Train traction motor-gear box coupling noise prediction method and system based on small vibration sample dataTechnical Field
The invention relates to the technical field of noise prediction and design of traction transmission systems of railway vehicles, in particular to a train traction motor-gear box coupling noise prediction method and system based on small vibration sample data.
Background
Coupling noise between the traction motor and the gearbox is a major source of rail vehicle noise. The strong coupling between the two exacerbates the noise making it extremely difficult to control. Meanwhile, accurately predicting the coupling noise between the traction motor and the gear box and considering the related characteristics thereof has been a difficulty in the design process of train noise. There is a certain correlation between the vibration signal and the coupling noise, and this correlation is just the basis for guiding the design of the noise of the whole train. Therefore, there is a need to fully study the coupling noise of traction motors and gearboxes and develop intelligent prediction methods and systems for coupling noise under small sample data of vibrations.
The existing solution is mainly to realize noise prediction of the traction transmission system through hardware equipment and software algorithms. The hardware equipment mainly comprises sensors, data collectors and the like and is used for collecting operation data of the traction transmission system. The software algorithm mainly comprises signal processing, feature extraction, pattern recognition and the like, and is used for analyzing and processing the collected data so as to realize noise control and fault diagnosis. Although this technique can be handled simply and quickly, it has some problems in practical applications. First, the traction motor and gearbox are in complex geometry with multi-source interference, and the collected available noise samples present a small and incomplete situation. Second, existing hardware devices and software algorithms often can only predict for a single type of noise, while for complex electromechanical coupling noise, decoupling and prediction often cannot be performed effectively. Finally, in terms of noise prediction, the existing technology often cannot realize high-precision and high-efficiency performance, which brings a certain risk to the reliability and safety of the vehicle.
The CN202110154558.5 discloses a gear box fault diagnosis method based on multi-source data fusion, which comprises the steps of firstly obtaining vibration signals, noise signals, temperature signals at a gear box bearing, displacement signals of a gear shaft and oil liquid data of gear box lubricating oil of a gear box, then respectively preprocessing the collected multi-source data, and finally diagnosing the fault of the gear box by utilizing a multi-sensor data fusion technology. According to the multi-sensor data fusion system, the multi-source data signals such as vibration, temperature, noise and displacement of the gear box are obtained, fault diagnosis is conducted on the multi-sensor data fusion technology, the functions of redundancy complementation of each sensor and a data source are fully exerted, fault information of the gear box can be expressed completely and clearly, and the positioning capability of faults of the gear box and the diagnosis capability of composite faults are improved. However, this method combines multiple source signals to diagnose faults, but has drawbacks. First, multiple acquisitions of the multi-source signal require precise time synchronization of the sensors and often require acquisition of a large number of data samples during the acquisition to support subsequent diagnostic tasks. Secondly, the method only uses multi-source data of the gearbox to carry out fault diagnosis on the multi-source data, and coupling influences generated by other electromechanical devices are not considered. Finally, the method uses conventional Back Propagation (BP) neural networks for feature extraction, respectively. The traditional back propagation neural network adopted in the characteristic extraction process is simple in structure, the super-parameter adjustment is difficult, and the diagnosis complexity and the calculation cost are increased. Therefore, how to effectively predict traction motor-gearbox coupling noise with low cost, high efficiency and high accuracy under small sample data is a problem to be solved in the current railway vehicle development.
Disclosure of Invention
Aiming at the defects that the reliable data of a traction motor-gear box is difficult to acquire, the electromechanical coupling noise is difficult to decouple and the accuracy of a noise prediction model is difficult to improve in the prior art, the invention provides a train traction motor-gear box coupling noise prediction method based on small vibration sample data, the coupling noise of the traction motor and the gearbox is predicted by utilizing the strong correlation between the coupling noise and the vibration signal and excellent decoupling capacity of the Bayes-attention-feedforward neural network model and using the vibration signal which is easy to acquire, so that the accurate prediction of the complex electromechanical coupling noise under different spatial positions is realized, and important front-end foundation and technology are provided for the forward design of the noise of the whole vehicle.
The invention solves the other technical problem of providing a train traction motor-gear box coupling noise prediction system for vibration small sample data.
