Train traction motor-gearbox coupling noise prediction method and system based on vibration small sample data

Abstract

The invention discloses a train traction motor-gearbox coupling noise prediction method and system based on vibration small sample data. The method comprises the steps that high correlation between coupling noise and vibration signals is jointly analyzed according to a square coherence-regression coefficient; establishing a mapping relation between the vibration signal and the coupling noise through a feedforward neural network; searching optimal hidden layer number, node number and learning rate configuration based on a Bayesian optimization algorithm of a tree structure, and extracting a mapping relation between vibration data and coupled noise data; and the attention mechanism calculates the weight of the feature data, constructs a Bayesian-attention-feedforward neural network model, performs enhancement and updating through iteration, and outputs a prediction result. According to the method, the Bayesian-attention-feedforward neural network coupling noise prediction model is constructed, the coupling noise is accurately predicted through a small number of vibration signals, and the technical problems that train traction motor-gearbox coupling noise prediction based on small sample data is poor in stability and low in accuracy are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of noise prediction and design of rail vehicle traction drive systems, and more specifically, to a method and system for predicting train traction motor-gearbox coupling noise based on small vibration sample data. Background Art

[0002] The coupling noise between the traction motor and the gearbox is the main source of noise in rail vehicles. The strong coupling between the two exacerbates the noise, making it extremely difficult to control. At the same time, accurately predicting the coupling noise between the traction motor and the gearbox and considering its related characteristics has always been a difficult point in the train noise design process. There is a certain correlation between the vibration signal and the coupling noise, and this correlation is the basis for guiding the entire train noise design. Therefore, it is necessary to conduct a comprehensive study on the coupling noise of the traction motor and the gearbox, and develop an intelligent prediction method and system for coupling noise under small vibration sample data.

[0003] The existing solutions mainly use hardware devices and software algorithms to achieve noise prediction of traction drive systems. Hardware devices mainly include sensors, data collectors, etc., which are used to collect the operating data of the traction drive system. Software algorithms mainly include signal processing, feature extraction, pattern recognition, etc., which are used to analyze and process the collected data to achieve noise control and fault diagnosis. Although this technology can be processed simply and quickly, there are still some problems in practical applications. First, the geometric environment in which the traction motor and gearbox are located is complex, with multiple sources of interference, and the available noise samples collected are small and incomplete. Secondly, the existing hardware devices and software algorithms can often only predict a single type of noise, and for complex electromechanical coupling noise, it is often impossible to effectively decouple and predict. Finally, the existing technology often cannot achieve high-precision and high-efficiency performance in noise prediction, which brings certain risks to the reliability and safety of the vehicle.

[0004] CN202110154558.5 discloses a gearbox fault diagnosis method based on multi-source data fusion, the steps include: first, obtaining the vibration signal, noise signal, temperature signal at the gearbox bearing, displacement signal of the gear shaft, and oil data of the gearbox lubricant; then pre-processing the collected multi-source data respectively; finally, diagnosing the gearbox fault using multi-sensor data fusion technology. This patent obtains multi-source data signals such as vibration, temperature, noise, and displacement of the gearbox, and performs fault diagnosis on the multi-sensor data fusion technology, giving full play to the redundant and complementary functions of each sensor and data source, and can express the fault information of the gearbox more completely and clearly, improving the positioning ability of the gearbox fault and the diagnosis ability of the composite fault. However, although this method integrates multi-source signals to diagnose faults, it also has shortcomings. First, the multiple acquisitions of multi-source signals require accurate time synchronization of each sensor, and a large number of data samples are often required during the acquisition process to support subsequent diagnostic work. Secondly, this method only uses the multi-source data of the gearbox to diagnose its faults, without considering the coupling effects of other electromechanical devices. Finally, the method uses the traditional back propagation (BP) neural network to extract features. However, the traditional back propagation neural network used in the patent for feature extraction has a simple structure, and hyperparameter adjustment is difficult, which increases the complexity of diagnosis and computational cost. Therefore, how to effectively predict the traction motor-gearbox coupling noise with low cost, high efficiency and high accuracy under small sample data is an urgent problem to be solved in the current development of rail vehicles. Summary of the invention

