Auxiliary decision-making method and system for steel rail corrugation grinding

Abstract

The invention discloses an auxiliary decision-making method and system for rail corrugation grinding. The method comprises the steps that S1, in-train noise data is obtained, specifically, real in-train noise signals generated in the train running process are collected, and the real in-train noise signals are collected by a collection system composed of a sound pressure sensor and a data collection instrument; s2, acquiring train running speed and rail corrugation data; s3, preprocessing in-vehicle noise data; s4, pre-training in-vehicle noise data; s5, a data label of steel rail corrugation grinding is constructed; s6, performing fine adjustment on the self-supervised contrast learning model by using steel rail corrugation grinding label data; and S7, in-vehicle noise data actually measured on site is input into the final steel rail corrugation grinding auxiliary decision-making model, and a steel rail grinding suggestion is obtained. According to the method, steel rail corrugation grinding suggestions can be given according to the in-vehicle noise, and the method has the advantages of being efficient, intelligent, active in maintenance and the like.

Description

Technical Field

[0001] The present invention relates to the field of railway engineering management, and in particular to an auxiliary decision-making method and system for rail corrugation grinding. Background Art

[0002] The scale and volume of railway lines are increasing year by year, and various diseases of the tracks are inevitable, which affect driving safety and riding comfort, and bring a series of problems such as vibration and noise. Rail corrugation is the most common type of damage phenomenon in the wheel-rail system. It is specifically manifested as the periodic wavy vertical unevenness along the longitudinal surface of the rail. It is the direct source of wheel-rail noise and internal noise of the train, and is also one of the main causes of damage to vehicle suspension components and breakage of rail fastener spring bars. Especially in the subway field, most of the vibration or noise problems complained by passengers are related to rail corrugation. The mechanism of rail corrugation is extremely complex and has not yet been determined. The most commonly used control measures are rail grinding or milling. According to the state of rail surface corrugation, large grinding vehicles or milling vehicles are used to remove rail surface corrugation in accordance with relevant standards.

[0003] The control of rail surface roughness is a very important engineering issue. Keeping the wheel-rail contact within reasonable limits is crucial for the stable operation of the wheel-rail system. Rail corrugation is an unavoidable problem in rail surface roughness control, but the current maintenance standards vary among metro operators. EN 13231-3:2012 specifies the limits of short-wave roughness that can remain on the rail after rail head correction and grinding. At the same time, it clearly states that this standard is not applicable to the correction of acoustic tracks. Another international standard, ISO 3095:2013, determines the limits of rail corrugation based on the passing noise of rolling stock, typically in the form of a 1 / 3 octave spectrum of rail surface roughness. However, the main purpose of ISO3095 is to obtain repeatable and comparable noise emission levels of different vehicles, and ISO3095 is based on ballasted track, so it is not applicable to the in-vehicle noise control of subway trains running on slab tracks in tunnels. According to field investigations, ISO 3095 may be too strict and increase economic costs. In the subway field, the current polishing acceptance standards are still in an unstandardized state.

[0004] At present, subway operation and maintenance mostly arrange rail grinding plans based on a combination of passenger complaints and periodic plans. However, this maintenance method is a passive strategy, and the grinding acceptance criteria do not take into account the noise inside the car. This has caused rails in some sections to be excessively and frequently ground, resulting in early rail replacement, which ultimately increases maintenance costs and is not conducive to the green development of rail transit. In summary, it is urgent to develop a rail corrugation grinding auxiliary decision-making method and system based on in-car noise, which can make rail grinding decisions based on the in-car noise status, thereby improving the intelligent level of railway operation and maintenance, and formulating maintenance plans scientifically and reasonably. Summary of the invention

[0005] The present invention aims to provide an auxiliary decision-making method for rail corrugation grinding to solve the above problems.

[0006] The technical solution of the present invention is: an auxiliary decision-making method for rail corrugation grinding, comprising:

[0007] S1. Acquiring in-vehicle noise data: Collecting the real in-vehicle noise signal generated during the operation of the train. The real in-vehicle noise signal is collected by a collection system consisting of a sound pressure sensor and a data acquisition instrument;

[0008] S2. Acquiring train running speed and rail corrugation data: collecting running speed data and real rail corrugation data during the running of the train. The real rail corrugation data is collected by the corrugation trolley, including wavelength and amplitude;

[0009] S3, preprocessing the in-vehicle noise data: remove the in-vehicle noise data when the train starts braking and stops at the station; determine the filtering range according to the wavelength distribution range of rail corrugation of different types of railways, use a digital filter to perform bandpass filtering on the original in-vehicle noise data to obtain filtered in-vehicle noise data, as shown in formulas (1) to (3); slice the filtered in-vehicle noise data to obtain an in-vehicle noise data set; perform random data enhancement on the in-vehicle noise data set to obtain an enhanced in-vehicle noise data set; wherein the random data enhancement includes adding Gaussian noise, random frequency domain masking, time reversal and horizontal flipping;

[0010]

[0011]

[0012] Where v max and v min are the maximum and minimum speeds of the train in this section, respectively, max and λ min are the maximum and minimum wavelengths of rail corrugation in this section, respectively, max and F minare the upper and lower limit frequencies of the bandpass filter, filter refers to the FIR digital bandpass filter, x(t) is the original in-car noise data, and f(t) is the in-car noise data after filtering;

[0013] S4, pre-training car interior noise data: use the self-supervised contrastive learning model to pre-train on the enhanced car interior noise dataset to extract low-dimensional embedding features of high-dimensional car interior noise data;