The aim of the invention is realized by the following technical scheme:
a train traction motor-gear box coupling noise prediction method based on vibration small sample data comprises the following steps:
s1, analyzing high correlation between coupling noise and vibration signals according to square coherence-regression coefficient combination, and converting complex data into a two-dimensional matrix on a frequency domain
S2, learning a two-dimensional matrix under a multi-task learning framework through a feedforward neural networkThe method comprises the steps of establishing a mapping relation between vibration signals in the horizontal direction, the vertical direction and the axial direction and coupling noise;
S3, searching the optimal hidden layer number, node number and learning rate configuration of the feedforward neural network by using a Bayesian optimization algorithm based on a tree structure, and extracting the mapping relation between vibration data and coupling noise data;
S31, selecting the hidden layer number L, the hidden layer node number N and the learning rate eta of the feedforward neural network as a feedforward neural network super-parameter optimization target, defining a super-parameter set of the feedforward neural network as x, defining an objective function value y of an updated super-parameter set x as y=f (x) =f (L, N, eta), and defining a configuration space as H= { [ L, N, eta ]1,[L,N,η]2,…,[L,N,η]n };
s32, calculating a super-parameter probability density function P (x|y) of a given objective function value according to the current super-parameter sample points, and constructing a Bayes optimization model based on a tree structure;
S33, using the expected improvement as an acquisition function EIy* (x), identifying the next sample point with the highest expected acquisition value;
s34, when the acquisition function EIy* (x) reaches the maximum value, the super-parameter set is regarded as an optimal super-parameter set x, expressed as:
Wherein L is the optimal hidden layer number, N is the optimal hidden layer node number, and eta is the optimal feedforward neural network learning rate;
s35, feeding back an optimal super-parameter set x to a feedforward neural network, and further training and optimizing;
s4, calculating the weight of key feature data by an attention mechanism, constructing a Bayes-attention-feedforward neural network model, and accurately capturing the coupling noise characteristics;
S41, calculating a characteristic relation between a vibration signal and coupling noise in input data in a feedforward neural network by using an attention mechanism, constructing a Bayes-attention-feedforward neural network model, and accurately capturing the characteristic of the coupling noise, wherein the process of calculating the characteristic si by using the attention mechanism can be expressed as follows:
Where tanh is the activation function,To note the weight of the mechanism, b is the bias,The output of the ith layer which is an implicit layer;
S42, converting si into an exponential form, and calculating the ratio of the current si to the sum of all attention score indexes to obtain an attention score betai, wherein the result betai of the attention scoring function can be expressed as:
wherein n is the number of input feature data;
S5, reinforcing and updating parameters and weight coefficients of the Bayes-attention-feedforward neural network model through iteration, and outputting a traction motor-gear box coupling noise prediction result, wherein the prediction result is expressed as:
In the middle ofCoupling noise sound pressure values predicted by the Bayesian optimization-attention mechanism-feedforward neural network model are ReLu as an activation function, and w and b' are respectively the weight and bias of an output layer;
s6, detecting signal abnormality according to the coupling noise prediction result, and predicting a plurality of related attributes of the traction motor-gear box, including fault type, fault position and severity.
Further, the squared coherence-regression coefficients are expressed as:
Wherein PQ_N (f) is the cross power spectral density of the vibration signal Q and the coupling noise signal N, PQ_Q (f) and PN_N (f) are the self power spectral densities of the vibration signal Q and the coupling noise signal N, respectively, Qi is the ith vibration amplitude, Ni is the ith coupling noise sound pressure value,Τ1+τ2 =1 for coupling noise sound pressure average value.
Further, two-dimensional matrixComprises a time T, a frequency F, a vibration signal V1 in the horizontal direction, a vibration signal V2 in the vertical direction, a vibration signal V3 in the axial direction, and a two-dimensional matrixExpressed as:
Wherein λ1 to λ5 are data sets T, F, V, V2, and V3, respectively, and superscripts 1 to n are input data sets.
Further, the process of feed forward neural network forward propagation is expressed as:
where ai denotes the output of the hidden layer i-th layer,The input of the ith layer in the hidden layer is represented, Wij and bi respectively represent the weight matrix and the bias of the ith layer in the hidden layer, q represents the activation function of the hidden layer, L is the hidden layer number of the feedforward neural network, and N is the hidden layer node number of the feedforward neural network.
Further, the probability density function p (x|y) of the hyper-parameters is expressed as:
where y represents a threshold value, and l (x) and g (x) represent density estimates where y is less than or equal to y*, respectively.
Further, the acquisition function EIy* (x) is expressed as:
p (y|x) is a posterior probability expressed as:
p (x) is a marginal likelihood function and p (y) is a distribution of objective function values.
Further, after the optimal super parameter set x is obtained, the forward propagation process of the feedforward neural network is updated as follows:
Further, the forward propagation LOSS function of the feedforward neural network is LOSS, and the formula of the LOSS function update parameter is:
In the middle ofAnd the weight of the j node of the i hidden layer in the updated feedforward neural network is represented.
