A Multi-Structure Data-Driven Method and System for Quantitative Identification of Wheel Flat Scars

Abstract

This invention discloses a multi-structure data-driven method and system for quantitative identification of wheel flats. It constructs a vehicle-track rigid-flexible coupling dynamic model, using synthetic flats as excitations for wheel out-of-roundness to obtain the dynamic response of the axle box under different working conditions. The vertical vibration acceleration of the axle box is processed in the time domain, frequency domain, and time-frequency domain to create sample sets with different structural forms. A multi-input convolutional neural network (MCNN) with appropriate structure and configuration parameters is constructed, using vehicle speed signals as network constraints. Sample sets of different structural forms and their combinations are fused with speed signals and input into the MCNN model for training. The accuracy and timeliness of the MCNN model for quantitative identification of wheel flats under different data structures are compared. Based on this invention, the optimal combination of sample input forms with the best recognition performance can be obtained. Furthermore, the resulting recognition model not only has good noise resistance but also exhibits better performance due to using speed as network constraint information, demonstrating overall superiority.

Description

Technical Field

[0001] This invention relates to the field of wheel flat spot recognition technology, specifically to a multi-structure data-driven method and system for quantitative recognition of wheel flat spots. Background Technology

[0002] Wheel flats are a common type of damage to railway vehicle wheel treads, and their occurrence is closely related to insufficient wheel-rail adhesion or excessive braking force leading to wheel slippage or skidding. Wheel flats cause severe periodic wheel-rail impact disturbances. If not addressed promptly, wheel flats will further develop and cause or exacerbate other wheel-rail damage, such as wheel polygonal deformation and rail corrugation, significantly affecting train operation quality and shortening the service life of vehicle and track components, even threatening traffic safety. Therefore, timely and accurate detection of wheel flats is crucial. Currently, the mainstream method for wheel flat detection relies on on-site engineers using specialized equipment to identify the condition of the wheel surface. This method offers high accuracy but is inefficient, often only allowing operation when the train is not in service, and is limited by environmental conditions and testing experience. Its low level of automation makes it difficult to meet the needs of daily maintenance. Therefore, developing accurate and efficient wheel flat condition monitoring methods is of significant guiding importance for research on its formation mechanism, degradation trend prediction, and control strategy formulation.

[0003] Domestic and international researchers have achieved intelligent diagnosis of wheel flats by analyzing various sensor signals collected during train service, such as vibration acceleration, noise, and wheel-rail forces. Among these, vibration acceleration signals possess high information richness and can more comprehensively capture wheel flat fault characteristics. Therefore, various data processing techniques are used to deeply analyze vibration acceleration signals, aiming to extract key features related to wheel flats. Existing wheel flat identification methods can be mainly divided into two categories: those based on traditional signal processing and those based on machine learning. Wheel flat fault identification methods based on traditional signal processing can uncover wheel flat fault features under strong noise interference. Although these methods can effectively extract fault features or find the linear relationship between wheel flat size and signal features, each existing method has certain limitations, and the feature extraction results lack in-depth comparison and analysis. Furthermore, traditional signal processing methods have limited ability to effectively extract data features and cannot automatically and efficiently explore hidden features in complex data.

[0004] In recent years, Machine Learning (ML) has been widely used in the field of wheel and rail damage detection due to its excellent performance in establishing complex mapping relationships between data and recognition results. ML-based wheel flat spot recognition methods can not only achieve qualitative recognition, distinguishing whether a wheel has a flat spot fault, but also further achieve quantitative recognition, accurately analyzing the specific size of the wheel flat spot. Depending on the method of processing the raw data, ML can be divided into one-dimensional and two-dimensional signal modes. In existing technologies, although complex nonlinear mapping relationships between sample sets and wheel flat spot sizes can be constructed using ML, there are few reports in the literature on which input form or input combination can achieve good recognition performance. Furthermore, most quantitative recognition research on wheel flat spots focuses on identifying a single size of the wheel flat spot. However, in actual vehicle service, the mapping relationship between the length and depth of the wheel flat spot is not singular and fixed. Therefore, quantitative recognition of wheel flat spots only for a single size cannot meet the needs of comprehensive and efficient wheel flat spot condition monitoring. Summary of the Invention

[0005] To address this, the present invention provides a multi-structure data-driven method and system for quantitative identification of wheel flats, which solves the problem that existing technologies have not found any research on what sample input form or input combination can achieve good identification performance, and that quantitative identification of wheel flats of only a single size cannot meet the needs of comprehensive and efficient wheel flat condition monitoring.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] According to a first aspect of the present invention, a multi-structure data-driven method for quantitative identification of wheel flat marks is proposed, the method comprising:

[0008] Based on the superposition of measured data on the surface roughness of wheel after turning and the wear data of ideal flat-scar wheel, a dataset of flat-scar wheel out-of-roundness under different flat-scar sizes is constructed.

[0009] The dataset of wheel out-of-roundness with different wheel out-of-roundness sizes is used as the wheel irregularity excitation and input into the constructed vehicle-track rigid-flexible coupling dynamic simulation model to obtain simulated axle box vibration acceleration data under various vehicle operating speed conditions.

[0010] The simulated axle box vibration acceleration data is sliced ​​and processed in the time domain, frequency domain, and time-frequency domain to obtain three different structural forms of simulated axle box vibration acceleration sample sets. The three different structural forms of simulated axle box vibration acceleration sample sets are then divided into training sets and test sets for different structural forms.

[0011] A multi-input convolutional neural network model is constructed, which includes a feature extraction module and a regression prediction module. The feature extraction module is used to extract features from input samples with different structural forms respectively. The regression prediction module is used to use the fusion results of sample features with different structural forms and their combination features with vehicle speed signals as different sample input forms, and outputs the quantitative identification result of wheel flat spot size.

[0012] The constructed multi-input convolutional neural network model is trained and tested using the training and test sets, respectively. The model recognition performance under different sample input formats is compared to obtain the optimal sample input format with the best recognition performance. Based on the obtained optimal sample input format and optimal model, quantitative identification of wheel flat scars is performed.

[0013] Furthermore, based on the superposition of measured wheel surface roughness data after turning and ideal flat-spot wheel wear data, a dataset of flat-spot wheel out-of-roundness under different flat-spot sizes is constructed, specifically including:

[0014] The method for measuring the surface roughness data of the wheel after turning includes: releasing the vehicle brake, placing the jack under the axle box to lift the wheel, rotating the wheel at a constant speed, and contacting the sensor probe of the wheel surface roughness measuring instrument BST to the wheel surface to measure the wheel diameter change, thereby obtaining the circumferential data of the wheel, and performing burr removal and curvature smoothing.

[0015] The measured data of unevenness on the wheel surface after turning is superimposed with the mathematical model of wear of the ideal wheel flat scar under different wheel flat scar sizes, and the data of random circumferential position is synthesized to form a flat scar wheel non-roundness dataset under different wheel flat scar sizes. The wheel flat scar size includes flat scar length and flat scar depth.

[0016] Furthermore, the methods for constructing the vehicle-track rigid-flexible coupling dynamic simulation model include:

[0017] The wheelset, rail, and track bed are all considered as flexible bodies. A finite element model of the track bed is established using the finite element method, and the rail is simulated using Timoshenko beams. The elastic treatment of the rail and track bed is realized in the dynamic model through the modal superposition method. Secondly, a rotating elastic wheelset model is established based on rotor dynamics theory, considering the gyroscopic effect and elastic vibration of the high-speed rotation of the wheelset, and a rigid-flexible coupled dynamic simulation model is constructed.

