Intelligent quantitative identification method for wheel non-circular roughness level based on axle box vibration

Abstract

The application provides a wheel non-circularity roughness level intelligent quantitative identification method based on axle box vibration, and belongs to the technical field of high-speed motor train unit state monitoring and intelligent diagnosis. The method comprises the following steps: step 1, sample pretreatment and sample set division; step 2, outlier sample elimination; step 3, quasi-stationary sample reservation; step 4, wavelength domain mapping and standardization processing; step 5, identification model construction and training; and step 6, multi-result fusion and state output. The application is driven by high-fidelity actual measurement data, a model training sample set facing real operation conditions is constructed, outlier and non-stationary samples are eliminated through a double-layer sample screening mechanism, sample features are aligned through time domain-wavelength domain mapping, and a neural network fusing convolution feature extraction and a multi-head attention mechanism is used to output wheel 1-30 order roughness level vectors, so that high-precision, strong-robust and on-line deployable quantitative identification of the wheel non-circularity state is realized.

Description

Technical Field

[0001] This invention relates to the field of high-speed train condition monitoring and intelligent diagnosis technology, and in particular to an intelligent quantitative identification method for wheel non-circular roughness level based on axle box vibration. Background Technology

[0002] High-speed trains operate under long-term high-speed and strongly coupled wheel-rail conditions, resulting in wheel tread wear phenomena such as eccentricity, flat spots, random non-circularity, and polygonal wear. Extensive operational experience shows that polygonal wear of wheels often exhibits high-order characteristics of 10 to 30, with the dominant orders concentrated in 14 to 16 and 18 to 28. The corresponding wavelength polygons easily excite high-frequency responses of 400 to 600 Hz at the axle box.

[0003] Existing wheel condition detection technologies are mainly divided into two categories: offline detection and online monitoring. Offline detection relies on portable wheel out-of-roundness measuring instruments, lasers, or structured light, which can accurately reconstruct wheel out-of-roundness. However, the detection is discontinuous, labor-intensive, and lacks timeliness, making it difficult to meet the real-time requirements of predictive maintenance for high-speed trains. Online monitoring typically uses response signals such as rail vibration, wheel-rail force, axle box vibration acceleration, or noise to invert the wheel out-of-roundness state. Among these, axle box vibration acceleration, due to its location under the primary suspension, close coupling with the wheelset, and sensitivity to high-frequency disturbances, has become an important information source for onboard identification of wheel out-of-roundness.

[0004] Existing research on wheel non-circularity identification related to axle box vibration can be broadly divided into two technical approaches: one is based on traditional signal processing methods, such as FFT (Fast Fourier Transform), order analysis, EEMD (Ensemble Empirical Mode Decomposition), VMD (Variational Mode Decomposition), improved frequency domain integration, and angular domain synchronous averaging; the other is based on machine learning or deep learning methods, such as 1D-CNN (One-Dimensional Convolutional Neural Network), 2D-CNN (Two-Dimensional Convolutional Neural Network), CNN-LSTM (Convolutional Neural Network-Long Short-Term Memory Network), and OOR-Net (Wheel Non-Circularity Network). The former has a certain degree of physical interpretability, but usually relies on manual feature engineering and empirical thresholds; the latter can automatically learn features, but often faces problems such as unstable training sample quality, insufficient generalization ability across working conditions, and model sensitivity to speed fluctuations.

[0005] In existing deep learning research, the closest approach to this invention is based on constructing the spectrum, envelope spectrum, or time-frequency image of axle box vibration signals, and using convolutional neural networks, recurrent neural networks, or combinations thereof to output the wheel out-of-roundness level or the amplitude of the polygonal principal order. Although this type of scheme achieves "end-to-end" recognition compared to traditional methods, it still generally suffers from the following problems: First, the training phase is highly dependent on simulation data; second, under actual working conditions, the distribution of features within the sample class is discrete due to factors such as uneven track, weld impact, and speed fluctuations; and third, most of them only output a single principal order or coarse classification result, and it is still difficult to directly provide roughness level vectors of orders 1 to 30.

[0006] In summary, the shortcomings of existing technologies are as follows: (1) Existing technologies still suffer from insufficient data authenticity and inadequate sample quality control in sample set construction. Currently, many deep learning-based wheel polygon recognition methods still adopt the method of "training with simulation data and testing with measured data" or training with a mixture of simulation data and a small amount of measured data. Since simulation models cannot fully represent nonlinear contact, structural response, environmental disturbances and sensor noise in real wheel-rail systems, there is a significant domain shift between the features of training data and actual application data, resulting in insufficient robustness and generalization ability of the model on measured datasets. On the other hand, under speed fluctuation conditions, the excitation frequency corresponding to the geometric features of the same polygon will shift. If existing methods directly rely on frequency domain features or spatial domain resampling, they often rely heavily on the accurate measurement of wheel radius and instantaneous speed, which can easily introduce cumulative errors and lead to divergent distribution of intra-class samples. In addition, factors such as uneven rail welds, uneven rail corrugation, and random irregularities can generate strong transient impacts in axle box vibration signals. Existing methods generally lack screening mechanisms for outliers and non-stationary samples, which can easily lead to interference components unrelated to wheel non-circularity being mixed into the training samples, thereby affecting the model's ability to learn effective non-circularity wear features.

