A convolution-fuzzy neural network-based active noise reduction algorithm for a sensor without secondary path

Abstract

This invention discloses an active noise reduction algorithm based on a convolutional-fuzzy neural network without secondary path sensors. The algorithm establishes a correlation model between noise sources at different locations within a vehicle and the noise in the target noise reduction area using a convolutional-fuzzy neural network. It calculates and fits a virtual error noise signal for the target noise reduction area, replacing the error signal obtained through secondary path sensors in traditional methods. This reduces the dependence of active noise reduction in large vehicle interior spaces on multiple secondary path sensors. Furthermore, the invention uses the virtual error noise signal as input to the fuzzy layer, iteratively calculating and identifying secondary paths, and outputting the optimal noise cancellation signal in real time, improving computational efficiency and enhancing the ability to handle nonlinear noise.

Description

Technical Field

[0001] This invention relates to the field of vehicle noise reduction control technology, and more specifically, to an active noise reduction algorithm for sensors without secondary paths based on a convolutional-fuzzy neural network. Background Technology

[0002] The noise from rail vehicles is complex, nonlinear, and multi-source, and can potentially cause physical and mental harm to passengers and drivers. Therefore, noise reduction in rail vehicles is extremely important. Active noise reduction technology uses electronic circuits and amplification equipment to generate sound with the same frequency but opposite phase as the noise, thereby canceling out the original noise and achieving the purpose of noise reduction. It aims to eliminate low-frequency noise.

[0003] In traditional active noise reduction control methods, the LMS algorithm is easy to implement, but it has poor robustness and low convergence accuracy in practical applications. The FxLMS algorithm with the introduction of the filtered -x signal and its improvements can improve the convergence speed or reduce the steady-state error. However, both of them are highly dependent on the performance, number and arrangement of the secondary path sensors in engineering practice.

[0004] CN202010862334.5 mentions a convolutional-fuzzy neural network method for actively controlling global spatial noise in vehicles. The method includes setting up multiple secondary paths around the vehicle's noise reduction zone; collecting residual noise signals from each secondary path; using a convolutional-fuzzy neural network for offline identification to obtain a secondary path model; simultaneously, using an adaptive active noise control algorithm for the secondary paths to online correct controller parameters; and finally outputting multi-directional noise cancellation signals. This invention uses a convolutional-fuzzy neural network to identify the inverse model of the object, providing an effective method for identifying nonlinear noise in the global space of vehicles. It utilizes the nonlinear approximation capability of the convolutional-fuzzy neural network to improve the identification accuracy of secondary paths; it employs an active feedback noise reduction system to establish a stable secondary path model; and it solves the problems of difficult and narrow bandwidth in global spatial noise control of vehicles. However, this method also suffers from problems related to the performance, quantity, and arrangement of secondary path sensors, and setting up multiple secondary path sensors also leads to increased costs, which is not conducive to the engineering practice of active noise reduction in large spaces inside vehicles. Summary of the Invention

[0005] The main technical problem to be solved by this invention is the difficulty in the engineering practice of vehicle noise reduction due to the dependence on multiple secondary path sensors in existing active noise reduction methods. This invention establishes a correlation model between noise sources at different locations and noise in the target noise reduction area, and provides an active noise reduction algorithm based on convolutional-fuzzy neural network without secondary path sensors.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] An active noise reduction algorithm for sensors without secondary paths based on a convolutional-fuzzy neural network is proposed. This algorithm combines a convolutional network (CNN) for computational fitting to obtain a virtual error signal to eliminate acoustic feedback and improve the correlation between the reference signal and the primary noise at the ideal noise reduction point. A fuzzy neural network (FNN) is used as the controller to iteratively calculate the output noise cancellation signal in real time, thereby improving computational speed and enhancing nonlinear active noise reduction capabilities. Specific steps include:

[0008] S1. Noise signals X(n) from multiple noise sources are collected using primary sensors.

[0009] X(n) = [x1(n), x2(n), ... x k (n)]

[0010] Where k is the number of primary sensors, x k (n) represents the noise signal collected by the k-th sensor;

[0011] S2. Determine the noise reduction point and collect the real noise signal r(n) using a sensor as the training signal for the convolutional network model. r(n) is denoted as...

[0012] r(n) = [r(n), r(n-1), ..., r(nl)]

[0013] Where l is the length of the noise signal sequence;

[0014] S3. The noise signals X(n) and r(n) from S1 are fed into the convolutional and pooling layers of the convolutional network for iterative calculation, and the virtual error signal o(n) is output through the activation function.

