Small-sample prediction method for the noise of train traction fans based on regression-neural network

Abstract

The present invention provides a small-sample prediction method for the noise of a train traction fan based on regression-neural network, which comprises the following steps: obtaining the influencing factors related to the noise of the fan of the train traction system, collecting several groups of independent influencing factor data, establishing a multiple regression model, obtaining the regression coefficients through the least squares method to obtain an empirical multiple regression model group, and screening out the abnormal point data in the influencing factor data; inputting the influencing factor data after screening into a fuzzy neural network; in the fuzzy hidden layer, calculating the membership degree value according to the fuzzy rules and then activating the membership degree value by using a multiplication function; calculating and outputting a predicted value through a double-layer fuzzy hidden layer. The present invention simplifies the model by reducing the redundant layers through the structure of the fuzzy neural network, reduces the calculation amount and improves the calculation speed, and can accurately predict under small-sample noise data, so as to realize the forward design of the noise of the train traction system fan and eliminate the chronic problem of the industry in the train system and parts field of manufacturing first and then reducing noise.

Description

Technical Field

[0001] The present invention relates to the field of train fan noise prediction, and particularly to a small-sample prediction method for the noise of a train traction fan based on regression-neural network. Background Art

[0002] As a cooling core component of the train electric traction system, the noise generated by the fan during operation is one of the main noise sources of the train, which not only has a great impact on the overall noise of the train, but also reduces the overall comfort of the train. Predicting the fan noise can provide a certain basis for train noise fault diagnosis and noise reduction, that is, targeted solutions to noise problems can be carried out according to the prediction results, and at the same time, ideas can be provided for the optimal design of the train electric traction system.

[0003] In the past, numerical analysis methods were often used to simulate and model the noise characteristics. The calculation process was lengthy, the computational resources consumed by complex and disordered non-linear noise were large, and the prediction accuracy was greatly affected by the established mathematical model. When using the time series method to predict the noise according to the changing trend of the noise over time, although the accuracy was improved and the calculation was simple, the input data was single, and it was difficult to comprehensively consider other factors of the fan, such as the influence of the fan structure on the noise, only based on the data before and after the time series for noise prediction.

[0004] The fan noise is affected by multiple factors. Therefore, for the noise prediction affected by multiple factors, a neural network can be considered, and the mapping relationship between different influencing factors and the actual noise value can be determined through the training of the neural network model. However, the noise prediction model based on the neural network usually requires a large amount of data for training to ensure the accuracy and stability of the model. The complexity and cost of collecting the fan noise data of the train electric traction system make it difficult to obtain a large-sample data set. Therefore, it is necessary to improve the noise prediction model for the case of small-sample data. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art. For this purpose, the present invention proposes a small-sample prediction method for the noise of a train traction fan based on regression-neural network, which includes the following steps:

[0006] S1: Obtain the influencing factors related to the noise of the train traction system fan, where the influencing factors include the number of blades, static pressure, flow rate, and rotational speed; collect several groups of independent influencing factor data, establish a multiple regression model, obtain the regression coefficients through the least squares method to get an empirical multiple regression model group, and screen out the abnormal point data in the influencing factor data;

[0007] S2: Input the influencing factor data after screening into the fuzzy neural network;

[0008] S3: After the membership value is calculated according to the fuzzy rules in the fuzzy hidden layer, the membership value is activated by the multiplication function;

[0009] S4: Calculate the output prediction value through the double-layer fuzzy hidden layer.

[0010] Preferably, the multiple regression model established in step S1 is:

[0011]

[0012] Among them, a0, a1, ...a n is the regression coefficient, a0 is the constant term; ε0, ε1, ...ε1 are the terms that satisfy the normal distribution N(0, σ 2 ), the random error vector x 11 ,…x 1n ,…,x l1 ...x ln is a set of sampled data, u1, u2,…, u1 are the noise output values under different relationship conditions respectively.

[0013] Preferably, the process of obtaining the regression coefficient in step S1 is:

[0014] The empirical multiple regression model group is obtained by the least squares method, and its expression is:

[0015]

[0016] in, is the regression coefficient, n is the number of influencing factors, and the correlation coefficient of each influencing factor is calculated through the above multiple regression formula.

[0017] Preferably, the step S2 is specifically as follows:

[0018] Multiple influencing factors input

[0019]

[0020] Among them, n is the number of influencing factors, k is the number of groups, and x 11 , x 12 ...x 1n ,...x k1 , x k2 ...x kn is a data set of k groups containing n influencing factors.