[0007] (2) Existing technologies still suffer from insufficient quantitative representation capabilities and limited global feature modeling capabilities in the design of recognition models. Existing methods typically focus on determining the presence of wheel polygons, identifying dominant orders, or estimating the magnitude of a single dominant order. They are unable to directly output complete roughness level vectors of wheel orders 1 to 30, thus failing to simultaneously achieve a unified quantitative representation of low-order and high-order non-circular features. Meanwhile, although pure convolutional networks have certain advantages in extracting local impact features, their receptive field is limited, and their ability to model long-range dependencies, cross-scale feature associations, and global importance allocation is insufficient, making it difficult to fully represent polygonal wear features under complex service conditions. On the other hand, while some complex hybrid networks improve expressive capabilities, their model parameter scale is large and computational overhead is high, which is not conducive to achieving a balance between high accuracy and high robustness in vehicle-mounted or edge computing scenarios. Summary of the Invention

[0008] This invention provides an intelligent quantitative identification method for the non-circular roughness level of wheels based on axle box vibration. The purpose is to: construct a model training sample set oriented towards real operating conditions using high-fidelity real-measured data as the driving force; eliminate outliers and non-stationary samples through a two-layer sample screening mechanism; achieve sample feature alignment through time-domain-wavelength domain mapping; and output wheel roughness level vectors of orders 1 to 30 through a neural network that integrates convolutional feature extraction and multi-head attention mechanism, thereby achieving high-precision, robust, and online-deployable quantitative identification of the non-circular state of wheels.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A smart quantitative identification method for non-circular roughness level of wheels based on axle box vibration includes: Step 1: Collect train axle box vibration signal, train speed signal and wheel out-of-roundness data, preprocess the wheel out-of-roundness data to generate a wheel roughness horizontal vector, segment the axle box vibration signal to generate initial vibration samples, and match the initial vibration samples with the wheel roughness horizontal vector to construct an initial sample set. Step 2: Combine the train speed signal to remove outliers from the initial sample set and generate the first selected sample set; Step 3: Perform quasi-stationary sample screening on the first screening sample set to generate a valid sample set; Step 4: Perform spectrum extraction, wavelength domain mapping, and resampling on the effective sample set to generate a wavelength domain normalized sample set; Step 5: Input the wavelength domain standardized sample set into the quantitative identification model for training to obtain the trained quantitative identification model; Step 6: Collect the axle box vibration signal and train speed signal of the train under test, obtain multiple roughness level prediction results of the wheel under test through the trained quantitative recognition model, fuse the multiple roughness level prediction results, and output the final roughness level vector of the wheel under test.

[0010] In this specification, in step 1, the preprocessing of the wheel out-of-roundness data is as follows: abnormal peak removal, curvature smoothing, detrending term and beginning and end closure correction. The corrected out-of-roundness data is then transformed to generate a horizontal vector of wheel roughness.

[0011] In this specification, in step 1, the axle box vibration signal is segmented according to a fixed time window to generate initial vibration samples. The initial vibration samples are matched with the corresponding wheel roughness horizontal vector based on the mileage node to complete the construction of the initial sample set.

[0012] In this specification, in step 2, the statistical feature vectors of each initial vibration sample in the initial sample set are extracted. The statistical feature vectors contain the speed fluctuation characteristics of the train speed signal within the corresponding time window. The statistical feature vectors are standardized, and the standardized statistical feature vectors are clustered using a clustering algorithm. Outliers are removed, and the remaining samples are retained to generate the first screening sample set.

[0013] In this specification, in step 2, the statistical feature vector is generated by combining the time-domain statistical features and frequency-domain statistical features of the initial vibration sample with the speed fluctuation features corresponding to the train speed signal.

[0014] In this specification, in step 3, each sample in the first screening sample set is divided into multiple continuous subframes, the power spectral density of each subframe is calculated, the spectral stability index and the main peak drift index of the sample are calculated based on the power spectral density, and samples that meet the quasi-stationary conditions are screened according to the preset threshold to generate an effective sample set.

[0015] In this specification, in step 4, the amplitude spectrum of the samples in the effective sample set is extracted and normalized. The average speed of the samples is calculated based on the train speed signal. The frequency dimension of the amplitude spectrum is mapped to the wavelength dimension. Non-uniform resampling is performed according to the wavelength interval. All samples are converted into sequences of uniform length to generate a wavelength domain standardized sample set.

[0016] In this specification, in step 5, the quantitative identification model is sequentially configured with a convolutional feature extraction module, an attention feature enhancement module, and a task perception output module; the convolutional feature extraction module extracts local features of wavelength domain standardized samples, the attention feature enhancement module models the global correlation features of wavelength domain standardized samples and assigns feature weights, and the task perception output module outputs the wheel roughness level prediction result.

[0017] In step 5 of this specification, using a single wheel as the dividing unit, all samples of the same wheel in the wavelength domain standardized sample set are divided into the same data subset, generating a training set, a validation set, and a test set respectively. The training set and validation set are used to complete the training and parameter optimization of the quantitative identification model, and the test set is used to verify the performance of the quantitative identification model.

[0018] In this specification, in step 6, for each order of the roughness level vector, all predicted values ​​corresponding to that order are statistically analyzed, and the median of the statistical values ​​is selected as the final value of that order, thus completing the fusion processing of multiple sets of roughness level prediction results.