[0018] Furthermore, the simulated axle box vibration acceleration data is sliced ​​and sampled, specifically including:

[0019] The simulated axle box vibration acceleration data is sliced ​​and resampled in units of a preset number of wheel rotations, with the overlap amount set as the scale corresponding to one wheel rotation.

[0020] Furthermore, the simulated axle box vibration acceleration data is sliced ​​and sampled, and then processed in the time domain, frequency domain, and time-frequency domain, specifically including:

[0021] The time-domain processing involves reducing or expanding the data points by interpolating the data segments at different vehicle speeds using cubic spline curves, and uniformly organizing each data segment into a preset number of data points.

[0022] Frequency domain processing involves performing Fast Fourier Transform and frequency domain sample normalization on data segments at different vehicle speeds, and combining each data point with its corresponding frequency information in the frequency domain to construct a frequency domain sample dataset with two channels. The frequency domain sample normalization process involves taking the first N data points as frequency domain samples when the minimum resolution of each frequency domain sample is different. The selection of N must ensure that the normalized frequency domain samples are effective frequency domain samples that occupy at least 99.99% of the frequency domain energy in the overall segment.

[0023] The time-frequency domain processing involves transforming the one-dimensional vertical vibration acceleration data of the axle box into a two-dimensional time-frequency graph using continuous wavelet transform, where the complex Morlet wavelet basis function is used during the continuous wavelet transform.

[0024] Furthermore, a multi-input convolutional neural network model is constructed, specifically including:

[0025] The feature extraction module includes a one-dimensional feature extraction module for extracting features from one-dimensional time-domain or frequency-domain samples and a two-dimensional feature extraction module for extracting features from two-dimensional time-frequency-domain samples.

[0026] The one-dimensional feature extraction module includes multiple one-dimensional convolutional layers, multiple one-dimensional pooling layers, and a one-dimensional global max pooling layer, with the one-dimensional convolutional layers and one-dimensional pooling layers arranged alternately; the two-dimensional feature extraction module includes multiple two-dimensional convolutional layers, multiple two-dimensional pooling layers, and a two-dimensional global max pooling layer, with the two-dimensional convolutional layers and two-dimensional pooling layers arranged alternately; the global max pooling layer performs max pooling on each feature map to generate a fixed-size output.

[0027] Furthermore, a multi-input convolutional neural network model is constructed, specifically including:

[0028] The regression prediction module includes an information fusion layer, two fully connected layers, and an output layer connected in sequence.

[0029] The information fusion layer connects a one-dimensional global max pooling layer and a two-dimensional global max pooling layer. The information fusion layer is used to fuse the feature extraction results and their combined features of one-dimensional time domain, one-dimensional frequency domain and two-dimensional time and frequency domain samples with the vehicle speed signal as different sample input forms, which are then input to the fully connected layer; the output layer outputs the quantitative identification result of wheel flats.

[0030] Furthermore, different sample input formats include individual sample features and combined sample features;

[0031] The individual sample features include time-domain sample features, frequency-domain sample features, and time-frequency-domain sample features;

[0032] The combined sample features include time-domain and frequency-domain sample combination features, time-domain and time-frequency-domain sample combination features, frequency-domain and time-frequency-domain sample combination features, and time-domain and frequency-domain sample combination features.

[0033] Furthermore, the constructed multi-input convolutional neural network model is trained and tested using the training and test sets, respectively, specifically including:

[0034] The Adam optimizer was used to train the network with a learning rate of 0.001 and mean squared error (MSE) as the loss function. Batch processing of samples was used for training. The mean absolute percentage error (MAPE) and coefficient of determination (R²) were used. 2 To comprehensively evaluate the model's performance using evaluation metrics, a 10-fold cross-validation method was used to verify the model's recognition performance.

[0035] According to a second aspect of the present invention, a multi-structure data-driven quantitative identification system for wheel flats is proposed, the system comprising:

[0036] The flat scar dataset construction module is used to construct a flat scar wheel out-of-roundness dataset under different flat scar sizes by superimposing measured wheel surface roughness data and ideal flat scar wheel wear data.

[0037] The simulation module is used to take the non-circularity dataset of wheel scuffs with different wheel scuff sizes as the wheel irregularity excitation and input it into the constructed vehicle-track rigid-flexible coupling dynamic simulation model to obtain simulated axle box vibration acceleration data under various vehicle operating speed conditions.

[0038] The sample set construction module is used to slice and sample the simulated axle box vibration acceleration data and perform time-domain, frequency-domain and time-frequency-domain processing to obtain three different structural forms of simulated axle box vibration acceleration sample sets. The three different structural forms of simulated axle box vibration acceleration sample sets are divided into training sets and test sets for different structural forms.

[0039] The model building module is used to build a multi-input convolutional neural network model. The multi-input convolutional neural network model includes a feature extraction module and a regression prediction module. The feature extraction module is used to extract features from input samples with different structural forms respectively. The regression prediction module is used to use the fusion results of sample features with different structural forms and their combination features with vehicle speed signals as different sample input forms, and output quantitative identification results of wheel flat spot size.

[0040] The model training and testing module is used to train and test the constructed multi-input convolutional neural network model using the training set and the test set respectively, compare the model recognition performance under different sample input forms, obtain the optimal sample input form with the best recognition performance, and perform quantitative recognition of wheel flat scars based on the obtained optimal sample input form and the optimal model.

[0041] This invention proposes a multi-structure data-driven method and system for quantitative identification of wheel flat marks. It integrates measured data of wheel surface irregularities with a mathematical model of flat mark wear to form a flat mark wheel out-of-roundness dataset. A vehicle-track rigid-flexible coupling dynamic model is constructed, using the synthesized flat mark as the excitation for wheel out-of-roundness to obtain the axle box dynamic response under different working conditions. The vertical vibration acceleration of the axle box is processed in the time domain, frequency domain, and time-frequency domain to create sample sets with different structural forms. A multi-input convolutional neural network (MCNN) with appropriate structure and configuration parameters is constructed, using vehicle speed signals as network constraints. Sample sets of different structural forms and their combinations are fused with speed signals and input into the MCNN model for training. The accuracy and timeliness of the MCNN model for quantitative identification of wheel flat marks under different data structures are compared. Based on this invention, the optimal combination of sample input forms with the best recognition performance can be obtained. Furthermore, the resulting recognition model not only has good noise resistance but also exhibits better performance due to using speed as network constraint information, demonstrating overall superiority. Attached Figure Description

[0042] To more clearly illustrate the embodiments of the present invention or the technical solutions in 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 merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0043] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0044] Figure 1 A flowchart of a multi-structure data-driven quantitative identification method for wheel flat scars provided in an embodiment of the present invention;

[0045] Figure 2 A flowchart illustrating the specific implementation of a multi-structure data-driven quantitative identification method for wheel flat scars provided in this embodiment of the invention;

[0046] Figure 3 This is the result of generating data on the surface roughness of a wheel with flat marks in a multi-structure data-driven quantitative identification method for wheel flat marks provided in an embodiment of the present invention.

[0047] Figure 4 A schematic diagram of the subway vehicle-track coupled dynamics model in a multi-structure data-driven quantitative identification method for wheel flats provided in an embodiment of the present invention;

[0048] Figure 5 This is the dynamic model verification result in a multi-structure data-driven quantitative identification method for wheel flat scars provided in an embodiment of the present invention;

[0049] Figure 6 This is a magnified view of a portion of the time-domain samples in a multi-structure data-driven quantitative identification method for wheel flat scars provided in an embodiment of the present invention.