[0021] Preferably, the step S3 specifically includes:

[0022] In the fuzzy hidden layer, the membership value is calculated according to the fuzzy rules and then the membership value is activated and defuzzified using the multiplication function. The activation result is

[0023]

[0024] j is the number of nodes in the fuzzy hidden layer, with a total of k nodes, usually set based on past experience. μ(·) is the membership function, and the Gaussian membership function is selected here. Its expression is

[0025]

[0026] c nj ,σ nj are the center and width of the membership function.

[0027] Preferably, the step S4 specifically includes:

[0028] Calculate the output prediction value through the double-layer fuzzy hidden layer

[0029]

[0030] where ω0, ω1...ω n are the fuzzy coefficients, and h j is the output of the fuzzy hidden layer, and its expression is Equation

[0031]

[0032] Define the cost function during the training of the FNN as

[0033]

[0034] In the formula, d is the expected output, y is the network prediction output, and e is the network prediction error. According to the cost function, the network parameters are updated using the stochastic gradient descent method, and the parameter update formula is as follows:

[0035]

[0036]

[0037]

[0038] In the formula, η and β represent the learning rate, and n represents the nth moment.

[0039] Compared with the prior art, the beneficial effects of the present invention are:

[0040] 1. In the case where it is difficult to ensure obtaining a large sample data set due to the complexity and cost of collecting fan noise data of the train electric traction system and it is impossible to obtain a large amount of noise data, the present invention reduces the number of redundant layers through the structure of the fuzzy neural network to simplify the model, reduces the amount of calculation and improves the calculation speed, and can accurately predict under small sample noise data;

[0041] 2. Implement the forward design of the fan noise of the train traction system to eliminate the chronic problem in the industry of manufacturing first and then reducing noise in the train system and components. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments recorded in the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0043] Figure 1 It is a block diagram of a noise prediction model based on multiple regression - double - layer fuzzy neural network provided by an embodiment of the present invention;

[0044] Figure 2 It is a flowchart of an algorithm based on multiple regression - double - layer fuzzy neural network provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0046] An embodiment of the present invention provides a small - sample prediction method for the noise of a train traction fan based on regression - neural network, which includes the following steps:

[0047] S1: Obtain the influencing factors related to the noise of the train traction system fan, collect several groups of independent influencing factor data, establish a multiple regression model, obtain the regression coefficients through the least - squares method to get an empirical multiple regression model group, and screen out the abnormal point data in the influencing factor data;

[0048] S2: Input the influencing factor data after screening into the fuzzy neural network;

[0049] S3: Calculate the membership degree value according to the fuzzy rules in the fuzzy hidden layer and then activate the membership degree value using the product function;

[0050] S4: Calculate and output the predicted value through the double - layer fuzzy hidden layer.

[0051] First, in order to obtain the influencing factors related to the fan noise of the train traction system, it is necessary to analyze the working principle and characteristics of the train electric traction system. The train electric traction system is a key mechatronic system in the whole vehicle that converts electrical energy into mechanical energy, drives the train to run and provides traction force, and is one of the main sources of train noise. The noise generated by the fan in the train electric traction system is the main contribution part of the train electric traction system, accounting for about 75%.

[0052] The fan noise mainly includes aerodynamic noise, mechanical noise and aerodynamic-solid coupling noise generated by the interaction of gas and solid elastic systems. Mechanical noise is caused by process defects during the installation or production of the fan and contributes less to the overall fan noise; while aerodynamic noise is the result of the superposition of rotational noise and vortex noise and is the main contributing noise of the fan.

[0053] The electric traction system is used for power distribution and voltage level change during train operation, and the fan is driven by an electric motor and consists of an air inlet, an air outlet, a moving impeller at the air inlet, a guiding fixed impeller and an air outlet. The fan is located inside the system cabinet for heat dissipation and cooling, which determines the safety and stability of the traction system. Therefore, the fan should have good noise performance to avoid energy loss and reduce efficiency.