[0019] In summary, the present invention has at least the following beneficial effects: (1) In view of the lack of systematic screening of outlier and non-stationary samples in the existing technology, which leads to the inclusion of interference components such as rail weld irregularities, rail corrugation irregularities, and random irregularities in the training samples, the present invention adopts a two-layer screening mechanism of "DBSCAN (density-based clustering algorithm) clustering outlier removal + spectral stability constraint to retain quasi-stationary samples", which can effectively remove low-quality samples affected by strong transient impacts, abnormal working condition fluctuations and obvious frequency detuning, improve the signal-to-noise ratio, intra-class consistency and learnability of the sample set, thereby enhancing the model's ability to extract effective features of wheel non-circularity.

[0020] (2) In view of the problem that existing technologies are prone to feature frequency shift under speed fluctuation conditions and are highly dependent on the accurate measurement of wheel radius and instantaneous speed, resulting in divergent distribution of sample features within the class, this invention uses a fixed time window to perform initial sample division of axle box vibration signal, and further maps the time domain features to the wavelength domain, and then forms a unified fixed-length input sequence through non-uniform resampling, thereby reducing the dependence on spatial domain resampling and accurate wheel diameter parameters, weakening the impact of frequency feature peak shift on recognition results under variable speed conditions, and improving the consistency of sample representation under different operating speed conditions and the stability of model recognition across operating conditions.

[0021] (3) To address the shortcomings of existing technologies, such as the insufficient ability of pure convolutional networks to model long-range dependencies, cross-scale correlations, and global importance allocation, and the difficulty of balancing accuracy, robustness, and engineering feasibility with complex hybrid networks, this invention constructs an AMCNN-Net (convolutional neural network with attention mechanism) model that integrates convolutional feature extraction and multi-head self-attention enhancement. The convolutional module extracts local harmonic and impact features, while the attention module models the global correlation between different wavelength segments, thereby achieving end-to-end deep modeling of the polygonal wear features of wheels. Experimental results show that, compared with benchmark models such as CNN, OOR-Net, and CNN-LSTM, the determination coefficient of this invention's model is significantly higher. The accuracy is improved by about 18%, and the mean square error (MSE) is reduced by about 12.4%, indicating that it has higher recognition accuracy and better robustness under complex service conditions. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the intelligent quantitative identification method for non-circular roughness level of wheels based on axle box vibration involved in this invention.

[0024] Figure 2 This is a schematic diagram of the AMCNN-Net structure involved in this invention.

[0025] Figure 3 This is a schematic diagram of a typical axle box vibration response involved in this invention.

[0026] Figure 4 This is a schematic diagram visualizing the clustering results involved in this invention.

[0027] Figure 5 This is a schematic diagram of the mapping and feature alignment effect of samples from the frequency domain to the wavelength domain (feature divergence and offset in the frequency domain) involved in this invention.

[0028] Figure 6 This is a schematic diagram of the mapping and feature alignment effect of samples from the frequency domain to the wavelength domain (feature consistency and alignment in the wavelength domain) involved in this invention.

[0029] Figure 7 This is a schematic diagram of the empirical structure and parameters of AMCNN-Net involved in this invention.

[0030] Figure 8 This is a schematic diagram of the polygon wear identification results (dominant order 13) of the empirical model involved in this invention.

[0031] Figure 9 This is a schematic diagram of the polygon wear identification results (dominant order 24) of the empirical model involved in this invention.

[0032] Figure 10 This is a schematic diagram of the polygon wear identification results (dominant orders 15 and 22) of the empirical model involved in this invention.

[0033] Figure 11 This is a schematic diagram of the polygon wear identification results (dominant orders 14 and 23) of the empirical model involved in this invention. Detailed Implementation

[0034] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0035] The following disclosure provides many different implementations or examples for carrying out different structures of the embodiments of the present invention. To simplify the disclosure of the embodiments of the present invention, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the embodiments of the present invention. Furthermore, reference numerals and / or reference letters may be repeated in different examples of the embodiments of the present invention; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various implementations and / or arrangements discussed.

[0036] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0037] like Figure 1 As shown, this embodiment provides an intelligent quantitative identification method for the non-circular roughness level of wheels based on axle box vibration, including: Step 1: Collect train axle box vibration signal, train speed signal and wheel out-of-roundness data, preprocess the wheel out-of-roundness data to generate a wheel roughness horizontal vector, segment the axle box vibration signal to generate initial vibration samples, and match the initial vibration samples with the wheel roughness horizontal vector to construct an initial sample set. Step 2: Combine the train speed signal to remove outliers from the initial sample set and generate the first selected sample set; Step 3: Perform quasi-stationary sample screening on the first screening sample set to generate a valid sample set; Step 4: Perform spectrum extraction, wavelength domain mapping, and resampling on the effective sample set to generate a wavelength domain normalized sample set; Step 5: Input the wavelength domain standardized sample set into the quantitative identification model for training to obtain the trained quantitative identification model; Step 6: Collect the axle box vibration signal and train speed signal of the train under test, obtain multiple roughness level prediction results of the wheel under test through the trained quantitative recognition model, fuse the multiple roughness level prediction results, and output the final roughness level vector of the wheel under test.