[0005] The main technical problem to be solved by the present invention is that in view of the shortcomings of the prior art that reliable data of traction motor-gearbox is difficult to obtain, electromechanical coupling noise is difficult to decouple, and the accuracy of noise prediction model is difficult to improve, the present invention provides a train traction motor-gearbox coupling noise prediction method based on small vibration sample data, which utilizes the strong correlation between coupling noise and vibration signal and the excellent decoupling ability of Bayesian-attention-feedforward neural network model, and uses easily obtained vibration signals to predict traction motor-gearbox coupling noise, thereby realizing accurate prediction of complex electromechanical coupling noise in different spatial positions, and providing an important prerequisite foundation and technology for the forward design of vehicle noise.

[0006] Another technical problem solved by the present invention is to provide a train traction motor-gearbox coupling noise prediction system based on small vibration sample data.

[0007] The purpose of the present invention is achieved through the following technical solutions:

[0008] A train traction motor-gearbox coupling noise prediction method based on small vibration sample data, comprising the following steps:

[0009] S1. Jointly analyze the high correlation between coupled noise and vibration signals based on the squared coherence-regression coefficient and convert the complex data into a two-dimensional matrix in the frequency domain

[0010] S2. Learning 2D matrices in a multi-task learning framework via a feedforward neural network Based on the characteristics, the mapping relationship between the vibration signals in the horizontal, vertical and axial directions and the coupled noise is established;

[0011] S3. Based on the tree-structured Bayesian optimization algorithm, find the optimal number of hidden layers, number of nodes and learning rate configuration of the feedforward neural network, and extract the mapping relationship between vibration data and coupled noise data;

[0012] S31. Select the number of hidden layers L, the number of hidden layer nodes N, and the learning rate η of the feedforward neural network as the hyperparameter optimization target of the feedforward neural network, define the hyperparameter set of the feedforward neural network as x, define the objective function value y for updating the hyperparameter set x as: y=f(x)=f(L,N,η), define the configuration space as H={[L,N,η]1,[L,N,η] 2 , …, [L, N, η] n};

[0013] S32. Calculate the probability density function P(x|y) of the hyperparameter of a given objective function value according to the current hyperparameter sample point, and construct a Bayesian optimization model based on a tree structure;

[0014] S33. Using expected improvement as acquisition function EI y* (x), identify the next sample point with the highest expected value of acquisition;

[0015] S34. When the acquisition function EI y* When (x) reaches the maximum value, the hyperparameter set is considered to be the optimal hyperparameter set x*, expressed as:

[0016]

[0017] Where L* is the optimal number of hidden layers, N* is the optimal number of hidden layer nodes, and η* is the optimal feedforward neural network learning rate;

[0018] S35. Feedback the optimal hyperparameter set x* to the feedforward neural network for further training and optimization;

[0019] S4. The attention mechanism calculates the weights of key feature data and constructs a Bayesian-attention-feedforward neural network model to accurately capture the characteristics of coupled noise;

[0020] S41. The attention mechanism is used in the feedforward neural network to calculate the characteristic relationship between the vibration signal and the coupling noise in the input data, and a Bayesian-attention-feedforward neural network model is constructed to accurately capture the coupling noise characteristics. The attention mechanism calculates the characteristic s i The process can be expressed as:

[0021]

[0022] Where tanh is the activation function, is the weight of the attention mechanism, b is the bias, is the output of the i-th hidden layer;

[0023] S42. i Convert to exponential form and calculate the current s i The ratio of the sum of all attention score indices to obtain the attention score β i , the result of the attention score function β i It can be expressed as:

[0024]

[0025] Where n is the number of input feature data;

[0026] S5. The parameters and weight coefficients of the Bayesian-attention-feedforward neural network model are enhanced and updated through iteration, and the prediction result of the traction motor-gearbox coupling noise is output, which is expressed as:

[0027]

[0028] In the formula is the coupled noise sound pressure value predicted by the Bayesian optimization-attention mechanism-feedforward neural network model, ReLu is the activation function, w and b' are the weight and bias of the output layer respectively;

[0029] S6. Perform signal anomaly detection based on the coupled noise prediction results, and predict multiple related attributes of the traction motor-gearbox, including fault type, fault location, and severity.

