, with its unique gating mechanism, can effectively handle long-term dependencies in sequential data. It is very suitable for analyzing time-series data such as train wheel tread vibration signals to achieve accurate identification and prediction of tread state. Therefore, this model is chosen for tread state identification and to build an efficient and accurate LSTM model.

[0077] Historical vibration signals from wheel operation are collected. After EEMD decomposition, signal reconstruction, and feature extraction, historical target feature sets corresponding to the historical vibration signals are obtained, including time-domain kurtosis, dominant frequency, dominant frequency amplitude, tread dB value, and wheel polygon order. Labels are assigned to each historical target feature set to characterize the presence of wheel polygon faults. Specifically, monitoring data from 328 wheel treads over the past month are collected as basic data, including 158 polygon fault cases and 170 normal data. These 328 cases are randomly shuffled and divided in a 7:3 ratio to construct training and validation sets composed of labeled historical target feature sets. The training and validation sets divided in this ratio can evaluate the performance of different hyperparameter combinations or different model structures in subsequent training, thereby selecting the optimal model configuration and avoiding overfitting or underfitting.

[0078] The initial diagnostic LSTM model was then trained using the training data. Training employed the Adam optimizer (β1=0.9, β2=0.999), an initial learning rate of 1e-3, a batch size of 64, and 130 epochs. The loss function was BCEWithLogitsLoss, a binary cross-entropy loss function with logits, which combines the sigmoid function and the binary cross-entropy loss function (BCE). The trained diagnostic LSTM model is obtained by using the Loss function. The performance of the trained diagnostic LSTM model is then validated using a validation set. The validation set evaluates the model's fault diagnosis effectiveness. If the validation results indicate that the trained diagnostic LSTM model meets the preset fine-tuning stopping conditions, it is directly designated as the target diagnostic LSTM model. If the validation results indicate that the trained diagnostic LSTM model does not meet the preset fine-tuning stopping conditions, the hyperparameters or model structure of the trained diagnostic LSTM model are fine-tuned based on the validation set's evaluation results. The fine-tuned diagnostic LSTM model is then designated as the target diagnostic LSTM model for subsequent real-time diagnosis of wheel polygonal faults.

[0079] In this embodiment, the input layer is used to convert the training data into a three-dimensional tensor, the LSTM layer is used to extract temporal features from the three-dimensional tensor, the dynamic discard layer is used to discard some of the temporal features, and the fully connected layer is used to obtain the fault probability value based on the discarded temporal features; wherein, the three-dimensional tensor includes the number of training samples in a single training iteration, the temporal length of a single training sample, and the feature dimension of a single time step.

[0080] The diagnostic LSTM model adopts a hierarchical progressive structure, consisting of an input layer, an LSTM layer, a dynamic dropout layer, and a fully connected layer. The input layer converts training data into a three-dimensional tensor adapted for model training. This three-dimensional tensor is specifically defined as (batch_size, time_steps, feature_dim), where batch_size is the number of samples in a single training iteration (configured to 64 in this embodiment to balance computational efficiency and gradient estimation stability), time_steps is the temporal length of a single training sample, and its value is determined by the signal sampling frequency and the fault feature period. For example, if the vibration signal sampling frequency is 12.8kHz and the fault impact period is 0.01s, 128 sampling points are needed per period. To capture multi-period features, time_steps is set to 1024 (covering 8 fault periods), suitable for low-frequency fault feature extraction. feature_dim is the feature dimension of a single time step, supporting multi-sensor signal fusion. The LSTM layer employs a two-layer hierarchical structure. The first LSTM layer has 64 hidden units and performs preliminary temporal feature extraction from the input 3D tensor, with an output dimension of (batch_size, seq_len, 64). The second LSTM layer has 32 hidden units and further mines the extracted temporal features, extracting higher-level temporal features. The final sequence representation is output with a dimension of (batch_size, 32) by setting return_sequences=False. The dynamic dropout layer is a Dropout layer with a dropout rate of 0.2, connected after each LSTM layer. During training, 20% of neurons are randomly disabled, discarding some extracted temporal features to effectively prevent overfitting. During inference, the dropout function is automatically disabled to ensure the diagnostic stability of the model. The fully connected layer is a fully connected classifier with an input dimension of 64 and an output dimension of 1. It is used to calculate the binary probability value representing the presence of a polygonal fault on the wheel based on the dropped temporal features using the Sigmoid activation function.

