Method for identifying wheel roughness of high-speed train

Abstract

The invention provides a method for identifying the wheel roughness of a high-speed train. The method comprises the following steps of: 1, acquiring power spectrums of axle box vibration accelerations of a plurality of wheels from a train as sample data; calculating an actual roughness curve of each wheel, and taking the actual roughness curve as a sample label; using the sample data and the sample labels to construct a wheel roughness identification set; step 2, based on the wheel roughness identification set, establishing a wheel roughness model by using a neural network structure; and step 3, predicting a roughness curve of the train by using the wheel roughness model.

Description

Technical Field

[0001] The invention relates to the technical field of railway safety monitoring, and in particular to a method for identifying the roughness of wheels of high-speed trains. Background Art

[0002] The interaction between wheels and tracks is characterized by complex transients, which makes the axle box vibration acceleration signal nonlinear and non-stationary. Therefore, a lot of research is devoted to breaking through this obstacle to realize the identification of wheel polygon characteristic parameters.

[0003] The existing identification of wheel polygon feature parameters mainly processes acceleration signals from three angles: time domain analysis, frequency domain analysis, and time-frequency domain, and establishes a model. Statistical indicators in the time domain can comprehensively present the basic characteristics and probability distribution of the signal, but when there is a fault in the mechanical equipment, the statistical indicators will show values ​​and probability distributions different from those of normal machinery, which will reduce the recognition accuracy and precision of the model; transforming the signal from the time domain to the frequency domain can well identify the wheel polygon, but the disadvantage of the constant time window of STFT (short-time Fourier transform) cannot be overcome, so that the collected signal cannot meet the time-frequency resolution at the same time; HHT (Hilbert-Huang transform) has good time-frequency resolution and adaptability, but HHT is often interfered by modal aliasing, which makes the frequency recognition inaccurate. Therefore, a method is needed to directly mine the characteristic information of the wheel polygon from the acceleration signal to achieve accurate recognition of the wheel polygon features. Summary of the invention

[0004] The purpose of the present invention is to provide a method for identifying the roughness of a high-speed train wheel, so as to shorten the calculation time of the wheel roughness and realize the accurate identification of the polygonal features of the wheel.

[0005] In order to achieve the above object, the present invention provides a method for identifying the roughness of a high-speed train wheel, comprising:

[0006] Step 1: obtaining the power spectra of the axle box vibration acceleration of several wheels on the train as sample data; calculating the actual roughness curve of each wheel, and using the actual roughness curve as a sample label; constructing a wheel roughness recognition set using the sample data and the sample label;

[0007] Step 2, based on the wheel roughness identification set, using a neural network structure to establish a wheel roughness model;

[0008] Step 3: using the wheel roughness model to predict the roughness curve of the train.

[0009] Optionally, in step 1, the method for obtaining the power spectrum of the axle box vibration acceleration includes:

[0010] Each wheel is sampled intermittently, and the axle box vibration acceleration signals of k groups of trains are collected on each wheel. The axle box vibration acceleration signals are Fourier transformed to obtain the power spectrum of the axle box vibration acceleration; wherein k≥2.

[0011] Optionally, a sensor provided on the train is used to intermittently collect axle box vibration acceleration signals for each wheel k times, the collection frequency of the sensor is 10000 Hz, and the duration of each sampling is 0.68 s.

[0012] Optionally, in step 1, the method for calculating the actual roughness curve of the wheel is as follows:

[0013]

[0014] Where, L θ is the roughness of the wheel polygon of order θ, θ is the order of the wheel polygon; is the radial deviation power of the wheel profile f θ The root mean square of the 1 / 3 octave band; r ref is the reference value of wheel roughness, taking r ref =1μm; v is the train speed, r is the theoretical radius of the wheel.

[0015] Optionally, in step 2, the wheel roughness identification set is divided into a training set and a validation set; the sample data in the training set includes a roughness identification vector composed of sample data collected on each wheel, and is extracted from the wheel roughness identification set at a ratio of 50%; the validation set is the roughness identification vector remaining after the training set is extracted from the wheel roughness identification set.

[0016] Optionally, the power spectrum of the axle box vibration acceleration collected from the same wheel in the training set is averaged, and the average is used as the mean sample data of the corresponding wheel in the training set, so that the mean sample data and the sample label of the corresponding wheel constitute the roughness recognition vector of the wheel in the training set, so that the number of roughness recognition vectors in the training set is the same as the number of collected wheels.