[0050] Figure 7 This is a partial frequency domain sample from a multi-structure data-driven quantitative identification method for wheel flat scars provided in an embodiment of the present invention;

[0051] Figure 8 This is an example of a frequency domain sample set in a multi-structure data-driven quantitative identification method for wheel flat scars provided in an embodiment of the present invention;

[0052] Figure 9 This is a partial time-frequency domain sample from a multi-structure data-driven quantitative identification method for wheel flat scars provided in an embodiment of the present invention;

[0053] Figure 10 This is a schematic diagram of the specific structure of the MCNN model in a multi-structure data-driven quantitative identification method for wheel flat scars provided in an embodiment of the present invention;

[0054] Figure 11 The test results of the MCNN model in a multi-structure data-driven quantitative identification method for wheel flat scars provided in an embodiment of the present invention;

[0055] Figure 12 The evolution of the MCNN model evaluation index and loss curve in a multi-structure data-driven quantitative identification method for wheel flat scars provided in this embodiment of the invention;

[0056] Figure 13 Comparative experimental results in a multi-structure data-driven quantitative identification method for wheel flat scars provided in an embodiment of the present invention;

[0057] Figure 14 This is a schematic diagram of a multi-structure data-driven quantitative identification system for wheel flat marks provided in an embodiment of the present invention. Detailed Implementation

[0058] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Multi-input Convolutional Neural Networks (MCNNs) have multiple input channels, enabling them to process data from different types or structures simultaneously. This architecture not only allows MCNNs to more adaptively handle heterogeneous information from multiple sources compared to traditional Convolutional Neural Networks (CNNs), but also significantly enhances the model's feature extraction capabilities and application generalization performance. The MCNN model established in this invention mainly consists of two parts: feature extraction and regression prediction.

[0060] (1) Feature extraction part

[0061] The feature extraction section consists of alternating convolutional layers, max-pooling layers, activation layers, and global max-pooling layers. Convolutional layers are the core component of a CNN. Each convolutional layer contains multiple filters responsible for extracting specific features from the input of the previous layer. These features are then encoded into feature maps. All neurons within a filter are directly connected to the input data points and perform weighted operations on these data points; all neurons in the same filter share a set of weight parameters. This weight-sharing mechanism effectively reduces the optimization complexity of the network, thereby accelerating the model training process. The mathematical model of a convolutional layer is as follows:

[0062]

[0063] In the formula, i and j are the position indices of the input and output data, respectively; represents the i-th input feature of layer n-1; f(·) is the activation function; M is the number of input feature maps; * represents the convolution operation; These are the weights of the convolution kernel; The bias matrix is ​​used; the activation function is used to introduce nonlinear factors, enabling the model to learn more complex feature patterns, suppress overfitting, and accelerate model convergence. This invention selects the common Corrected Linear Unit (ReLU) as the activation function, and its mathematical expression is as follows:

[0064] f(x) = max{0,x} (2)

[0065] In the formula, x represents the feature information input to the nonlinear activation layer operation.

[0066] Pooling layers typically follow activation layers and serve to reduce feature dimensionality. By using a specific stride to downsample within the local receptive domain, model parameters and computational scale are reduced, and network robustness is enhanced. This invention employs max pooling for downsampling, which effectively reduces the number of model parameters and imparts feature shift invariance. The specific computational process of max pooling is as follows:

[0067]

[0068] In the formula, The values ​​of neurons in the (n+1)th layer after pooling operations; Let t be the value of the t-th neuron in the i-th feature matrix of the n-th layer; t∈[(j-1)W+1,jW], where W is the width of the pooling region. The global max-pooling layer performs max pooling on each feature map to generate a fixed-size output, reducing the input dimension of the fully connected layer. The specific computation process is as follows:

[0069]

[0070] In the formula, `t` is the output value of the `i`-th feature map in layer `n+1` after global max pooling, and `T` is the total number of neurons in the `i`-th feature map in layer `n`. Compared to traditional pooling layers, global max pooling layers provide a more efficient way to extract features and capture global optimization information, and have demonstrated superior performance in various application scenarios.

[0071] (2) Regression Prediction Section

[0072] The regression prediction part consists of an information fusion layer, a fully connected layer, and an output layer. The information fusion layer fuses the feature information extracted by the CNN with the vehicle speed signal and inputs it to the fully connected layer. The fully connected layer has non-linear transformation and feature interaction capabilities, thereby increasing the model's feature extraction and application generalization performance. Its mathematical expression is as follows:

[0073] Y k =f(W k X k-1 +B k (5)

[0074] In the formula, k is the index of the network layer; Y k For output; X k-1 It is an expanded one-dimensional eigenvector; W k B is the weighting coefficient. k This is a bias term.

[0075] The first embodiment of this invention provides a method for quantitative identification of wheel flat marks driven by multi-structure data. The following is a combination of... Figure 1 and Figure 2 Please provide an explanation.

[0076] This invention proposes a method for quantitative identification of wheel flats based on multi-structure data. This method establishes a rigid-flexible coupling dynamic model of a subway vehicle-track system, obtains axle box vibration responses under different operating conditions, creates axle box vibration signal sample sets with different data structure representations, constructs an MCNN model to map the correlation between wheel flats and axle box acceleration signals, discusses and analyzes the recognition performance of the MCNN model under different data structure inputs, and ultimately achieves intelligent quantitative identification of wheel flats driven by multi-structure data. The process is as follows: Figure 2 As shown, it can be divided into 5 parts: (1) Production of flat-scar wheel data; (2) Construction of rigid-flexible coupling dynamic simulation model; (3) Creation of multi-structure data sample set; (4) Design and training of MCNN architecture; (5) Quantitative identification and performance analysis of wheel flat scars.

[0077] like Figure 1 As shown, in step S100, based on the superposition of measured wheel surface roughness data after turning and ideal flat-scar wheel wear data, a flat-scar wheel out-of-roundness dataset under different wheel flat-scar sizes is constructed.

[0078] The above steps specifically include: the method for measuring the surface roughness data of the wheel after turning includes: releasing the vehicle brake, placing the jack under the axle box to lift the wheel, rotating the wheel at a uniform speed, and contacting the sensor probe of the wheel surface roughness measuring instrument BST with the wheel surface to measure the wheel diameter change, thereby obtaining the circumferential data of the wheel, and performing burr removal and curvature smoothing processing; superimposing the measured surface roughness data of the wheel after turning with the ideal wheel flat scar wear mathematical model under different wheel flat scar sizes by random circumferential position data, and synthesizing the flat scar wheel non-roundness dataset under different wheel flat scar sizes, wherein the wheel flat scar size includes flat scar length and flat scar depth.

[0079] Data creation for flat-scarred wheel:

[0080] Specifically, the data production for flat-scar wheels mainly consists of two parts: measurement of the actual surface roughness of the wheel after turning and production of ideal flat-scar wear data. To simulate the measured surface roughness of the flat-scar wheel, the collected surface roughness data of the wheel after turning is linearly combined with the ideal flat-scar wear data.

[0081] The process of measuring wheel surface roughness on-site is as follows: The vehicle brakes are released, a jack is placed under the axle box to lift the wheel, and the wheel is rotated at a constant speed. The sensor probe of the BST wheel surface roughness measuring instrument is then brought into contact with the wheel surface to measure the wheel diameter change, thereby obtaining circumferential data. The obtained circumferential data may contain some interference signals caused by small wheel surface defects. Therefore, the measured wheel data needs to be processed by burr removal and curvature smoothing before it can be used to simulate the boundary conditions of the wheel in the dynamic model.

