Wheel polygon abrasion quantitative identification method based on generative adversarial network
By using the axle box vibration acceleration signal to quantitatively identify wheel polygon wear based on the method of generating an adversarial network, the problem of quantitative identification of wheel polygon wear in the prior art is solved, and the safety and comfort of train operation are improved.
Patent Information
- Application Number
- CN202410883917.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-03
- Publication Date
- 2024-11-29
AI Technical Summary
The prior art is difficult to realize quantitative identification of wheel polygon wear, resulting in deterioration of wheel-rail relationship, increasing noise, and shortening of steel spring service life, affecting the operational comfort and safety of trains.
Using a method based on the generative adversarial network, the generative adversarial network is designed and trained by obtaining the axle box vibration acceleration signal, and the time domain vibration data is used to quantitatively identify the order and roughness of the wheel polygon. The input to the generation adversarial network is axle box vibration data. , outputs the wear roughness level of the wheel polygon.
It realizes accurate identification of wheel polygon wear, improves the accuracy of wheel status monitoring and fault diagnosis, and enhances the quality and safety of train operation.
Smart Images

Figure CN119046746A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of subway trains, and in particular to a method for quantitatively identifying wheel polygon wear based on a generative adversarial network. Background Art
[0002] Wheel out-of-round damage is common in subway vehicles, among which the typical form of damage is wheel polygon wear. Wheel polygon wear refers to the special phenomenon of periodic uneven wear of train wheels along the circumference. The overall wheel roughness level of domestic subways is relatively high, and some wheels have obvious high-order polygons. Wheel polygon wear will deteriorate the wheel-rail relationship, increase the force between the wheel and the rail, and cause increased noise. Under the excitation of wheel polygon wear, it will stimulate the resonance of the steel spring, thereby sharply shortening the service life of the steel spring. At the same time, high-order wheel polygon wear will cause vibration of vehicle components and track structures, affecting the comfort and safety of train operation. When the wheel has polygon wear, wheel turning is generally used to turn the wheel polygon away. Wheel turning is the most effective control measure for managing wheel polygons. If the state of wheel polygon wear can be mastered in real time, it will have a very important guiding significance for the reasonable arrangement of wheel turning. Intelligent monitoring of wheel polygon wear is the development trend of intelligent operation and maintenance of subway vehicles in the future. Research on quantitative identification methods of wheel polygon wear is of great significance to ensuring the quality and safety of vehicle operation.
[0003] Dynamic detection has the great advantage of high detection efficiency and is a commonly used method in vehicle status detection. Dynamic detection of wheel polygon wear can be divided into two methods: trackside method and vehicle-mounted method according to the different operating positions of the sensors. The trackside method refers to the installation of detection systems along the track and in the area around the track to collect and analyze various signals generated by the wheels passing by. The vehicle-mounted method refers to the installation of sensors on the train, which can realize continuous and comprehensive measurement of the wheels, and identify wheel faults by analyzing and processing the signals collected by the sensors.
[0004] The measurement accuracy of the trackside detection method is easily affected by factors such as the track structure, sensor layout location and layout method, and it is difficult to perform uninterrupted measurements. In addition, the trackside detection method cannot realize online monitoring of wheel faults. The wheel polygon detection based on the vehicle-mounted method is mainly based on real-time monitoring of the axle box vibration acceleration, and the wheel polygon is identified through fault diagnosis methods. At present, most of the research on wheel polygon faults is qualitative identification of wheel polygons, while the quantitative identification of wheel polygon wear amplitude is more instructive for intelligent train operation and maintenance. Summary of the invention
[0005] In order to solve the problems existing in the prior art, the purpose of the present invention is to provide a method for quantitatively identifying wheel polygon wear based on a generative adversarial network. The present invention can realize state monitoring and fault diagnosis of wheel polygons.
[0006] To achieve the above object, the technical solution adopted by the present invention is: a quantitative identification method of wheel polygon wear based on a generative adversarial network, comprising the following steps:
[0007] Step 1, obtaining polygonal wear data of the wheel, and judging whether the wheel has polygonal wear according to the wear data;
[0008] Step 2: Based on the subway vehicle-track rigid-flexible coupling dynamics model, simulate the influence of axle box vibration under different polygonal excitations to obtain the axle box vibration acceleration signal;
[0009] Step 3, preparing a vibration acceleration data sample set according to the axle box vibration acceleration signal;
[0010] Step 4: Design and train a generative adversarial network using the acceleration data sample set, and use the generative adversarial network to quantitatively identify the wheel polygon order and roughness level. The input of the generative adversarial network is time domain vibration data, and the output is the roughness level of multiple orders of the wheel polygon.