[0038] In some embodiments, in step 1, the preprocessing of the wheel out-of-roundness data includes abnormal peak removal, curvature smoothing, detrending term and beginning-end closure correction, and the corrected out-of-roundness data is transformed to generate a wheel roughness horizontal vector.

[0039] In some embodiments, in step 1, the axle box vibration signal is segmented according to a fixed time window to generate initial vibration samples, and the initial vibration samples are matched with the corresponding wheel roughness horizontal vector according to the mileage node to complete the construction of the initial sample set.

[0040] In some embodiments, in step 2, statistical feature vectors of each initial vibration sample in the initial sample set are extracted. The statistical feature vectors contain the speed fluctuation characteristics of the train speed signal within the corresponding time window. The statistical feature vectors are standardized, and the standardized statistical feature vectors are clustered using a clustering algorithm. Outliers are removed, and the remaining samples are retained to generate the first screening sample set.

[0041] In some embodiments, in step 2, the statistical feature vector is generated by combining the time-domain statistical features and frequency-domain statistical features of the initial vibration sample with the speed fluctuation features corresponding to the train speed signal.

[0042] In some embodiments, in step 3, each sample in the first screening sample set is divided into multiple continuous subframes, the power spectral density of each subframe is calculated, the spectral stability index and the main peak drift index of the sample are calculated based on the power spectral density, and samples that meet the quasi-stationary conditions are screened according to a preset threshold to generate an effective sample set.

[0043] In some embodiments, in step 4, the amplitude spectrum of the samples in the effective sample set is extracted and normalized by performing transformation processing. Based on the average speed of the samples calculated from the train speed signal, the frequency dimension of the amplitude spectrum is mapped to the wavelength dimension. Non-uniform resampling is performed according to the wavelength interval, and all samples are converted into sequences of uniform length to generate a wavelength domain standardized sample set.

[0044] In some embodiments, in step 5, the quantitative identification model is sequentially configured with a convolutional feature extraction module, an attention feature enhancement module, and a task perception output module; the convolutional feature extraction module extracts local features of wavelength domain standardized samples, the attention feature enhancement module models the global correlation features of wavelength domain standardized samples and assigns feature weights, and the task perception output module outputs the wheel roughness level prediction result.

[0045] In some embodiments, in step 5, using a single wheel as the dividing unit, all samples of the same wheel in the wavelength domain standardized sample set are divided into the same data subset, and training set, validation set and test set are generated respectively. The training set and validation set are used to complete the training and parameter optimization of the quantitative identification model, and the test set is used to verify the performance of the quantitative identification model.

[0046] In some embodiments, in step 6, for each order of the roughness level vector, all predicted values ​​corresponding to that order are statistically analyzed, and the median of the statistical values ​​is selected as the final value of that order, thus completing the fusion processing of multiple sets of roughness level prediction results.

[0047] The technical concept of this invention is as follows: This invention is based on measured axle box vibration signals and train speed signals. First, signal preprocessing and sample label matching are completed to construct an initial sample set. Then, the samples are purified through a two-layer mechanism of DBSCAN clustering to remove outliers and spectral stability constraints to screen quasi-stationary samples. The temporal features of the samples are mapped to the wavelength domain and non-uniformly resampled to generate a standardized input sequence. The sequence is then input into a self-developed AMCNN-Net model to complete feature extraction and enhancement training. Finally, the prediction results of multiple samples are fused to output a complete roughness horizontal vector of wheel order 1 to 30, realizing intelligent quantitative identification of the non-circular state of the wheel.

[0048] Step 1: Sample preprocessing and sample set partitioning.

[0049] Out-of-roundness is measured on the wheel under test, and the original out-of-roundness sequence is subjected to abnormal peak removal, curvature smoothing, detrending term removal, and beginning and end closure correction. Then, the roughness level vectors of order 1 to 30 are obtained by discrete Fourier transform and used as labels. At the same time, the vibration signal of the target axle box and the train speed signal are collected simultaneously, and the vibration sampling frequency is preferably 5000Hz. Then, the original axle box vibration signal is divided into initial samples according to a fixed time window, preferably with a window length of 2 seconds, and each vibration sample is matched with the corresponding roughness label to form a sample set.

[0050] Step 2: First-level screening – outlier sample removal.

[0051] For each initial sample obtained in step 1, a statistical feature vector is constructed (including kurtosis, peak factor, impulse factor, spectral entropy, spectral flatness, main peak energy ratio, and vehicle speed standard deviation within the time window). After standardizing each feature dimension, the DBSCAN clustering algorithm is used to perform unsupervised clustering analysis on the initial samples. Noise points and outliers identified during the clustering process are removed, thereby eliminating low-quality samples affected by strong transient shocks, significant frequency detuning, or abnormal operating condition fluctuations.

[0052] Step 3: Second-level screening – retention of quasi-stationary samples.

[0053] The samples retained in step 2 are further screened for quasi-stationarity. Specifically, each sample is divided into multiple consecutive subframes, and the power spectral density of each subframe is estimated using the Welch method. Based on this, the cosine similarity between the normalized spectra of adjacent subframes is calculated as an index of spectral stability, and the standard deviation of the main peak frequency of each subframe is calculated as an index of main peak drift. According to a preset quantile threshold or an adaptive threshold, only samples with high spectral similarity and small main peak drift are retained to obtain an effective sample set that meets the quasi-stationarity condition.