[0082] Extensive experimental data and references indicate that the initial formation stage of wheel allergies (new allergies) is short-lived, often rapidly deteriorating and developing into wear-type allergies (old allergies). This is because during continuous vehicle service, the sharp edges of new allergies are quickly rounded by impact loads, evolving into old allergies. Therefore, this invention conducts quantitative identification research based on a mathematical model of old allergies. The cosine function proposed by Lyon is commonly used to approximate the mathematical form of old allergies on wheels, as shown in the following formula:

[0083]

[0084]

[0085] In the formula, Δr represents the change in wheel diameter; D r The depth of the wheel tread flat spot is represented by 's'; 's' represents the displacement along the length of the flat spot; L f Indicates the length of the abrasion; D max Representative in L f Under the given conditions, the maximum depth that the wheel flat scar can achieve while maintaining a cosine shape; R represents the wheel radius.

[0086] In actual vehicle service, the length and depth of old flat marks are not a one-to-one mapping relationship. Often, flat marks of the same length correspond to multiple depths due to different wheel wear levels. Researchers have conducted in-depth studies on the multiple mapping relationships between wheel flat mark length and depth. The setting range of flat mark length and depth in the references of this invention, combined with formula (6), designs the flat mark dimensions as shown in Table 1. The multiple corresponding working conditions of flat mark length and depth can better reflect the actual wheel flat mark state.

[0087] Table 1. Dimensions of Simulated Wheel Flat Scars

[0088]

[0089] Figure 3 This document demonstrates the process of creating partial data on the out-of-roundness of wheel slabs with flat marks. It involves overlaying measured surface roughness data of the wheel after turning with a mathematical model characterizing the wear features of an ideal wheel slab with flat marks at random circumferential positions. The process of creating out-of-roundness data for a 100mm long and 0.1mm deep wheel slab is shown as an example. Figure 3 (a) Data on the surface roughness of the wheel after turning. Figure 3 (b) represents ideal flat scar wear data. Figure 3 (c) to make Figure 3 (a) and Figure 3 (b) The out-of-roundness data of the wheels used as input to the dynamic model after synthesis. Figure 3 (d) shows another composite result of the measured wheel surface roughness data after turning and the wheel flat spot wear data (the circumferential position of the wheel flat spot changes). The process for generating the other flat spot wheel out-of-roundness data is the same and will not be described in detail.

[0090] like Figure 1 As shown, in step S200, the dataset of wheel out-of-roundness under different wheel out-of-roundness sizes is used as the wheel's irregularity excitation and input into the constructed vehicle-track rigid-flexible coupling dynamic simulation model to obtain simulated axle box vibration acceleration data under various vehicle operating speed conditions.

[0091] In this embodiment, the method for constructing the vehicle-track rigid-flexible coupling dynamic simulation model includes: considering the wheelset, rail, and track bed as flexible bodies; establishing a finite element model of the track bed using the finite element method, simulating the rail using Timoshenko beams, and realizing the elastic treatment of the rail and track bed in the dynamic model through modal superposition; secondly, establishing a rotating elastic wheelset model based on rotor dynamics theory, considering the gyroscopic effect and elastic vibration of the high-speed rotation of the wheelset, and constructing a rigid-flexible coupling dynamic simulation model.

[0092] Construction of rigid-flexible coupling dynamic simulation model:

[0093] Specifically, measured axle box vibration acceleration data under wheel flat scar excitation suffers from problems such as a small sample size and low variety. Therefore, this invention establishes a dynamic simulation model to obtain the axle box dynamic response under various working conditions.

[0094] Since a purely rigid body model cannot effectively simulate the high-frequency interaction between the wheel and rail caused by wheel flats, the wheelset, rail, and track bed are all considered as flexible bodies, constructing a rigid-flexible coupled dynamic numerical model. A finite element model of the track bed is established using the finite element method, and the rail is simulated using Timoshenko beams. The elastic treatment of the rail and track bed is achieved in the dynamic model through modal superposition. Secondly, a rotating elastic wheelset model is established based on rotor dynamics theory, considering the gyroscopic effect and elastic vibration of the wheelset during high-speed rotation, to achieve a more realistic rigid-flexible coupled dynamic simulation.

[0095] The established dynamic model of the metro vehicle-track rigid-flexible coupling is as follows: Figure 4As shown, the synthesized flat-scar wheel sample set was used as the wheel irregularity excitation for the subway vehicle-track rigid-flexible coupling dynamic model, and the track irregularity excitation adopted the US five-level spectrum. The vehicle operating speed was set to 40–120 km / h, with an interval of 10 km / h, and the sampling frequency was 5000 Hz. To verify the reliability of the established dynamic model, the measured wheel out-of-roundness was input into the dynamic simulation model. Figure 5 The results show a comparison between measured and simulated axle box vibration acceleration data. It can be seen that the simulated signal has a high degree of consistency with the measured data in both the time and frequency domains, indicating that the established simulation model has high accuracy.

[0096] like Figure 1 As shown, in step S300, the simulated axle box vibration acceleration data is sliced ​​and sampled, and processed in the time domain, frequency domain and time-frequency domain to obtain three different structural forms of simulated axle box vibration acceleration sample sets. The three different structural forms of simulated axle box vibration acceleration sample sets are then divided into training sets and test sets for different structural forms.

[0097] The above steps specifically include: resampling the simulated axle box vibration acceleration data in slices based on a preset number of wheel rotations, with the overlap amount set as the scale corresponding to one wheel rotation.

[0098] In this example, the time-domain processing involves reducing or expanding the data points of data segments at different vehicle speeds using cubic spline curve interpolation, and uniformly organizing each data segment into a preset number of data points.

[0099] Frequency domain processing involves performing Fast Fourier Transform and frequency domain sample normalization on data segments at different vehicle speeds, and combining each data point with its corresponding frequency information in the frequency domain to construct a frequency domain sample dataset with two channels. The frequency domain sample normalization process involves taking the first N data points as frequency domain samples when the minimum resolution of each frequency domain sample is different. The selection of N must ensure that the normalized frequency domain samples are effective frequency domain samples that occupy at least 99.99% of the frequency domain energy in the overall segment.

[0100] The time-frequency domain processing involves transforming the one-dimensional vertical vibration acceleration data of the axle box into a two-dimensional time-frequency graph using continuous wavelet transform, where the complex Morlet wavelet basis function is used during the continuous wavelet transform.

[0101] Creating different input sample sets:

[0102] Specifically, wheel flats cause periodic impact disturbances, meaning that an impact caused by the flats occurs every time the wheel rotates once. Therefore, this invention uses the number of wheel rotations as a reference benchmark to slice the axle box vibration response data output by the simulation model. Emphasis must be placed on the slice unit size, i.e., the number of wheel rotations, to ensure that the slice data accurately and representatively reflects the wheel flats characteristics. Using one wheel rotation as the slicing unit may result in model accuracy being affected by other excitation sources during vehicle service. Furthermore, due to factors such as track conditions and wheel-rail contact, the wheel flats characteristics may be weaker at a particular wheel rotation, thus reducing the model's ability to recognize wheel flats. However, while using multiple rotations as the slicing unit can reduce the influence of external factors on the axle box dynamic response to some extent, it leads to a smaller sample size and increases the risk of model overfitting. Based on extensive experiments and literature review, this invention resamples the slices every four wheel rotations, with the overlap set to the scale corresponding to one wheel rotation.