Further, signal anomaly detection is carried out on the prediction result of the coupling noise of the traction motor and the gearbox, when the coupling noise obtained through prediction is an anomaly signal, multi-task learning is carried out, meanwhile, the fault type, the fault position and the severity degree of the traction motor and the gearbox are predicted, and targeted maintenance is carried out according to faults of different types, different positions and different severity degrees, so that intelligent early warning is realized.
A train traction motor-gearbox coupling noise prediction system based on vibration small sample data, comprising:
The original data acquisition and analysis module acquires original data and performs data processing and analysis;
The feedforward neural network module comprises 1 input layer, L hidden layers and 1 output layer, learns the characteristics of input data and establishes a mapping relation between a vibration signal and coupling noise;
the Bayesian optimization module is used for carrying out super-parameter optimization on the feedforward neural network model based on a Tree structure Parsen estimator (Tree-structured Parzen Estimator), searching an optimal super-parameter set, and extracting the mapping relation between vibration data and coupling noise data;
the attention mechanism module is used for calculating the weight of the key characteristic data and accurately capturing the coupling noise characteristics;
And the multidimensional visualization module is used for carrying out visualization processing on coupling noise sound pressure values obtained by predicting vibration data at different sensor positions, wherein the visualization processing comprises two-dimensional visualization, three-dimensional visualization and modal visualization, and fitting conditions between the two are clearly shown.
Compared with the prior art, the beneficial effects are that:
The invention analyzes the vibration data and the coupling noise data of the traction motor-gear box of the train and extracts key characteristics. And establishing a feedforward neural network model by utilizing the correlation between the vibration data and the coupling noise data, and accurately predicting the coupling noise generated by the traction motor and the gear box at different spatial positions through vibration signals which are easy to acquire. In the feedforward neural network, the super-parameter selection of the neural network is pre-trained by using Bayesian optimization, and the optimum super-parameter is approximated. An attention mechanism is introduced in the data training process, and the characteristics of complex noise are further noted and calculated, so that accurate noise prediction is realized.
The method improves the accuracy of coupling noise prediction under the small vibration sample data set, predicts the coupling noise of key train components by utilizing the time-frequency domain vibration signals, and provides an effective technical approach for early control and reducing train noise. In addition, the present invention provides an important tool for evaluating noise performance in selecting train components, designing the train as a whole, and implementing manufacturing processes.
Drawings
Fig. 1 is a flowchart of a method for predicting coupling noise of a traction motor and a gearbox of a train based on vibration small sample data provided in embodiment 1.
Fig. 2 is a diagram of a train traction motor-gearbox coupling noise prediction system based on vibration small sample data according to embodiment 2.
Fig. 3 is a block diagram of a train traction motor-gearbox coupling noise prediction system based on vibration small sample data provided in embodiment 2.
Fig. 4 is a schematic diagram of experimental data provided in example 3.
Fig. 5 is a schematic diagram of a train traction motor-gearbox coupling noise prediction result based on vibration small sample data provided in embodiment 3.
Fig. 6 is a schematic diagram of a prediction error of coupling noise of a traction motor and a gearbox of a train based on vibration small sample data provided in embodiment 3.
Detailed Description
The present invention is further illustrated and described below with reference to examples, which are not intended to be limiting in any way.
Example 1
The embodiment provides a train traction motor-gear box coupling noise prediction method based on vibration small sample data, as shown in fig. 1, comprising the following steps:
a train traction motor-gear box coupling noise prediction method based on vibration small sample data comprises the following steps:
s1, analyzing high correlation between coupling noise and vibration signals according to square coherence-regression coefficient combination, and converting complex data into a two-dimensional matrix on a frequency domain
S11, carrying out coherence analysis on the processed vibration and coupling noise sample data by utilizing a square coherence-regression coefficient, and determining a vibration signal as a main association parameter for predicting subsequent coupling noise, wherein the square coherence-regression coefficient is expressed as:
Wherein PQ_N (f) is the cross power spectral density of the vibration signal Q and the coupling noise signal N, PQ_Q (f) and PN_N (f) are the self power spectral densities of the vibration signal Q and the coupling noise signal N, respectively, Qi is the ith vibration amplitude, Ni is the ith coupling noise sound pressure value,Τ1+τ2 =1 for coupling noise sound pressure average value.