[0081] The current target feature set, after data cleaning, data imputation, and truncation, is converted into a three-dimensional tensor format that meets the input requirements of the target diagnosis LSTM model. This tensor format includes the number of samples in a single inference, the temporal length of a single sample, and the feature dimension of a single time step. The converted feature data is then input into the target diagnosis LSTM model that has been trained and validated and meets the preset accuracy requirements. The model sequentially passes through the input layer, LSTM layer, dynamic dropout layer, and fully connected layer. The LSTM layer extracts the temporal features corresponding to the vibration signal. After the dynamic dropout layer suppresses overfitting, the fully connected layer and the sigmoid activation function output the corresponding binary classification probability value. Based on this probability value, it is determined whether the wheel corresponding to the current vibration signal has a polygonal fault. Finally, the wheel tread fault diagnosis result corresponding to the current vibration signal is output.

[0082] The beneficial effects of this application are as follows: This application collects the current vibration signal of the wheel during operation, performs modal decomposition on the current vibration signal based on the EEMD decomposition method to obtain multiple IMF components, and selects the target IMF component containing the most tread impact features from the IMF components based on the maximum similarity principle; performs signal reconstruction based on the target IMF component, and performs feature extraction based on the reconstructed signal to obtain the current target feature set; inputs the current target feature set into the target diagnosis LSTM model to output the fault diagnosis result corresponding to the current vibration signal; wherein, the training data of the target diagnosis LSTM model is the historical target feature set corresponding to the historical vibration signal and the label corresponding to the historical target feature set used to characterize whether there is a polygonal fault in the wheel. Therefore, this application utilizes the EEMD decomposition method to perform modal decomposition on the current vibration signal, which effectively suppresses modal aliasing and improves the accuracy of signal analysis. Based on the principle of maximum similarity, it selects the target IMF component containing the most tread impact features and reconstructs the signal, concentrating the vibration components related to tread impact in the original signal and reducing vibration interference caused by non-tread factors. This makes the reconstructed signal closer to the actual operating state of the wheel, providing high-quality data support for subsequent feature extraction. Extracting the current target feature set related to the wheel polygon fault from the reconstructed signal accurately quantifies fault-related characteristics, providing a representative and discriminative basis for fault diagnosis. Inputting the current target feature set into the target diagnosis LSTM model leverages the LSTM model's advantages in processing time-series data and the learned correlation between historical target feature sets and fault labels, enabling effective diagnosis of wheel polygon faults. Furthermore, the target diagnosis LSTM model trained based on historical target feature sets and related labels corresponding to historical vibration signals can fully explore long-term dependencies and fault characteristic patterns in the data, improving the accuracy and reliability of fault diagnosis. Simultaneously, the entire process requires minimal manual intervention, achieving automated signal processing, feature extraction, and fault diagnosis, thus improving the efficiency of fault diagnosis.

[0083] See Figure 2 As shown in the figure, this application discloses a wheel polygon fault diagnosis device based on EEMD and LSTM, including: The component acquisition module 11 is used to collect the current vibration signal when the wheel is running, perform modal decomposition on the current vibration signal based on the EEMD decomposition method to obtain multiple IMF components, and select the target IMF component containing the most tread impact features from the IMF components based on the maximum similarity principle. The feature extraction module 12 is used to reconstruct the signal based on the target IMF component and extract features based on the reconstructed signal to obtain the current target feature set; The result output module 13 is used to input the current target feature set into the target diagnosis LSTM model to output the fault diagnosis result corresponding to the current vibration signal; wherein, the training data of the target diagnosis LSTM model is the historical target feature set corresponding to the historical vibration signal and the label corresponding to the historical target feature set used to characterize whether there is a wheel polygon fault.

[0084] Furthermore, embodiments of this application also provide an electronic device. Figure 3 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0085] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Specifically, it may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the wheel polygon fault diagnosis method based on EEMD and LSTM disclosed in any of the foregoing embodiments.

[0086] In this embodiment, the power supply 23 is used to provide operating voltage for various hardware devices on the electronic device; the communication interface 24 can create a data transmission channel between the electronic device and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0087] The processor 21 may include one or more processing cores, such as a quad-core processor or an octa-core processor. The processor 21 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 21 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, the processor 21 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, the processor 21 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0088] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored on it include operating system 221, computer program 222 and data 223, etc., and the storage method can be temporary storage or permanent storage.