[0103] Considering that the simulation model operates at speeds of 40–120 km / h (in 10 km / h intervals), and the length of the data points corresponding to each cycle of wheel rotation is inconsistent, the simulation signal can be processed in the time domain, frequency domain, and time-frequency domain respectively, namely Cubic Spline Interpolation (CSI), Fast Fourier Transform (FFT), and Continuous Wavelet Transform (CWT), to unify the sample scale and explore the impact of different data structure forms on the wheel flat spot recognition results.

[0104] During time-domain processing, data segments at different speeds are reduced or expanded using CSI. Each data segment is uniformly organized into 2048 data points (the median number of data points at speeds from 40 to 120 km / h, and data lengths that are powers of 2 are more suitable for MCNN training). Figure 6 The time-domain waveforms of the axle box vibration response during one wheel rotation are shown under extreme speed conditions (minimum speed 40 km / h and maximum speed 120 km / h). It can be seen that the data after CSI processing still reflects the time-domain characteristics of the original signal quite well.

[0105] In frequency domain processing, the use of overlapping slices based on the circumference of four wheels results in differences in the time scale of each data segment, leading to varying frequency resolutions after FFT processing. The frequency resolution Δf is related to the sampling duration T of the signal being analyzed. s The relationship is as follows:

[0106]

[0107] Therefore, using the conventional data point normalization method with a set maximum cutoff frequency will still result in inconsistent data point counts for each sample segment. Furthermore, considering the rich frequency-domain variations in axle box vibration acceleration signals containing wheel flattening features, using simple mathematical interpolation for data point normalization may lead to the loss of crucial features.

[0108] In summary, this invention proposes a specific frequency domain data normalization method, which involves fixing the first N data points as frequency domain samples when the minimum resolution of each frequency domain sample is different. At a sampling frequency of 5000Hz, after FFT processing, according to the sampling theorem, data within the range of 0–2500Hz are all valid. However, not all frequencies within this range are practically meaningful for the quantitative identification of wheel flats. In fact, some frequencies correspond to very small amplitudes, almost zero, and these frequencies may be redundant for quantitative identification. Therefore, to ensure the representativeness of the selected frequency domain data, this invention refers to the 3σ principle in statistics, defining frequency domain segments with an energy proportion exceeding 99.90% as "effective frequencies" from the perspective of energy distribution. The formula for the frequency domain energy proportion is as follows:

[0109]

[0110] In the formula, T e E represents the proportion of energy in the frequency domain. total and E N These represent the energy of the complete frequency domain sample and the energy of the normalized frequency domain sample, respectively; M is the number of points in the frequency domain sample set; and A(m) is the amplitude corresponding to the m-th frequency data point.

[0111] This invention is based on a maximum vehicle speed of 120 km / h, taking the first 768 data points as the normalization reference N. At this point, the normalized frequency domain samples account for as much as 99.99% of the energy in the overall segment. Moreover, this selection ensures that under other speed conditions, the energy proportion of the normalized frequency domain samples remains above 99.90%, that is, the normalized frequency domain sample set at different speeds is an "effective frequency domain".

[0112] Figure 7 The results show the axle box dynamic response after frequency domain sample normalization and its time-domain waveform obtained by inverse fast Fourier transform under the conditions of a minimum speed of 40 km / h and a maximum speed of 120 km / h. Figure 7 As shown in (a) to (b), the above processing methods not only maintain the frequency domain variation characteristics of the axle box vibration acceleration signal, but also retain most of the characteristic frequencies related to wheel flats and their integer multiples. Furthermore, Figure 7 (c) to (d) show that the time-domain waveform obtained by the inverse fast Fourier transform of the normalized frequency domain samples has good consistency with the original signal. In summary, this shows that the set normalization reference N0 can include most of the meaningful frequency domain data.

[0113] Although the frequency domain data normalization method used in this invention can unify the size of the frequency domain sample set, the frequency domain meaning represented by each point differs due to the different speeds corresponding to each sample set. For example... Figure 8 As shown, under the same operating conditions of a minimum speed of 40 km / h and a maximum speed of 120 km / h, for ease of explanation, we take the first 10 data points in the sample set as an example. The frequency ranges represented by the first 10 data points are approximately 0–10 Hz and 0–30 Hz, respectively. These data points represent different frequency information, but when they are input into MCNN, each data point is forced to correspond sequentially, resulting in the loss of corresponding frequency domain information. To solve this problem, the frequency information corresponding to each data point in the frequency domain (the dashed box in the figure) is combined with it to construct a frequency domain sample dataset with two channels.

[0114] In the time-frequency domain processing, the one-dimensional vertical vibration acceleration data of the axle box is transformed into a two-dimensional time-frequency plot using CWT. For any function g(t) in L2(R) space, CWT is defined as:

[0115]

[0116]

[0117] In the formula, ψ a,τ (t) is obtained by scaling and translating the mother wavelet ψ(t); a is the scaling factor; τ is the translation factor; the symbol <> is the inner product operator; ψ(t) is the complex conjugate of ψ(t). The key to CWT lies in the selection of wavelet basis functions. This invention uses the complex Morlet wavelet, which has better adaptability. The specific calculation process is existing technology and will not be described here.

[0118] The final processing result is as follows Figure 9 As shown, taking the time-frequency domain processing results at speeds of 40 and 120 km / h as examples, the frequency range is selected as 0–1500 Hz, and it is converted into a three-channel RGB image with a size of 128*128*3. The sample sets at other speeds and scar sizes are similar and will not be described in detail.

[0119] The above processing methods create three different sample sets from the overlapping sliced ​​wheel flat axle box dynamic response data: time domain, frequency domain, and time-frequency domain. Table 2 lists the specific number of samples for one of these data structures. The other two data structures have the same number of samples, totaling 33,417. The number of wheel flat sample sets with different speeds and sizes is evenly distributed, with 80% of the total sample set used as the training set and the remaining 20% ​​as the test set.

[0120] Table 2. Sample set of wheel flat scars in time domain / frequency domain / time-frequency domain

[0121]

[0122]

[0123] like Figure 1 As shown, in step S400, a multi-input convolutional neural network model is constructed. The multi-input convolutional neural network model includes a feature extraction module and a regression prediction module. The feature extraction module is used to extract features from input samples with different structural forms respectively. The regression prediction module is used to use the fusion results of sample features with different structural forms and their combined features with the vehicle speed signal as different sample input forms, and output the quantitative identification result of the wheel flat spot size.

[0124] In this embodiment, the feature extraction module includes a one-dimensional feature extraction module for extracting features from one-dimensional time-domain or frequency-domain samples and a two-dimensional feature extraction module for extracting features from two-dimensional time-frequency-domain samples. The one-dimensional feature extraction module includes multiple one-dimensional convolutional layers, multiple one-dimensional pooling layers, and a one-dimensional global max pooling layer, with the one-dimensional convolutional layers and one-dimensional pooling layers arranged alternately. The two-dimensional feature extraction module includes multiple two-dimensional convolutional layers, multiple two-dimensional pooling layers, and a two-dimensional global max pooling layer, with the two-dimensional convolutional layers and two-dimensional pooling layers arranged alternately. Max pooling is performed on each feature map by the global max pooling layer to generate a fixed-size output.