[0014] Among them, the self-supervised contrastive learning model includes an improved frequency-domain enhanced skip convolution module and a multi-head attention autoregressive module;

[0015] Among them, the frequency domain enhanced skip convolution module g enc Encode the high-dimensional in-vehicle noise data into a low-dimensional embedding space, as shown in formula (4);

[0016] Multi-head attention autoregressive module g ar Use the embedded vector to predict future time steps, as shown in formula (5);

[0017] z t =g enc (x t )(4)

[0018] c t =g ar (z ≤t )(5)

[0019] Where z t is the embedding vector encoded by the frequency domain enhanced skip convolution module, c t is the context vector output by the multi-head attention autoregressive module;

[0020] S5, constructing data labels for rail corrugation grinding;

[0021] S6. Fine-tune the self-supervised contrastive learning model using rail corrugation grinding label data;

[0022] The optimal parameters of the self-supervised contrastive learning model obtained on the in-car noise dataset are used as pre-training parameters;

[0023] Build a classification head model and combine it with the frozen self-supervised contrastive learning model to form a fine-tuning model;

[0024] Train on the labeled data training set of rail corrugation grinding to obtain the parameters of the fine-tuning model;

[0025] The model was validated on the labeled data validation set of rail corrugation and grinding, and the optimal fine-tuning model parameters were selected to obtain the final rail corrugation and grinding auxiliary decision model.

[0026] S7. In the final rail corrugation and grinding auxiliary decision model, the on-site measured in-vehicle noise data is input to obtain rail grinding recommendations.

[0027] Preferably, in step S4,

[0028] The improved frequency domain enhanced skip convolution module includes frequency domain enhanced and skip convolution neural network;

[0029] Among them, frequency domain enhancement obtains frequency domain data by performing fast Fourier transform on the interior noise, as shown in formula (6); only the result in the real domain is retained, and then a zero padding operation is performed to obtain a vector with the same length as the interior noise data slice; finally, the interior noise data and the frequency domain enhancement data are spliced ​​together as the input vector;

[0030]

[0031] Where X[k] is the frequency domain result, x[n] is the time domain input signal, and N is the number of sampling points of the signal;

[0032] The skip-connected convolutional neural network includes an input layer, a convolutional layer, an activation layer, and an output layer connected in sequence, as well as a skip-connected intermediate layer; the skip connection is to directly add the output of a layer in the convolutional neural network to the output of a subsequent layer.

[0033] Preferably, in step S4,

[0034] The multi-head attention autoregressive module consists of a multi-head attention layer, a layer normalization layer, and a fully connected layer connected in sequence, where the attention score is determined by formulas (7) to (8), and the output of the fully connected layer is determined by formula (9);

[0035]

[0036] Where Q is the query matrix, K is the key matrix, V is the corresponding value matrix, and d k is the dimension of the key;

[0037] MultiHead(Q,K,V)=Concat(head1,...,head h )W O

[0038] where head i =Attention(QW i Q ,KW i K ,VW i V )(8)

[0039] Where W i Q, W i Q and W i Q are the linear projection matrices corresponding to Q, K, and V, respectively. Q, K, and V are projected into different subspaces, and then the attention scores are calculated in each subspace. O It is the matrix of linear projection after concatenating all attention calculation results, and h refers to the number of heads;

[0040] FFN(x)=max(0,xW1+b1)W2+b2(9)

[0041] Where FFN(x) is the output of the fully connected layer, W1 and W2 are the weights of the two linear transformations, b1 and b2 are the corresponding bias terms, and max(.) represents the ReLU function.

[0042] Preferably, in step S5,

[0043] According to the train running speed data, the in-vehicle noise data and the rail corrugation data are aligned in the spatial domain; the rail corrugation data is processed to obtain the roughness within a certain wavelength range, which is determined by formula (10);

[0044]

[0045] Where L r is the rail surface roughness value, in dB; r rms is the effective value of the rail surface corrugation amplitude, in μm; r0 is the reference value;

[0046] The relationship between rail roughness and in-vehicle noise is established based on the in-vehicle noise prediction model, and the rail roughness limit is inferred from the in-vehicle noise limit.

[0047] Define the labels corresponding to the in-vehicle noise dataset according to the rail roughness limit;

[0048] The labeled data of rail corrugation grinding are divided into training set and validation set.

[0049] Preferably, in step S3, random data enhancement is performed on the in-vehicle noise dataset, specifically including:

[0050] Add Gaussian noise to the in-vehicle noise dataset with the first probability, and the size of the Gaussian noise is controlled by the signal-to-noise ratio, as shown in formula (11);

[0051]

[0052] Where SNR is the signal-to-noise ratio, P s is the power of the signal, P n is the power of the noise;

[0053] A certain frequency band in the frequency domain of the vehicle interior noise data set is set to zero with the second probability, and then the time domain data is obtained by inverse Fourier transform, as shown in formula (12);

[0054]

[0055] Performing time-reversal on the in-vehicle noise dataset with a second probability;

[0056] The interior noise dataset is numerically flipped with the second probability.