[0054] Step 4: Wavelength domain mapping and normalization.

[0055] The samples retained in step 3 are subjected to a Fast Fourier Transform (FFT) to extract the amplitude spectrum within a preset analysis frequency band and perform normalization. Then, based on the average running speed corresponding to each sample, the frequency points are mapped to physical wavelengths, satisfying the condition: physical wavelength = average running speed / frequency point. To balance the fidelity of key features with computational efficiency, the mapped wavelength domain spectrum is non-uniformly resampled according to different wavelength intervals. Preferably, a smaller sampling interval is used in the 80–180 mm wavelength range, and a larger sampling interval is used in the longer wavelength range. Finally, all samples are uniformly converted into a fixed-length wavelength domain sequence as model input samples.

[0056] Step 5: Identification model construction and training.

[0057] The wavelength-domain standardized samples obtained in step 4 are input into the AMCNN-Net polygon wear quantitative identification model for training. The AMCNN-Net includes a convolutional feature extraction module, an attention feature enhancement module, and a task-aware output module. Specifically, the convolutional feature extraction module extracts local harmonic and impact features from the samples; the attention feature enhancement module models the long-range correlation between different wavelength bands and adaptively assigns weights to key features; and the task-aware output module outputs the wheel roughness horizontal vectors from order 1 to 30.

[0058] Step 6: Multi-result fusion and state output.

[0059] For each order of the prediction results of multiple samples of the same wheel, statistics are performed separately, and the median of each order is taken as the final recognition result of that order, so as to obtain the final roughness level vector of the wheel from order 1 to 30.

[0060] Example 1: A complete wheel turning cycle tracking test was conducted on a high-speed train with an operating speed of 250 km / h under actual track conditions. Data acquisition covered one complete wheel turning cycle. The test train consisted of 8 cars, with a nominal wheel rolling circle diameter of 920 mm. An initial measured sample set for model training and validation was constructed by simultaneously acquiring wheel out-of-roundness data and axle box vibration signals. Subsequently, a quasi-stationary sample set was obtained using a two-layer screening method combining DBSCAN clustering outlier removal and spectral stability constraints. Finally, the quasi-stationary sample set was input into the recognition model to obtain quantitative identification results of wheel polygonal wear.

[0061] Step 1: After acquiring the original out-of-roundness data, abnormal peak removal, curvature smoothing, detrending, and beginning-end closure correction are performed to obtain a closed roughness curve. Then, the wheel roughness level vectors of orders 1 to 30 are extracted using Discrete Fourier Transform as training labels for the neural network. This wheel out-of-roundness measurement is only used for label construction during the training phase and does not need to be executed during online application. Simultaneously, accelerometers are deployed at the target axle box during train operation, and train speed signals are collected synchronously. The vibration signal sampling frequency is set to 5000Hz. Subsequently, the original axle box vibration signal is divided into candidate samples according to a fixed time window; after analysis, the time window is set to 2 seconds. Finally, each vibration sample is matched with the wheel roughness level vector at the corresponding mileage node to form an initial sample set.

[0062] Step 2: For each initial sample obtained in Step 1, construct a statistical feature vector (including kurtosis, peak factor, impulse factor, spectral entropy, spectral flatness, main peak energy ratio, and standard deviation of vehicle speed within the window). After Z-score standardization of each feature dimension, use the DBSCAN clustering algorithm to perform unsupervised clustering analysis on the initial samples. Noise points and outliers identified during the clustering process are removed, thus eliminating low-quality samples affected by rail weld irregularities, rail corrugation irregularities, random irregularities, strong transient impacts, significant frequency detuning, or abnormal operating condition fluctuations. This step improves the signal-to-noise ratio and intra-class consistency of the sample set. To verify the effectiveness of the proposed sample screening method in identifying outliers, Figure 3 The time-domain waveforms of a typical axle box vibration response are presented.

[0063] In some embodiments, the wavelength domain normalized sample set construction process is used to process the axle box vibration signal and generate a wavelength domain normalized sample set that can be used for model training. The process is divided into six steps: parameter input, four-step progressive screening and transformation, and result output, as detailed below: To begin, enter the preset parameters: When the process starts, all basic parameters are pre-entered and set, including: Original axle box vibration acceleration signal Train speed signal Axle box vibration signal sampling frequency Number of sample subframes Minimum threshold for quantiles used in spectral stability screening With the maximum threshold and the neighborhood radius parameter of the DBSCAN clustering algorithm. and minimum sample size parameter .

[0064] Step 1. Preprocessing and initial sample splitting: 1. The input axle box vibration acceleration signal and vehicle speed signal are preprocessed together to complete signal cleaning and synchronization correction, resulting in a preprocessed axle box vibration acceleration signal. With vehicle speed signal The expression is: preprocessing ; 2. The preprocessed axle box vibration acceleration signal is segmented according to fixed rules to generate several initial axle box vibration acceleration samples. All initial axle box vibration acceleration samples together constitute the initial axle box vibration acceleration sample set. The expression is: , For the first Bar box vibration acceleration sample, For sample number, This represents the initial number of axle box vibration acceleration samples.