[0017] Optionally, in step 2, a wheel roughness model is constructed using a generalized regression neural network structure, and the generalized regression neural network structure includes:

[0018] The input layer is the input end of the wheel roughness model and can import input data into the neural network structure; the sample data vector in the validation set is used as input data; wherein the sample data vector in the validation set is composed of all the sample data in the validation set;

[0019] The pattern layer uses the training set to construct nodes. Each node corresponds to a set of roughness recognition vectors in the training set. The distance of the Gaussian function between the sample data in the validation set input by the input layer and the sample data of a set of roughness recognition vectors in the training set can be calculated to obtain the calculation results of the pattern layer, which can be expressed as:

[0020]

[0021] In the formula, p i is the distance of the Gaussian function of the i-th node in the pattern layer, X m is the sample data vector in the validation set, X n is the sample data vector in the training set, σ is the hyperparameter;

[0022] The summation layer is used to normalize the calculation results of the pattern layer and calculate the prediction results, which can be expressed as:

[0023]

[0024] In the formula, s D is the arithmetic sum of the Gaussian function distances of all nodes in the pattern layer, s N is the weighted sum of the Gaussian function distances of all nodes in the pattern layer, y i is the sample label in the i-th roughness identification vector in the training set; y′ i is the prediction result of the i-th element in the output vector obtained based on the input data;

[0025] The output layer can output the prediction results in the summation layer.

[0026] Optionally, in step 2, the Euclidean distance between the prediction result output by the neural network structure and the sample label in the verification set is taken as the error of the wheel roughness model, and the error of the wheel roughness model can be changed by adjusting the hyperparameters; when the error of the wheel roughness model is minimized, the hyperparameters are the optimal parameters of the wheel roughness model, and the wheel roughness model constructed using the optimal parameters can be used to predict the roughness curve of the wheel.

[0027] Optionally, in step 3, k data are collected for the same wheel to obtain k groups of power spectra of axle box vibration acceleration, and the power spectra of k axle box vibration acceleration are input into the wheel roughness model to obtain k groups of prediction results; the k groups of prediction results are averaged, and the mean of the prediction results is used as the predicted roughness curve of the wheel.

[0028] Compared with the prior art, the technical solution of the present invention has at least the following beneficial effects:

[0029] The present invention obtains the power spectrum of the axle box vibration acceleration signal from the axle box vibration acceleration signal, uses the power spectrum of the axle box vibration acceleration signal as sample data, uses the wheel roughness curve as a sample label reflecting the wheel polygon feature information, and establishes the relationship between the sample data and the sample label through machine learning to obtain the wheel roughness model, thereby realizing accurate recognition of the wheel polygon feature; the present invention simplifies the method for obtaining the wheel roughness curve, and there is no need to collect a variety of train operation parameters to calculate the wheel roughness of different orders separately, and there is no need to measure various parameters of the high-speed train and model the vehicle in the process of establishing the wheel roughness model, so that the identification method is simple and easy to collect and process data.

[0030] The present invention establishes a wheel roughness model by adopting a GRNN neural network structure, and does not need to iterate the model through a large amount of data, thereby reducing the data processing amount of the wheel roughness model and shortening the processing time of the wheel roughness model. At the same time, the present invention averages the power spectrum of the axle box vibration acceleration signal of the same wheel in the training set, so that the number of roughness identification vectors in the training set is the same as the number of sampled wheels, and by reducing the number of roughness identification vectors in the training set and thus reducing the number of nodes in the neural network structure, the calculation time of the wheel roughness model is further shortened.

[0031] The present invention intermittently collects multiple axle box vibration acceleration power spectra on a wheel and inputs the wheel roughness model to obtain multiple wheel roughness curves, calculates the mean of the multiple wheel roughness curves, and uses the mean of the wheel roughness curve as the final wheel roughness curve of the wheel to eliminate the influence of vibration impact during vehicle operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 The figure is a flow chart of the method for identifying the roughness of high-speed train wheels according to the present invention.

[0033] Figure 2 This is the collection of axle box vibration acceleration signals in the high-speed train wheel roughness identification method of the present invention.

[0034] Figure 3 It is the axle box vibration acceleration power spectrum obtained in the high-speed train wheel roughness identification method of the present invention.

[0035] Figure 4 It is the wheel roughness curve obtained in the method for identifying the wheel roughness of a high-speed train of the present invention.

[0036] Figure 5 This is a diagram of the prediction results of the wheel roughness model output in the high-speed train wheel roughness identification method of the present invention.

