CN-116013238-B - Active noise reduction algorithm of sensor without secondary path based on convolution-fuzzy neural network

Abstract

The invention discloses a convolution-fuzzy neural network-based active noise reduction algorithm without secondary path sensor, which establishes a correlation model between noise sources at different positions of a vehicle and noise of a target noise reduction area through the convolution-fuzzy neural network, and calculating and fitting to obtain a virtual error noise signal of the target noise reduction area to replace the error signal obtained by a secondary path sensor, thereby reducing the dependence of large-space active noise reduction in the vehicle on a plurality of secondary path sensors. Meanwhile, the virtual error noise signal is used as the input of the fuzzy layer, the secondary path identification is completed through iterative calculation, the optimal noise elimination signal is output in real time, the calculation efficiency is improved, and the processing capacity of nonlinear noise is enhanced.

Inventors

• LI TAO
• WU XIAOTING
• WANG MINQI
• HE YUYAO
• YANG JUN
• DING RONGJUN
• LUO ZHUHUI
• XIAO LEI
• MEI WENQING
• HE JING

Assignees

• 湖南工业大学

Dates

Publication Date

20260505

Application Date

20221102

Claims (8)

1. 1. A convolutional-fuzzy neural network-based active noise reduction algorithm without secondary path sensor, characterized by comprising the steps of: S1, collecting noise signals X (n) at a noise source as Where k is the number of primary sensors, Noise signals acquired for the kth sensor; S2, determining noise reduction points, collecting noise signals as ideal signals r (n) trained by a convolutional network model, wherein r (n) is expressed as Where l is the noise signal sequence length; s3, enabling noise signals X (n) and r (n) in the S1 to enter a convolution layer and a pooling layer of a convolution network to carry out iterative computation, wherein a training function E (n) of the convolution network is as follows: the output virtual error signal o (n) is: Where f (-) is the activation function, ω (n), b (n) are weights of the convolutional network, are network offsets, where the weights ω (n) of the convolutional network and the network offsets b (n) are updated according to the training function E (n) as: ; s4, inputting the virtual error signal o (n) output by the convolution layer into a fuzzy neural network, calculating a membership value, activating the membership value by using a multiplication function, and outputting a fuzzy layer output value u (n) as Wherein j is the number of nodes of the fuzzy layer, H nodes are altogether, lambda j (n) is the result of the output signal of the convolution network after the output signal passes through the membership function of the fuzzy layer, and p 0 and p 1 are fuzzy coefficients; S5, filtering a noise signal acquired by a target noise reduction point through a primary path P (z) to obtain a desired signal d (n), and obtaining a noise reduction signal y (n) obtained by passing a fuzzy layer output value u (n) through a secondary path H (z), wherein an objective function J (n) of an active noise reduction algorithm is obtained 。
2. 2. The active noise reduction algorithm of a no-secondary-path sensor based on a convolutional-fuzzy neural network of claim 1, wherein in step S1 Is expanded into L is the noise signal sequence length.
3. 3. The convolutional-fuzzy neural network-based no-secondary-path sensor active noise reduction algorithm of claim 1, wherein the convolutional network in step S3 uses a leak-Relu function as an activation function f (θ), expressed as Wherein, theta is a function independent variable, a is a number of intervals (1, + -infinity).
4. 4. The active noise reduction algorithm of the no-secondary-path sensor based on the convolution-fuzzy neural network according to claim 1, wherein the activation result of o (n) into the fuzzy layer in the step S4 is that Wherein j is the number of fuzzy layer nodes, H nodes are total, and mu j is a membership function.
5. 5. The convolutional-fuzzy neural network-based no-secondary-path sensor active noise reduction algorithm of claim 4, wherein the membership function is expressed as Wherein c (n) and σ (n) are the center and width of the membership function, respectively.
6. 6. The active noise reduction algorithm of a no secondary path sensor based on a convolutional-fuzzy neural network according to claim 1, wherein the desired signal d (n) in step S5 is: 。
7. 7. The active noise reduction algorithm of a non-secondary path sensor based on a convolutional-fuzzy neural network according to claim 1, wherein the noise reduction signal y (n) in step S5 is: 。
8. 8. The active noise reduction algorithm of a no-secondary-path sensor based on a convolution-fuzzy neural network according to claim 1, wherein the objective function J (n) in step S5 is updated by a gradient descent method in real-time training of the fuzzy neural network, and the fuzzy coefficient p (n), the center c (n) of the membership function, and the width σ (n) of the membership function are respectively Where α and β are fuzzy network learning rates.