[0011] As a further improvement of the present invention, the step 1 is specifically as follows:
[0012] The obtained wheel polygon rectangular coordinate data is processed by using the out-of-roundness data processing method to obtain the wheel polygon order representation. The relationship between the wheel polygon order O and the wheel polygon wavelength λ is as follows:
[0013]
[0014] When the wheel polygonal wear exceeds a certain value of the roughness threshold curve, it is considered that the wheel has polygonal wear.
[0015] As a further improvement of the present invention, the step 3 is specifically as follows:
[0016] First, the axle box vibration acceleration signal is preprocessed, including abnormal data removal and data filtering, and then the data is cut and regularized.
[0017] As a further improvement of the present invention, during data filtering, according to the subway wheel polygon order range O and the subway vehicle running speed v, the axle box vibration signal frequency range f is determined by the following formula:
[0018]
[0019] Bandpass filtering is used to filter the axle box vibration acceleration signal.
[0020] As a further improvement of the present invention, the data is cut and regularized as follows:
[0021] Determine the vibration signal cutting window length l and set the sample set data length n: By calculating the displacement length of the wheel turning three wheel circumferences, that is, the cutting window length l, the vibration signal is adaptively cut according to the time it takes for the wheel to turn three circles at different speeds, and the vibration data length n of the wheel turning three wheel circumferences is used as the standard sample set data length, where:
[0022]
[0023] Where r is the wheel radius; v is the vehicle speed; Fs is the vibration acceleration sampling frequency. It is a floor rounding operation;
[0024] Sample data length cutting and regularization: The sample data cutting method adopts a cutting window length of l and a step length of l / 3. The resampling method is used to regularize the data lengths at different speeds, and the sample data of different lengths are regularized to data length n.
[0025] As a further improvement of the present invention, in step 4, the generative adversarial network includes a generative network and a discriminant network. The generative network is used to find the mapping relationship between the vertical vibration acceleration of the axle box and the polygonal wear of the wheel. The vibration acceleration data sample set obtained in step 3 is input into the generative network. After nonlinear mapping of the generative model, the wheel polygonal wear roughness level is output. The discriminant network guides the parameter update strategy of the generative network and the discriminant network based on the loss between the generated polygonal wear data and the actual polygonal wear data. The generated polygonal wear data and the actual polygonal wear data are input into the discriminant network. The discriminant model returns the adversarial loss and the discriminant loss, and the parameters of the generative network and the discriminant network are updated in reverse.
[0026] As a further improvement of the present invention, the generation network is composed of a feature extraction layer and a fully connected layer. The feature extraction layer of the generation network is formed by stacking a one-dimensional convolution layer, a batch normalization layer and a ReLU activation layer structure. The one-dimensional convolution layer is used for axle box acceleration data feature extraction. The batch normalization layer is used to make the data distribution consistent, reduce internal covariate transfer and speed up network training. The ReLU activation layer is a typical activation function used to increase the nonlinear fitting ability of the model. The fully connected layer of the generation network is composed of a linear layer and random dropout. The linear layer is used to establish a mapping relationship between the features obtained by the feature extraction layer and the wheel polygon wear, and the random dropout is used to suppress network overfitting.
[0027] As a further improvement of the present invention, in the discriminant network, the discriminant network is stacked by a one-dimensional convolutional layer, a spectral normalization layer, a random dropout layer, a flattening layer, and a linear layer, and the spectral normalization layer is used to ensure the Lipschitz continuity of the network.
[0028] As a further improvement of the present invention, the objective function of the generative adversarial network is as follows:
[0029]
[0030] Where x is the input axle box vibration signal, p x is the data distribution of the input vibration signal, y is the actual wheel order data, and p y is the actual wheel order data distribution, ||·||2 is the Euclidean norm, and λ is the weight parameter for balancing the adversarial loss and the root mean square loss.
[0031] As a further improvement of the present invention, R 2 Evaluate the similarity between the actual values and the generated predicted values, R 2 is defined as follows:
[0032]
[0033] Where n is the length of the output data of the input-generated adversarial network, y i is the ith actual value in the actual data, is the i-th predicted value in the predicted data, is the average value of the actual data.