[0065] Step 2. DBSCAN clustering screening (first-level screening): Initial axle box vibration acceleration sample set The process for performing unsupervised clustering and outlier removal is as follows: 1. For the first sample in the initial sample set k Bar box vibration acceleration sample Calculate and construct the first k Multidimensional statistical eigenvectors of axle box vibration acceleration samples f k The expression is: ; For ravine, As the peak factor, For pulse factor, For spectral entropy, For spectral flatness, The proportion of energy at the main peak. The standard deviation of vehicle speed within the time window; 2. Multidimensional statistical eigenvectors of all axle box vibration acceleration samples f k Z-score standardization is performed along the feature dimension to eliminate dimensional differences. The expression is: ; The feature matrix is ​​composed of all multidimensional statistical eigenvectors. In the A vector composed of features along each dimension and They are respectively The mean and standard deviation, For the first The result after standardization of each feature dimension For feature dimensions.

[0066] 3. The DBSCAN clustering algorithm with preset parameters is used to cluster the standardized feature vectors, and a clustering label is assigned to each sample. The expression is: label ; The standardized feature matrix, The neighborhood radius parameter, The minimum number of samples parameter is used; OpDBSCAN specifies that DBSCAN clustering calculations should be performed. 4. Remove samples marked as noise / outliers, and retain only samples with normal cluster labels to form the first selected sample set. .

[0067] Step 3. Spectral stability constraint screening (second-level screening): For the first screening sample set Perform quasi-stationarity verification and retain stable and valid samples. The procedure is as follows: 1. Regarding Each sample in Divide it into a preset number of consecutive subframes. for The sample numbers were assigned, and the power spectrum of each subframe was calculated using the Welch method, followed by normalization of the power spectrum. 2. Based on the normalized power spectrum of each subframe, calculate the standard deviation of the main peak frequency and the spectral similarity index of the samples. Among these, the main peak drift index... The expression is: , For the first The first sample Subframe main peak frequency Indicates execution sample The standard deviation of the dominant peak frequency across all subframes is calculated to measure the degree of frequency drift over time. The index is obtained by calculating the spectral similarity. ; 3. Based on the preset quantile threshold, the first selected sample set is... Perform a second round of filtering; the filtering criteria expression is: and Only samples that meet the conditions are retained to form a valid sample set. .

[0068] Step 4. Frequency-wavelength domain mapping and non-uniform resampling: For the effective sample set The process of performing domain transformation and standardization to generate usable samples for the model is as follows: 1. Regarding Each sample in Perform a Fast Fourier Transform (FFT). for The sample number is used to obtain the frequency domain amplitude spectrum of the sample, expressed as: ; This indicates that for each sample The complex spectrum obtained by performing a fast Fourier transform. For its frequency The amplitude (i.e., the complex modulus) at a given point reflects the energy distribution of the sample across its frequency components.

[0069] 2. Combined with the average vehicle speed corresponding to this sample This maps frequency points in the frequency domain to physical wavelengths, completing the conversion from the frequency domain to the wavelength domain. The feature transformation is expressed as: ; 3. Perform fixed-length non-uniform resampling on the wavelength domain spectrum, and output a uniform length. A one-dimensional sequence =1200 The expression is: , Indicates resampling; 4. All resampled sequences together constitute the final wavelength-domain normalized sample set. .

[0070] End, output result: Once the process is complete, the final wavelength domain normalized sample set will be returned and output. This serves as the input data for the subsequent wheel non-circularity recognition model.

[0071] To further verify the rationality of feature construction and clustering discrimination, the samples were mapped to a low-dimensional feature space composed of time-domain impact features, spectral structure features, and operating condition fluctuation features. The clustering results were then visualized in three dimensions using the PCA (Principal Component Analysis) method. Figure 4 As shown in the figure. The results indicate that normal samples form well-defined high-density clusters within the principal component space, while outlier samples are mainly distributed around the periphery of the main clusters and in local low-density areas, exhibiting obvious discrete characteristics. This demonstrates that the constructed multidimensional statistical features can effectively distinguish between normal and outlier samples, and DBSCAN can effectively utilize the inconsistencies in the local density distribution of samples to achieve automatic segmentation. Figure 4 The tracking results of typical abnormal samples further show that samples with characteristics such as "amplitude attenuation" and "frequency detuning" all fall outside the main cluster of normal samples and are identified as outliers, while the target samples are stably distributed in the high-density area of ​​normal samples, indicating that the method has a good ability to identify both significant and latent abnormal samples.

[0072] Step 3: Further quasi-stationarity screening is performed on the samples retained in Step 2. Specifically, each sample is divided into 3 consecutive subframes, and the power spectral density of each subframe is estimated using the Welch method. Based on this, the cosine similarity between the normalized spectra of adjacent subframes is calculated, and the standard deviation of the main peak frequency of each subframe is calculated as an indicator of main peak drift. According to the preset 90th quantile threshold, only samples with high spectral similarity and small main peak drift are retained to obtain an effective sample set that meets the quasi-stationarity condition.