[0057] Preferably, in S4, pre-training is performed on the enhanced in-vehicle noise dataset using a self-supervised contrastive learning model, specifically including:

[0058] The weight parameters in the self-supervised contrastive learning model established in step S4 are updated. The parameter update uses the Adam optimizer, which is defined as formula (13);

[0059] v t =β1v t-1 +(1-β1)g t

[0060] s t =β2s t-1 +(1-β2)g t ⊙g t (13)

[0061] In the formula, v t 、v t-1 are the momentum of the gradient at time t and t-1 respectively; s t 、s t-1 are the second moments of the gradient at time t and t-1 respectively; β1 and β2 are non-negative weighting parameters; g t represents the gradient at time t;

[0062] The loss function uses infoNCE, which is defined as in formulas (14) to (15);

[0063]

[0064] Where W k is a linear weight matrix, i.e. t Use a simple linear transformation to predict future observations x t+k ;

[0065] Use weight decay and gradient clipping to regularize the model, use early stopping mechanism to prevent model overfitting, and stop iteration when the validation set loss no longer decreases within the preset number of steps;

[0066] The input of the frequency domain enhanced skip convolution module is the high-dimensional vector of the concatenation of the in-car noise data and the frequency domain enhanced data, and the output is the low-dimensional embedding vector z;

[0067] The input of the multi-head attention autoregressive module is a low-dimensional embedding vector z, and the output is a vector c containing contextual information.

[0068] Preferably, in step S5, the in-vehicle noise prediction model specifically includes:

[0069] The prediction model includes a train body finite element model and a train interior acoustic cavity finite element model, wherein the train body finite element model includes the chassis, side walls, end walls, roof and floor components, and the tunnel finite element model is connected by boundary elements to obtain a tunnel-car body finite element-boundary element model;

[0070] The model excitation mainly includes the secondary suspension force excitation and the wheel-rail noise excitation, among which the wheel-rail noise is the field measured data; the model excitation is the input of the prediction model, and the output of the prediction model is the estimated value of the vehicle interior noise;

[0071] The secondary suspension force excitation is calculated by considering the rail corrugation excitation with different roughness in the vehicle-track rigid-flexible coupling dynamic model;

[0072] The vehicle-track coupled dynamics model includes the vehicle body, frame, axle box, flexible wheelset, rails and substructure, which is used to calculate the secondary suspension force and consider corrugation excitations of different wavelengths and amplitudes.

[0073] Preferably, in step S6, classifying the head model specifically includes:

[0074] The classification head model is a neural network model consisting of an input layer, a fully connected layer, an activation layer, and an output layer connected in sequence;

[0075] The input layer is the context vector c encoded by the pre-trained self-supervised contrastive learning model, and the output layer is the labels corresponding to different rail corrugation grinding;

[0076] The loss function uses cross entropy, which is defined as formula (16);

[0077] H(p,q)=-∑ x (p(x)logq(x)+(1-p(x))log(1-q(x)))(16)

[0078] In the formula, probability distribution p is the expected output, and probability distribution q is the actual output;

[0079] During the fine-tuning process, only the network parameters in the classification head model are updated, and the frozen pre-trained model network parameters are not updated.

[0080] An auxiliary decision system for rail corrugation grinding, comprising:

[0081] The data acquisition module is used to acquire the in-car noise data, which is the sound pressure change data inside the car during the train operation;

[0082] A data preprocessing module, used for bandpass filtering the in-vehicle noise data, and dividing the in-vehicle noise data into in-vehicle noise data sets according to the window length and step length;

[0083] A label acquisition module is used to divide the in-vehicle noise dataset into a labeled dataset and an unlabeled dataset according to the rail corrugation test data;

[0084] Model pre-training module, which is used to build a self-supervised contrastive learning model and train it on an unlabeled in-car noise dataset to obtain the optimal pre-training parameters;

[0085] The model fine-tuning module is used to build a classification head model, train and verify it on a labeled in-vehicle noise dataset, and obtain the final rail corrugation grinding auxiliary decision model;

[0086] The decision module is used to input the in-vehicle noise data into the trained rail corrugation grinding auxiliary decision model to obtain rail grinding suggestions;

[0087] Data visualization module, used to manage in-vehicle noise data and visualize rail corrugation grinding to assist decision making.

[0088] An electronic device comprises a memory and a processor, the processor and the memory communicate with each other, the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute an auxiliary decision method for rail corrugation grinding as claimed in any one of claims 1 to 8.

[0089] The beneficial effects of the present invention are:

[0090] Compared with the existing method of measuring rail corrugation using a corrugation trolley or an electronic straightness ruler, the present invention directly uses the noise data inside the vehicle to judge the rail corrugation status of the line, which has the advantages of low cost and high efficiency.

[0091] Compared with the existing rail grinding acceptance standard, the present invention can give the rail roughness limit value by considering the in-vehicle noise amplitude, and can also give rail grinding suggestions, which is a more economical, scientific and intelligent rail maintenance strategy;

[0092] Compared with the existing intelligent detection technology for rail corrugation, the present invention first performs self-supervised comparative learning on a large amount of unlabeled in-vehicle noise data, and then performs fine-tuning on a customized rail corrugation and grinding label dataset, solving the technical problem of expensive acquisition of field data sample labels. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 A flow chart of an auxiliary decision-making method for rail corrugation grinding provided by an embodiment of the present invention;

[0094] Figure 2 A schematic diagram of an on-site test of in-vehicle noise provided by an embodiment of the present invention;

[0095] Figure 3 A diagram of measured data of in-vehicle noise provided by an embodiment of the present invention;

[0096] Figure 4 A sample data graph of a random data enhancement method provided by an embodiment of the present invention;

[0097] Figure 5 A schematic diagram of the structure of a frequency domain enhanced skip convolution module provided by an embodiment of the present invention;

[0098] Figure 6 A schematic diagram of multi-head attention score calculation provided by an embodiment of the present invention;

[0099] Figure 7 A schematic diagram of the structure of a self-supervised contrastive learning model provided by an embodiment of the present invention;

[0100] Figure 8 A schematic diagram of the change of a loss curve during a pre-training process provided by an embodiment of the present invention;

[0101] Fig. 9 A topological diagram of a vehicle-track rigid-flexible coupling analysis model provided in an embodiment of the present invention;

[0102] Fig.10 A schematic diagram of a rail corrugation grinding label provided by an embodiment of the present invention;

[0103] Fig.11 A schematic diagram of a fine-tuning model structure provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0104] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. The implementation methods of the present invention are not limited thereto.