[0034] The beneficial effects of the present invention are:
[0035] The invention is based on the quantitative identification of wheel polygon wear of the vertical acceleration of the train axle box. Firstly, complete axle box acceleration data is obtained, and a polygon wear quantitative identification model of Wasserstein generative adversarial networks (WGAN) based on Wasserstein distance is established. The axle box acceleration time domain data is selected as the input feature of the model, and the output is the wear roughness level corresponding to the wheel polygon order. The wheel fault condition is judged by the wear roughness level corresponding to the wheel polygon order obtained by the output, thereby realizing the state monitoring and fault diagnosis of the wheel polygon. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 This is a flowchart of quantitative identification of polygonal wear of subway wheels based on WGAN in an embodiment of the present invention;
[0037] Figure 2 This is a schematic diagram of the out-of-roundness result of a wheel measured in an embodiment of the present invention;
[0038] Figure 3 This is a subway wheel threshold curve diagram in an embodiment of the present invention;
[0039] Figure 4Schematic diagram of a subway vehicle-track rigid-flexible coupling dynamics model in an embodiment of the present invention;
[0040] Figure 5 This is a WGAN process framework diagram in an embodiment of the present invention;
[0041] Figure 6 An example diagram of a 12-order wheel polygon constructed in an embodiment of the present invention;
[0042] Figure 7 12th-order and 13th-order wheel polygon example diagrams constructed in the embodiments of the present invention;
[0043] Figure 8 This is a diagram of model training and testing results in an embodiment of the present invention;
[0044] Fig. 9 This is a schematic diagram of quantitative identification results of simulation data of 12th-order dominant working conditions in an embodiment of the present invention;
[0045] Fig.10 This is a schematic diagram of quantitative identification results of 12th-order and 13th-order dominant operating condition data in an embodiment of the present invention;
[0046] Fig.11 It is a schematic diagram of the measured vertical vibration acceleration of the axle box in an embodiment of the present invention;
[0047] Fig.12 Schematic diagram of quantitative recognition results of polygonal wheel measured data in an embodiment of the present invention;
[0048] Fig.13 This is a schematic diagram of quantitative identification results of measured data of a normal wheel in an embodiment of the present invention;
[0049] Fig.14 Schematic diagram of multiple test results in an embodiment of the present invention. DETAILED DESCRIPTION
[0050] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.
[0051] Example
[0052] This embodiment proposes a quantitative identification method for polygonal wear of subway wheels based on a generative adversarial network (WGAN). The method identifies the order and amplitude of polygonal wear of wheels based on the vertical vibration acceleration signal of the train axle box. Figure 1 The quantitative identification process of polygonal wear of subway wheels based on WGAN can be divided into the following five parts: ① Acquisition of wheel polygonal wear data; ② Acquisition of axle box vibration acceleration data; ③ Preparation of vibration data sample set; ④ WGAN structure design and training; ⑤ Quantitative identification of wheel polygonal wear roughness level; the details are as follows:
[0053] 1. Judgment basis of wheel polygonal wear:
[0054] The wheel circumferential out-of-roundness data obtained through mathematical model construction and on-site measurement are often rectangular coordinate data. The rectangular coordinates of the out-of-roundness data are as follows: Figure 2 However, it is often difficult to directly see the wear data corresponding to each order of the wheel polygon from the rectangular coordinate data. Therefore, it is necessary to process the acquired wheel polygon rectangular coordinate data. In this embodiment, the data is processed by the out-of-roundness data processing method to obtain the wheel polygon order representation, as shown in FIG. Figure 3 The relationship between the wheel polygon order O and the wheel polygon wavelength λ can be determined by formula (1).
[0055]
[0056] The ISO-3095 standard contains a rail corrugation roughness limit curve, but the wavelength of the rail corrugation roughness threshold curve only covers 30-400mm. For a subway wheel with a diameter of 840mm, the 30-400mm wavelength only covers the 7-87 order wheel polygon. In EN 15313, for trains running at a speed of less than or equal to 120km / h, it is stipulated that the wheel diameter runout value cannot exceed 0.5mm. The 0.5mm diameter runout value is converted into a roughness level. The threshold of the 1st-order wheel polygon wear roughness value is set to 50.97dB. At the same time, referring to the ISO-3095 standard, the 7th-order wheel polygon (corresponding to a wavelength of 377mm) wear roughness threshold is set to 16.58dB. Increase the subway wheel polygon linear roughness threshold of the wheel polygon wavelength from 377mm to 2639mm. In order to determine whether the subway wheel contains polygonal wear, the subway wheel polygon roughness threshold curve is proposed with reference to the high-speed train wheel polygon roughness threshold proposed in ISO-3095. When the wheel polygon wear exceeds the roughness threshold curve by 6dB, it is considered that the wheel has polygonal wear. Figure 3 The broken line with large fluctuations in the middle is the out-of-roundness order data of a certain measured wheel, which obviously exceeds the roughness threshold curve by 6dB. At this time, the wheel has obvious polygonal wear.