S12, after square coherence-regression coefficient joint analysis, preliminary data screening is carried out on the original vibration signals, high coherence vibration data with MQ_N being more than 0.8 is obtained, and the coupling noise signals and the vibration signals are converted into a two-dimensional matrix on the frequency domainComprises a time T, a frequency F, a vibration signal V1 in the horizontal direction, a vibration signal V2 in the vertical direction, a vibration signal V3 in the axial direction, and a two-dimensional matrixExpressed as:
Wherein λ1 to λ5 are data sets T, F, V, V2, and V3, respectively, and superscripts 1 to n are input data sets.
S2, learning a two-dimensional matrix under a multi-task learning framework through a feedforward neural networkThe method is characterized in that a mapping relation between vibration signals in the horizontal direction, the vertical direction and the axial direction and coupling noise is established, and the forward propagation process of the feedforward neural network is expressed as follows:
where ai denotes the output of the hidden layer i-th layer,The input of the ith layer in the hidden layer is represented, Wij and bi respectively represent the weight matrix and the bias of the ith layer in the hidden layer, q represents the activation function of the hidden layer, L is the hidden layer number of the feedforward neural network, and N is the hidden layer node number of the feedforward neural network.
S3, searching the optimal hidden layer number, node number and learning rate configuration of the feedforward neural network by using a Bayesian optimization algorithm based on a tree structure, and extracting the mapping relation between vibration data and coupling noise data;
S31, selecting the hidden layer number L, the hidden layer node number N and the learning rate eta of the feedforward neural network as a feedforward neural network super-parameter optimization target, defining a super-parameter set of the feedforward neural network as x, defining an objective function value y of an updated super-parameter set x as y=f (x) =f (L, N, eta), and defining a configuration space as H= { [ L, N, eta ]1,[L,N,η]2,…,[L,N,η]n };
S32, calculating a super-parameter probability density function P (x|y) of a given objective function value according to a current super-parameter sample point, and constructing a Bayesian optimization model based on a tree structure, wherein the super-parameter probability density function P (x|y) is expressed as:
where y represents a threshold value, and l (x) and g (x) represent density estimates where y is less than or equal to y*, respectively.
S33, using the expected improvement as an acquisition function EIy* (x), identifying the next sample point with the highest expected acquisition value, wherein the acquisition function EIy* (x) is expressed as:
p (y|x) is a posterior probability expressed as:
p (x) is a marginal likelihood function and p (y) is a distribution of objective function values. Defining γ as a constructor error adjustment function, γ=p (y < y x), p (x) =γl (x) + (1- γ) g (x), the acquisition function EIy* (x) is further expressed as:
The formula is an iterative process that optimizes and maximizes g (x)/l (x), the better the super-parameter set performance when EIy* (x) is near the maximum. The expression to the left of the equation is proportional to the expression to the right.
S34, when the acquisition function EIy* (x) reaches the maximum value, the super-parameter set is regarded as an optimal super-parameter set x, expressed as:
wherein L is the optimal hidden layer number, N is the optimal hidden layer node number, and eta is the optimal feedforward neural network learning rate.
After the optimal super parameter set x is obtained, the forward propagation process of the feedforward neural network is updated as follows:
establishing a LOSS function as LOSS by using the output value obtained by forward propagation and the actual sample label, wherein the formula of the LOSS function updating parameter is as follows:
In the middle ofAnd the weight of the j node of the i hidden layer in the updated feedforward neural network is represented.
S4, calculating the weight of key feature data by an attention mechanism, constructing a Bayes-attention-feedforward neural network model, and accurately capturing the coupling noise characteristics;
S41, calculating a characteristic relation between a vibration signal and coupling noise in input data in a feedforward neural network by using an attention mechanism, constructing a Bayes-attention-feedforward neural network model, and accurately capturing the characteristic of the coupling noise, wherein the process of calculating the characteristic si by using the attention mechanism can be expressed as follows:
Where tanh is the activation function,To note the weight of the mechanism, b is the bias,The output of the ith layer which is an implicit layer;
S42, converting si into an exponential form, and calculating the ratio of the current si to the sum of all attention score indexes to obtain an attention score betai, wherein the result betai of the attention scoring function can be expressed as:
wherein n is the number of input feature data;
S5, reinforcing and updating parameters and weight coefficients of the Bayes-attention-feedforward neural network model through iteration, and outputting a traction motor-gear box coupling noise prediction result, wherein the prediction result is expressed as:
In the middle ofCoupling noise sound pressure values predicted by the Bayesian optimization-attention mechanism-feedforward neural network model are ReLu as an activation function, and w and b' are respectively the weight and bias of an output layer;
S6, carrying out signal anomaly detection according to the coupling noise prediction result, and simultaneously predicting a plurality of related attributes of the traction motor-gear box, including fault types, fault positions and severity, wherein the method comprises the steps of carrying out signal anomaly detection on the coupling noise prediction result of the traction motor-gear box, carrying out multi-task learning when the coupling noise obtained by prediction is an anomaly signal, and simultaneously predicting the fault types, the fault positions and the severity of the traction motor-gear box, carrying out targeted maintenance according to faults of different types, different positions and different severity, and realizing intelligent early warning.