[0105] Example 1

[0106] like Figure 1 As shown, an auxiliary decision-making method for rail corrugation grinding includes:

[0107] Step 1: Measure the sound pressure signal inside the carriage generated during the train operation, such as Figure 2As shown; IEPE type high-precision sound pressure sensor is used, 24AD high-precision data acquisition system is used for data acquisition, sampling frequency is set to 20000Hz, and the sound pressure sensor is placed in the area corresponding to the bogie position in the same carriage as the train speed sensor, so as to ensure the spatial matching of train running mileage and in-car noise data;

[0108] Step 2: Collect the running speed data recorded by the train ATP system and convert it into ASCII format; during the window time, use a corrugation trolley to test the corrugation of the rail surface in some sections to obtain the rail surface roughness data;

[0109] Step 3: Remove the noise data inside the train when the train starts braking and stops at the station, such as Figure 3 As shown in the figure, the filtering range is determined according to the wavelength distribution range of rail corrugation of different types of railways, the wavelength range of subway rail corrugation is 30 to 300 mm, and the original in-vehicle noise data is band-pass filtered using a FIR digital filter, as shown in formulas (1) to (3), and the stopband attenuation is not less than 60 dB, thereby obtaining the filtered in-vehicle noise data; the filtered in-vehicle noise data is sliced, and the window length is 10240 data points with an overlap of 50%, thereby obtaining the in-vehicle noise data set;

[0110]

[0111] Where v max and v min are the maximum and minimum speeds of the train in this section, respectively, max and λ min are the maximum and minimum wavelengths of rail corrugation in this section, respectively, max and F min are the upper and lower limit frequencies of the bandpass filter, filter refers to the FIR digital bandpass filter, x(t) is the original in-car noise data, and f(t) is the in-car noise data after filtering.

[0112] Step 4: Perform random data enhancement on the in-vehicle noise dataset, and add Gaussian noise to the in-vehicle noise dataset with a certain probability. The size of the Gaussian noise is controlled by the signal-to-noise ratio, as shown in formula (4); wherein the probability is preferably 20%.

[0113]

[0114] Where SNR is the signal-to-noise ratio, P s is the power of the signal, P n is the power of the noise;

[0115] A certain frequency band in the frequency domain of the vehicle interior noise data set is set to zero with a certain probability, and then the time domain data is obtained by using the inverse Fourier transform, as shown in formula (5); wherein the probability is preferably 50%.

[0116]

[0117] The in-vehicle noise dataset is time-reversed with a certain probability, wherein the probability is preferably 50%.

[0118] The vehicle interior noise data set is numerically flipped with a certain probability, wherein the probability is preferably 50%.

[0119] The enhanced in-car noise dataset is obtained, such as Figure 4 As shown;

[0120] Step 5: Construct frequency domain enhanced skip convolution module g enc ,like Figure 5 As shown; frequency domain enhanced skip convolution module g enc Encode the high-dimensional in-vehicle noise data into a low-dimensional embedding space, as shown in formula (6);

[0121] z t =g enc (x t )(6)

[0122] Where z t is the embedding vector encoded by the frequency domain enhanced skip convolution module;

[0123] The frequency domain enhancement skip convolution module is mainly composed of two parts: frequency domain enhancement and skip convolution neural network. The frequency domain enhancement obtains frequency domain data by performing fast Fourier transform on the in-vehicle noise, as shown in formula (7); only the result in the real domain is retained, and then zero padding is performed to obtain a vector with the same length as the in-vehicle noise data slice; finally, the in-vehicle noise data and the frequency domain enhancement data are concatenated together as the input vector;

[0124]

[0125] Where X[k] is the frequency domain result, x[n] is the time domain input signal, and N is the number of sampling points of the signal;

[0126] The skip-connected convolutional neural network includes an input layer, a convolutional layer, an activation layer, and an output layer connected in sequence, as well as an intermediate layer of skip connections; wherein the skip connection is to directly add the output of a layer in the convolutional neural network to the output of a subsequent layer;

[0127] Step 6: Construct a multi-head attention autoregressive module g ar , this module uses the embedded vector z to predict future time steps, as shown in formula (8);

[0128] c t =g ar (z ≤t )(8)

[0129] Where c t is the context vector output by the multi-head attention autoregressive module;

[0130] The multi-head attention autoregressive module is composed of a multi-head attention layer, a layer normalization layer, and a fully connected layer connected in sequence, where the attention score is determined by formulas (9) to (10), as follows: Figure 6 As shown, the output of the fully connected layer is determined by formula (11);

[0131]

[0132] Where Q is the query matrix, K is the key matrix, V is the corresponding value matrix, and d k is the dimension of the key;

[0133] MultiHead(Q,K,V)=Concat(head1,...,head h )W O

[0134] where head i =Attention(QW i Q ,KW i K ,VW i V )(10)

[0135] Where W i Q , W i Q and W i Q are the linear projection matrices corresponding to Q, K, and V, respectively. Q, K, and V are projected into different subspaces, and then the attention scores are calculated in each subspace. O It is the matrix of linear projection after concatenating all attention calculation results, and h refers to the number of heads;

[0136] FFN(x)=max(0,xW1+b1)W2+b2(11)

[0137] Where FFN(x) is the output of the fully connected layer, W1 and W2 are the weights of the two linear transformations, b1 and b2 are the corresponding bias terms. max(.) represents the ReLU function;

[0138] Step 7: Assemble the frequency domain enhanced skip convolution module and the multi-head attention autoregressive module to form a self-supervised contrastive learning model, such as Figure 7As shown in the figure; the input of the frequency domain enhanced skip convolution module is the high-dimensional vector concatenated with the in-car noise data and the frequency domain enhanced data, and the output is the low-dimensional embedding vector z; the input of the multi-head attention autoregressive module is the low-dimensional embedding vector z, and the output is the vector c containing context information.

