An auxiliary decision method and system for rail corrugation grinding

Abstract

The application discloses an auxiliary decision-making method and system for rail corrugation grinding, and the method comprises the following steps: S1, acquiring in-vehicle noise data: collecting real in-vehicle noise signals generated in the process of train operation, wherein the real in-vehicle noise signals are collected by a collection system composed of a sound pressure sensor and a data collector; S2, acquiring train running speed and rail corrugation data; S3, preprocessing in-vehicle noise data; S4, pre-training in-vehicle noise data; S5, constructing data labels for rail corrugation grinding; S6, fine-tuning a self-supervised contrast learning model by using rail corrugation grinding label data; and S7, inputting in-situ measured in-vehicle noise data into the final rail corrugation grinding auxiliary decision-making model to obtain rail grinding suggestions. The method can give rail corrugation grinding suggestions according to in-vehicle noise, and has the advantages of high efficiency, intelligence, active maintenance and the like.

Description

Technical Field

[0001] This invention relates to the field of railway engineering management, specifically to an auxiliary decision-making method and system for rail corrugation grinding. Background Technology

[0002] With the increasing scale and capacity of the railway network, various track defects are inevitable, affecting driving safety and passenger comfort, and causing problems such as vibration and noise. Rail corrugation, as one of the most common damage phenomena in wheel-rail systems, manifests as periodic, wavy vertical irregularities on the longitudinal surface of the rail. It is a direct source of wheel-rail noise and train interior noise, and one of the main causes of damage to vehicle suspension components and breakage of rail fasteners. Especially in the subway sector, most passenger complaints about vibration or noise are related to rail corrugation. The mechanism of rail corrugation is extremely complex and remains inconclusive. Currently, the most common remedial measure is rail grinding or milling. Depending on the corrugation condition of the rail surface, large grinding or milling machines are used to remove the corrugations according to relevant standards.

[0003] Controlling rail surface roughness is a crucial engineering issue. Keeping wheel-rail contact within reasonable limits is essential for the stable operation of the wheel-rail system. Rail corrugation is an unavoidable problem in rail surface roughness control; however, current maintenance standards vary among metro operators. EN 13231-3:2012 specifies the permissible limits for short-wave irregularities remaining on the rail after railhead finishing and grinding. It also explicitly states that this standard is not applicable to acoustic track finishing. Another international standard, ISO 3095:2013, determines rail corrugation limits based on vehicle noise, typically providing a 1 / 3 octave band spectrum of rail surface roughness. However, ISO 3095's primary purpose is to obtain repeatable and comparable emitted noise levels from different vehicles, and since it is based on ballasted track, it is not suitable for noise control inside metro trains operating on slab tracks in tunnels. Based on field investigations, ISO 3095 may be overly stringent and increase economic costs. In the subway industry, the current polishing and acceptance standards are still in a state of flux.

[0004] Currently, subway maintenance largely relies on a combination of passenger complaints and periodic plans to schedule rail grinding. However, this approach is a reactive strategy, and the grinding acceptance standards do not consider in-vehicle noise levels. This leads to excessively frequent grinding of rails in certain sections, resulting in premature rail replacements, increased maintenance costs, and hindering the green development of rail transit. Therefore, there is an urgent need to develop a rail grinding auxiliary decision-making method and system based on in-vehicle noise levels. This method would provide rail grinding decisions based on in-vehicle noise conditions, thereby improving the intelligence level of railway maintenance and enabling the scientific and rational formulation of maintenance plans. 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 this invention is: an auxiliary decision-making method for rail corrugation grinding, comprising:

[0007] S1. Acquire in-vehicle noise data: Collect real in-vehicle noise signals generated during train operation. The real in-vehicle noise signals are collected by a collection system consisting of a sound pressure sensor and a data acquisition instrument.

[0008] S2. Acquire train speed and rail corrugation data: Collect train speed data and actual rail corrugation data during train operation. The actual rail corrugation data is collected by the corrugation trolley, including wavelength and amplitude.