[0030] Further, the squared coherence-regression coefficient is expressed as:

[0031]

[0032] Where P Q_N (f) is the cross power spectrum density of the vibration signal Q and the coupled noise signal N, P Q_Q (f) and P N_N (f) are the auto-power spectral densities of the vibration signal Q and the coupled noise signal N, respectively. i is the ith vibration amplitude, Ni is the sound pressure value of the ith coupling noise, is the average sound pressure of the coupling noise, τ1+τ2=1.

[0033] Furthermore, the two-dimensional matrix Including time T, frequency F, vibration signal V1 in the horizontal direction, vibration signal V2 in the vertical direction, vibration signal V3 in the axial direction, two-dimensional matrix It is expressed as:

[0034]

[0035] Among them, λ1~λ5 are the data sets of T, F, V1, V2, and V3 respectively, and the superscripts 1~n are the input data sets.

[0036] Furthermore, the forward propagation process of the feedforward neural network is expressed as:

[0037]

[0038] In the formula, a i represents the output of the i-th hidden layer, represents the input of the i-th layer in the hidden layer, W ij and b i They represent the weight matrix and bias of the i-th layer in the hidden layer respectively, q represents the activation function of the hidden layer, L is the number of hidden layers of the feedforward neural network, and N is the number of hidden layer nodes of the feedforward neural network.

[0039] Furthermore, the probability density function p(x|y) of the hyperparameter is expressed as:

[0040]

[0041] In the formula, y* represents the threshold, l(x) and g(x) represent y is less than or greater than or equal to y. * The density estimate of .

[0042] Furthermore, the acquisition function EI y* (x) is expressed as:

[0043]

[0044] p(y|x) is the posterior probability, expressed as:

[0045]

[0046] p(x) is the marginal likelihood function, and p(y) is the distribution of the target function value.

[0047] Furthermore, after obtaining the optimal hyperparameter set x*, the forward propagation process of the feedforward neural network is updated as follows:

[0048]

[0049] Furthermore, the loss function of the forward propagation of the feedforward neural network is LOSS, and the formula for updating the loss function parameters is expressed as:

[0050]

[0051] In the formula Represents the weight of the jth node in the i-th hidden layer of the updated feedforward neural network.

[0052] Furthermore, signal anomaly detection is performed on the traction motor-gearbox coupling noise prediction results. When the predicted coupling noise is an abnormal signal, multi-task learning is performed to simultaneously predict the fault type, fault location, and severity of the traction motor-gearbox. Targeted maintenance is performed based on faults of different types, locations, and severities to achieve intelligent early warning.

[0053] A train traction motor-gearbox coupling noise prediction system based on small vibration sample data, comprising:

[0054] Raw data collection and analysis module, collects raw data and performs data processing and analysis;

[0055] The feedforward neural network module includes 1 input layer, L hidden layers and 1 output layer, which learns the characteristics of the input data and establishes the mapping relationship between the vibration signal and the coupled noise;

[0056] The Bayesian optimization module optimizes the hyperparameters of the feedforward neural network model based on the Tree-structured ParzenEstimator, finds the optimal hyperparameter set, and extracts the mapping relationship between vibration data and coupled noise data;

[0057] The attention mechanism module calculates the weights of key feature data and accurately captures the characteristics of coupled noise;

[0058] The multi-dimensional visualization module visualizes the coupled noise sound pressure values ​​predicted by the vibration data at different sensor positions, including two-dimensional visualization, three-dimensional visualization, and modal visualization, and clearly shows the fitting between the two.