[0073] Step 4: Perform a Fast Fourier Transform on the samples retained in Step 3, extract the amplitude spectrum within the preset analysis frequency band, and complete the normalization process; then, based on the average value of the velocity signal within the current time window, map the frequency points to wavelengths to eliminate feature offset, such as... Figure 5 and Figure 6 As shown in Table 1, to balance the fidelity of key features with computational efficiency, the mapped wavelength domain spectrum is non-uniformly resampled according to different wavelength ranges. A smaller sampling interval is used in the 80–180 mm wavelength range to preserve the high-order polygonal short-wavelength features; a larger sampling interval is used in the longer wavelength range to compress redundant background information. After resampling, all samples are uniformly converted into a fixed-length wavelength domain sequence; in this embodiment, the samples are uniformly a one-dimensional input sequence of length 1200. Figure 5 The mapping and feature alignment effect of samples from the frequency domain to the wavelength domain (feature divergence and offset in the frequency domain). Figure 6 This describes the mapping and feature alignment effect of samples from the frequency domain to the wavelength domain (feature consistency and alignment in the wavelength domain).

[0074] Table 1. Non-uniform resampling schemes ; Step 5: To avoid data leakage and ensure the authenticity of the model evaluation results, the sample set is preferably divided by wheels rather than individual samples. That is, all samples obtained from the same wheel within the same turning cycle are grouped into the same data subset to avoid samples from the same source being distributed in the training set, validation set, and test set simultaneously. In a preferred embodiment, a large number of axle box vibration samples are extracted for each valid wheel, and the specific sample distribution statistics are shown in Table 2; among them, a portion of the wheels are randomly selected to construct the test set, and the samples corresponding to the remaining wheels are used for the training set and validation set.

[0075] Table 2. Statistical Table of Sample Set Distribution ; The wavelength-domain standardized samples obtained in step 4 are input into the polygonal wear quantitative identification model AMCNN-Net for training. The AMCNN-Net consists of a convolutional feature extraction module, an attention feature enhancement module, and a task-aware output module, as follows: Figure 2 and Figure 7 As shown in the diagram. The feature extraction module, located at the front end of the model, consists of multiple one-dimensional convolutional layers, activation layers, and pooling layers. It is used to extract local harmonic and impact features from the samples. In a preferred embodiment, the number of convolutional layers is 4, the kernel size is 3, and the number of kernels is 16, 16, 32, and 64 respectively. The attention feature enhancement module is located in the middle layer, preferably employing a multi-head self-attention structure. It models the long-range dependencies between different positions in the wavelength domain sequence and adaptively assigns weights to key features. In a preferred embodiment, this module consists of two parallel sub-modules, each containing a multi-head attention layer and a feedforward network. The multi-head attention layer has 4 attention heads, each with a dimension of 64. The task-aware output module is located at the back end of the model, preferably integrating a global average pooling layer, a fully connected layer, and a Dropout layer. It outputs the roughness horizontal vectors of the wheel from order 1 to 30, achieving end-to-end quantitative identification of polygonal wear on the wheel. During training, the Adam optimizer can be used, with mean squared error or log-mean squared error as the loss function.

[0076] Step 6: In practical applications, offline out-of-roundness measurement of the wheel is unnecessary. Only real-time acquisition of axle box vibration and velocity signals is required. Following steps 1 to 4, the raw signals are windowed, sample-filtered, and standardized in the wavelength domain. The standardized samples are then input into the trained AMCNN-Net model to obtain the predicted roughness levels of orders 1 to 30 for each window. For multiple predictions obtained for the same wheel within the same detection period, statistical fusion is performed according to the order. The median is preferably used as the final recognition value for each order, thus obtaining the final roughness level vectors of orders 1 to 30 for the wheel. The empirical model polygonal wear recognition results are shown below. Figures 8 to 11 As shown, Figure 8 The dominant order is 13. Figure 9 The dominant order is 24. Figure 10 The dominant orders are 15 and 22. Figure 11 The dominant orders are 14 and 23.

[0077] Implementation Results: The identification method constructed using the above implementation method can directly output the 1st to 30th order roughness horizontal vectors of wheels based on on-site measured data from a complete turning cycle. Compared with methods that only perform anomaly detection, dominant order identification, or single frame estimation, this invention can simultaneously characterize low-order and high-order polygonal wear states, exhibiting stronger quantitative capabilities. Comparative experimental results show that, compared with benchmark models such as CNN, OOR-Net, and CNN-LSTM, AMCNN-Net has a higher coefficient of determination. The accuracy is improved by about 18%, and the mean square error (MSE) is reduced by about 12.4%. In the single-order wear identification task, the prediction error of the dominant order is controlled within 0.8 dBre1μm. In the mixed-order wear identification task, the error difference with the optimal benchmark model is only 0.07 dBre1μm. This shows that the present invention has high identification accuracy and strong robustness under complex service conditions.

[0078] The embodiments described above are for illustrative purposes only and are not intended to limit the invention. Therefore, any changes in numerical values ​​or substitutions of equivalent elements should still fall within the scope of this invention.

[0079] The above detailed description will enable those skilled in the art to understand that the present invention can indeed achieve the aforementioned objectives and has complied with the provisions of the Patent Law.

[0080] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention. The above descriptions are merely preferred embodiments of the invention and are not intended to limit the invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention.

[0081] It should be noted that the above description of the process is for illustrative purposes only and does not limit the scope of this specification. Those skilled in the art can make various modifications and changes to the process under the guidance of this specification. However, these modifications and changes remain within the scope of this specification.

[0082] The basic concepts have been described above. Obviously, for those skilled in the art who have read this application, the above disclosure is merely illustrative and does not constitute a limitation of this application. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and corrections to this application. Such modifications, improvements, and corrections are suggested in this application, and therefore, such modifications, improvements, and corrections still fall within the spirit and scope of the exemplary embodiments of this application.