[0009] S3. Preprocessing in-vehicle noise data: Remove in-vehicle noise data during train start-up, braking, and when the train is stopped at the station; determine the filtering range according to the distribution range of rail corrugation wavelengths of different types of railways, and 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 dataset; perform random data augmentation on the in-vehicle noise dataset to obtain an augmented in-vehicle noise dataset; among which, random data augmentation includes adding Gaussian noise, random frequency domain masking, time reversal, and numerical inversion;

[0010]

[0011]

[0012] In the formula v max and v min These are the maximum and minimum speeds of the trains operating in this section, λ and λ are respectively. max and λ min These are the maximum and minimum wavelengths of rail corrugation in this section, F. max and F minThese are the upper and lower limits of the bandpass filter, respectively. "filter" refers to the FIR digital bandpass filter. "x(t)" represents the original in-vehicle noise data, and "f(t)" represents the filtered in-vehicle noise data.

[0013] S4. Pre-trained in-vehicle noise data: The self-supervised contrastive learning model is pre-trained on the enhanced in-vehicle noise dataset to extract low-dimensional embedding features from the high-dimensional in-vehicle 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 Predict future time steps using embedding vectors, as shown in formula (5);

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

[0018] c t =g ar (z ≤t (5)

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

[0020] S5. Construct data tags 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-vehicle noise dataset are used as pre-training parameters.

[0023] A classification head model is constructed and combined with a frozen self-supervised contrastive learning model to form a fine-tuned model;

[0024] The parameters of the fine-tuned model were obtained by training on the labeled data training set of rail corrugation grinding.

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

[0026] S7. In the final rail corrugation grinding auxiliary decision model, input the actual measured in-vehicle noise data to obtain rail grinding suggestions.

[0027] Preferably, in step S4,

[0028] The improved frequency-domain enhanced skip-connected convolution module includes frequency-domain enhanced and skip-connected convolutional neural networks;

[0029] Among them, frequency domain enhancement is obtained by performing a fast Fourier transform on the in-vehicle noise to obtain frequency domain data, as shown in formula (6); only the real number domain results are 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 frequency domain enhancement data are concatenated together as the input vector.

[0030]

[0031] In the formula, 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] A skip-connected convolutional neural network consists of an input layer, a convolutional layer, an activation layer, and an output layer connected in sequence, as well as intermediate layers with skip connections; where a skip connection is to directly add the output of one 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 is composed of a multi-head attention layer, a layer standardization layer, and a fully connected layer connected in sequence. 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] In the formula, Q is the query matrix, K is the key matrix, V is the corresponding value matrix, and d k 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] In the formula W i QW i Q and W i Q These are the linear projection matrices corresponding to Q, K, and V, respectively. Q, K, and V are projected onto different subspaces, and then the attention score is calculated in each subspace; W O It is a matrix that is a linear projection of all attention calculation results, where h refers to the number of heads;

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

[0041] In the formula, 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] Based on the train 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);

[0044]

[0045] In the formula L r This refers to the surface roughness value of the rail, expressed in dB; r rms r0 is the effective value of the corrugation amplitude on the rail surface, 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 derived by using the in-vehicle noise limit.

[0047] The labels for the in-vehicle noise dataset are defined based on the rail roughness limits.

[0048] The label data for rail corrugation grinding was divided into a training set and a validation set.

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

[0050] Gaussian noise is added to the in-vehicle noise dataset with the first probability. The magnitude of the Gaussian noise is controlled by the signal-to-noise ratio, as shown in formula (11).

[0051]

[0052] In the formula, SNR is the signal-to-noise ratio, and P s It is the power of the signal, P n It is the power of the noise;

[0053] The frequency band of the in-vehicle noise dataset 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] Time reversal is performed on the in-vehicle noise dataset with the second probability;

[0056] The in-vehicle noise dataset is numerically flipped with a second probability.

[0057] Preferably, in S4, a self-supervised contrastive learning model is pre-trained on the enhanced in-vehicle noise dataset, 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, defined as in 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 The momentum of the gradient at times t and t-1 are respectively; s t s t-1 These are the second moments of the gradients at times t and t-1, respectively; β1 and β2 are non-negative weighting parameters; g t Let represent the gradient at time t;

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

[0063]

[0064] In the formula W k It is a linear weight matrix, that is, for c t Predict future observations x using a simple linear transformation. t+k ;

[0065] The model is regularized using weight decay and gradient clipping, and an early stopping mechanism is used to prevent overfitting. The iteration stops when the validation set loss no longer decreases within a preset number of steps.

[0066] The input to the frequency domain enhanced skip convolution module is a high-dimensional vector concatenated with the in-vehicle noise data and the frequency domain enhanced data, and the output is a low-dimensional embedding vector z.