[0059] Compared with the prior art, the beneficial effects are:

[0060] The present invention analyzes the vibration data and coupling noise data of the train traction motor-gearbox and extracts key features. A feedforward neural network model is established using the correlation between the vibration data and the coupling noise data. The coupling noise generated by the traction motor and the gearbox at different spatial positions is accurately predicted through the easily accessible vibration signal. In the feedforward neural network, Bayesian optimization is used to pre-train the hyperparameter selection of the neural network to approximate the optimal hyperparameter. An attention mechanism is introduced in the data training process to further focus on and calculate the characteristics of complex noise, thereby achieving accurate noise prediction.

[0061] The method of the present invention improves the accuracy of coupled noise prediction under small sample data sets of vibration, and uses time-frequency domain vibration signals to predict coupled noise of key train components, providing an effective technical approach for early control and reduction of train noise. In addition, the present invention provides an important tool for evaluating noise performance in the selection of train components, design of the entire train, and implementation of manufacturing. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 A flow chart of a method for predicting train traction motor-gearbox coupling noise based on small vibration sample data provided in Example 1.

[0063] Figure 2 A diagram of the structure of a train traction motor-gearbox coupling noise prediction system based on small vibration sample data provided in Example 2.

[0064] Figure 3 A block diagram of a train traction motor-gearbox coupling noise prediction system based on small vibration sample data provided in Example 2.

[0065] Figure 4 This is a schematic diagram of experimental data provided for Example 3.

[0066] Figure 5 Schematic diagram of the train traction motor-gearbox coupling noise prediction results based on small vibration sample data provided in Example 3.

[0067] Figure 6 Schematic diagram of train traction motor-gearbox coupling noise prediction error based on small vibration sample data provided in Example 3. DETAILED DESCRIPTION

[0068] The present invention will be further explained and illustrated below in conjunction with the embodiments, but the specific embodiments do not limit the present invention in any form.

[0069] Example 1

[0070] This embodiment provides a train traction motor-gearbox coupling noise prediction method based on small vibration sample data, such as Figure 1, the steps include:

[0071] A train traction motor-gearbox coupling noise prediction method based on small vibration sample data, comprising the following steps:

[0072] S1. Jointly analyze the high correlation between coupled noise and vibration signals based on the squared coherence-regression coefficient and convert the complex data into a two-dimensional matrix in the frequency domain

[0073] S11. Use the square coherence-regression coefficient to perform coherence analysis on the processed vibration and coupled noise sample data, and determine the vibration signal as the main correlation parameter for subsequent coupled noise prediction. The square coherence-regression coefficient is expressed as:

[0074]

[0075] Where P Q_N (f) is the cross power spectrum density of the vibration signal Q and the coupled noise signal N, P Q_Q (f) and P N_N (f) are the auto-power spectral densities of the vibration signal Q and the coupled noise signal N, respectively. i is the ith vibration amplitude, N i is the sound pressure value of the ith coupling noise, is the average sound pressure of the coupling noise, τ1+τ2=1.

[0076] S12. After the squared coherence-regression coefficient joint analysis, the original vibration signal was preliminarily screened and M Q_N High coherence vibration data greater than 0.8, converting the coupled noise signal and vibration signal into a two-dimensional matrix in the frequency domain Including time T, frequency F, vibration signal V1 in the horizontal direction, vibration signal V2 in the vertical direction, vibration signal V3 in the axial direction, two-dimensional matrix It is expressed as:

[0077]

[0078] Among them, λ1~λ5 are the data sets of T, F, V1, V2, and V3 respectively, and the superscripts 1~n are the input data sets.