[0083] Furthermore, this application uses specific terms to describe its embodiments. For example, "an embodiment," "one embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic related to at least one embodiment of this application. Therefore, it should be emphasized and noted that "an embodiment," "one embodiment," or "an alternative embodiment" mentioned twice or more in different positions in this specification do not necessarily refer to the same embodiment. In addition, certain features, structures, or characteristics in one or more embodiments of this application can be appropriately combined.

[0084] It should be noted that, in order to simplify the description of the present application and thus aid in the understanding of one or more embodiments of the invention, the foregoing description of the embodiments of the present application sometimes combines multiple features into a single embodiment, drawing, or description thereof. However, this approach of the present application should not be construed as reflecting an intention that the claimed subject matter requires more features than expressly recited in each claim. Rather, the subject of the invention should possess fewer features than in any single embodiment described above.

Claims

1. A method for intelligent quantitative identification of wheel non-circular roughness level based on axle box vibration, characterized in that, include: Step 1: Collect train axle box vibration signal, train speed signal and wheel out-of-roundness data, preprocess the wheel out-of-roundness data to generate a wheel roughness horizontal vector, segment the axle box vibration signal to generate initial vibration samples, and match the initial vibration samples with the wheel roughness horizontal vector to construct an initial sample set. Step 2: Combine the train speed signal to remove outliers from the initial sample set and generate the first selected sample set; Step 3: Perform quasi-stationary sample screening on the first screening sample set to generate a valid sample set; Step 4: Perform spectrum extraction, wavelength domain mapping, and resampling on the effective sample set to generate a wavelength domain normalized sample set; Step 5: Input the wavelength domain standardized sample set into the quantitative identification model for training to obtain the trained quantitative identification model; Step 6: Collect the axle box vibration signal and train speed signal of the train under test, obtain multiple roughness level prediction results of the wheel under test through the trained quantitative recognition model, fuse the multiple roughness level prediction results, and output the final roughness level vector of the wheel under test. In step 2, the statistical feature vectors of each initial vibration sample in the initial sample set are extracted. The statistical feature vectors contain the speed fluctuation features of the train speed signal within the corresponding time window. The statistical feature vectors are standardized, and the standardized statistical feature vectors are clustered using a clustering algorithm. Outliers are removed, and the remaining samples are retained to generate the first screening sample set. In step 3, each sample in the first screening sample set is divided into multiple continuous subframes, the power spectral density of each subframe is calculated, the spectral stability index and the main peak drift index of the sample are calculated based on the power spectral density, and samples that meet the quasi-stationary conditions are screened according to the preset threshold to generate an effective sample set. In step 5, the quantitative identification model is set up with a convolutional feature extraction module, an attention feature enhancement module, and a task perception output module in sequence. The convolutional feature extraction module extracts the local features of the wavelength domain standardized samples, the attention feature enhancement module models the global correlation features of the wavelength domain standardized samples and assigns feature weights, and the task perception output module outputs the prediction result of the wheel roughness level.

2. The intelligent quantitative identification method for non-circular roughness level of wheels based on axle box vibration according to claim 1, characterized in that, In step 1, the preprocessing of the wheel out-of-roundness data includes abnormal peak removal, curvature smoothing, detrending term and beginning-end closure correction. The corrected out-of-roundness data is then transformed to generate a horizontal vector of wheel roughness.

3. The intelligent quantitative identification method for non-circular roughness level of wheels based on axle box vibration according to claim 1, characterized in that, In step 1, the axle box vibration signal is segmented according to a fixed time window to generate initial vibration samples. The initial vibration samples are matched with the corresponding wheel roughness horizontal vector based on the mileage node to complete the construction of the initial sample set.

4. The intelligent quantitative identification method for non-circular roughness level of wheels based on axle box vibration according to claim 1, characterized in that, In step 2, the statistical feature vector is generated by combining the time-domain statistical features and frequency-domain statistical features of the initial vibration sample with the speed fluctuation features corresponding to the train speed signal.

5. The intelligent quantitative identification method for non-circular roughness level of wheels based on axle box vibration according to claim 1, characterized in that, In step 4, the amplitude spectrum of the samples in the effective sample set is extracted and normalized. Based on the average speed of the samples calculated from the train speed signal, the frequency dimension of the amplitude spectrum is mapped to the wavelength dimension. Non-uniform resampling is performed according to the wavelength interval, and all samples are converted into sequences of uniform length to generate a wavelength domain standardized sample set.

6. The intelligent quantitative identification method for non-circular roughness level of wheels based on axle box vibration according to claim 1, characterized in that, In step 5, using a single wheel as the dividing unit, all samples of the same wheel in the wavelength domain standardized sample set are divided into the same data subset, generating training set, validation set and test set respectively. The training set and validation set are used to complete the training and parameter optimization of the quantitative identification model, and the test set is used to verify the performance of the quantitative identification model.

7. The intelligent quantitative identification method for non-circular roughness level of wheels based on axle box vibration according to claim 1, characterized in that, In step 6, for each order of the roughness level vector, all predicted values ​​corresponding to that order are statistically analyzed, and the median of the statistical values ​​is selected as the final value for that order, thus completing the fusion processing of multiple sets of roughness level prediction results.

Images