[0057] 2. Acquisition of axle box vibration acceleration signal based on dynamic simulation:
[0058] To study the response of the axle box vibration acceleration under the simulated working conditions, it is necessary to establish a complete vehicle and track model for simulation calculation. The dynamic model that considers the flexibility of the wheelset and the flexibility of the rail can more accurately reflect the vibration response of the axle box than the rigid model. This embodiment uses the subway vehicle-track rigid-flexible coupling dynamic model to simulate the influence of the axle box vibration under different polygonal excitations, such as Figure 4 shown.
[0059] 3. Preparation of vibration acceleration data sample set:
[0060] First, the axle box vibration acceleration signal is preprocessed, including abnormal data removal and data filtering. In terms of data filtering, according to the subway wheel polygon order range O and the subway vehicle running speed v, the axle box vibration signal frequency range f can be determined by the following formula:
[0061]
[0062] The maximum actual operating speed of a subway vehicle is generally 120 km / h. In this embodiment, the wheel polygon order range O only considers the first 40 orders. The frequency range of the axle box vibration signal can be determined to be 1-505 Hz by formula (2). Therefore, the vibration signal filter frequency is set to 1-500 Hz, and a bandpass filter is used to filter the axle box vibration acceleration signal.
[0063] Considering the frequent start and stop of subway vehicles, the influence of speed on the mapping relationship between the axle box vibration response and the wheel out-of-roundness should be considered. However, the time taken for the wheel to roll over a wheel circumference displacement at different speeds will lead to inconsistent data lengths, making it difficult to train the neural network. Therefore, the data needs to be cut and regularized. The specific steps are as follows:
[0064] (1) Determine the vibration signal cutting window length l and set the sample set data length n. Calculate the displacement length l of the wheel rotating three wheel circumferences. The calculation method of l is shown in formula (3). According to the time it takes for the wheel to rotate three circles at different speeds, the vibration signal is adaptively cut. When the wheel rotates the same length of displacement, due to the same data sampling frequency, data cutting at different operating speeds will result in different vibration signal cutting window lengths, which will make the sample set data lengths different, making it difficult to train the neural network model. In order to meet the consistency requirements of the data input length of the WGAN network model, the vibration data length n of the wheel rotating three wheel circumferences is used as the standard sample set data length.
[0065]
[0066] Where r is the wheel radius in meters; v is the vehicle speed in kilometers per hour; Fs is the vibration acceleration sampling frequency in Hz. This is a floor operation.
[0067] (2) Sample data length cutting and regularization. The sample data cutting method uses a cutting window length of l, a step length of l / 3, and an overlap rate of 66% between two adjacent data samples. Since the displacement data lengths of the wheel after three turns are inconsistent at different speeds, this embodiment uses a resampling method to regularize the data lengths at different speeds, and regularizes the sample data of different lengths to data length n.
[0068] 4. WGAN model and structure design:
[0069] Basic theory of WGAN: Generative adversarial networks (GAN) are a popular neural network model. GAN mainly consists of two models, namely the generative model and the discriminative model. The generative model is responsible for generating data similar to the real data, and the discriminative model is responsible for judging whether the given data is real data or data generated by the generative model. The core of GAN is the confrontation and game between the generative model and the discriminative model. Its objective function is as follows:
[0070]
[0071] Where x is the actual data distribution P r Samples of, z is the noise data distribution P z , D(`) is the output of the discriminant model, G(·) is the output of the generative model, and E(·) is the mathematical expectation operation.
[0072] Traditional GANs have problems with gradient vanishing and model collapse during training, which makes it difficult for training to converge. Therefore, a generative adversarial network (WGAN) based on Wasserstein distance is proposed, which uses Wasserstein distance instead of JS divergence to measure the distribution distance between real samples and generated samples. The mathematical expression of Wasserstein distance between two distributions is as follows:
[0073]
[0074] Where f(x) is the conversion function, P r is the actual data distribution, P g To generate data distribution, L represents Lipschitz continuity.