[0089] The operating system 221 manages and controls the various hardware devices and computer programs 222 on the electronic device to enable the processor 21 to perform calculations and processing on the massive amounts of data 223 in the memory 22. It can be Windows, Unix, Linux, etc. The computer program 222, in addition to including a computer program capable of performing the EEMD and LSTM-based wheel polygon fault diagnosis method disclosed in any of the foregoing embodiments, may further include computer programs capable of performing other specific tasks. The data 223 may include data received by the electronic device from external devices, as well as data collected by its own input / output interface 25.

[0090] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned disclosed method for diagnosing polygonal wheel faults based on EEMD and LSTM. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0091] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0092] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application. The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly in hardware, software modules executed by a processor, or a combination of both. The software module may be located in random access memory (RAM), memory, read-only memory (ROM), electrically programmable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), register, hard disk, removable disk, CD-ROM (Compact Disc Read-Only Memory), or any other form of storage medium known in the art.

[0093] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0094] The foregoing has provided a detailed description of a method, apparatus, device, and medium for diagnosing polygonal wheel faults based on EEMD and LSTM. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only intended to help understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for diagnosing polygonal wheel faults based on EEMD and LSTM, characterized in that, include: The current vibration signal of the wheel during operation is collected, and the current vibration signal is decomposed into multiple IMF components based on the EEMD decomposition method. The target IMF component containing the most tread impact features is selected from the IMF components based on the maximum similarity principle. The signal is reconstructed based on the target IMF components, and the current target feature set is obtained by feature extraction based on the reconstructed signal. The current target feature set is input into the target diagnosis LSTM model to output the fault diagnosis result corresponding to the current vibration signal; wherein, the training data of the target diagnosis LSTM model is the historical target feature set corresponding to the historical vibration signal and the label corresponding to the historical target feature set used to characterize whether there is a wheel polygon fault.

2. The wheel polygon fault diagnosis method based on EEMD and LSTM according to claim 1, characterized in that, The current vibration signal is modally decomposed using the EEMD decomposition method to obtain multiple IMF components, including: White noise is added to the current vibration signal in multiple batches to obtain the noise-added signal for each batch; The noisy signal is decomposed using EMD to obtain multiple initial IMF components and decomposition residuals with different serial numbers in each batch; wherein the serial number indicates the position of the initial IMF component in each batch. The initial IMF components of all batches with the same sequence number are averaged to obtain the final IMF component corresponding to each sequence number.

3. The wheel polygon fault diagnosis method based on EEMD and LSTM according to claim 2, characterized in that, The addition of white noise to the current vibration signal in multiple batches includes: Determine the standard deviation of the current vibration signal, and determine the target amplitude of the white noise based on the standard deviation; Multiple batches of white noise, each with the target amplitude, are added to the current vibration signal.

4. The wheel polygon fault diagnosis method based on EEMD and LSTM according to claim 2, characterized in that, The process of selecting the target IMF component containing the most tread impact features from the IMF components based on the maximum similarity principle includes: The similarity between the final IMF component corresponding to each sequence number and the current vibration signal is determined respectively; IMF components that meet preset conditions are selected from the final IMF components and summed up to obtain the target IMF component.

5. The wheel polygon fault diagnosis method based on EEMD and LSTM according to claim 4, characterized in that, The step of determining the similarity between the final IMF component corresponding to each sequence number and the current vibration signal includes: The correlation coefficient between the final IMF component corresponding to each sequence number and the current vibration signal is determined, and the correlation coefficient is determined as the similarity between the final IMF component corresponding to each sequence number and the current vibration signal.

6. The wheel polygon fault diagnosis method based on EEMD and LSTM according to claim 4, characterized in that, The step of selecting IMF components that meet preset conditions from each of the final IMF components and accumulating them as the target IMF component includes: The final IMF components corresponding to each sequence number are sorted in descending order of similarity to obtain the IMF component sequence. IMF components that meet the preset similarity criteria are selected from the IMF component sequence, and the IMF components that meet the preset criteria are accumulated. The accumulated result is taken as the target IMF component that contains the most tread impact features.