Description

Active noise reduction algorithm of sensor without secondary path based on convolution-fuzzy neural network Technical Field The invention relates to the technical field of vehicle noise reduction control, in particular to a convolution-fuzzy neural network-based active noise reduction algorithm without a secondary channel sensor. Background The noise of the railway vehicle is complex, nonlinear and multi-source, and is likely to cause physical and psychological injury to passengers and drivers, so that the noise elimination of the railway vehicle is extremely important, and the active noise reduction technology is a technology for eliminating the original noise by utilizing electronic circuits and sound amplifying equipment to generate sounds with the same phase and opposite phase as the frequency of the noise so as to achieve the purpose of noise reduction, and aims at eliminating low-frequency noise. In the traditional active noise reduction control method, the LMS algorithm is easy to realize, but has poor robustness and lower convergence accuracy in the actual application scene, and the FxLMS algorithm introducing the filtered-x signal and improvement thereof can improve the convergence speed or reduce steady-state errors, but both depend on the performance, the quantity and the arrangement of the secondary path sensor in engineering practice. The CN202010862334.5 is a convolution-fuzzy neural network method for actively controlling global spatial noise of a vehicle, which comprises the steps of arranging a plurality of secondary paths around a noise reduction area of the vehicle, collecting noise residual signals of the secondary paths, carrying out off-line identification by adopting the convolution-fuzzy neural network to obtain a secondary path model, simultaneously carrying out on-line correction on controller parameters by using a self-adaptive active noise control algorithm of the secondary paths, and finally outputting multi-azimuth noise cancellation signals. The method uses the convolution-fuzzy neural network for identifying the inverse model of the object, provides an effective method for identifying the nonlinear noise in the global space of the vehicle, improves the identification precision of the secondary path by utilizing the nonlinear approximation capability of the convolution-fuzzy neural network to the function, establishes a stable secondary path model by adopting an active feedback noise elimination system, and solves the problems of difficult control and narrow frequency band of the global space noise of the vehicle. However, the method also has the problems of dependence on the performance, the number, the arrangement and the like of the secondary path sensors, and the arrangement of a plurality of secondary path sensors also leads to cost rise, which is unfavorable for the engineering practice of large-space active noise reduction in the vehicle. Disclosure of Invention The invention aims to solve the main technical problem that the practice of vehicle noise reduction engineering is difficult due to the dependence on a multi-level channel sensor in the existing active noise reduction method, establishes a correlation model between noise sources at different positions and noise in a target noise reduction area, and provides an active noise reduction algorithm of a non-secondary channel sensor based on a convolution-fuzzy neural network. The aim of the invention is realized by the following technical scheme: An active noise reduction algorithm of a no-secondary-path sensor based on a convolution-fuzzy neural network is used for obtaining a virtual error signal by combining a convolution network (CNN) to perform calculation fitting so as to eliminate acoustic feedback, and the correlation between a reference signal and primary noise at an ideal noise reduction point is improved. And with the Fuzzy Neural Network (FNN) as a controller, the method iteratively calculates and outputs noise elimination signals in real time, enhances the nonlinear active noise reduction capability while improving the calculation speed, and specifically comprises the following steps: s1, collecting noise signals X (n) at a plurality of noise sources as X(n)=[x1(n),x2(n),…xk(n)] Where k is the number of primary sensors and x k (n) is the noise signal acquired by the kth sensor; s2, determining noise reduction points, and collecting real noise signals r (n) through a sensor to serve as signals trained by a convolutional network model, wherein r (n) is expressed as r(n)=[r(n),r(n-1),...r(n-l)] Where l is the noise signal sequence length; s3, performing iterative computation on the noise signals X (n) and r (n) in S1 entering a convolution layer and a pooling layer of a convolution network, and outputting a virtual error signal o (n) as an activation function o(n)=f(∑X(n)*ω(n)+b(n)) Where f (·) is the activation function, ω (n) is the weight of the convolutional network, b (n) is the network bias; S4, inputt