[0125] In this embodiment, the regression prediction module includes an information fusion layer, two fully connected layers, and an output layer connected in sequence. The information fusion layer is connected to a one-dimensional global max pooling layer and a two-dimensional global max pooling layer. The information fusion layer is used to fuse the feature extraction results and their combined features of one-dimensional time domain, one-dimensional frequency domain, and two-dimensional time-frequency domain samples with the vehicle speed signal as different sample input forms, which are then input to the fully connected layer. The output layer outputs the quantitative identification results of wheel flats.

[0126] In this embodiment, different sample input formats include individual sample features and combined sample features; the individual sample features include time-domain sample features, frequency-domain sample features, and time-frequency-domain sample features; the combined sample features include time-domain and frequency-domain sample combination features, time-domain and time-frequency-domain sample combination features, frequency-domain and time-frequency-domain sample combination features, and time-domain and frequency-domain sample combination features.

[0127] MCNN architecture design and training:

[0128] Different axle box vibration acceleration signals have their own advantages in wheel flat spot recognition. To explore the impact of different data structure forms and their combinations on the recognition effect of wheel flat spots, this invention designs MCNN models with 7 different data structure inputs, such as... Figure 10 As shown.

[0129] The model's feature extraction section consists of alternating convolutional and pooling layers, designed to automatically extract flattened fault features under different inputs. Notably, the feature extraction sections for different structural forms use the same neural network structure parameters, followed by a global max pooling layer to compress each channel into a single value, reducing the network's computational complexity. The regression prediction section fuses input features from different structural forms through an information fusion layer. Since time-domain, frequency-domain, and time-frequency-domain signals often lack velocity information, velocity signals are introduced as composite features in the MCNN regression prediction section to improve the quantitative detection capability of MCNN for wheel encoding. Finally, two fully connected layers enhance the model's nonlinear expressive power, and an output layer ultimately achieves quantitative identification of the length and depth of wheel flattened faults.

[0130] After extensive parameter tuning experiments, the structural parameters of MCNN were determined, as shown in Table 3. Furthermore, zero-padding was employed in the convolutional operations to maintain the spatial size of the feature maps. To prevent overfitting, an early stopping mechanism was introduced, with a patience set to 50, meaning training stopped when the test set loss no longer decreased within 50 epochs.

[0131] Table 3 MCNN Structure Parameters

[0132]

[0133] like Figure 1 As shown, in step S500, the constructed multi-input convolutional neural network model is trained and tested using the training set and test set respectively, and the model recognition performance under different sample input forms is compared to obtain the optimal sample input form with the best recognition performance. Based on the obtained optimal sample input form and optimal model, the wheel flat scar quantitative recognition is performed.

[0134] The above steps specifically include: training the network using the Adam optimizer with a learning rate of 0.001 and using mean squared error (MSE) as the loss function; training using batch samples; using mean absolute percentage error (MAPE) and coefficient of determination (R²) as evaluation metrics to comprehensively evaluate the model performance; and using ten-fold cross-validation to verify the model's recognition performance.

[0135] Quantitative identification and performance comparison of wheel flat marks:

[0136] In this embodiment, the "Adam" optimizer is used to train the network with a learning rate of 0.001. Mean Squared Error (MSE) is efficiently calculated and exhibits superior performance; therefore, it is used as the loss function, as shown in the following formula:

[0137]

[0138] In the formula, n is the sample size; y i This represents the true value of the i-th sample. Let be the predicted value of the i-th sample. To avoid gradient vanishing and gradient explosion, batch processing of samples is used for training, with a batch size of 128.

[0139] The mean absolute percentage error (MAPE) and the coefficient of determination (R-squared Coefficient) are used. 2 The evaluation index is used to comprehensively evaluate the model's performance, and the corresponding expression is as follows:

[0140]

[0141]

[0142] In the formula, For y i The average value of MAPE is the difference between the predicted and the true values, representing the average relative error; R2 measures how well the model fits the data. The closer MAPE converges to 0 and the closer R2 is to 1, the higher the quantitative recognition accuracy of MCNN for wheel flats.

[0143] MCNN is built on the Keras deep learning library based on Python 3.7.16. The computer hardware configuration is an i7-10700K processor, 16GB of memory, and a Windows 10 system. The performance of various models in recognizing flat tire scars under different structural combinations was compared, the optimal model structure was selected, and finally, quantitative recognition of the length and depth of flat tire scars was achieved.

[0144] (1) Comparison of model training results and recognition performance

[0145] During model training, different sample set partitioning methods and different initialization weights can all affect the final recognition results. To eliminate interference from random terms, a ten-fold cross-validation method was used to observe model performance. Table 4 shows the average recognition performance index of ten-fold cross-validation after different data structure forms and combinations are input into the MCNN network. It can be seen that the recognition performance of multiple input structure forms is better than that of a single input structure form. However, this advantage is accompanied by increased time consumption, especially when the input structure form contains two-dimensional data, the time consumption increases significantly compared to other combinations.

[0146] In multi-input structures, the recognition performance is optimal when the input structure consists of time-domain, frequency-domain, and time-frequency-domain combinations. Its R... 2The MCNN model achieved a score of 0.9978, with a MAPE of approximately 1.947%. The average processing time per sample for this input format was 0.2353 ms, slightly higher than other input formats, but still meeting the timeliness requirements for online monitoring of wheel flats. In summary, the MCNN model (hereinafter referred to as TFTF-MCNN for convenience, where T represents the time domain and F represents the frequency domain) exhibited the best overall recognition performance when the input structure was combined into time domain, frequency domain, and time-frequency domain.

[0147] Table 4 Comparison of Recognition Performance for Different Input Formats

[0148]

[0149] To further analyze the performance of the TFTF-MCNN network, Figure 11 Specifically, the results of the 10-fold cross-validation of the TFTF-MCNN model are presented, and its R... 2 The values ​​fluctuated between 0.9975 and 0.9983, and the MAPE distribution ranged from 1.726% to 2.127%, indicating that the model has good robustness across various data subsets. Furthermore, the processing time for a single sample did not exceed 0.24 ms, meeting the timeliness requirements for online monitoring of wheel flats.

[0150] Taking the first experiment, which had the worst performance, as an example, we analyze the evolution of various evaluation indicators during model training, such as... Figure 12 As shown, due to the reasonable architecture design of TFTF-MCNN, no overfitting problem occurred. The loss curves of the training and test sets gradually converged to 0 with the increase of training epochs, and R² gradually increased with the decrease of loss, eventually converging to around 1; MAPE also gradually decreased with the decrease of loss, eventually converging to around 1.9%. Due to the introduction of the Early Stopping mechanism, when the number of training epochs reached 456, the test set loss had accumulated for 50 training epochs and no longer improved, so the model terminated training. Considering the overall results of ten-fold cross-validation and the changes in the specific evaluation metrics and loss curves of the first experiment, it is further shown that the TFTF-MCNN model exhibits good robustness when facing different data subsets.

[0151] (2) Model generalization performance analysis

[0152] To further verify the generalization performance of the proposed quantitative identification method for wheel flats, based on the discrete flat size range set in Table 2, the rigid-flexible coupling dynamic model of the metro vehicle-track established in this invention was used to randomly simulate and calculate the vertical vibration acceleration of the axle box under 10 random velocities and wheel flat size excitation. The corresponding data were processed in the time domain, frequency domain, and time frequency domain and then input into the trained TFTF-MCNN model for identification and analysis.