[0037] Figure 6The present invention is a schematic diagram of processing for obtaining a wheel roughness curve by using a wheel roughness model in the method for identifying wheel roughness of a high-speed train. DETAILED DESCRIPTION

[0038] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] In the description of the present invention, it should be noted that the terms "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0040] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0041] The present invention provides a method for identifying the roughness of high-speed train wheels. The power spectrum of axle box vibration acceleration is used as sample data, a vehicle wheel roughness model is established using a neural network structure, and the relationship between the axle box vibration acceleration signal and the wheel roughness curve is calculated, which is convenient for practical engineering applications.

[0042] like Figure 1 As shown, the method for identifying the roughness of high-speed train wheels of the present invention comprises the following specific steps:

[0043] Step 1: Collect sample data and sample labels.

[0044] The sensors installed on the train intermittently collect the axle box vibration acceleration signal of the train on each wheel, and transmit the train axle box vibration acceleration signal to the data processing platform, which performs Fourier transform on the axle box vibration acceleration signal to generate the power spectrum of the axle box vibration acceleration, such as Figure 2As shown, the power spectrum of the axle box vibration acceleration is taken as sample data.

[0045] Specifically, the sensor is used to intermittently collect the axle box vibration acceleration signal of each wheel k times, wherein the collection frequency of the sensor is 10000 Hz, the duration of each sampling is 0.68 s, and k≥2.

[0046] In a preferred embodiment, the same wheel is sampled 30 times intermittently, that is, k=30.

[0047] The actual roughness curve of each wheel is calculated and plotted, and the actual roughness curve is used as the sample label of the sample data of the corresponding wheel (i.e., the power spectrum of the axle box vibration acceleration). The calculation method of the actual roughness of the wheel is shown in Formula 1:

[0048]

[0049] Where, L θ is the roughness of the wheel polygon of order θ, θ is the order of the wheel polygon, and its value range is 1-23, where each wheel of order corresponds to a roughness, such as Figure 3 As shown; is the radial deviation power of the wheel profile f θ The root mean square of the 1 / 3 octave band; r ref is the reference value of wheel roughness, usually r ref =1μm.

[0050]

[0051] In the formula, f θ is the radial deviation power of the wheel profile, v is the train speed, and r is the theoretical radius of the wheel.

[0052] like Figure 3 As shown in the figure, the wheel roughness curve is drawn according to the 1-23 order wheel roughness calculated by formula 1-2 for the same wheel, and the wheel roughness curve is used as the sample label of the corresponding wheel sample data. Then, the wheel roughness recognition set A is established according to the sample data and sample labels, which can be expressed as:

[0053]

[0054] In the formula, is the sample data of the jth wheel at the kth sampling time (i.e., the power spectrum of the axle box vibration acceleration); T j is the sample label corresponding to the jth wheel (i.e., the actual roughness curve of the wheel), which is a 23-order vector; is the roughness identification vector obtained for the j-th wheel at the k-th sampling.

[0055] Step 2: Use machine learning to establish a wheel roughness model.

[0056] The wheel roughness identification set A is divided into a training set B1 and a verification set B2, wherein the sample data in the training set B1 should include the roughness identification vectors formed by the sample data collected on each wheel, and are extracted from the wheel roughness identification set A at a ratio of 50%, which can be expressed as Formula 4; the verification set B2 is the roughness identification vectors remaining after the wheel roughness identification set A extracts the training set B1, which can be expressed as Formula 5, so that the number of roughness identification vectors in the training set B1 is the same as the number of roughness identification vectors in the verification set B2, and the combination of the training set B1 and the verification set B2 can constitute the wheel roughness identification set A.

[0057]

[0058] In the formula, x n is the sample data of the nth group of roughness identification vectors in the training set B1, y n is the sample label in the nth group of roughness identification vectors in the training set; x m is the sample data of the mth group of roughness identification vectors in the validation set B2, y m is the sample label of the mth group of roughness identification vectors in the validation set B2.

[0059] The wheel roughness model is constructed using a GRNN (generalized regression) neural network structure. The neural network structure includes an input layer, a pattern layer, a summation layer, and an output layer. The input layer can use sample data in the validation set (i.e., the power spectrum of the axle box vibration acceleration) as input data.

[0060] The input layer is the input end of the wheel roughness model, which can import input data into the neural network structure. m as input data.

[0061] The training set is used to construct nodes in the pattern layer, and the training set is used as the structural parameter of the neural network. Each node corresponds to a group of rough recognition vectors in the training set, so that the number of nodes is the same as the number of sample data in the training set.