7. The wheel polygon fault diagnosis method based on EEMD and LSTM according to claim 1, characterized in that, The signal reconstruction based on the target IMF component includes: The target IMF components are summed at the corresponding time points to obtain the reconstructed signal.

8. The wheel polygon fault diagnosis method based on EEMD and LSTM according to claim 1, characterized in that, The target features in the current target feature set are any one of the following features: the time-domain kurtosis of the current vibration signal, the dominant frequency of the wheel tread, the amplitude of the dominant frequency, the dB value of the tread, and the polygon order of the wheel.

9. The wheel polygon fault diagnosis method based on EEMD and LSTM according to claim 8, characterized in that, The step of extracting features from the reconstructed signal to obtain the current target feature set includes: Determine the signal mean of the reconstructed signal, and use the signal mean and the reconstructed signal to obtain the time-domain kurtosis of the current vibration signal; The reconstructed signal after removing the mean is subjected to Fourier transform to obtain a frequency domain signal, and the amplitude spectrum of the frequency domain signal is generated. The target frequency point with the largest amplitude in the amplitude spectrum is found, and the frequency and amplitude of the target frequency point are determined as the main frequency and main frequency amplitude of the wheel tread, respectively. Determine the fault characteristic spectrum of the wheel, and determine the theoretical high-order spectrum at each high-order number based on the axle curing characteristic spectrum of the wheel tread and each high-order fault. If it is determined that there is a fault impact on the wheel tread based on the theoretical high-order spectrum, then the fault impact is quantified to obtain the tread dB value. Multiple spectral lines to be verified are generated at each order using the fault feature spectral number, and target spectral lines that meet preset prominence conditions are selected from the multiple spectral lines to be verified. The polygon order of the wheel is determined based on the spectral amplitude of the target spectral lines.

10. The wheel polygon fault diagnosis method based on EEMD and LSTM according to claim 9, characterized in that, The determination of the fault characteristic spectrum of the wheel includes: If the current vibration signal sampling scenario is speed tracking sampling, then the number of wheel sampling revolutions is determined as the fault characteristic spectrum of the wheel; If the current vibration signal is sampled using an astronomical clock, the fault characteristic spectrum of the wheel is determined based on the wheel speed, the number of signal sampling points, and the sampling frequency.

11. The wheel polygon fault diagnosis method based on EEMD and LSTM according to claim 9, characterized in that, If a fault impact is determined to exist on the wheel tread based on the aforementioned higher-order spectral lines, then the fault impact is quantified, including: If there is a prominent spectral line in the amplitude spectrum within a first preset range to the left and right of the theoretical higher-order spectral line at the current higher order, then the average value of all spectral lines within a second preset range to the left and right of the prominent spectral line is determined, and a check value is determined based on the average value. If the amplitude of the prominent spectral line is greater than the check value, then it is determined that the current tread higher-order spectral line is prominent and there is a wheel tread fault. The number of high-order numbers for all wheel tread faults is counted. If the number of high-order numbers is greater than a preset threshold, it is determined that there is a fault impact on the wheel tread, and the fault impact is quantified.

12. The wheel polygon fault diagnosis method based on EEMD and LSTM according to claim 9, characterized in that, The quantification of the fault impact to obtain the tread dB value includes: Obtain the product between the rotational speed of the axle containing the wheel and the diameter of the axle containing the wheel; The ratio between the impact value of the fault impact and the product is obtained, and the logarithm of the ratio is taken to obtain the tread dB value.

13. The wheel polygon fault diagnosis method based on EEMD and LSTM according to claim 9, characterized in that, The step of selecting target spectral lines that meet preset prominence conditions from multiple spectral lines to be verified, and determining the polygon order of the wheel based on the spectral amplitude of the target spectral lines, includes: If the amplitude of the current spectral line to be verified at the current order to be verified is greater than the preset prominence threshold, then the current spectral line to be verified is determined to be the target spectral line that meets the preset prominence condition, so as to obtain the spectral line amplitude of the target spectral line at the current order to be verified. The order with the largest spectral amplitude of the target spectral line among all orders to be verified is determined as the polygon order of the wheel.