[0153] The recognition results are shown in Table 5. It can be seen that the TFTF-MCNN model still has high recognition accuracy and can quantitatively estimate the size of wheel flats relatively well. The overall relative error does not exceed 2.6%. The mean relative errors for wheel flat length and depth are approximately 2.020% and 2.095%, respectively, and the relative errors of the two are quite close. In summary, this indicates that TFTF-MCNN still has high recognition accuracy even when facing specific working conditions not present in the training set.

[0154] Table 5. Validation results of MCNN generalization performance

[0155]

[0156]

[0157] (3) Influence of random noise and velocity factor

[0158] The actual operating environment is quite complex, and axle box vibration signals are often interfered with by the vibration of other components, resulting in a large amount of noise. To simulate the impact of vehicle vibration behavior and track conditions on data acquisition under real-world conditions, all samples were subjected to Gaussian white noise with three different signal-to-noise ratios (-4dB, 0dB, and 4dB) to simulate noise interference in real-world scenarios. The model was then retrained and tested, and ablation experiments were conducted to demonstrate the superiority of fusing speed signals as constraint information into the TFTF-MCNN model. To control experimental variables, the structural parameters of the TFTF-MCNN model remained unchanged.

[0159] In addition, to minimize the impact of random factors on the identification results, Figure 13 The data in the table are all averages after 10-fold cross-validation. The results show that even under -4dB noise, the R-value of the TFTF-MCNN model with velocity constraint information is [not specified]. 2 The values ​​remained above 0.96 and the MAPE value did not exceed 8%, indicating that the proposed TFTF-MCNN model still possesses good recognition performance and feature extraction capabilities in high-noise environments. Furthermore, under all experimental conditions, compared to the TFTF-MCNN model without velocity information, the TFTF-MCNN model with velocity constraint information exhibited better noise resistance at a given signal-to-noise ratio, specifically reflected in a higher R0 value. 2 The higher speed and lower MAPE values ​​further illustrate the superiority of using speed as network constraint information in improving the accuracy of wheel flat spot recognition.

[0160] Compared with existing technologies, this invention provides a multi-structure data-driven quantitative identification method for wheel flat marks. It integrates measured data of wheel surface irregularities with a mathematical model of flat mark wear to form a flat mark wheel out-of-roundness dataset. A vehicle-track rigid-flexible coupling dynamic model is constructed, using the synthesized flat mark as the excitation for wheel out-of-roundness to obtain the axle box dynamic response under different working conditions. The vertical vibration acceleration of the axle box is processed in the time domain, frequency domain, and time-frequency domain to create sample sets with different structural forms. A multi-input convolutional neural network (MCNN) with appropriate structure and configuration parameters is constructed, using vehicle speed signals as network constraints. Sample sets of different structural forms and their combinations are fused with speed signals and input into the MCNN model for training. The accuracy and timeliness of the MCNN model for quantitative identification of wheel flat marks under different data structures are compared. Based on this invention, the optimal combination of sample input forms with the best recognition performance can be obtained. Furthermore, the resulting recognition model not only has good noise resistance but also exhibits better performance due to using speed as network constraint information, demonstrating overall superiority.

[0161] Corresponding to the above-disclosed method for quantitative identification of wheel flats driven by multi-structure data, this invention also discloses a system for quantitative identification of wheel flats driven by multi-structure data, such as... Figure 14 As shown, it specifically includes:

[0162] The flat scar dataset construction module is used to construct a flat scar wheel out-of-roundness dataset under different flat scar sizes by superimposing measured wheel surface roughness data and ideal flat scar wheel wear data.

[0163] The simulation module is used to take the non-circularity dataset of wheel scuffs with different wheel scuff sizes as the wheel irregularity excitation and input it into the constructed vehicle-track rigid-flexible coupling dynamic simulation model to obtain simulated axle box vibration acceleration data under various vehicle operating speed conditions.

[0164] The sample set construction module is used to slice and sample the simulated axle box vibration acceleration data and perform time-domain, frequency-domain and time-frequency-domain processing to obtain three different structural forms of simulated axle box vibration acceleration sample sets. The three different structural forms of simulated axle box vibration acceleration sample sets are divided into training sets and test sets for different structural forms.

[0165] The model building module is used to build a multi-input convolutional neural network model. The multi-input convolutional neural network model includes a feature extraction module and a regression prediction module. The feature extraction module is used to extract features from input samples with different structural forms respectively. The regression prediction module is used to use the fusion results of sample features with different structural forms and their combination features with vehicle speed signals as different sample input forms, and output quantitative identification results of wheel flat spot size.

[0166] The model training and testing module is used to train and test the constructed multi-input convolutional neural network model using the training set and the test set respectively, compare the model recognition performance under different sample input forms, obtain the optimal sample input form with the best recognition performance, and perform quantitative recognition of wheel flat scars based on the obtained optimal sample input form and the optimal model.

[0167] It should be noted that for a detailed description of the multi-structure data-driven quantitative identification system for wheel flats provided in the embodiments of the present invention, please refer to the relevant description of the multi-structure data-driven quantitative identification method for wheel flats provided in the embodiments of this application, which will not be repeated here.

[0168] In addition, embodiments of the present invention also provide an electronic device, the device comprising: a processor and a memory; the memory being used to store one or more program instructions; the processor being used to execute one or more program instructions to perform the steps of a multi-structure data-driven quantitative identification method for wheel flat scars as described in any of the preceding embodiments.

[0169] It should be noted that for a detailed description of an electronic device provided in the embodiments of the present invention, please refer to the relevant description of a multi-structure data-driven quantitative identification method for wheel flat scars provided in the embodiments of this application, which will not be repeated here.

[0170] In addition, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of a long text summarization method as described in any of the preceding embodiments.

[0171] It should be noted that for a detailed description of the computer-readable storage medium provided in the embodiments of the present invention, please refer to the relevant description of the multi-structure data-driven quantitative identification method for wheel flat scars provided in the embodiments of this application, which will not be repeated here.

[0172] In this embodiment of the invention, the processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0173] The various methods, steps, and logic diagrams disclosed in the embodiments of this invention can be implemented or executed. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The processor reads information from the storage medium and, in conjunction with its hardware, completes the steps of the above methods.

[0174] The storage medium can be memory, such as volatile memory or non-volatile memory, or may include both volatile and non-volatile memory.

[0175] Among them, non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory.

[0176] Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (Synchlink DRAM, SLDRAM), and direct memory bus RAM (DRRAM).