[0015] o(n) = f(∑X(n)*ω(n)+b(n))

[0016] Where f(·) is the activation function, ω(n) is the weight of the convolutional network, and b(n) is the network bias;

[0017] S4. Input the virtual error signal o(n) output by the convolutional layer into the fuzzy neural network, calculate the membership value according to the fuzzy rules, activate the membership value using the multiplication function, and output the fuzzy layer output value u(n).

[0018]

[0019] Where j is the number of nodes in the fuzzy layer, with a total of H nodes, and λ j (n) represents the output of the convolutional network after passing through the membership function of the fuzzy layer, where p0 and p1 are fuzzy coefficients.

[0020] S5. The desired signal d(n) is obtained by filtering the noise signal collected at the target denoising point through the primary path P(z). The denoised signal y(n) is obtained by passing the fuzzy layer output value u(n) through the secondary path H(z). The objective function J(n) of the active denoising algorithm is then obtained.

[0021]

[0022] Furthermore, the aforementioned sensor is a microphone or transducer.

[0023] Furthermore, in step S1, x k (n) expands to x k (n)=[x k (n),x k (n-1),…x k [(nl)], where l is the length of the noise signal sequence.

[0024] Furthermore, to avoid the problem of neurons not learning in the negative interval caused by the ReLU activation function, a commonly used activation function in convolutional network models, the leaky-ReLU function is used as the activation function f(θ) in the convolutional network in step S3, expressed as:

[0025]

[0026] Where θ is the independent variable of the function, and a is a parameter in the interval (1,+∞).

[0027] Furthermore, in step S3, the training function E(n) of the convolutional network is...

[0028] E(n) = 1 / 2(r(n) - o(n)) 2 .

[0029] Furthermore, the weights ω(n) and network bias b(n) of the convolutional network are updated according to the training function E(n).

[0030]

[0031]

[0032] Furthermore, the activation result of o(n) entering the fuzzy layer in step S4 is:

[0033] λ j (n)=μ j (o(n)), j = 1, 2, ..., H

[0034] Where j is the number of nodes in the fuzzy layer, with a total of H nodes, μ j This is a membership function.

[0035] Furthermore, the membership function is expressed as

[0036]

[0037] Here, c(n) and σ(n) are the center and width of the membership function, respectively.

[0038] Furthermore, the desired signal d(n) in step S5 is:

[0039] d(n) = r(n) * P(z).

[0040] Furthermore, the noise-reduced signal y(n) in step S5 is:

[0041] y(n) = u(n) * H(z).

[0042] Furthermore, due to the time-varying characteristics of the noise signal, the objective function J(n) in step S5 is updated using gradient descent during the real-time training of the fuzzy neural network. That is, the denoised signal y(n) approaches the desired signal d(n) infinitely, and the objective function J(n) approaches zero. The updated fuzzy coefficients p(n), the center c(n) of the membership function, and the width σ(n) of the membership function are respectively...

[0043]

[0044]

[0045]

[0046] Where α and β are the learning rates of the fuzzy network.

[0047] Compared with existing technologies, the beneficial effects are:

[0048] This invention uses convolutional networks to estimate and analyze the relationship between noise sources at different locations and the ideal noise reduction point. Specifically, it leverages the translation invariance of convolutional layers to extract noise signal features and model them, while sharing weights to reduce computational load and improve speed. Multiple primary noise signals are fitted and estimated through a convolutional network to obtain an accurate virtual error signal o(n) that closely approximates the real error noise signal. This virtual error signal o(n) replaces the noise signal collected by secondary path sensors (microphones / voices), reducing the system's dependence on the number and performance of secondary path sensors (microphones / voices). Then, through the self-learning adjustment of parameters using a fuzzy neural network, a noise cancellation signal is output, improving noise reduction capability and achieving optimal noise reduction effect. Attached Figure Description

[0049] Figure 1 This is a block diagram illustrating the principle of an active noise reduction algorithm for sensors (microphones) without secondary paths based on a convolutional-fuzzy neural network.

[0050] Figure 2 A schematic diagram illustrating the implementation of active noise reduction for trains without secondary path sensors (microphones / voices);

[0051] Figure 3 The results of the active noise reduction experiment in the train driver's cab in Example 3;

[0052] Where 'a' represents the graph of the noise signal at the target noise reduction point and the virtual error signal calculated for fitting.