[0062] Each node is used to calculate the distance between the radial basis function (i.e., Gaussian function) of the sample data in the validation set of the input layer and the sample data of a set of roughness identification vectors in the training set (i.e., the Euclidean distance between the input data and the sample data vectors in the training set), and output the Gaussian calculation result of the node, which can be expressed as:

[0063]

[0064] X n =(x1, x2, x3, ..., x n ) (Formula 7)

[0065] X m =(x1, x2, x3, ..., x m ) (Formula 8)

[0066] In the formula, p i is the distance of the Gaussian function of the i-th node in the pattern layer, X m is the sample data vector in the validation set, X n is the sample data vector in the training set, σ is a hyperparameter that controls the width of the Gaussian function, thereby changing the accuracy of the neural network structure in predicting sample labels based on sample data.

[0067] The summation layer is used to normalize the calculation results of the pattern layer, and perform arithmetic summation on the distance of the Gaussian function (i.e., the Gaussian calculation result) as the overall contribution of the sample data in the training set to the output result of the neural network structure, as shown in Formula 9; at the same time, the weighted sum of the distance of the Gaussian function is calculated based on each sample label in the training set, and the contribution of the sample data in the training set to the output result of the neural network structure is adjusted, as shown in Formula 10; and the weighted average of the weighted sum is performed to calculate the prediction result, as shown in Formula 11.

[0068]

[0069] In the formula, s D is the arithmetic sum of the distances of the Gaussian functions of all nodes in the pattern layer, s N is the weighted sum of the distances of the Gaussian functions of all nodes in the pattern layer, y i is the sample label in the i-th roughness identification vector in the training set; y′ i is the prediction result of the i-th element in the output vector obtained based on the input data (i.e., the sample data vector in the validation set), that is, the prediction result y′ obtained after the i-th element in the sample data vector of the validation set is calculated by the neural network structure i .

[0070] The output layer can output the prediction result in the summation layer; wherein the prediction result is a predicted wheel roughness curve output after a group of sample data is input into the wheel roughness model.

[0071] The Euclidean distance between the prediction result output by the neural network structure and the sample label in the validation set is used as the error of the wheel roughness model, and the error of the wheel roughness model can be changed by adjusting the hyperparameter. When the error of the wheel roughness model is the smallest, the hyperparameter is the optimal parameter of the wheel roughness model, and the wheel roughness model constructed using the optimal parameter can be used to predict the wheel roughness curve.

[0072] In a preferred embodiment, the training set is simplified to reduce the number of nodes in the pattern layer. The power spectrum of the axle box vibration acceleration collected from the same wheel in the training set is averaged, and the average is used as the average sample data of the corresponding wheel in the training set, so that the average sample data and the sample label of the corresponding wheel constitute the roughness recognition vector of the wheel in the training set, and the number of roughness recognition vectors in the training set is the same as the number of collected wheels. By reducing the number of roughness recognition vectors in the training set and thus reducing the number of nodes in the pattern layer, the inference time consumed by the wheel roughness model to iterate each node in the pattern layer is reduced.

[0073] Step 3: Use the wheel roughness model to predict the wheel roughness curve.

[0074] The wheel roughness model with the optimal parameters substituted in step 2 is put into practical use. There is no need to measure a large number of parameters of the high-speed train and obtain wheel roughness of different orders one by one. The wheel roughness curve acquisition method is simplified. The sensor installed on the train re-collects the same wheel k times, and k groups of axle box vibration acceleration signals are sent to the data processing platform. The data processing platform processes the acceleration signals and generates the power spectrum of k groups of axle box vibration acceleration. Figure 5 As shown in , the power spectra of k groups of axle box vibration acceleration are input into the input layer of the wheel roughness model to obtain k groups of prediction results. Figure 6 As shown, k groups of prediction results are averaged, and the average of the prediction results is used as the predicted roughness curve of the wheel to eliminate the influence of vibration and impact during vehicle operation.

[0075] In summary, the present invention obtains the power spectrum of the axle box vibration acceleration signal from the axle box vibration acceleration signal, calculates the wheel roughness curve, constructs a wheel roughness identification set, and establishes a wheel roughness model based on the wheel roughness identification set through the GRNN neural network structure, thereby achieving accurate identification of the wheel polygonal features, simplifying the method for obtaining the wheel roughness curve, and shortening the calculation time of the wheel roughness.

[0076] Although the content of the present invention has been described in detail through the above preferred embodiments, it should be appreciated that the above description should not be considered as a limitation of the present invention. After reading the above content, it will be apparent to those skilled in the art that various modifications and substitutions of the present invention will occur. Therefore, the protection scope of the present invention should be limited by the appended claims.