[0067] The input to 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 an in-vehicle acoustic cavity finite element model. The train body finite element model includes the underframe, side walls, end walls, roof plate and floor components. The tunnel finite element model is connected using boundary element methods to obtain the tunnel-train body finite element-boundary element model.

[0070] The model excitation mainly includes the secondary suspension force excitation and the wheel-rail noise excitation, where 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 in-vehicle noise.

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

[0072] The vehicle-track coupled dynamics model includes the car body, frame, axle boxes, flexible wheelsets, rails, and substructure. It is used to calculate the secondary suspension force, taking into account wave excitation with different wavelengths and amplitudes.

[0073] Preferably, in step S6, the classification 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 a context vector c encoded by a pre-trained self-supervised contrastive learning model, and the output layer is a label corresponding to different rail corrugations.

[0076] The loss function is cross-entropy, defined as in 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 fine-tuning, 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-making system for rail corrugation grinding includes:

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

[0082] The data preprocessing module is used to perform bandpass filtering on the in-vehicle noise data and to segment the in-vehicle noise data into in-vehicle noise datasets according to the window length and step size.

[0083] The label acquisition module is used to divide the in-vehicle noise dataset into labeled and unlabeled datasets based on the rail corrugation test data.

[0084] The model pre-training module is used to build a self-supervised contrastive learning model and train it on an unlabeled in-vehicle noise dataset to obtain the optimal pre-training parameters.

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

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

[0087] The data visualization module is used to manage in-vehicle noise data and visualize the auxiliary decision-making for rail corrugation grinding.

[0088] An electronic device includes a memory and a processor, the processor and the memory communicating with each other, the memory storing program instructions executable by the processor, and the processor calling the program instructions to execute an auxiliary decision-making method for rail corrugation grinding as claimed in any one of claims 1-8.

[0089] The beneficial effects of this invention are as follows:

[0090] Compared with existing methods that use corrugation trolleys or electronic straightness gauges to measure rail corrugation, this invention directly uses in-vehicle noise data to determine the rail corrugation status, which has the advantages of low cost and high efficiency.

[0091] Compared with existing rail grinding and acceptance standards, this invention can take into account the noise amplitude inside the vehicle to give rail roughness limits, and can also give rail grinding suggestions, which is a more economical, scientific and intelligent rail maintenance strategy.

[0092] Compared with existing intelligent detection technologies for rail corrugation, this invention first performs self-supervised comparative learning on a large amount of unlabeled in-vehicle noise data, and then fine-tunes it on a custom rail corrugation grinding label dataset, thus solving the technical problem of expensive acquisition of field data sample labels. Attached Figure Description

[0093] Figure 1 A flowchart of an auxiliary decision-making method for rail corrugation grinding is provided in an embodiment of the present invention;

[0094] Figure 2 This is a schematic diagram of an in-vehicle noise field test provided by an embodiment of the present invention;

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

[0096] Figure 4 A sample data graph of applying a random data augmentation method is provided in an embodiment of the present invention;

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

[0098] Figure 6 This is a schematic diagram of multi-head attention score calculation provided in an embodiment of the present invention;

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

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

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

[0102] Figure 10 This is a schematic diagram of a steel rail corrugation grinding label provided in an embodiment of the present invention;

[0103] Figure 11 This is a schematic diagram of a fine-tuning model structure provided in an embodiment of the present invention. Detailed Implementation

[0104] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. The embodiments 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 train carriage during operation, such as... Figure 2As shown; an IEPE-type high-precision sound pressure sensor is used, and the data acquisition uses a 24AD high-precision data acquisition system with a sampling frequency of 20000Hz. The sound pressure sensor is placed in the same area as the bogie position in the same carriage as the train speed sensor, thereby ensuring spatial matching between the train's running mileage and the noise data inside the carriage.

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

[0109] Step 3: Remove in-car noise data during train start-up, braking, and stops at stations, such as... Figure 3 As shown; the filtering range is determined according to the distribution range of rail corrugation wavelength of different types of railways. The rail corrugation wavelength range of the subway is 30-300mm. The original in-vehicle noise data is bandpass filtered using FIR digital filters, as shown in formulas (1) to (3). The stopband attenuation is not less than 60dB, and the filtered in-vehicle noise data is obtained. The filtered in-vehicle noise data is sliced, with a window length of 10240 data points and 50% overlap, to obtain the in-vehicle noise dataset.