[0079] S2. Learning 2D matrices in a multi-task learning framework via a feedforward neural network The mapping relationship between the vibration signals in the horizontal, vertical and axial directions and the coupled noise is established based on the characteristics. The forward propagation process of the feedforward neural network is expressed as:

[0080]

[0081] In the formula, ai represents the output of the $i$-th layer of the hidden layer, represents the input of the $i$-th layer in the hidden layer, $W$ ij and $b$ i respectively represent the weight matrix and bias of the $i$-th layer in the hidden layer, $q$ represents the activation function of the hidden layer, $L$ is the number of hidden layers of the feedforward neural network, and $N$ is the number of nodes in the hidden layer of the feedforward neural network.

[0082] S3. Use the tree-structured Bayesian optimization algorithm to find the optimal number of hidden layers, number of nodes, and learning rate configuration of the feedforward neural network, and extract the mapping relationship between vibration data and coupling noise data;

[0083] S31. Select the number of hidden layers $L$, number of nodes $N$, and learning rate $\eta$ of the feedforward neural network as the hyperparameter optimization objectives of the feedforward neural network. Define the hyperparameter set of the feedforward neural network as $x$, and define the objective function value $y$ for updating the hyperparameter set $x$ as: $y = f(x) = f(L, N, \eta)$. Define the configuration space as $H=\{[L,N,\eta]_1,[L,N,\eta]$ 2 , \ldots, [L,N,\eta]$ n \};

[0084] S32. Calculate the probability density function $P(x|y)$ of the hyperparameters for the given objective function value according to the current hyperparameter sample points, and construct a tree-structured Bayesian optimization model. The probability density function $p(x|y)$ of the hyperparameters is expressed as:

[0085]

[0086] In the formula, $y^*$ represents the threshold, and $l(x)$ and $g(x)$ respectively represent the density estimates when $y$ is less than or greater than or equal to $y$ * of.

[0087] S33. Use the expected improvement as the acquisition function $EI$ y* (x), and identify the next sample point with the highest acquisition expected value; The acquisition function $EI$ y* (x) is expressed as:

[0088]

[0089] $p(y|x)$ is the posterior probability, expressed as:

[0090]

[0091] $p(x)$ is the marginal likelihood function, and $p(y)$ is the distribution of the objective function value. Define $\gamma$ as the constructor error adjustment function, $\gamma = p(y \lt y^*)$, $p(x)=\gamma l(x)+(1 - \gamma)g(x)$, and the acquisition function $EI$ y* (x) is further expressed as:

[0092]

[0093] This formula is an iterative process of optimizing g(x) / l(x) and maximizing g(x) / l(x). y* The closer (x) is to the maximum value, the better the performance of the hyperparameter set. ∝ means that the expression on the left side of the equation is proportional to the expression on the right side.

[0094] S34. When the acquisition function EI y* When (x) reaches the maximum value, the hyperparameter set is considered to be the optimal hyperparameter set x*, expressed as:

[0095]

[0096] Where L* is the optimal number of hidden layers, N* is the optimal number of hidden layer nodes, and η* is the optimal feedforward neural network learning rate.

[0097] After obtaining the optimal hyperparameter set x*, the forward propagation process of the feedforward neural network is updated as follows:

[0098]

[0099] A loss function is established using the output value obtained by forward propagation and the actual sample label as LOSS. The formula for updating the parameters of the loss function is expressed as:

[0100]

[0101] In the formula Represents the weight of the jth node in the i-th hidden layer of the updated feedforward neural network.

[0102] S4. The attention mechanism calculates the weights of key feature data and constructs a Bayesian-attention-feedforward neural network model to accurately capture the characteristics of coupled noise;

[0103] S41. The attention mechanism is used in the feedforward neural network to calculate the characteristic relationship between the vibration signal and the coupling noise in the input data, and a Bayesian-attention-feedforward neural network model is constructed to accurately capture the coupling noise characteristics. The attention mechanism calculates the characteristic s i The process can be expressed as:

[0104]

[0105] Where tanh is the activation function, is the weight of the attention mechanism, b is the bias, is the output of the i-th hidden layer;