[0054] Rotational noise is generated by the periodic impact of air particles by the rotating blades, which is related to the number of fan blades and the rotational speed, and is particularly prominent at high speeds. The frequency f1 is calculated by Equation (1). Because rotational noise has discrete characteristics, it is also called discrete noise, and the fundamental frequency noise intensity is the strongest, and its fundamental frequency is also called the blade passing frequency. Vortex noise is generated by the interaction of the airflow with the impeller, the fan inlet guide cover, etc. when the blade rotates. Its frequency band is wide and there is no obvious single-peak characteristic.

[0055] The frequency f2 is calculated by Equation (2)

[0056] f1 = nZi / 60 i = 1,2,3… (1)

[0057] f2 = S r wi / L i = 1,2,3… (2)

[0058] Where, Z is the number of blades; n is the rotational speed; i is the harmonic, and when i = 1, it is the fundamental frequency; in Equation (2), S r is the Strouhal number, and its value is generally between 0.14 and 0.20; w is the relative velocity between the gas and the blade; L is the projection width of the object's front surface perpendicular to the velocity plane.

[0059] When evaluating the noise performance of the fan, the specific A-weighted sound pressure level is used for evaluation and comparison, and its calculation formula is

[0060] L SA = LA -10lg(Qp 2 ) + 19.8 (3)

[0061] where L SA is the A-weighted sound pressure level ratio, with the unit of dB; L A is the A-weighted sound pressure level, with the unit of dB(A); Q is the flow rate and p is the pressure.

[0062] It can be seen from equations (1) and (2) that the fan noise is closely related to the number of blades and their rotational speed. It can be seen from equation (3) that the noise performance of the fan is also related to the flow rate and pressure.

[0063] By analyzing this embodiment, the number of blades, static pressure, flow rate, and rotational speed, these 4 influencing factors are selected as the input values for predicting the fan noise of the train electric traction system.

[0064] The specific process for obtaining the regression coefficients in the embodiments of the present invention is as follows:

[0065] Using multiple regression to further screen the inputs of the prediction model to avoid individual outliers in the input data. l sets of independent data are obtained from the fan noise acquisition data, and a multiple regression model can be established

[0066]

[0067] where a0, a1,..., a n are the regression coefficients, a0 is the constant term; ε0, ε1,..., ε1 are random error vectors that satisfy the normal distribution N(0, σ 2 ). The random error vector, x 11 ,..., x 1n ,..., x l1 ...x ln are 1 set of sampled data, and u1, u2,..., u1 are the noise output values under different relationship conditions respectively. According to equation (4), a system of equations is established, and through the least squares method, an empirical multiple regression model group is obtained, and its expression is as shown in (5)

[0068]

[0069] where are the regression coefficients to be obtained, n is the number of influencing factors, and the correlation coefficients of each influencing factor are calculated through the multiple regression in equation (5).

[0070] Due to the different influences of different combinations of influencing factors on the actual noise level, which are non-linear. Each influencing factor has a certain relationship with the noise. In the embodiments of the present invention, a fuzzy neural network (FNN, Fuzzy Neural Network) with strong non-linear mapping ability is used to establish the mapping relationship between different influencing factors and the noise value, and the gradient descent method is used to optimize the loss function. Since it is difficult and costly to collect the fan noise data of the train electric traction system, resulting in few data samples, a double-layer fuzzy neural network except for input and output is specially selected. In the embodiments of the present invention, by applying fuzzification and defuzzification to the fuzzy hidden layer, the number of layers is reduced and the generalization ability of the network is improved.

[0071] As Figure 1 shown in the block diagram of the noise prediction model based on multiple regression - double-layer fuzzy neural network provided by the embodiments of the present invention, assuming that multiple influencing factors are input

[0072]

[0073] Among them, n is the number of types of influencing factors, k is the number of groups, x 11 , x 12 …x 1n ,...x k1 , x k2 ...x kn are k groups of data sets containing n influencing factors.

[0074] In the fuzzy hidden layer, after calculating the membership degree value according to the fuzzy rule, the product function is used to activate the membership degree value for defuzzification, and its activation result is

[0075]

[0076] j is the number of nodes in the fuzzy hidden layer, with a total of k nodes, usually set according to past experience. μ(·) is the membership function, and the Gaussian membership function is selected here, and its expression is as shown in Equation (8)

[0077]

[0078] c nj , σ nj are the center and width of the membership function.