[0110]

[0111] In the formula v max and v min These are the maximum and minimum speeds of the trains operating in this section, λ and λ are respectively. max and λ min These are the maximum and minimum wavelengths of rail corrugation in this section, F. max and F min These are the upper and lower limits of the bandpass filter, respectively. "filter" refers to the FIR digital bandpass filter. "x(t)" represents the original in-vehicle noise data, and "f(t)" represents the filtered in-vehicle noise data.

[0112] Step 4: Perform random data augmentation on the in-vehicle noise dataset by adding Gaussian noise to the in-vehicle noise dataset with a certain probability. The magnitude 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] In the formula, SNR is the signal-to-noise ratio, and P s It is the power of the signal, P n It is the power of the noise;

[0115] With a certain probability, a certain frequency band in the frequency domain of the in-vehicle noise dataset is set to zero, and then the time domain data is obtained by 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; preferably, the probability is 50%.

[0118] The in-vehicle noise dataset is numerically flipped with a certain probability; preferably, the probability is 50%.

[0119] The enhanced in-vehicle 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 in formula (6);

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

[0122] In the formula z t It is the embedded vector after being encoded by the frequency domain enhanced skip convolution module;

[0123] The frequency domain enhancement jump-connected convolution module is mainly composed of two parts: frequency domain enhancement and jump-connected convolutional neural network. Frequency domain enhancement obtains frequency domain data by performing fast Fourier transform on the in-vehicle noise, as shown in formula (7); only the real number domain result 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] In the formula, 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] A skip-connected convolutional neural network consists of an input layer, a convolutional layer, an activation layer, and an output layer connected in sequence, as well as intermediate layers with skip connections; where a skip connection is to directly add the output of one layer in the convolutional neural network to the output of a subsequent layer.

[0127] Step 6: Construct the 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] In the formula c t It is the context vector output by the multi-head attention autoregressive module;

[0130] The multi-head attention autoregressive module consists of a multi-head attention layer, a layer normalization layer, and a fully connected layer connected sequentially. The attention score is determined by formulas (9) to (10), such as... Figure 6 As shown, the output of the fully connected layer is determined by formula (11);

[0131]

[0132] In the formula, Q is the query matrix, K is the key matrix, V is the corresponding value matrix, and d k 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] In the formula W i Q W i Q and W i Q These are the linear projection matrices corresponding to Q, K, and V, respectively. Q, K, and V are projected onto different subspaces, and then the attention score is calculated in each subspace; W O It is a matrix that is a linear projection of all attention calculation results, where h refers to the number of heads;

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

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

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

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

[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 The momentum of the gradient at times t and t-1 are respectively; s t s t-1 β1 and β2 are the second moments of the gradient at times t and t-1, respectively; β1 and β2 are non-negative weighting parameters, typically taken as 0.9 and 0.999, respectively; g t This represents the gradient at time t.

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

[0144]

[0145] In the formula W k It is a linear weight matrix, that is, for c t Predict future observations x using a simple linear transformation. t+k ;

[0146] The model is regularized using weight decay and gradient clipping, and an early stopping mechanism is used to prevent overfitting. The iteration stops when the validation set loss no longer decreases within a preset number of steps.

[0147] Step 9: Establish a vehicle-track rigid-flexible coupling analysis model. The model topology diagram is as follows: Figure 9As shown, the model mainly consists of a car body, frame, axle boxes, wheelsets, rails, and substructure. The secondary suspension connecting the car body and frame and the primary suspension connecting the frame and axle boxes are simulated using linear spring force elements. The axle boxes and wheelsets are connected by hinges with only rotational freedom around the wheel axle. Fasteners are simulated using rushing force elements. The car body, frame, axle boxes, and the geodetic coordinate system are connected by hinges with 6 degrees of freedom.

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

[0149]

[0150] In the formula, E is the elastic modulus; δ is the penetration amount; μ is Poisson's ratio; x l (y) represents the leading edge boundary of the wheel-rail contact patch along the longitudinal axis; ε is a correction factor, taken as 0.55; R W The radius of the wheel's rolling circle;

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

[0152] Table 1 Parameters of the 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 Level 5 spectrum to simulate medium- and long-wave track irregularities, with a safety factor of 0.25 and a wavelength range of 3–80 m. The definition of the American Level 5 spectrum is as follows:

[0155]

[0156] In the formula, S v (φ) represents the power spectral density of track irregularities, in m³ / s. 2 / (1 / m); φ is the spatial frequency, in units of 1 / m; Av is the roughness constant, in units of m; φ v1 φ v2 The cutoff frequency is expressed in units of 1 / m.