[0139] Step 8: Use the self-supervised contrastive learning model to pre-train on the enhanced in-vehicle noise dataset, such as Figure 8 As shown in Figure 1, in contrastive learning, the number of time steps predicted in the future is set to 12, and the number of negative samples is set to 10. The SNR after adding Gaussian noise is 30 dB, and the frequency domain masking width is 100 Hz. The parameter update uses the Adam optimizer, defined as in formula (12), the initial learning rate is set to 0.0001, the batch size is set to 32, and a total of 300 epochs are trained.

[0140] v t =β1v t-1 +(1-β1)g t

[0141] s t =β2s t-1 +(1-β2)g t ⊙g t (12)

[0142] In the formula, v t 、v t-1 are the momentum of the gradient at time t and t-1 respectively; s t 、s t-1 are the second moments of the gradient at time t and t-1 respectively; β1 and β2 are non-negative weighting parameters, usually 0.9 and 0.999; g t represents the gradient at time t.

[0143] The loss function uses infoNCE, which is defined as in formulas (13) to (14);

[0144]

[0145] Where W k is a linear weight matrix, i.e. t Use a simple linear transformation to predict future observations x t+k ;

[0146] Use weight decay and gradient clipping to regularize the model, use early stopping mechanism to prevent model overfitting, and stop iteration when the validation set loss no longer decreases within the preset number of steps;

[0147] Step 9: Establish a vehicle-track rigid-flexible coupling analysis model. The model topology diagram is as follows: Fig. 9As shown in the figure, the model is mainly composed of a car body, a frame, an axle box, a wheel set, a rail and a lower foundation. The secondary suspension connecting the car body and the frame and the primary suspension connecting the frame and the axle box are simulated by linear spring force elements. The axle box and the wheelset are connected by a hinge with only rotational freedom around the wheel axle. The fasteners are simulated by bushing force elements. The car body, frame, axle box and the geodetic coordinate system are connected by a hinge with 6 degrees of freedom.

[0148] The wheelset is simulated by a flexible wheelset, the rail is simulated by a Timoshenko beam, the wheel-rail tangential adhesion-creep state is simulated by the FASTSIM algorithm, the wheel-rail normal contact is based on the non-Hertz contact theory, and the Kik-Piotrowski algorithm based on virtual penetration is used. The wheel-rail contact force and contact stress are:

[0149]

[0150] Where E is the elastic modulus; δ is the penetration; μ is the Poisson's ratio; x l (y) is the front edge of the wheel-rail contact patch along the longitudinal axis; ε is the correction coefficient, which is 0.55; R W is the radius of the wheel rolling circle;

[0151] Taking the subway type B car as an example, the parameter values ​​of the vehicle-track rigid-flexible coupling analysis model are shown in the following table;

[0152] Table 1 Parameters of vehicle-track rigid-flexible coupling model

[0153]

[0154] Step 10: Calculate the magnitude of the secondary suspension force of the train under different rail roughness conditions, use the American five-level spectrum to simulate medium and long wave track irregularities, take a safety factor of 0.25, and select a wavelength range of 3 to 80 m. The American five-level spectrum is defined as follows:

[0155]

[0156] In the formula, S v (φ) is the track irregularity power spectrum density, unit is m 2 / (1 / m); φ is the spatial frequency, the unit is 1 / m; Av is the roughness constant, the unit is m; φ v1 ,φ v2 is the cutoff frequency, the unit is 1 / m;

[0157] The Sato spectrum is used to simulate the shortwave track irregularity, and the wavelength range is selected from 0.01 to 3 m. The Sato spectrum is defined as follows:

[0158]

[0159] Where Ω is the spatial frequency, unit is 1 / m; A is the wheel-rail surface roughness coefficient, value is 4.15×10 -8 ~5.0×10 -7 ;

[0160] The mathematical expression for simulating rail corrugation is as follows;

[0161]

[0162] In the formula, Z cor (t) is the amplitude of the corrugation irregularity; a is the depth of the corrugation; λ is the wavelength of the corrugation;

[0163] The constructed track irregularities with different rail roughness are input into the vehicle-track rigid-flexible coupling analysis model for calculation to obtain the secondary suspension force.