14. The wheel polygon fault diagnosis method based on EEMD and LSTM according to claim 1, characterized in that, The step of inputting the current target feature set into the target diagnosis LSTM model includes: The current target feature set is cleaned and / or its length is processed to obtain a processed current target feature set, which is then input into the target diagnostic LSTM model.

15. The wheel polygon fault diagnosis method based on EEMD and LSTM according to claim 14, characterized in that, The step of performing data cleaning and / or data length processing on the current target feature set to obtain the processed current target feature set includes: Replace the outliers in the current target feature set with preset feature values ​​to obtain the cleaned current target feature set; Determine the baseline target features from the cleaned current target feature set or from each target feature in the current target feature set; Data padding and / or data truncation are performed on the cleaned current target feature set or the target features in the current target feature set other than the benchmark target feature according to the length of the benchmark target feature, so as to obtain the current target feature set after length processing; The cleaned current target feature set or the length-processed current target feature set is determined as the processed current target feature set.

16. The wheel polygon fault diagnosis method based on EEMD and LSTM according to claim 15, characterized in that, Determining the baseline target feature from each target feature in the cleaned current target feature set includes: Determine the number of measurement points for each target feature in the current target feature set after cleaning, and determine the target feature with the largest number of measurement points as the baseline target feature; Accordingly, the step of filling the target features in the cleaned current target feature set, excluding the benchmark target feature, with data according to the length of the benchmark target feature includes: The target features in the current cleaned target feature set, excluding the baseline target features, are determined as target features to be filled. The preset filling feature value is filled into the target feature to be filled according to the length of the reference target feature.

17. The wheel polygon fault diagnosis method based on EEMD and LSTM according to any one of claims 1 to 16, characterized in that, Before inputting the current target feature set into the target diagnosis LSTM model, the method further includes: Construct an initial diagnostic LSTM model that includes an input layer, an LSTM layer, a dynamic dropout layer, and a fully connected layer; Collect historical vibration signals of the wheel during operation, obtain the historical target feature set corresponding to the historical vibration signals, and obtain the label corresponding to the historical target feature set to characterize whether there is a polygonal fault in the wheel. Construct training data and verification set including the historical target feature set and the label corresponding to the historical target feature set. The initial diagnostic LSTM model is trained using the training data to obtain the trained diagnostic LSTM model. The trained diagnostic LSTM model is validated using the validation set. If the validation result indicates that the trained diagnostic LSTM model meets the preset fine-tuning stopping condition, then the trained diagnostic LSTM model is determined as the target diagnostic LSTM model. If the validation result indicates that the trained diagnostic LSTM model does not meet the preset fine-tuning stopping condition, then the trained diagnostic LSTM model is fine-tuned, and the fine-tuned diagnostic LSTM model is determined as the target diagnostic LSTM model.

18. The wheel polygon fault diagnosis method based on EEMD and LSTM according to claim 17, characterized in that, The input layer is used to convert the training data into a three-dimensional tensor. The LSTM layer is used to extract temporal features from the three-dimensional tensor. The dynamic discarding layer is used to discard some of the temporal features. The fully connected layer is used to obtain the fault probability value based on the discarded temporal features. The three-dimensional tensor includes the number of training samples in a single training iteration, the temporal length of a single training sample, and the feature dimension of a single time step.

19. A wheel polygon fault diagnosis device based on EEMD and LSTM, characterized in that, include: The component acquisition module is used to collect the current vibration signal when the wheel is running, perform modal decomposition on the current vibration signal based on the EEMD decomposition method to obtain multiple IMF components, and select the target IMF component containing the most tread impact features from the IMF components based on the maximum similarity principle. The feature extraction module is used to reconstruct the signal based on the target IMF components, and to extract features from the reconstructed signal to obtain the current target feature set. The result output module is used to input the current target feature set into the target diagnosis LSTM model to output the fault diagnosis result corresponding to the current vibration signal; wherein, the training data of the target diagnosis LSTM model is the historical target feature set corresponding to the historical vibration signal and the label corresponding to the historical target feature set used to characterize whether there is a wheel polygon fault.

20. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor is configured to execute the computer program to implement the steps of the wheel polygon fault diagnosis method based on EEMD and LSTM as described in any one of claims 1 to 18.

21. A computer-readable storage medium, characterized in that, Used to store a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the wheel polygon fault diagnosis method based on EEMD and LSTM as described in any one of claims 1 to 18.

Images