[0106] S42. iConvert to exponential form and calculate the current s i The ratio of the sum of all attention score indices to obtain the attention score β i , the result of the attention score function β i It can be expressed as:

[0107]

[0108] Where n is the number of input feature data;

[0109] S5. The parameters and weight coefficients of the Bayesian-attention-feedforward neural network model are enhanced and updated through iteration, and the prediction result of the traction motor-gearbox coupling noise is output, which is expressed as:

[0110]

[0111] In the formula is the coupled noise sound pressure value predicted by the Bayesian optimization-attention mechanism-feedforward neural network model, ReLu is the activation function, w and b' are the weight and bias of the output layer respectively;

[0112] S6. Perform signal anomaly detection based on the coupling noise prediction results, and predict multiple related attributes of the traction motor-gearbox, including fault type, fault location and severity. Specifically: perform signal anomaly detection on the traction motor-gearbox coupling noise prediction results. When the predicted coupling noise is an abnormal signal, perform multi-task learning and predict the fault type, fault location and severity of the traction motor-gearbox at the same time. Perform targeted maintenance according to faults of different types, locations and severity to achieve intelligent early warning.

[0113] Example 2

[0114] This embodiment provides a train traction motor-gearbox coupling noise prediction system based on small vibration sample data, such as Figure 2 , 3 ,include:

[0115] Raw data collection and analysis module, collects raw data and performs data processing and analysis;

[0116] The feedforward neural network module consists of 1 input layer, L hidden layers, and 1 output layer. It learns the characteristics of the input data and establishes the mapping relationship between the vibration signal and the coupled noise.

[0117] Bayesian optimization module: Bayesian optimization selects the Tree-structured Parzen Estimator to optimize the hyperparameters of the feedforward neural network model, find the optimal hyperparameter set, and extract the mapping relationship between vibration data and coupled noise data;

[0118] The attention mechanism module calculates the weights of key feature data and accurately captures the characteristics of coupled noise;

[0119] The multi-dimensional visualization module visualizes the coupled noise sound pressure values ​​predicted by the vibration data at different sensor positions, including two-dimensional visualization, three-dimensional visualization, and modal visualization, and clearly shows the fitting between the two.

[0120] This embodiment can also be used for prediction of traction motor-gearbox coupling noise under other variables.

[0121] Example 3

[0122] This embodiment provides a specific example of a method for predicting train traction motor-gearbox coupling noise based on small vibration sample data. Figure 4-6 ,include:

[0123] The experiment was carried out on a certain model of traction motor-gearbox. The experimental setup included 8 accelerometers and 9 noise collection microphones. The accelerometers were placed at various vibration locations on the traction motor-gearbox, and the positions corresponding to different vibration accelerometers were marked as V1-V8. The noise collection microphones were placed at various locations near the traction motor-gearbox, designated as N1-N9. The accelerometers and noise collection microphones respectively collected vibration signals and coupled noise signals in the time domain, and then Fourier transform processing and analysis were performed on the central control platform and LMS Test.Lab14A software to obtain the frequency domain data of the vibration signals and coupled noise. The change curves of the coupled noise and vibration signals in the frequency domain are shown as follows: Figure 4 As shown in (a) and (b).

[0124] The horizontal (X), vertical (Y) and axial (Z) vibration data of the 8 acceleration sensors of the traction motor and gearbox are extracted at equal intervals as prediction inputs through downsampling processing, and a total of 1,600 sets of small sample data sets on the correlation between frequency, vibration data and coupling noise data are obtained. The Bayesian-attention-feedforward neural network noise prediction model uses 95% of the data for training, and the remaining 5% is used for prediction. The overall steps of this embodiment are the same as those of Embodiment 1. After multiple iterations, the coupling noise prediction results are output.

[0125] Figure 5 Predict the sound pressure value of the coupled noise of the traction motor-gearbox V1-V8 acceleration sensor at the N1 noise collection point. Figure 5 It can be seen that after evaluating 80 sets of data, the Bayesian optimization-attention-feedforward neural network has a high accuracy in predicting the coupling noise between the traction motor and the gearbox.

