Distributed multichannel active noise control system and method based on time domain neural network

Abstract

The invention discloses a distributed multi-channel active noise control system and method based on a time domain neural network. The distributed multi-channel active noise control system comprises a reference microphone, a loudspeaker, an error microphone and a controller. The reference microphone is close to a noise source, the loudspeakers are located in a noise reduction area, and each loudspeaker is provided with an error microphone and a channel controller. The controller is formed by connecting a self-adaptive controller and a neural network controller in series. The reference microphone and the error microphone are connected with the input end of the adaptive controller, and the loudspeaker is connected with the output end of the neural network controller. The self-adaptive controller adopts a distributed control strategy and estimates primary channel impulse response in real time through a self-adaptive algorithm, and the neural network controller models a secondary channel inverse model offline through a secondary channel measurement and controllable neural network training method. The system and the method can effectively reduce modeling errors and crosstalk interference, track noise source position changes in real time, and improve noise reduction performance in a complex acoustic environment.

Description

Technical Field

[0001] The present invention relates to the research field of noise control technology, and in particular to a decentralized multi-channel active noise control system and method based on a time domain neural network. Background Art

[0002] With the rapid development of modern industry and transportation, noise pollution has become a problem that seriously affects human health and quality of life. As an effective noise control method, active noise control (ANC) technology has been widely used in the field of low-frequency noise elimination. Compared with single-channel systems, multi-channel ANC systems can achieve better noise elimination effects in a larger space by configuring multiple speakers and error microphones. However, the control strategy of multi-channel ANC systems still faces many challenges in practical applications.

[0003] In multi-channel ANC systems, centralized control methods excel in noise suppression and stability, but their high computational complexity limits their practical applications. In contrast, decentralized control algorithms have been widely used in practical engineering due to their high computational efficiency. However, the inherent secondary channel coupling problem in decentralized systems may lead to a decrease in control accuracy and even affect the stability of the entire system when some channels diverge.

[0004] In recent years, the introduction of deep learning technology has provided new ideas for the development of ANC systems. Researchers have tried to apply neural networks to ANC systems to solve the limitations of traditional methods in dealing with nonlinear problems. For example, convolutional recurrent networks are used to estimate the spectrum of the cancellation signal directly from the reference signal and implement multi-channel expansion in multi-channel ANC systems (H. Zhang and DL Wang, A deep learning method to multi-channel active noise control, Proc. of INTERSPEECH 2021, Brno, Czechia, 2021). However, the above methods have two main limitations: one is that it is difficult to achieve real-time control, and the other is that they cannot adapt to changes in the location of the sound source.

[0005] To solve the problem of tracking moving sound sources, researchers proposed a hybrid system combining an adaptive filter and a time-domain convolutional recurrent network, separating the nonlinear part of the secondary speaker through a neural network (D. Chen, L. Cheng, D. Yao, J. Li, and Y. Yan, A secondary path decoupled active noise control algorithm based on deep learning, IEEE Signal Proc. Lett., 29 (2022), 234-238). Another study used two gated convolutional recurrent networks to model the inverse impulse response of the secondary path, so that the adaptive controller only needs to track the main path (J. Park, JH Choi, Y. Kim, and JH Hang, HAD-ANC: A Hybrid system comprising an adaptive filter and deep neural networks for active noise control, INTERSPEECH, 2023). These methods have improved the adaptability of the system to a certain extent, but they are mainly aimed at single-channel systems and fail to give full play to the advantages of multi-channel systems in terms of control range.