[0164] Step 11, establish a prediction model for in-vehicle noise, which mainly consists of a train body finite element model composed of components such as the chassis, side walls, end walls, roof and floor, and a finite element model for the in-vehicle acoustic cavity, and use boundary elements to connect the tunnel finite element model to obtain a tunnel-car body finite element-boundary element model;

[0165] The model excitation mainly includes secondary suspension force excitation and wheel-rail noise excitation, among which the wheel-rail noise is the field measured data; the secondary suspension force excitation is calculated by considering the rail corrugation excitation of different roughness in the vehicle-track rigid-flexible coupling dynamic model;

[0166] Input the calculated secondary suspension force and the wheel-rail noise data measured on site into the in-vehicle noise prediction model to obtain the in-vehicle noise level;

[0167] Through multiple calculations, the rail roughness limit is inferred based on the noise inside the train;

[0168] Step 12: align the in-vehicle noise data and the rail corrugation data in the spatial domain according to the train running speed data; process the rail corrugation data to obtain the roughness within the wavelength range of 5 to 315 mm, which is determined by formula (20);

[0169]

[0170] Where L r is the rail surface roughness value, in dB; r rms is the effective value of the rail surface corrugation amplitude, in μm; r0 is the reference value, generally 1 μm;

[0171] Define the labels corresponding to the in-vehicle noise dataset according to the rail roughness limit, such as Fig.10 As shown,

[0172] The labels corresponding to the in-vehicle noise dataset are defined according to the rail roughness limit. The label format is {no corrugation*no need to polish; no corrugation*need to polish; corrugation with a wavelength of 30-100mm*need to polish; corrugation with a wavelength of 30-100mm*no need to polish; corrugation with a wavelength of 100-300mm*need to polish; corrugation with a wavelength of 100-300mm*no need to polish; corrugation with a wavelength of 30-100mm and 100-300mm*need to polish; corrugation with a wavelength of 30-100mm and 100-300mm**no need to polish}, a total of 8 labels;

[0173] The label data of the rail corrugation grinding is divided into a training set and a validation set in a ratio of 8:2;

[0174] Step 13: Build a classification head model, which is a neural network model consisting of an input layer, a fully connected layer, an activation layer, and an output layer connected in sequence;

[0175] The input layer is the context vector c encoded by the pre-trained self-supervised contrastive learning model, and the output layer is the corresponding 8 rail corrugation grinding labels;

[0176] The loss function uses cross entropy, which is defined as formula (21);

[0177] H(p,q)=-Σ x (p(x)logq(x)+(1-p(x))log(1-q(x)))(21)

[0178] In the formula, probability distribution p is the expected output, and probability distribution q is the actual output;

[0179] Step 14: Use the rail corrugation grinding label data to fine-tune the self-supervised contrastive learning model, such as Fig.11 As shown in the figure, during the fine-tuning process, only the network parameters in the classification head model are updated, and the frozen pre-trained model network parameters are not updated;

[0180] Firstly, the optimal parameters of the self-supervised contrastive learning model obtained on the in-car noise dataset are used as pre-training parameters;

[0181] Secondly, the model is trained on the labeled data training set of rail corrugation grinding to obtain the parameters of the fine-tuning model;

[0182] Finally, the model is verified on the labeled data validation set of rail corrugation and grinding, and the optimal fine-tuning model parameters are selected to obtain the final rail corrugation and grinding auxiliary decision-making model.

[0183] Step 14: Input the on-site measured in-vehicle noise data into the final rail corrugation grinding auxiliary decision model to obtain rail grinding recommendations.

[0184] In S6, during fine-tuning, the network weight parameters of the pre-trained model are no longer updated, and only the parameters of the classification head model are updated. Freezing is because the pre-trained model has already learned a relatively suitable parameter in advance, and only the parameters of the classification head need to be fine-tuned. If it is not frozen, it is equivalent to finding the appropriate parameters again, which will waste computing resources and fail to reflect the advantages of pre-training.

[0185] Example 2

[0186] On the basis of Example 1, this Example 2 provides an auxiliary decision system for rail corrugation grinding, which corresponds to the auxiliary decision method for rail corrugation grinding, and specifically includes:

[0187] A data acquisition module is used to acquire in-car noise data, wherein the in-car noise data is the sound pressure change data inside the car during the train operation;

[0188] A data preprocessing module, used for performing bandpass filtering on the in-vehicle noise data, and dividing the in-vehicle noise data into in-vehicle noise data sets according to a window length and a step length;

[0189] A label acquisition module is used to divide the in-vehicle noise dataset into a labeled dataset and an unlabeled dataset according to the rail corrugation test data;

[0190] A model pre-training module, used to construct a self-supervised contrastive learning model and train it on the unlabeled in-vehicle noise dataset to obtain optimal pre-training parameters;

[0191] A model fine-tuning module is used to construct a classification head model, and to train and verify it on the labeled in-vehicle noise dataset to obtain a final rail corrugation grinding auxiliary decision model;

[0192] The decision module is used to input the in-vehicle noise data into the trained rail corrugation and grinding auxiliary decision model to obtain rail grinding suggestions.

[0193] Data visualization module, used to manage in-vehicle noise data and visualize rail corrugation grinding to assist decision making.

[0194] For specific details, please refer to the description of an auxiliary decision-making method for rail corrugation grinding, which will not be repeated here.

[0195] Example 3

[0196] Embodiment 3 of the present invention provides an electronic device, including a memory and a processor, the processor and the memory communicate with each other, the memory stores program instructions executable by the processor, the processor calls the program instructions to execute an auxiliary decision method for rail corrugation grinding, the auxiliary decision method for rail corrugation grinding includes the following process steps:

[0197] Acquiring in-car noise data, wherein the in-car noise data is the sound pressure data inside the car generated during the operation of the train;

[0198] Acquiring train running speed data and rail corrugation data, wherein the rail corrugation data is measured by a corrugation trolley;

[0199] Preprocess and randomly enhance the in-vehicle noise data to obtain the in-vehicle noise dataset;

[0200] The rail roughness limit is determined using the in-vehicle noise prediction model, and the in-vehicle noise dataset is divided into an unlabeled dataset and a labeled dataset according to the rail roughness limit.