[0053] b is the error diagram between the primary noise signal r(n) and the virtual error signal o(n) at the target noise reduction point.

[0054] c represents the time-domain denoising result.

[0055] d represents the noise reduction result in the frequency domain. Detailed Implementation

[0056] The following examples further explain and clarify the invention, but the specific examples do not limit the invention in any way.

[0057] Example 1

[0058] This embodiment provides an active noise reduction algorithm for sensors without secondary paths based on convolutional-fuzzy neural networks, including the following steps:

[0059] S1. Noise signals X(n) from multiple noise sources are collected using primary sensors.

[0060] X(n) = [x1(n), x2(n), ... x k (n)]

[0061] Where k is the number of primary sensors, x k (n) represents the noise signal collected by the k-th sensor, x k (n) can be extended to x k (n)=[x k (n),x k (n-1),…x k [(nl)], where l is the length of the noise signal sequence.

[0062] S2. Determine the noise reduction point and set up a sensor to collect real noise signals as the ideal signal r(n) for training the convolutional network model. r(n) is denoted as...

[0063] r(n) = [r(n), r(n-1), ..., r(nl)]

[0064] Where l is the length of the noise signal sequence;

[0065] S3. Input the noise signals X(n) and r(n) in S1 into the convolutional and pooling layers of the convolutional network for iterative calculation, and output the virtual error signal o(n) through the ReLU activation function.

[0066] Constructing a virtual error signal o(n) actually involves using a convolutional neural network to build a relationship between the reference signals collected by multiple primary sensors and the primary noise signal at the target noise reduction point. This replaces the secondary sensors, eliminates the acoustic feedback problem, avoids algorithm divergence, and does not "pollute" the primary noise signal.

[0067] S4. The virtual error signal o(n) output from the convolutional layer is fed into the fuzzy neural network. The fuzzy neural network is then used for online tracking control of the uncertain nonlinear system to improve control performance. o(n) enters the fuzzy layer, and membership values ​​are calculated according to fuzzy rules. A multiplication function is then used to activate the membership values, and the output value u(n) of the fuzzy layer is given.

[0068]

[0069] Where j is the number of nodes in the fuzzy layer, with a total of H nodes, and p0 and p1 are fuzzy coefficients.

[0070] The activation result of o(n) entering the fuzzy layer is:

[0071] λ j (n)=μ j (o(n)), j = 1, 2, ..., H

[0072] Where j is the number of nodes in the fuzzy layer, with a total of H nodes, μ j Membership function μ j Represented as

[0073]

[0074] Here, c(n) and σ(n) are the center and width of the membership function, respectively.

[0075] S5. The desired signal d(n) is obtained by filtering the noise signal collected at the target denoising point through the primary path P(z). The denoised signal y(n) is obtained by passing the fuzzy layer output value u(n) through the secondary path H(z). The objective function J(n) of the active denoising algorithm is then obtained.

[0076]

[0077] The desired signal d(n) is:

[0078] d(n) = r(n) * P(z).

[0079] The noise-reduced signal y(n) is:

[0080] y(n) = u(n) * H(z).

[0081] Due to the time-varying characteristics of the noise signal, the objective function J(n) in step S5 is updated using gradient descent during the real-time training of the fuzzy neural network. That is, the denoised signal y(n) approaches the desired signal d(n) infinitely, and the objective function J(n) approaches zero. The updated fuzzy coefficients p(n), the center c(n) of the membership function, and the width σ(n) of the membership function are respectively...

[0082]

[0083]

[0084]

[0085] Example 2

[0086] This embodiment provides an active noise reduction algorithm for sensors without secondary paths based on convolutional-fuzzy neural networks, such as... Figure 1 As shown, it includes the following steps:

[0087] S1. Noise signals X(n) from multiple noise sources are collected using primary sensors.

[0088] X(n) = [x1(n), x2(n), ... x k (n)]

[0089] Where k is the number of primary sensors, x k (n) represents the noise signal collected by the k-th sensor, x k (n) can be extended to x k (n)=[x k (n),x k (n-1),…x k [(nl)], where l is the length of the noise signal sequence.

[0090] S2. Determine the noise reduction point and set up a sensor to collect real noise signals as the ideal signal r(n) for training the convolutional network model. r(n) is denoted as...

[0091] r(n) = [r(n), r(n-1), ..., r(nl)]

[0092] Where l is the length of the noise signal sequence;

[0093] S3. The noise signals X(n) and r(n) from S1 are fed into the convolutional and pooling layers of the convolutional network for iterative calculation, and the virtual error signal o(n) is output through the activation function.