[0075] The framework of the quantitative identification method of wheel polygon wear roughness level proposed in this embodiment is as follows: Figure 5 As shown, it consists of a generative model and a discriminative model, where the solid line is the forward propagation path of the model and the dotted line is the reverse propagation path of the model.
[0076] The purpose of the generative network is to find the mapping relationship between the vertical vibration acceleration of the axle box and the polygonal wear of the wheel. The discriminant network guides the parameter update strategy of the generative network based on the loss between the generated polygonal wear data and the actual polygonal wear data. The mean square error (MSE) between the actual wheel order data and the generated order data is used as the loss function to update the parameters in the generative network. The mean square error is used to evaluate the expected value of the square of the difference between the predicted value and the actual value of the data. The smaller the mean square error value, the better the accuracy of the data. Its mathematical model is as follows:
[0077]
[0078] In the formula, n represents the data length of a sample, y i is the actual value of the ith value in the sample, is the predicted value of the sample at the i-th time.
[0079] The mathematical expression of the objective function of the wheel polygon wear quantitative identification method based on WGAN in this embodiment is:
[0080]
[0081] Where x is the input axle box vibration signal, p x is the data distribution of the input vibration signal, y is the actual wheel order data, and p y is the actual wheel order data distribution, ||·||2 is the Euclidean norm, and λ is the weight parameter for balancing the adversarial loss and the root mean square loss, which is set to 0.01 in this embodiment.
[0082] WGAN model structure design:
[0083] The network model of this embodiment is a wheel polygon wear regression prediction model. Different from the wheel polygon classification or identification model, the wheel polygon wear roughness level regression prediction model can realize the quantitative identification of the wheel polygon order and roughness level. The input of the generation network is the time domain vibration data. Considering the influence of the vehicle running speed on the axle box vibration signal, the number of input channels is 2, and the output is the roughness level of the first 40 orders of the wheel polygon. The generation network consists of a feature extraction layer and a fully connected layer. The feature extraction layer of the generation network is stacked with a one-dimensional convolution layer (Conv1d), a batch normalization layer (BatchNormalization, BN) and a ReLU activation layer structure. The one-dimensional convolution layer is used for feature extraction, the batch normalization layer is to make the data distribution consistent, reduce the internal covariate transfer and speed up the network training speed, and the ReLU activation layer is a typical activation function that can increase the nonlinearity of the network. Spectral Normalization (SN) is used in the discriminant network to ensure the 1-Lipschitz continuity of the network. The structures of the generation network and the discriminant network are shown in Table 1 and Table 2 respectively.
[0084] Table 1 Generate network specific structure and parameters
[0085]
[0086] Table 2 Determine the specific structure and parameters of the dual network
[0087]
[0088] During the training process, the WGAN polygon wear roughness level quantitative identification model proposed in this embodiment uses R 2 Evaluates the similarity between actual values and the predicted values generated by the model. 2 The closer it is to 1, the more similar the actual value is to the predicted value, that is, the more accurate the wheel polygon wear roughness predicted by the model is. 2 is defined as follows:
[0089]
[0090] Where n is the length of the output data of the input WGAN model, y i is the ith actual value in the actual data, is the i-th predicted value in the predicted data, is the average value of the actual data.
[0091] Quantitative identification of wheel polygon wear roughness level based on simulation signals: This embodiment uses a numerical simulation method to obtain a vibration acceleration sample set and a wheel polygon wear sample set, and verifies and analyzes the effectiveness of the WGAN-based wheel polygon wear quantitative identification method proposed in this embodiment.
[0092] Wheel polygon construction and dataset construction:
[0093] Wheel polygon, also known as wheel periodic out-of-roundness, refers to wheel wear in which the change in wheel radius shows periodic fluctuations. It is generally believed that polygons with an order of less than 9 are low-order polygons, and polygons with an order of greater than 9 are high-order polygons.
[0094] In the actual field of subway vehicles, wheel polygons are mostly 5-8 order polygons (low-order polygons) and 12-20 order polygons (high-order polygons). Therefore, low-order wheel polygons and high-order wheel polygons are mainly considered in the simulation analysis and verification. In order to simplify the wheel polygon mathematical model, this embodiment only considers the simulation wheel polygon composed of the superposition of the sine functions represented by the first 40 orders. The wheel out-of-roundness mathematical model constructor is as follows:
[0095]
[0096] Where x is the circumferential position of the wheel, i is the polygonal order of the wheel, is the phase of the i-th order polygon, A i is the amplitude of the i-th order polygon, and r is the wheel radius.