[0177] The storage media described in the embodiments of the present invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0178] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this invention can be implemented using a combination of hardware and software. When applied as software, the corresponding functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of computer programs from one place to another. Storage media can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0179] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A method for quantitative identification of wheel flat marks driven by multi-structure data, characterized in that, The method includes: Based on the superposition of measured data on the surface roughness of wheel after turning and the wear data of ideal flat-scar wheel, a dataset of flat-scar wheel out-of-roundness under different flat-scar sizes is constructed. The dataset of wheel out-of-roundness with different wheel out-of-roundness sizes is used as the wheel irregularity excitation and input into the constructed vehicle-track rigid-flexible coupling dynamic simulation model to obtain simulated axle box vibration acceleration data under various vehicle operating speed conditions. The simulated axle box vibration acceleration data is sliced ​​and processed in the time domain, frequency domain, and time-frequency domain to obtain three different structural forms of simulated axle box vibration acceleration sample sets. The three different structural forms of simulated axle box vibration acceleration sample sets are then divided into training sets and test sets for different structural forms. A multi-input convolutional neural network model is constructed, which includes a feature extraction module and a regression prediction module. The feature extraction module is used to extract features from input samples with different structural forms respectively. The regression prediction module is used to use the fusion results of sample features with different structural forms and their combination features with vehicle speed signals as different sample input forms, and outputs the quantitative identification result of wheel flat spot size. The constructed multi-input convolutional neural network model is trained and tested using the training set and test set respectively, and the model recognition performance under different sample input forms is compared. The optimal sample input form with the best recognition performance is obtained, and the wheel flat scar is quantitatively identified based on the obtained optimal sample input form and optimal model. Based on the superposition of measured data on surface roughness of turned wheels and wear data of ideal flat-scar wheels, a dataset of non-roundness of flat-scar wheels under different flat-scar sizes is constructed, specifically including: The method for measuring the surface roughness data of the wheel after turning includes: releasing the vehicle brake, placing the jack under the axle box to lift the wheel, rotating the wheel at a constant speed, and contacting the sensor probe of the wheel surface roughness measuring instrument BST to the wheel surface to measure the wheel diameter change, thereby obtaining the circumferential data of the wheel, and performing burr removal and curvature smoothing. The measured data of unevenness on the wheel surface after turning are superimposed with the mathematical model of wear of ideal wheel flat marks under different wheel flat mark sizes, and the data of random circumferential position are combined to synthesize the flat mark wheel non-roundness dataset under different wheel flat mark sizes. The wheel flat mark size includes flat mark length and flat mark depth. The methods for constructing vehicle-track rigid-flexible coupling dynamic simulation models include: The wheelset, rail, and track bed are all considered as flexible bodies. A finite element model of the track bed is established using the finite element method, and the rail is simulated using Timoshenko beams. The elastic treatment of the rail and track bed is realized in the dynamic model through the modal superposition method. Secondly, a rotating elastic wheelset model is established based on rotor dynamics theory, considering the gyroscopic effect and elastic vibration of the high-speed rotation of the wheelset, and a rigid-flexible coupled dynamic simulation model is constructed. The simulated axle box vibration acceleration data is sliced ​​and sampled, specifically including: The simulated axle box vibration acceleration data is sliced ​​and resampled in units of a preset number of wheel rotations, with the overlap amount set as the scale corresponding to one wheel rotation.

2. The method for quantitative identification of wheel flat marks driven by multi-structure data according to claim 1, characterized in that, The simulated axle box vibration acceleration data is sliced ​​and sampled, and then processed in the time domain, frequency domain, and time-frequency domain, specifically including: The time-domain processing involves reducing or expanding the data points by interpolating the data segments at different vehicle speeds using cubic spline curves, and uniformly organizing each data segment into a preset number of data points. Frequency domain processing involves performing Fast Fourier Transform and frequency domain sample normalization on data segments at different vehicle speeds, and combining each data point with its corresponding frequency information in the frequency domain to construct a dual-channel frequency domain sample dataset. The frequency domain sample normalization process involves fixing the first few samples while maintaining a different minimum resolution for each frequency domain sample. 10 data points are used as frequency domain samples The selection must ensure that the regular frequency domain samples are all effective frequency domain samples that occupy at least 99.99% of the frequency domain energy in the whole segment; The time-frequency domain processing involves transforming the one-dimensional vertical vibration acceleration data of the axle box into a two-dimensional time-frequency graph using continuous wavelet transform, where the complex Morlet wavelet basis function is used during the continuous wavelet transform.

3. The method for quantitative identification of wheel flat marks driven by multi-structure data according to claim 1, characterized in that, Constructing a multi-input convolutional neural network model specifically includes: The feature extraction module includes a one-dimensional feature extraction module for extracting features from one-dimensional time-domain or frequency-domain samples and a two-dimensional feature extraction module for extracting features from two-dimensional time-frequency-domain samples. The one-dimensional feature extraction module includes multiple one-dimensional convolutional layers, multiple one-dimensional pooling layers, and a one-dimensional global max pooling layer, with the one-dimensional convolutional layers and one-dimensional pooling layers arranged alternately; the two-dimensional feature extraction module includes multiple two-dimensional convolutional layers, multiple two-dimensional pooling layers, and a two-dimensional global max pooling layer, with the two-dimensional convolutional layers and two-dimensional pooling layers arranged alternately; the global max pooling layer performs max pooling on each feature map to generate a fixed-size output.

4. The method for quantitative identification of wheel flat marks driven by multi-structure data according to claim 1, characterized in that, Constructing a multi-input convolutional neural network model specifically includes: The regression prediction module includes an information fusion layer, two fully connected layers, and an output layer connected in sequence. The information fusion layer connects a one-dimensional global max pooling layer and a two-dimensional global max pooling layer. The information fusion layer is used to fuse the feature extraction results and their combined features of one-dimensional time domain, one-dimensional frequency domain and two-dimensional time and frequency domain samples with the vehicle speed signal as different sample input forms, which are then input to the fully connected layer; the output layer outputs the quantitative identification result of wheel flats.

5. The method for quantitative identification of wheel flat marks driven by multi-structure data according to claim 1, characterized in that, Different sample input formats include individual sample features and combined sample features; The individual sample features include time-domain sample features, frequency-domain sample features, and time-frequency-domain sample features; The combined sample features include time-domain and frequency-domain sample combination features, time-domain and time-frequency-domain sample combination features, frequency-domain and time-frequency-domain sample combination features, and time-domain and frequency-domain sample combination features.

6. The method for quantitative identification of wheel flat marks driven by multi-structure data according to claim 1, characterized in that, The constructed multi-input convolutional neural network model is trained and tested using the training and test sets, respectively, specifically including: The network was trained using the Adam optimizer with a learning rate of 0.001 and mean squared error (MSE) as the loss function. Batch processing of samples was used for training. Mean absolute percentage error (MAPE) and the coefficient of determination were employed. R 2 To comprehensively evaluate the model's performance using evaluation metrics, a 10-fold cross-validation method was used to verify the model's recognition performance.

7. A multi-structure data-driven quantitative identification system for wheel flat marks, characterized in that, The system includes: The flat scar dataset construction module is used to construct a flat scar wheel out-of-roundness dataset under different flat scar sizes by superimposing measured wheel surface roughness data and ideal flat scar wheel wear data. The simulation module is used to take the non-circularity dataset of wheel scuffs with different wheel scuff sizes as the wheel irregularity excitation and input it into the constructed vehicle-track rigid-flexible coupling dynamic simulation model to obtain simulated axle box vibration acceleration data under various vehicle operating speed conditions. The sample set construction module is used to slice and sample the simulated axle box vibration acceleration data and perform time-domain, frequency-domain and time-frequency-domain processing to obtain three different structural forms of simulated axle box vibration acceleration sample sets. The three different structural forms of simulated axle box vibration acceleration sample sets are divided into training sets and test sets for different structural forms. The model building module is used to build a multi-input convolutional neural network model. The multi-input convolutional neural network model includes a feature extraction module and a regression prediction module. The feature extraction module is used to extract features from input samples with different structural forms respectively. The regression prediction module is used to use the fusion results of sample features with different structural forms and their combination features with vehicle speed signals as different sample input forms, and output quantitative identification results of wheel flat spot size. The model training and testing module is used to train and test the constructed multi-input convolutional neural network model using the training set and the test set respectively, compare the model recognition performance under different sample input forms, obtain the optimal sample input form with the best recognition performance, and perform quantitative recognition of wheel flat scars based on the obtained optimal sample input form and the optimal model.

Images