[0126] Among the evaluation indicators of the model, the mean absolute error (MAE) measures the gap between the predicted value and the true value, the root mean square error (RMSE) reflects the error size and generalization ability of the model, and the mean absolute percentage error (MAPE) is a further standardization of MAE. Therefore, MAE, RMSE and MAPE are used as evaluation indicators of model accuracy and effect, and their calculation formulas are as follows:

[0127]

[0128] Where Y represents the true value of the traction motor-gearbox noise sample, Represents the predicted value output by the noise prediction model based on Bayesian optimization-attention-feedforward neural network. Figure 6 As shown, the coupled noise prediction error value of the traction motor-gearbox V1-V8 acceleration sensor at the N1 noise collection point is visualized through a line graph.

[0129] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A train traction motor-gearbox coupling noise prediction method based on small vibration sample data, characterized in that the steps include: S1. Jointly analyze the high correlation between coupled noise and vibration signals based on the squared coherence-regression coefficient, and convert the complex small sample data into a two-dimensional matrix in the frequency domain S2. Learning 2D matrices in a multi-task learning framework via a feedforward neural network Based on the characteristics, the mapping relationship between the vibration signals in the horizontal, vertical and axial directions and the coupled noise is established; S3. Based on the tree-structured Bayesian optimization algorithm, find the optimal number of hidden layers, number of nodes and learning rate configuration of the feedforward neural network, and extract the mapping relationship between vibration data and coupled noise data; S31. Select the number of hidden layers L, the number of hidden layer nodes N, and the learning rate η of the feedforward neural network as the hyperparameter optimization target of the feedforward neural network, define the hyperparameter set of the feedforward neural network as x, define the objective function value y for updating the hyperparameter set x as: y=f(x)=f(L,N,η), define the configuration space as H={[L,N,η] 1 ,[L,N,η] 2 , …, [L, N, η] n }; S32. Calculate the probability density function P(x|y) of the hyperparameter of a given objective function value according to the current hyperparameter sample point, and construct a Bayesian optimization model based on a tree structure; S33. Using expected improvement as acquisition function EI y* (x), identify the next sample point with the highest expected value of acquisition; S34. When the acquisition function EI y* When (x) reaches the maximum value, the hyperparameter set is considered to be the optimal hyperparameter set x*, expressed as: Where L* is the optimal number of hidden layers, N* is the optimal number of hidden layer nodes, and η* is the optimal feedforward neural network learning rate; S35. Feedback the optimal hyperparameter set x* to the feedforward neural network for further training and optimization; S4. The attention mechanism calculates the weights of key feature data and constructs a Bayesian-attention-feedforward neural network model to accurately capture the characteristics of coupled noise; S41. The attention mechanism is used in the feedforward neural network to calculate the characteristic relationship between the vibration signal and the coupling noise in the input data, and a Bayesian-attention-feedforward neural network model is constructed to accurately capture the coupling noise characteristics. The attention mechanism calculates the characteristic s i The process can be expressed as: Where tanh is the activation function, is the weight of the attention mechanism, b is the bias, is the output of the i-th hidden layer; S42. i Convert to exponential form and calculate the current s i The ratio of the sum of all attention score indices to obtain the attention score β i , the result of the attention score function β i It can be expressed as: Where n is the number of input feature data; S5. The parameters and weight coefficients of the Bayesian-attention-feedforward neural network model are enhanced and updated through iteration, and the prediction result of the traction motor-gearbox coupling noise is output, which is expressed as: In the formula is the coupled noise sound pressure value predicted by the Bayesian optimization-attention mechanism-feedforward neural network model, ReLu is the activation function, w and b' are the weight and bias of the output layer respectively; S6. Perform signal anomaly detection based on the coupled noise prediction results, and predict multiple related attributes of the traction motor-gearbox, including fault type, fault location, and severity.