[0157] The Sato spectrum was used to simulate shortwave orbit irregularities, with a wavelength range of 0.01–3 m. The Sato spectrum was defined as follows:

[0158]

[0159] In the formula, Ω is the spatial frequency, with units of 1 / m; A is the wheel-rail surface roughness coefficient, with a value of 4.15 × 10⁻⁶. -8 ~5.0×10 -7 ;

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

[0161]

[0162] In the formula, Z cor (t) represents the amplitude of the corrugation unevenness; a represents the depth of the corrugation; λ represents 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 an in-vehicle noise prediction model. The prediction model mainly consists of a train body finite element model and an in-vehicle acoustic cavity finite element model composed of components such as the underframe, side walls, end walls, roof, and floor. The tunnel finite element model is connected using boundary element methods to obtain the tunnel-vehicle body finite element-boundary element model.

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

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

[0167] Multiple calculations were performed to infer the rail roughness limit based on the noise level inside the vehicle.

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

[0169]

[0170] In the formula L r This refers to the surface roughness value of the rail, expressed in dB; r rms r0 is the effective value of the corrugation amplitude on the rail surface, in μm; r0 is a reference value, generally taken as 1 μm.

[0171] The labels corresponding to the in-vehicle noise dataset are defined based on the aforementioned rail roughness limits, such as... Figure 10 As shown,

[0172] Based on the rail roughness limit definition, the labels corresponding to the in-vehicle noise dataset are defined. The label format is {no corrugation *no grinding required; no corrugation *grinding required; corrugation with wavelength 30-100mm *grinding required; corrugation with wavelength 30-100mm *no grinding required; corrugation with wavelength 100-300mm *grinding required; corrugation with wavelength 100-300mm *no grinding required; corrugation with wavelengths of 30-100mm and 100-300mm *grinding required; corrugation with wavelengths of 30-100mm and 100-300mm **no grinding required}, 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: Construct the 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 a context vector c encoded by a pre-trained self-supervised contrastive learning model, and the output layer is the label corresponding to the eight types of rail corrugation grinding.

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

[0177] H(p,q)=-∑ 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: Fine-tune the self-supervised contrastive learning model using rail corrugation grinding label data, such as... Figure 11 As shown, 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] First, the optimal parameters of the self-supervised contrastive learning model obtained on the in-vehicle noise dataset are used as pre-training parameters.

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

[0182] Finally, the model was validated on the label data validation set of rail corrugation grinding, and the optimal fine-tuning parameters were selected to obtain the final rail corrugation grinding auxiliary decision model.

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

[0184] In S6, during fine-tuning, the network weight parameters of the pre-trained model are no longer updated; only the parameters of the classification head model are updated. This freezing is because the pre-trained model has already learned a suitable set of parameters, so only the parameters of the classification head need to be fine-tuned. If the parameters weren't frozen, it would be equivalent to searching for suitable parameters again, which would waste computational resources and fail to demonstrate the advantages of pre-training.

[0185] Example 2

[0186] Based on Example 1, Example 2 provides an auxiliary decision-making system for rail corrugation grinding. This auxiliary decision-making system corresponds to the aforementioned auxiliary decision-making method for rail corrugation grinding, and specifically includes:

[0187] The data acquisition module is used to acquire in-vehicle noise data, which is the sound pressure change data inside the carriage during train operation.

[0188] The data preprocessing module is used to perform bandpass filtering on the in-vehicle noise data and to segment the in-vehicle noise data into in-vehicle noise datasets according to the window length and step size.

[0189] The label acquisition module is used to divide the in-vehicle noise dataset into labeled and unlabeled datasets based on the rail corrugation test data.

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

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

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

[0193] The data visualization module is used to manage in-vehicle noise data and visualize the auxiliary decision-making for rail corrugation grinding.

[0194] For specific details, please refer to the description in the section on auxiliary decision-making methods 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, which communicate with each other. The memory stores program instructions that can be executed by the processor. The processor calls the program instructions to execute an auxiliary decision-making method for rail corrugation grinding. The auxiliary decision-making method for rail corrugation grinding includes the following steps:

[0197] Acquire in-vehicle noise data, which is the sound pressure data inside the carriage generated during train operation;

[0198] Train speed data and rail corrugation data are acquired, wherein the rail corrugation data is measured by a corrugation trolley;

[0199] Preprocessing and random data augmentation of in-vehicle noise data yields an in-vehicle noise dataset.