[0079] The predicted value is calculated through the double-layer fuzzy hidden layer

[0080]

[0081] Among them, ω0, ω1...ω n are the fuzzy coefficients, h j is the output of the fuzzy hidden layer, and its expression is Equation

[0082]

[0083] Define the cost function during FNN training as

[0084]

[0085] where \(d\) is the expected output, \(y\) is the network predicted output, and \(e\) is the network prediction error. According to the cost function, the network parameters are updated using the stochastic gradient descent method, and the parameter update formulas are as follows:

[0086]

[0087]

[0088]

[0089] In Equations (12)-(14), \(\eta\) and \(\beta\) represent the learning rate, and \(n\) represents the \(n\)th moment.

[0090] In summary, Equations (6) to (14) constitute the train electric traction system fan noise prediction model diagram based on multiple regression and double-layer fuzzy neural network; Figure 2 It is the algorithm flowchart provided by the embodiment of the present invention based on multiple regression - double-layer fuzzy neural network. With multiple factors affecting the fan noise level of the train electric traction system as inputs, calculations are performed through the fuzzy inference, defuzzification, and self-learning capabilities of the FNN, and finally the relationship between multiple influencing factors and the fan noise level is output to complete the prediction of the fan noise of the train electric traction system.

[0091] Through analysis, the embodiment of the present invention selects 4 influencing factors, namely the number of blades, static pressure, flow rate, and rotational speed, as the input values for predicting the fan noise of the train electric traction system. Then, the FNN is used as the simplest network structure of the prediction model to avoid wasting computing resources to handle small-sample fan noise data. By experimentally verifying the model, the method based on multiple regression - double-layer fuzzy neural network can accurately predict under small-sample train electric traction system fan noise data, can achieve the best optimization after iterative cycles, and has a fast calculation speed.

[0092] The above embodiments are only used to illustrate the technical method of the present invention and not to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical method of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A small-sample prediction method for the noise of a train traction fan based on regression-neural network, characterized in that It includes the following steps: S1: Obtain the influencing factors related to the fan noise of the train traction system. The influencing factors include the number of blades, static pressure, flow rate, and rotational speed; collect several groups of independent influencing factor data, establish a multiple regression model, obtain the regression coefficients through the least squares method to get an empirical multiple regression model group, and screen out the abnormal point data in the influencing factor data; S2: Input the influencing factor data after screening into the fuzzy neural network; S3: In the fuzzy hidden layer, after calculating the membership degree value according to the fuzzy rules, use the product function to activate the membership degree value; S4: Calculate and output the noise prediction value through the double-layer fuzzy hidden layer, specifically including: Calculate and output the noise prediction value through the double-layer fuzzy hidden layer wherein is the fuzzy coefficient, j is the number of nodes in the fuzzy hidden layer, with a total of k nodes, which are usually set based on past experience, is the output of the fuzzy hidden layer, and its expression is Equation Define the cost function during the FNN training as wherein, d is the expected output, y is the network prediction output, e is the network prediction error; according to the cost function, the network parameters are updated by the stochastic gradient descent method, and the parameter update formula is as follows: In the formula, η , β represents the learning rate, n represents the n moment.

2. The small sample prediction method for the train traction fan noise based on regression-neural network according to claim 1, characterized in that, The multiple regression model established in step S1 is: wherein, is the regression coefficient, is the constant term; is a random error vector satisfying the normal distribution N ( 0 , ), ,… ,…, … is for sampling l A set of data, which are the noise output values under different relationship conditions respectively.

3. The small-sample prediction method for the train traction fan noise based on regression-neural network according to claim 2, wherein The process of obtaining the regression coefficients in step S1 is: Obtain an empirical multiple regression model group through the least squares method, and its expression is Among them, is the regression coefficient to be obtained, n is the number of influencing factors, and the correlation coefficients of each influencing factor are calculated through the above multiple regression.

4. The small-sample prediction method for the train traction fan noise based on regression-neural network according to claim 1, wherein Step S2 is specifically: Input of multiple influencing factors Among them, n is the number of types of influencing factors, k is the number of groups, is the dataset of k groups containing n influencing factors.

5. The small-sample prediction method for the train traction fan noise based on regression-neural network according to claim 1, wherein Step S3 specifically includes: In the fuzzy hidden layer, after calculating the membership degree value according to the fuzzy rules, use the product function to activate the membership degree value for defuzzification, and its activation result is j is the number of nodes in the fuzzy hidden layer, with a total of k nodes, which are usually set based on past experience. is the membership function. Here, the Gaussian membership function is selected, and its expression is are the center and width of the membership function.

Images