[0097] The wheel polygon working conditions with only a single dominant order are set as low-order wheel polygons (5-8 orders) and high-order wheel polygons (12-17 orders), with an amplitude of 0.02mm-0.1mm and an increment of 0.02mm, for a total of 50 working conditions. The wheel polygon working conditions formed by the superposition of two dominant orders are set as two low-order wheel polygon combinations (5th and 6th orders), low-order and high-order wheel polygon combinations (8th and 14th orders), and two high-order wheel polygon combinations (12th and 13th orders), for a total of 48 working conditions. The wheel polygon out-of-round setting working conditions are shown in Table 3. Figure 6 For a wheel polygon dominated only by order 12, the wear amplitude of the 12th-order wheel polygon is 0.06 mm. Figure 7 There are two dominant wheel polygons of order 12 and 13, and the wear amplitudes of the wheel polygons of order 12 and 13 are both 0.03 mm.
[0098] Table 3 Simulation conditions
[0099]
[0100] The constructed wheel polygon data is input into the rigid-flexible coupling dynamics model to obtain the simulated axle box vibration acceleration data. However, the simulated acceleration data needs to be processed to obtain the axle box vibration acceleration sample set that can be directly input into the WGAN network.
[0101] The method for constructing the simulation data sample set is as follows: first, the vertical vibration acceleration signal of the axle box is preprocessed and processed using a bandpass filter, with the cutoff frequencies of the bandpass filter set to 1 Hz and 500 Hz respectively. The data sample set is obtained according to the sample set production method proposed in the article. The simulation sample set has a data volume of 10,976, of which 80% is used as a training set and 20% is used as a test set.
[0102] WGAN model training and result analysis:
[0103] The generator and discriminator of the WGAN model are trained using the “RMSprop” optimizer, and the learning rates of the generator and discriminator are set to 0.001 and 10 respectively. -5 . In order to avoid gradient diffusion and gradient explosion, batch sample technology is used for training, and the batch size is set to 64. In order to accelerate model training and improve the generalization performance of the model, random dropout technology is used, and the random dropout of this WGAN model is set to 0.15. The WGAN model is built on the pytorch framework based on Python 3.8. The computer hardware configuration is i7-10700 K processor 16GB memory Windows 11 system.
[0104] In order to speed up the convergence of the model, the method of preheating the model is adopted, that is, the discriminant model is trained 25 times before formal training, and then the WGAN model is trained. The WGAN model training method adopts the generative model training 3 times and the discriminant model training 1 time alternately. Figure 8 The figure is the quantitative prediction result of wheel polygon wear by WGAN on the simulation data sample set. From the output results, we can see that when the WGAN model is trained 500 times, the model training is sufficient and R 2 Stable at around 0.95, the model recognition effect is good. In addition, the model is trained and tested with fewer training rounds. 2 All of them can reach a high level, indicating that the model converges quickly. Figure 8 It can be seen that in the first 100 rounds of training, the test set R 2 Both are higher than the training set R 2 This is because the discriminant model was trained 25 times before the formal training of the model and the discriminant model is smaller than the generative model, which causes the discriminant model to converge faster than the generative model.
[0105] Some prediction orders are as follows Fig. 9 and Fig.10 As shown, from Fig. 9 and Fig.10 It can be seen that the WGAN model can accurately identify the dominant order of wheel polygons and their roughness values, and the average error of the roughness value of the dominant order of polygons is 0.936dB. There are certain errors in the roughness identification of other orders, which is mainly because the amplitude of wear of other polygon orders is set to 0mm. Since the corresponding non-dominant order roughness values obtained by the wheel out-of-roundness processing program are not the same, it is acceptable to have certain errors in the roughness values of other orders when verifying the generalization performance of the model.
[0106] Field test data verification:
[0107] In order to further verify the effectiveness of the method for quantitatively identifying the wear and roughness level of wheel polygons proposed in this embodiment in practical engineering applications, the proposed method was verified using actual measured data collected on a domestic subway.
[0108] Field test: A vibration acceleration test was conducted on a domestic subway vehicle to obtain the axle box vibration acceleration. The axle box vibration test uses a professional measuring device HBM multi-channel data acquisition system to collect axle box vibration acceleration data containing polygonal wear characteristics. At the same time, the circumferential roughness of the corresponding wheels of the train undergoing the vibration acceleration test is measured using a wheel roughness measuring instrument. The wheel circumferential roughness is measured.