[0201] Constructing a self-supervised contrastive learning model and pre-training it on the unlabeled dataset to obtain pre-trained model parameters;

[0202] Build a classification head model, and fine-tune it on a labeled dataset using frozen pre-trained parameters to obtain the final rail corrugation grinding auxiliary decision model;

[0203] Input the in-vehicle noise data to be detected into the trained rail corrugation and grinding auxiliary decision-making model to obtain rail grinding suggestions.

[0204] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. An auxiliary decision-making method for rail corrugation grinding, characterized in that: include: S1. Acquiring in-vehicle noise data: Collecting the real in-vehicle noise signal generated during the operation of the train. The real in-vehicle noise signal is collected by a collection system consisting of a sound pressure sensor and a data acquisition instrument; S2. Acquiring train running speed and rail corrugation data: collecting running speed data and real rail corrugation data during the running of the train. The real rail corrugation data is collected by the corrugation trolley, including wavelength and amplitude; S3, pre-processing the in-vehicle noise data: removing the in-vehicle noise data when the train starts braking and stops at the station; determining the filtering range according to the wavelength distribution range of rail corrugation of different types of railways, using a digital filter to perform band-pass filtering on the original in-vehicle noise data, and obtaining the filtered in-vehicle noise data, as shown in formulas (1) to (3); The filtered in-vehicle noise data is sliced ​​to obtain an in-vehicle noise data set; the in-vehicle noise data set is randomly enhanced to obtain an enhanced in-vehicle noise data set; wherein the random data enhancement includes adding Gaussian noise, random frequency domain masking, time reversal and horizontal flipping; Where v max and v min are the maximum and minimum speeds of the train in this section, respectively, max and λ min are the maximum and minimum wavelengths of rail corrugation in this section, respectively, max and F min are the upper and lower limit frequencies of the bandpass filter, filter refers to the FIR digital bandpass filter, x(t) is the original in-vehicle noise data, and f(t) is the in-vehicle noise data after filtering; S4, pre-training car interior noise data: use the self-supervised contrastive learning model to pre-train on the enhanced car interior noise dataset to extract low-dimensional embedding features of high-dimensional car interior noise data; Among them, the self-supervised contrastive learning model includes an improved frequency-domain enhanced skip convolution module and a multi-head attention autoregressive module; Among them, the frequency domain enhanced skip convolution module g enc Encode the high-dimensional in-vehicle noise data into a low-dimensional embedding space, as shown in formula (4); Multi-head attention autoregressive module g ar Use the embedded vector to predict future time steps, as shown in formula (5); z t =g enc (x t )(4) c t =g ar (z ≤t )(5) Where z t is the embedding vector encoded by the frequency domain enhanced skip convolution module, c t is the context vector output by the multi-head attention autoregressive module; S5, constructing data labels for rail corrugation grinding; S6. Fine-tune the self-supervised contrastive learning model using rail corrugation grinding label data; The optimal parameters of the self-supervised contrastive learning model obtained on the in-car noise dataset are used as pre-training parameters; Build a classification head model and combine it with the frozen self-supervised contrastive learning model to form a fine-tuning model; Train on the labeled data training set of rail corrugation grinding to obtain the parameters of the fine-tuning model; The model was validated on the labeled data validation set of rail corrugation and grinding, and the optimal fine-tuning model parameters were selected to obtain the final rail corrugation and grinding auxiliary decision model. S7. In the final rail corrugation and grinding auxiliary decision model, the on-site measured in-vehicle noise data is input to obtain rail grinding recommendations.

2. A decision-making aid method for rail corrugation grinding according to claim 1, characterized in that: In step S4, The improved frequency domain enhanced skip convolution module includes frequency domain enhanced and skip convolution neural network; Among them, frequency domain enhancement obtains frequency domain data by performing fast Fourier transform on the interior noise, as shown in formula (6); only the result in the real domain is retained, and then a zero padding operation is performed to obtain a vector with the same length as the interior noise data slice; finally, the interior noise data and the frequency domain enhancement data are spliced ​​together as the input vector; Where X[k] is the frequency domain result, x[n] is the time domain input signal, and N is the number of sampling points of the signal; The skip-connected convolutional neural network includes an input layer, a convolutional layer, an activation layer, and an output layer connected in sequence, as well as a skip-connected intermediate layer; the skip connection is to directly add the output of a layer in the convolutional neural network to the output of a subsequent layer.

3. A decision-making aid method for rail corrugation grinding according to claim 1, characterized in that: In step S4, The multi-head attention autoregressive module is composed of a multi-head attention layer, a layer normalization layer, and a fully connected layer connected in sequence, where the attention score is determined by formulas (7) to (8), and the output of the fully connected layer is determined by formula (9); Where Q is the query matrix, K is the key matrix, V is the corresponding value matrix, and d k is the dimension of the key; MultiHead(Q,K,V)=Concat(head1,...,head h )W O where head i =Attention(QW i Q ,KW i K ,VW i V )(8) Where W i Q , W i Q and W i Q are the linear projection matrices corresponding to Q, K, and V, respectively. Q, K, and V are projected into different subspaces, and then the attention scores are calculated in each subspace. O It is the matrix of linear projection after concatenating all attention calculation results, and h refers to the number of heads; FFN(x)=max(0,xW1+b1)W2+b2(9) Where FFN(x) is the output of the fully connected layer, W1 and W2 are the weights of the two linear transformations, b1 and b2 are the corresponding bias terms, and max(.) represents the ReLU function.