[0094] o(n) = f(∑X(n)*ω(n)+b(n))

[0095] Where f(·) is the activation function, ω(n) is the weight of the convolutional network, and b(n) is the network bias.

[0096] To avoid the problem of neurons not learning in the negative region caused by the ReLU activation function, which is commonly used in convolutional network models, a leaky-ReLU function is introduced as the activation function f(θ), denoted as:

[0097]

[0098] Where θ is the independent variable of the function, and a is a parameter in the interval (1,+∞).

[0099] In the convolutional network module, its training function E(n) is:

[0100] E(n) = 1 / 2(r(n) - o(n)) 2

[0101] The weights ω(n) and network bias b(n) of the convolutional network are updated according to the training function E(n).

[0102]

[0103]

[0104] Constructing a virtual error signal o(n) actually involves using a convolutional neural network to build a relationship between the reference signals collected by multiple primary sensors and the primary noise signal at the target noise reduction point. This replaces the secondary sensors, eliminates the acoustic feedback problem, avoids algorithm divergence, and does not "pollute" the primary noise signal.

[0105] S4. The virtual error signal o(n) output from the convolutional layer is fed into the fuzzy neural network. The fuzzy neural network is then used for online tracking control of the uncertain nonlinear system to improve control performance. o(n) enters the fuzzy layer, and membership values ​​are calculated according to fuzzy rules. A multiplication function is then used to activate the membership values, and the output value u(n) of the fuzzy layer is given.

[0106]

[0107] Where j is the number of nodes in the fuzzy layer, with a total of H nodes, and p0 and p1 are fuzzy coefficients.

[0108] The activation result of o(n) entering the fuzzy layer is:

[0109] λ j (n)=μ j (o(n)), j = 1, 2, ..., H

[0110] Where j is the number of nodes in the fuzzy layer, with a total of H nodes, μ j Membership function μ j Represented as

[0111]

[0112] Here, c(n) and σ(n) are the center and width of the membership function, respectively.

[0113] S5. The desired signal d(n) is obtained by filtering the noise signal collected at the target denoising point through the primary path P(z). The denoised signal y(n) is obtained by passing the fuzzy layer output value u(n) through the secondary path H(z). The objective function J(n) of the active denoising algorithm is then obtained.

[0114]

[0115] The desired signal d(n) is:

[0116] d(n) = r(n) * P(z).

[0117] The noise-reduced signal y(n) is:

[0118] y(n) = u(n) * H(z).

[0119] Due to the time-varying characteristics of the noise signal, the objective function J(n) in step S5 is updated using gradient descent during the real-time training of the fuzzy neural network. That is, the denoised signal y(n) approaches the desired signal d(n) infinitely, and the objective function J(n) approaches zero. The updated fuzzy coefficients p(n), the center c(n) of the membership function, and the width σ(n) of the membership function are respectively...

[0120]

[0121]

[0122]

[0123] Where α and β are the learning rates of the fuzzy network.

[0124] Example 3

[0125] This embodiment is an active noise reduction experiment for trains without secondary path sensors, such as... Figure 2As shown, firstly, sensors are arranged in the train cabin at a set height and position. Sensors 1-3 are primary sensors, collecting signals from the primary noise source; sensor 4 collects the training signal for the convolutional-fuzzy neural network target; and sensor 5 is a test sensor, detecting noise signals emitted by the secondary sound source. Secondly, the sensors are connected to the controller via a signal conditioner; the secondary sound source is implemented using a speaker, which is also connected to the controller. Then, the active noise reduction system based on the convolutional-fuzzy neural network without secondary path sensors uses an NI Ethernet chassis as the controller hardware. The active noise reduction algorithm based on the convolutional-fuzzy neural network without secondary path sensors is written in C language, compiled into a dynamic link library, and called by the LabVIEW source program; the board is configured to simulate input and output. Finally, the primary noise signals collected by sensors 1-3 are input into the system to calculate the virtual error signal o(n), which is compared with the secondary path noise signal r(n) collected by sensor 4 at the target noise reduction area, and finally the noise cancellation signal y(n) is calculated.

[0126] A real-vehicle noise reduction experiment was conducted in the train driver's cab. The primary noise signals collected by sensors 1-3 were input into the system to calculate the virtual error signal o(n). This was compared with the primary noise signal r(n) at the target noise reduction point collected by sensor 4. According to the algorithm described in Example 2, the noise cancellation signal y(n) was finally calculated, and the result is as follows. Figure 3 As shown.