[0109] The vibration test uses a sampling frequency of 5000 Hz to collect the vertical vibration acceleration of the axle box when the train is running at speeds of 64 km / h, 75 km / h, and 80 km / h. Fig.11 This is the measured vertical vibration signal of the axle box of a subway wheel at a speed of 75km / h. During the data preprocessing process, the influence of rail welding joints, rail corrugation, etc. on the axle box vibration acceleration signal is removed. The acquired data samples are regularized using data processing methods. A total of 3031 data samples are obtained, and all the acquired data samples are randomly divided into 2420 data samples as training sets and 611 data samples as test sets.
[0110] Result analysis: The obtained measured data sample set is input into the WGAN model for training. Fig.12 and Fig.13The recognition effect of normal wheel and polygonal wheel data is demonstrated. When recognizing normal wheels, the wheel roughness values of each order obtained are all below the roughness threshold line, and the recognition result is a normal wheel, which is consistent with the actual result. When the model recognizes polygonal wheels, the wheel roughness value exceeds the wheel roughness threshold line by 6dB, and the recognition result is a polygonal wheel. When the wheel roughness value exceeds the roughness threshold line by 6dB, it is considered to be the dominant order of the wheel polygon. The average error level of the roughness value recognition of the dominant order of the wheel polygon in the measured data results is 1.36dB. Overall, the recognition method proposed in this embodiment has good recognition effects on the wheel polygon order and amplitude.
[0111] In order to evaluate the reliability and applicability of the model in identifying wheel polygons, this embodiment uses multiple tests to observe the test results, thereby avoiding the influence of the randomness of a single test result. By conducting 10 tests on the measured sample set, the average error of the dominant order of the 10 tests and the test time of a single sample are obtained. The results are as follows: Fig.14 The average error level of the roughness of the dominant order of the wheel polygon in 10 tests is 1.52dB, and the average time consumed for single sample recognition is about 0.25ms. This shows that the model has good robustness and high recognition efficiency.
[0112] This example studies the quantitative identification method of wheel polygon wear roughness level based on axle box vibration acceleration for subway wheel polygon wear. The main conclusions are as follows:
[0113] (1) A quantitative identification method for wheel polygon wear roughness level based on WGAN is proposed. The time domain data of the vertical vibration acceleration of the axle box is regularized, and the time domain vibration acceleration is used as input and the wheel polygon roughness value is used as output. At the same time, three roughness threshold curve levels of subway wheels are set. WGAN is used to quantitatively identify the wheel polygon wear order and roughness level, and the wheel polygon wear status is judged by the roughness threshold curve.
[0114] (2) Based on the rigid-flexible coupling dynamics model, a simulated axle box vibration data sample set is obtained and trained with the WGAN model. The obtained model has high recognition accuracy. 2 The generalization of the model was verified, and it was found that the recognition effect of the main order and its roughness value of the model was relatively good, and the average error of the main order roughness value obtained was 0.936dB.
[0115] (3) The axle box vibration data sample set obtained by field measurement was regularized and input into the WGAN model. The average error level of the roughness of the dominant order of the wheel polygon in 10 tests was 1.524 dB, and the average time required for single sample identification was about 0.25 ms.
[0116] The above-mentioned embodiments only express the specific implementation of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.
Claims
1. A quantitative identification method for wheel polygon wear based on generative adversarial network, characterized in that: The following steps are involved: Step 1, obtaining wheel polygon wear data, and judging whether the wheel has polygon wear according to the wear data; Step 2: Based on the subway vehicle-track rigid-flexible coupling dynamics model, simulate the influence of axle box vibration under different polygonal excitations to obtain the axle box vibration acceleration signal; Step 3, preparing a vibration acceleration data sample set according to the axle box vibration acceleration signal; Step 4: Design and train a generative adversarial network using the acceleration data sample set, and use the generative adversarial network to quantitatively identify the wheel polygon order and roughness level. The input of the generative adversarial network is time domain vibration data, and the output is the roughness level of multiple orders of the wheel polygon.
2. The method for quantitative identification of wheel polygon wear based on generative adversarial network according to claim 1 is characterized in that: The step 1 is specifically as follows: The obtained wheel polygon rectangular coordinate data is processed by using the out-of-roundness data processing method to obtain the wheel polygon order representation. The relationship between the wheel polygon order O and the wheel polygon wavelength λ is as follows: When the wheel polygonal wear exceeds a certain value of the roughness threshold curve, it is considered that the wheel has polygonal wear.