[0200] The in-vehicle noise prediction model is used to determine the rail roughness limit. Based on the rail roughness limit, the in-vehicle noise dataset is divided into an unlabeled dataset and a labeled dataset.

[0201] A self-supervised contrastive learning model is constructed and pre-trained on the unlabeled dataset to obtain the pre-trained model parameters.

[0202] A classification head model is constructed and fine-tuned 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 grinding auxiliary decision model to obtain rail grinding suggestions.

[0204] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. An auxiliary decision-making method for grinding and polishing rail corrugations, characterized in that, include: S1. Acquire in-vehicle noise data: Collect real in-vehicle noise signals generated during train operation. The real in-vehicle noise signals are collected by a collection system consisting of a sound pressure sensor and a data acquisition instrument. S2. Acquire train speed and rail corrugation data: Collect train speed data and actual rail corrugation data during train operation. The actual rail corrugation data is collected by the corrugation trolley, including wavelength and amplitude. S3. Preprocessing in-vehicle noise data: Remove in-vehicle noise data when the train starts and brakes and stops at the station; determine the filtering range according to the distribution range of rail corrugation wavelength of different types of railways, and use digital filters to perform bandpass filtering on the original in-vehicle noise data to obtain 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 dataset; random data augmentation is then performed on the in-vehicle noise dataset to obtain an augmented in-vehicle noise dataset; the random data augmentation includes adding Gaussian noise, random frequency domain masking, time reversal, and numerical flipping. In the formula v max and v min These are the maximum and minimum speeds of the trains operating in this section, λ and λ are respectively. max and λ min These are the maximum and minimum wavelengths of rail corrugation in this section, F. max and F min These are the upper and lower limits of the bandpass filter, respectively. "filter" refers to the FIR digital bandpass filter. "x(t)" represents the original in-vehicle noise data, and "f(t)" represents the filtered in-vehicle noise data. S4. Pre-trained in-vehicle noise data: The self-supervised contrastive learning model is pre-trained on the enhanced in-vehicle noise dataset to extract low-dimensional embedding features from the high-dimensional in-vehicle 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 Predict future time steps using embedding vectors, as shown in formula (5); z t =g enc (x t )(4) c t =g ar (z ≤t )(5) In the formula z t It is the embedding vector after being encoded by the frequency domain enhanced skip convolution module, c t It is the context vector output by the multi-head attention autoregressive module; S5. Construct data tags 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-vehicle noise dataset are used as pre-training parameters. A classification head model is constructed and combined with a frozen self-supervised contrastive learning model to form a fine-tuned model; The parameters of the fine-tuned model were obtained by training on the labeled data training set of rail corrugation grinding. The model was validated on the label data validation set of rail corrugation grinding, and the optimal fine-tuning parameters were selected to obtain the final rail corrugation grinding auxiliary decision model. S7. In the final rail corrugation grinding auxiliary decision model, input the actual measured in-vehicle noise data to obtain rail grinding suggestions.

2. The auxiliary decision-making method for rail corrugation grinding according to claim 1, characterized in that, In step S4, The improved frequency-domain enhanced skip-connected convolution module includes frequency-domain enhanced and skip-connected convolutional neural networks; Among them, frequency domain enhancement is obtained by performing a fast Fourier transform on the in-vehicle noise to obtain frequency domain data, as shown in formula (6); only the real number domain results are 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 frequency domain enhancement data are concatenated together as the input vector. In the formula, 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; A skip-connected convolutional neural network consists of an input layer, a convolutional layer, an activation layer, and an output layer connected in sequence, as well as intermediate layers with skip connections; where a skip connection is to directly add the output of one layer in the convolutional neural network to the output of a subsequent layer.

3. The auxiliary decision-making 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 standardization layer, and a fully connected layer connected in sequence. The attention score is determined by formulas (7) to (8), and the output of the fully connected layer is determined by formula (9). In the formula, Q is the query matrix, K is the key matrix, V is the corresponding value matrix, and d k 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) In the formula W i Q W i Q and W i Q These are the linear projection matrices corresponding to Q, K, and V, respectively. Q, K, and V are projected onto different subspaces, and then the attention score is calculated in each subspace; W O It is a matrix that is a linear projection of all attention calculation results, where h refers to the number of heads; FFN(x)=max(0,xW1+b1)W2+b2 (9) In the formula, 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. The auxiliary decision-making method for rail corrugation grinding according to claim 1, characterized in that, In step S5, Based on the train 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); In the formula L r This refers to the surface roughness value of the rail, expressed in dB; r rms r0 is the effective value of the corrugation amplitude on the rail surface, 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 derived by using the in-vehicle noise limit. The labels for the in-vehicle noise dataset are defined based on the rail roughness limits. The label data for rail corrugation grinding was divided into a training set and a validation set.