Example 2
The present embodiment provides a train traction motor-gearbox coupling noise prediction system based on vibration small sample data, as shown in fig. 2 and 3, including:
The original data acquisition and analysis module acquires original data and performs data processing and analysis;
the feedforward neural network module consists of 1 input layer, L hidden layers and 1 output layer, learns the characteristics of input data and establishes a mapping relation between a vibration signal and coupling noise;
The Bayesian optimization module is used for performing super-parameter optimization on the feedforward neural network model by using a Bayesian optimization selection Tree structure Parsen estimator (Tree-structured Parzen Estimator), searching an optimal super-parameter set, and extracting a mapping relation between vibration data and coupling noise data;
the attention mechanism module is used for calculating the weight of the key characteristic data and accurately capturing the coupling noise characteristics;
And the multidimensional visualization module is used for carrying out visualization processing on coupling noise sound pressure values obtained by predicting vibration data at different sensor positions, wherein the visualization processing comprises two-dimensional visualization, three-dimensional visualization and modal visualization, and fitting conditions between the two are clearly shown.
The embodiment can also be used for traction motor-gearbox coupling noise prediction under other variables.
Example 3
The present embodiment provides a specific example of a train traction motor-gearbox coupling noise prediction method based on vibration small sample data, as shown in fig. 4 to 6, including:
Experiments were performed on a model traction motor-gearbox, with an experimental set-up comprising 8 acceleration sensors and 9 noise pickup microphones. The acceleration sensor is placed at each vibration location of the traction motor-gearbox, and the positions corresponding to the different vibration acceleration sensors are marked as V1-V8. Noise collecting microphones are placed at various locations, designated N1-N9, near the traction motor-gearbox. The acceleration sensor and the noise acquisition microphone acquire vibration signals and coupling noise signals in the time domain respectively, and then Fourier transformation processing and analysis are carried out in the central control platform and LMS test.Lab14A software, so that frequency domain data of the vibration signals and the coupling noise are obtained. The curves of the coupling noise and vibration signal in the frequency domain are shown in fig. 4 (a) and (b), respectively.
Horizontal (X), vertical (Y) and axial (Z) vibration data of the traction motor and the 8 acceleration sensors of the gear case are extracted as prediction inputs in equal intervals by a downsampling process, and a total of 1600 sets of small sample data sets concerning correlations between frequency, vibration data and coupling noise data are obtained. The bayesian-attention-feedforward neural network noise prediction model is trained using 95% of the data, with the remaining 5% being used for prediction. The overall steps of this embodiment are the same as those of embodiment 1, and after a plurality of iterations, the coupling noise prediction result is output.
Fig. 5 is a predicted sound pressure value of coupling noise of the traction motor-gear box V1-V8 acceleration sensor at the N1 noise collection point. As can be seen from fig. 5, the bayesian optimization-attention-feed forward neural network has high accuracy in predicting the coupling noise between the traction motor and the transmission through the evaluation of the 80 sets of data.
In the evaluation index of the model, the Root Mean Square Error (RMSE) represents the error size and generalization capability of the model by taking the difference between the average absolute error (MAE) measurement predicted value and the true value into consideration, and the average absolute percentage error (MAPE) is further standardized by the MAE. Therefore, MAE, RMSE and MAPE are taken as evaluation indexes of model precision and effect, and the calculation formula is as follows:
wherein Y represents the true value of the traction motor-gearbox noise sample,Representing predicted values output by a noise prediction model based on a bayesian optimization-attention-feedforward neural network. As shown in fig. 6, the coupling noise prediction error values of the traction motor-gear box V1-V8 acceleration sensors at the N1 noise acquisition point are visualized by a line graph.
It is to be understood that the above examples of the present invention are provided by way of illustration only and not by way of limitation of the embodiments of the present invention. Other variations or modifications of the above teachings will be apparent to those of ordinary skill in the art. It is not necessary here nor is it exhaustive of all embodiments. Any modification, equivalent replacement, improvement, etc. which come within the spirit and principles of the invention are desired to be protected by the following claims.