3. The wheel polygon wear quantitative identification method based on generative adversarial network according to claim 1 is characterized in that: The step 3 is as follows: First, the axle box vibration acceleration signal is preprocessed, including abnormal data removal and data filtering, and then the data is cut and regularized.
4. The method for quantitatively identifying wheel polygon wear based on a generative adversarial network according to claim 3 is characterized in that: During data filtering, according to the subway wheel polygon order range O and the subway vehicle running speed v, the axle box vibration signal frequency range f is determined by the following formula: Bandpass filtering is used to filter the axle box vibration acceleration signal.
5. The method for quantitative identification of wheel polygon wear based on generative adversarial network according to claim 4 is characterized in that: The data is cut and organized as follows: Determine the vibration signal cutting window length l and set the sample set data length n: By calculating the displacement length of the wheel turning three wheel circumferences, that is, the cutting window length l, the vibration signal is adaptively cut according to the time it takes for the wheel to turn three circles at different speeds, and the vibration data length n of the wheel turning three wheel circumferences is used as the standard sample set data length, where: Where r is the wheel radius; v is the vehicle speed; Fs is the vibration acceleration sampling frequency. This is the floor rounding operation; Sample data length cutting and regularization: The sample data cutting method adopts a cutting window length of l and a step length of l / 3. The resampling method is used to regularize the data lengths at different speeds, and the sample data of different lengths are regularized to data length n.
6. The method for quantitative identification of wheel polygon wear based on generative adversarial network according to claim 5 is characterized in that: In step 4, the generative adversarial network includes a generative network and a discriminant network. The generative network is used to find the mapping relationship between the vertical vibration acceleration of the axle box and the polygonal wear of the wheel. The vibration acceleration data sample set obtained in step 3 is input into the generative network. After the nonlinear mapping of the generative model, the wheel polygonal wear roughness level is output. The discriminant network guides the parameter update strategy of the generative network and the discriminant network based on the loss between the generated polygonal wear data and the actual polygonal wear data. The generated polygonal wear data and the actual polygonal wear data are input into the discriminant network. The discriminant model returns the adversarial loss and the discriminant loss, and the parameters of the generative network and the discriminant network are reversely updated.
7. The method for quantitatively identifying wheel polygon wear based on a generative adversarial network according to claim 6 is characterized in that: The generation network consists of a feature extraction layer and a fully connected layer. The feature extraction layer of the generation network is formed by stacking a one-dimensional convolution layer, a batch normalization layer and a ReLU activation layer structure. The one-dimensional convolution layer is used for axle box acceleration data feature extraction. The batch normalization layer is used to make the data distribution consistent, reduce internal covariate transfer and speed up network training. The ReLU activation layer is a typical activation function used to increase the nonlinear fitting ability of the model. The fully connected layer of the generation network consists of a linear layer and random dropout. The linear layer is used to establish a mapping relationship between the features obtained by the feature extraction layer and the wheel polygon wear, and the random dropout is used to suppress network overfitting.
8. The method for quantitatively identifying wheel polygon wear based on a generative adversarial network according to claim 6, characterized in that: In the discriminant network, the discriminant network is stacked by a one-dimensional convolution layer, a spectral normalization layer, a random dropout layer, a flattening layer, and a linear layer, and the spectral normalization layer is used to ensure the Lipschitz continuity of the network.
9. The method for quantitative identification of wheel polygon wear based on generative adversarial network according to claim 6, characterized in that: The objective function of the generative adversarial network is as follows: Where x is the input axle box vibration signal, p x is the data distribution of the input vibration signal, y is the actual wheel order data, and p y is the actual wheel order data distribution, ||`||2 is the Euclidean norm, and λ is the weight parameter for balancing the adversarial loss and the root mean square loss.
10. The method for quantitatively identifying wheel polygon wear based on a generative adversarial network according to claim 6, characterized in that: Using R 2 Evaluate the similarity between the actual values and the generated predicted values, R 2 is defined as follows: Where n is the length of the output data of the input-generated adversarial network, y i is the ith actual value in the actual data, is the i-th predicted value in the predicted data, is the average value of the actual data.
Citation Information
Cited By
Vehicle-mounted wheel polygon abrasion identification method based on multi-wheel cooperation
CN122046215A
A method for identifying polygonal wear on vehicle wheels based on multi-wheel collaboration
CN122046215B
Wheel non-roundness roughness level intelligent quantitative identification method based on axle box vibration
CN122388448A
Intelligent quantitative identification method for wheel non-circular roughness level based on axle box vibration
CN122388448B