5. A method for auxiliary decision-making in rail corrugation grinding according to claim 1, characterized in that, In step S3, random data augmentation is performed on the in-vehicle noise dataset, specifically including: Gaussian noise is added to the in-vehicle noise dataset with the first probability. The magnitude of the Gaussian noise is controlled by the signal-to-noise ratio, as shown in formula (11). In the formula, SNR is the signal-to-noise ratio, and P s It is the power of the signal, P n It is the power of the noise; The frequency band of the in-vehicle noise dataset 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). Time reversal is performed on the in-vehicle noise dataset with the second probability; The in-vehicle noise dataset is numerically flipped with a second probability.

6. A method for auxiliary decision-making in rail corrugation grinding according to claim 1, characterized in that, In step S4, a self-supervised contrastive learning model is pre-trained 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, defined as in 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 The momentum of the gradient at times t and t-1 are respectively; s t s t-1 These are the second moments of the gradients at times t and t-1, respectively; β1 and β2 are non-negative weighting parameters; g t Let represent the gradient at time t; The loss function is infoNCE, defined as in formulas (14) to (15); In the formula W k It is a linear weight matrix, that is, for c t Predict future observations x using a simple linear transformation. t+k ; The model is regularized using weight decay and gradient clipping, and an early stopping mechanism is used to prevent overfitting. The iteration stops when the validation set loss no longer decreases within a preset number of steps. The input to the frequency domain enhanced skip convolution module is a high-dimensional vector concatenated with the in-vehicle noise data and the frequency domain enhanced data, and the output is a low-dimensional embedding vector z. The input to 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 method for auxiliary decision-making in 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 an in-vehicle acoustic cavity finite element model. The train body finite element model includes the underframe, side walls, end walls, roof plate and floor components. The tunnel finite element model is connected using boundary element methods to obtain the tunnel-train body finite element-boundary element model. The model excitation mainly includes the secondary suspension force excitation and the wheel-rail noise excitation, where 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 in-vehicle noise. The secondary suspension force excitation is calculated by considering rail corrugation excitation with different roughnesses in the vehicle-track rigid-flexible coupling dynamic model. The vehicle-track coupled dynamics model includes the car body, frame, axle boxes, flexible wheelsets, rails, and substructure. It is used to calculate the secondary suspension force, taking into account wave excitation with different wavelengths and amplitudes.

8. A method for auxiliary decision-making in 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 a context vector c encoded by a pre-trained self-supervised contrastive learning model, and the output layer is a label corresponding to different rail corrugations. The loss function is cross-entropy, defined as in 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 fine-tuning, 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 grinding and polishing rail corrugations, characterized in that, include: The data acquisition module is used to acquire in-vehicle noise data, which is the sound pressure change data inside the carriage during train operation. The data preprocessing module is used to perform bandpass filtering on the in-vehicle noise data and to segment the in-vehicle noise data into in-vehicle noise datasets according to the window length and step size. The label acquisition module is used to divide the in-vehicle noise dataset into labeled and unlabeled datasets based on the rail corrugation test data. The model pre-training module is used to build a self-supervised contrastive learning model and train it on an unlabeled in-vehicle noise dataset to obtain the optimal pre-training parameters. The model fine-tuning module is used to build a classification head model and train and validate it on a labeled in-vehicle noise dataset to obtain the final rail corrugation grinding auxiliary decision model. The decision module is used to input in-vehicle noise data into the trained rail corrugation grinding auxiliary decision model to obtain rail grinding suggestions; The data visualization module is used to manage in-vehicle noise data and visualize the auxiliary decision-making for rail corrugation grinding.

10. An electronic device comprising a memory and a processor, the processor and the memory communicating with each other, the memory storing program instructions executable by the processor, the processor calling the program instructions to execute an auxiliary decision-making method for rail corrugation grinding as claimed in any one of claims 1-8.

Images