2. The train traction motor-gearbox coupling noise prediction method based on small vibration sample data according to claim 1 is characterized in that: The squared coherence-regression coefficient is expressed as: Where P Q_N (f) is the cross power spectrum density of the vibration signal Q and the coupled noise signal N, P Q_Q (f) and P N_N (f) are the auto-power spectral densities of the vibration signal Q and the coupled noise signal N, respectively. i is the ith vibration amplitude, N i is the i-th coupling noise sound pressure value, N is the average coupling noise sound pressure value, τ1+τ2=1.

3. The train traction motor-gearbox coupling noise prediction method based on small vibration sample data according to claim 1 is characterized in that: Two-dimensional matrix Including time T, frequency F, vibration signal V1 in the horizontal direction, vibration signal V2 in the vertical direction, vibration signal V3 in the axial direction, two-dimensional matrix It is expressed as: Among them, λ1~λ5 are the data sets of T, F, V1, V2, and V3 respectively, and the superscripts 1~n are the input data sets.

4. The train traction motor-gearbox coupling noise prediction method based on small vibration sample data according to claim 1 is characterized in that: The forward propagation process of the feedforward neural network is expressed as: In the formula, a i represents the output of the i-th hidden layer, represents the input of the i-th layer in the hidden layer, W ij and b i They represent the weight matrix and bias of the i-th layer in the hidden layer respectively, q represents the activation function of the hidden layer, L is the number of hidden layers of the feedforward neural network, and N is the number of hidden layer nodes of the feedforward neural network.

5. The train traction motor-gearbox coupling noise prediction method based on small vibration sample data according to claim 1 is characterized in that: The probability density function P(x|y) of the hyperparameter is expressed as: In the formula, y* represents the threshold, l(x) and g(x) represent y is less than or greater than or equal to y. * The density estimate of .

6. The train traction motor-gearbox coupling noise prediction method based on small vibration sample data according to claim 1 is characterized in that: Acquisition function EI y* (x) is expressed as: p(y|x) is the posterior probability, expressed as: p(x) is the marginal likelihood function, and p(y) is the distribution of the target function value.

7. The train traction motor-gearbox coupling noise prediction method based on small vibration sample data according to claim 1 is characterized in that: After obtaining the optimal hyperparameter set x*, the forward propagation process of the feedforward neural network is updated as follows:

8. The train traction motor-gearbox coupling noise prediction method based on small vibration sample data according to claim 1 is characterized in that: The loss function of the forward propagation of the feedforward neural network is LOSS, and the formula for updating the parameters of the loss function is expressed as: In the formula Represents the weight of the jth node in the i-th hidden layer of the updated feedforward neural network.

9. The train traction motor-gearbox coupling noise prediction method based on small vibration sample data according to claim 1 is characterized in that: Signal anomaly detection is performed on the prediction results of the traction motor-gearbox coupling noise. When the predicted coupling noise is an abnormal signal, multi-task learning is performed to simultaneously predict the fault type, fault location, and severity of the traction motor-gearbox. Targeted maintenance is performed based on faults of different types, locations, and severities to achieve intelligent early warning.

10. A train traction motor-gearbox coupling noise prediction system based on small vibration sample data, characterized in that: include: Raw data collection and analysis module, collects raw data and performs data processing and analysis; The feedforward neural network module includes 1 input layer, L hidden layers and 1 output layer, which learns the characteristics of the input data and establishes the mapping relationship between the vibration signal and the coupled noise; The Bayesian optimization module optimizes the hyperparameters of the feedforward neural network model based on the Tree-structured Parzen Estimator, finds the optimal hyperparameter set, and extracts the mapping relationship between vibration data and coupled noise data; The attention mechanism module calculates the weights of key feature data and accurately captures the characteristics of coupled noise; The multi-dimensional visualization module performs one-dimensional to multi-dimensional visualization of the coupled noise sound pressure values ​​predicted by the vibration data at different sensor positions, and clearly displays the fit between the two.

Images