Claims

1. A method for identifying the roughness of high-speed train wheels, characterized in that: include: Step 1, obtaining power spectra of axle box vibration accelerations of several wheels on a train as sample data; Calculate the actual roughness curve of each wheel and use the actual roughness curve as a sample label; construct a wheel roughness identification set using the sample data and the sample label; Step 2, based on the wheel roughness identification set, using a neural network structure to establish a wheel roughness model; Step 3: using the wheel roughness model to predict the roughness curve of the train.

2. The method for identifying the roughness of high-speed train wheels according to claim 1, characterized in that: In step 1, the method for obtaining the power spectrum of the axle box vibration acceleration includes: Each wheel is sampled intermittently, and the axle box vibration acceleration signals of k groups of trains are collected on each wheel. The axle box vibration acceleration signals are Fourier transformed to obtain the power spectrum of the axle box vibration acceleration; wherein k≥2.

3. The method for identifying the roughness of high-speed train wheels according to claim 2, characterized in that: The sensor installed on the train is used to intermittently collect the axle box vibration acceleration signal of each wheel k times. The collection frequency of the sensor is 10000 Hz, and the duration of each sampling is 0.68 s.

4. The method for identifying the roughness of high-speed train wheels according to claim 1, characterized in that: In step 1, the method for calculating the actual roughness curve of the wheel is as follows: Where, L θ is the roughness of the wheel polygon of order θ, θ is the order of the wheel polygon; is the radial deviation power of the wheel profile f θ The root mean square of the 1 / 3 octave band; r ref is the reference value of wheel roughness, taking r ref =1μm; v is the train speed, r is the theoretical radius of the wheel.

5. The method for identifying the roughness of high-speed train wheels according to claim 1, characterized in that: In step 2, the wheel roughness identification set is divided into a training set and a validation set; the sample data in the training set includes a roughness identification vector composed of sample data collected on each wheel, and is extracted from the wheel roughness identification set at a ratio of 50%; the validation set is the roughness identification vector remaining after the training set is extracted from the wheel roughness identification set.

6. The method for identifying the roughness of high-speed train wheels according to claim 5, characterized in that: The power spectrum of the axle box vibration acceleration collected from the same wheel in the training set is averaged, and the average is used as the mean sample data of the corresponding wheel in the training set, so that the mean sample data and the sample label of the corresponding wheel constitute the roughness recognition vector of the wheel in the training set, so that the number of roughness recognition vectors in the training set is the same as the number of collected wheels.

7. The method for identifying the roughness of high-speed train wheels according to claim 5, characterized in that: In step 2, a wheel roughness model is constructed using a generalized regression neural network structure, and the generalized regression neural network structure includes: The input layer is the input end of the wheel roughness model and can import input data into the neural network structure; the sample data vector in the validation set is used as input data; wherein the sample data vector in the validation set is composed of all the sample data in the validation set; The pattern layer uses the training set to construct nodes. Each node corresponds to a set of roughness recognition vectors in the training set. The distance of the Gaussian function between the sample data in the validation set input by the input layer and the sample data of a set of roughness recognition vectors in the training set can be calculated to obtain the calculation results of the pattern layer, which can be expressed as: In the formula, p i is the distance of the Gaussian function of the i-th node in the pattern layer, X m is the sample data vector in the validation set, X n is the sample data vector in the training set, σ is the hyperparameter; The summation layer is used to normalize the calculation results of the pattern layer and calculate the prediction results, which can be expressed as: In the formula, s D is the arithmetic sum of the Gaussian function distances of all nodes in the pattern layer, s N is the weighted sum of the Gaussian function distances of all nodes in the pattern layer, y i is the sample label in the i-th roughness identification vector in the training set; y′ i is the prediction result of the i-th element in the output vector obtained based on the input data; The output layer can output the prediction results in the summation layer.

8. The method for identifying the roughness of high-speed train wheels according to claim 7, characterized in that: In step 2, the Euclidean distance between the prediction result output by the neural network structure and the sample label in the verification set is taken as the error of the wheel roughness model, and the error of the wheel roughness model can be changed by adjusting the hyperparameters; when the error of the wheel roughness model is the smallest, the hyperparameters are the optimal parameters of the wheel roughness model, and the wheel roughness model constructed using the optimal parameters can be used to predict the roughness curve of the wheel.

9. The method for identifying the roughness of high-speed train wheels according to claim 8, characterized in that: In step 3, k data are collected for the same wheel to obtain k groups of axle box vibration acceleration power spectra, and the power spectra of the k axle box vibration accelerations are input into the wheel roughness model to obtain k groups of prediction results; the k groups of prediction results are averaged, and the mean of the prediction results is used as the predicted roughness curve of the wheel.

Images