4. A decision-making aid method for rail corrugation grinding according to claim 1, characterized in that: In step S5, According to the train running speed data, the in-vehicle noise data and the rail corrugation data are aligned in the spatial domain; the rail corrugation data are processed to obtain the roughness within a certain wavelength range, which is determined by formula (10); Where L r is the rail surface roughness value, in dB; r rms is the effective value of the rail surface corrugation amplitude, in μm; r0 is the reference value; The relationship between rail roughness and in-vehicle noise is established based on the in-vehicle noise prediction model, and the rail roughness limit is inferred from the in-vehicle noise limit. Define the labels corresponding to the in-vehicle noise dataset according to the rail roughness limit; The labeled data of rail corrugation grinding are divided into training set and validation set.

5. A decision-making aid method for rail corrugation grinding according to claim 1, characterized in that: In step S3, random data enhancement is performed on the in-vehicle noise dataset, specifically including: Add Gaussian noise to the in-vehicle noise dataset with the first probability, and the size of the Gaussian noise is controlled by the signal-to-noise ratio, as shown in formula (11); Where SNR is the signal-to-noise ratio, P s is the power of the signal, P n is the power of the noise; A certain frequency band in the frequency domain of the vehicle interior noise data set is set to zero with the second probability, and then the time domain data is obtained by inverse Fourier transform, as shown in formula (12); Performing time-reversal on the in-vehicle noise dataset with a second probability; The interior noise dataset is numerically flipped with the second probability.

6. A decision-making aid method for rail corrugation grinding according to claim 1, characterized in that: In step S4, a self-supervised contrastive learning model is used to perform pre-training on the enhanced in-vehicle noise dataset, specifically including: The weight parameters in the self-supervised contrastive learning model established in step S4 are updated. The parameter update uses the Adam optimizer, which is defined as formula (13); v t =β1v t-1 +(1-β1)g t s t =β2s t-1 +(1-β2)g t ⊙g t (13) In the formula, v t 、v t-1 are the momentum of the gradient at time t and t-1 respectively; s t 、s t-1 are the second moments of the gradient at time t and t-1 respectively; β1 and β2 are non-negative weighting parameters; g t represents the gradient at time t; The loss function uses infoNCE, which is defined as in formulas (14) to (15); Where W k is a linear weight matrix, that is, for c t Use a simple linear transformation to predict future observations x t+k ; Use weight decay and gradient clipping to regularize the model, use early stopping mechanism to prevent model overfitting, and stop iteration when the validation set loss no longer decreases within the preset number of steps; The input of the frequency domain enhanced skip convolution module is the high-dimensional vector of the concatenation of the in-car noise data and the frequency domain enhanced data, and the output is the low-dimensional embedding vector z; The input of the multi-head attention autoregressive module is a low-dimensional embedding vector z, and the output is a vector c containing contextual information.

7. A decision-making aid method for rail corrugation grinding according to claim 1, characterized in that: In step S5, the in-vehicle noise prediction model specifically includes: The prediction model includes a train body finite element model and a train interior acoustic cavity finite element model, wherein the train body finite element model includes the chassis, side walls, end walls, roof and floor components, and the tunnel finite element model is connected by boundary elements to obtain a tunnel-car body finite element-boundary element model; The model excitation mainly includes the secondary suspension force excitation and the wheel-rail noise excitation, among which the wheel-rail noise is the field measured data; the model excitation is the input of the prediction model, and the output of the prediction model is the estimated value of the vehicle interior noise; The secondary suspension force excitation is calculated by considering the rail corrugation excitation with different roughness in the vehicle-track rigid-flexible coupling dynamic model; The vehicle-track coupled dynamics model includes the vehicle body, frame, axle box, flexible wheelset, rails and substructure, which is used to calculate the secondary suspension force and consider corrugation excitations of different wavelengths and amplitudes.

8. A decision-making aid method for rail corrugation grinding according to claim 1, characterized in that: In step S6, the classification head model specifically includes: The classification head model is a neural network model consisting of an input layer, a fully connected layer, an activation layer, and an output layer connected in sequence; The input layer is the context vector c encoded by the pre-trained self-supervised contrastive learning model, and the output layer is the labels corresponding to different rail corrugation grinding; The loss function uses cross entropy, which is defined as formula (16); H(p,q)=-∑ x (p(x)logq(x)+(1-p(x))log(1-q(x)))(16) In the formula, probability distribution p is the expected output, and probability distribution q is the actual output; During the fine-tuning process, only the network parameters in the classification head model are updated, and the frozen pre-trained model network parameters are not updated.

9. An auxiliary decision-making system for rail corrugation grinding, characterized in that: include: The data acquisition module is used to acquire the in-car noise data, which is the sound pressure change data inside the car during the train operation; A data preprocessing module, used for bandpass filtering the in-vehicle noise data, and dividing the in-vehicle noise data into in-vehicle noise data sets according to the window length and step length; A label acquisition module is used to divide the in-vehicle noise dataset into a labeled dataset and an unlabeled dataset according to the rail corrugation test data; Model pre-training module, which is used to build a self-supervised contrastive learning model and train it on an unlabeled in-car noise dataset to obtain the optimal pre-training parameters; The model fine-tuning module is used to build a classification head model, train and verify it on a labeled in-vehicle noise dataset, and obtain the final rail corrugation grinding auxiliary decision model; The decision module is used to input the in-vehicle noise data into the trained rail corrugation grinding auxiliary decision model to obtain rail grinding suggestions; Data visualization module, used to manage in-vehicle noise data and visualize rail corrugation grinding to assist decision making.

10. An electronic device comprising a memory and a processor, the processor and the memory communicate with each other, the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute an auxiliary decision method for rail corrugation grinding according to any one of claims 1 to 8.

Images