[0127] like Figure 3 As shown in (a), the noise signal r(n) of the target noise reduction point collected by the primary sensor 4 and the obtained virtual error signal o(n) show that the calculated and fitted virtual error signal o(n) basically fits the primary noise signal r(n) collected by the sensor 4.

[0128] like Figure 3 (b) shows the error between the virtual error signal o(n) and the primary noise signal r(n) at the target noise reduction point. It can be seen that the error is between (-1, 1) (in dB), that is, the reference signal is strongly correlated with the primary noise signal at the target noise reduction point.

[0129] like Figure 3 (c) shows the time-domain denoising results. It can be seen that the denoising effect is good except for the noise increase at a few points. The overall noise reduction is about 2.2dB, and the highest noise reduction can reach 8.2dB.

[0130] like Figure 3 The frequency domain analysis shown in (d) also shows a significant noise reduction effect, with the noise reduction effect reaching up to 18dB near the peak at 150Hz to 260Hz.

[0131] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. An active noise reduction algorithm for sensors without secondary paths based on convolutional-fuzzy neural networks, characterized by the following steps: include: S1. Collect noise signals at the noise source. X(n) for in, k It is the number of primary sensors. This represents the noise signal collected by the k-th sensor. S2. Determine the noise reduction point and collect the noise signal as the ideal signal for training the convolutional network model. r(n) , r(n) Represented as in, l It is the length of the noise signal sequence; S3. Combine the noise signal X(n) from S1 with... r(n) The convolutional and pooling layers of the convolutional network undergo iterative computation; the training function of the convolutional network... E(n) for: Output virtual error signal o(n) for: in, f (•) It is an activation function. ω (n) b(n) These are the weights of the convolutional network, and these are the network biases. The weights of the convolutional network... ω (n) and network bias b(n) Based on the training function E(n) Updated to: ； S4. Convert the virtual error signal output by the convolutional layer. o(n) The fuzzy neural network is fed into the system, membership values ​​are calculated, and a multiplication function is used to activate the membership values. The output value of the fuzzy layer is then output. u(n) for Where j represents the number of nodes in the fuzzy layer, with a total of H nodes. λ j (n) This is the result of the convolutional network output signal passing through the membership function of the fuzzy layer. p 0 and p 1 is the fuzzy coefficient; S5. Through the primary pathway P(z) The desired signal is obtained by filtering the noise signal collected at the target noise reduction point. d ( n ), output values ​​of the blur layer u(n) The noise-reduced signal obtained after passing through the secondary path H(z) y ( n The objective function of the active noise reduction algorithm is obtained. J(n) for 。 2. The active noise reduction algorithm for sensors without secondary paths based on convolutional-fuzzy neural networks according to claim 1, characterized in that, In step S1 Expand to , l It is the length of the noise signal sequence.

3. The active noise reduction algorithm for sensors without secondary paths based on convolutional-fuzzy neural networks according to claim 1, characterized in that, In step S3, the leaky-ReLU function is used as the activation function in the convolutional network. f(θ) , represented as in, θ The independent variable of the function, a It is a parameter in the interval (1, +∞).

4. The active noise reduction algorithm for sensors without secondary paths based on convolutional-fuzzy neural networks according to claim 1, characterized in that, In step S4 o(n) The activation result of entering the blur layer is in, j The total number of nodes in the fuzzy layer is [number]. H 1 node μ j This is a membership function.

5. The active noise reduction algorithm for sensors without secondary paths based on convolutional-fuzzy neural networks according to claim 4, characterized in that, Membership function is represented as in, c ( n )and σ ( n ) are the center and width of the membership function, respectively.

6. The active noise reduction algorithm for sensors without secondary paths based on convolutional-fuzzy neural networks according to claim 1, characterized in that, Desired signal in step S5 d ( n )for: 。 7. The active noise reduction algorithm for sensors without secondary paths based on convolutional-fuzzy neural networks according to claim 1, characterized in that, Noise reduction signal in step S5 y ( n )for: 。 8. The active noise reduction algorithm for sensors without secondary paths based on convolutional-fuzzy neural networks according to claim 1, characterized in that, In step S5, the objective function J(n) is updated using gradient descent during the real-time training of the fuzzy neural network, and the fuzzy coefficients... p(n) The center of membership function c ( n ) and the width of the membership function σ ( n ) are respectively in, α and β It is the learning rate of a fuzzy network.

Images