[0006] Although neural network algorithms have made significant progress in solving the nonlinear modeling problem of loudspeakers and noise signals, there is still an overlooked problem in the filter-x type algorithms used in traditional decentralized structures, that is, the inverse modeling of the secondary path is usually approximated by a finite impulse response linear solution. This approximation will introduce significant modeling errors in a reverberant environment, and the crosstalk effect between channels further degrades the control performance of the system. Therefore, how to achieve efficient and accurate inverse modeling of the secondary path in a multi-channel ANC system while solving nonlinear problems and crosstalk effects is still a difficult problem that needs to be solved in the current technical field. Summary of the invention

[0007] In a reverberant environment, traditional adaptive controllers are estimated to introduce modeling errors, resulting in a decrease in system control performance, and the crosstalk phenomenon between multiple channels further exacerbates the performance degradation. In order to solve the above technical problems, the main purpose of the present invention is to provide a decentralized multi-channel active noise control system and design method based on a time domain neural network, which uses the decoupled neural network least mean square algorithm (DecNet-LMS) proposed by the present invention.

[0008] The main purpose of the present invention is achieved through the following technical solutions:

[0009] A decentralized multi-channel active noise control system and design method based on a time domain neural network, the system comprising: a reference microphone, a loudspeaker, an error microphone, and a controller; wherein:

[0010] The reference microphone is arranged at one end close to the noise source;

[0011] The loudspeaker is arranged near the area to be noise-reduced, and is used to output a control signal to offset the noise signal, and the error microphone is arranged at the rear end of each loudspeaker;

[0012] The controller includes an adaptive controller and a neural network controller, which are used to process the signals of the reference microphone and the error microphone and generate the control signal;

[0013] The reference microphone and the error microphone are connected to the input of the adaptive controller;

[0014] The adaptive controller is connected to the neural network controller, and the output end of the neural network controller is connected to the speaker.

[0015] Further, the neural network controller is used to offline model the inverse model of the secondary channel and suppress the crosstalk between channels;

[0016] The adaptive controller is used for estimating the transfer function of the primary channel in real time.

[0017] Furthermore, the neural network controller includes the following modules:

[0018] The secondary channel measurement module is used to collect the transfer function of the secondary channel through the system identification method to provide basic data for the neural network model training;

[0019] The training module uses the collected data to train the neural network and offline model the secondary channel inverse model with nonlinear mapping capabilities;

[0020] The neural network control module is used to deploy the trained neural network in the system and calculate the output control signal in real time to drive the speaker to achieve effective noise cancellation.

[0021] Furthermore, the adaptive controller comprises the following modules:

[0022] Initialization module, used to configure the initial parameters of the system, including adaptive step size μ, filter order L, and compensation delay τ, to ensure the convergence and stability of the system;

[0023] A local adaptive updating module, used to update the coefficients of the adaptive controller according to real-time input data to adapt to environmental changes;

[0024] The filter output module is used to filter the input noise signal and output it to the neural network controller.

[0025] Furthermore, the local adaptive update module of the adaptive controller uses a distributed structure, and a sub-controller is set corresponding to each error microphone. Each sub-controller independently updates the coefficient according to the residual signal received by the corresponding error microphone, and the coefficient update is realized by one of the following methods: minimum mean square error algorithm; normalized minimum mean square error algorithm; wherein the selection of the algorithm is optimized based on the statistical characteristics of the input noise and the real-time requirements of the system.

[0026] Furthermore, the input and output configurations of the adaptive controller and the neural network controller are as follows:

[0027] The input of the adaptive controller includes: a noise signal from the reference microphone and a residual signal from the error microphone;

[0028] The output of the adaptive controller is a signal after being filtered by the controller and serves as one of the inputs of the neural network controller;

[0029] The input of the neural network controller includes: the filtered output signal from the adaptive controller, the secondary channel transfer function data collected by the secondary channel measurement module, and the residual signal fed back by the error microphone;

[0030] The output of the neural network controller is a control signal for driving the speaker, which is used to generate a reverse sound wave with a phase opposite to that of the noise signal to achieve effective noise cancellation.

[0031] Furthermore, the secondary channel measurement module collects the transfer function of the secondary channel through the following steps:

[0032] Outputting a preset test signal through the speaker, and collecting a response signal of the secondary channel using the error microphone;

[0033] Calculate the transfer function of the secondary channel based on the system identification method;

[0034] The calculated transfer function data is stored in the secondary channel measurement module as basic data for neural network training and used for offline modeling of the inverse model of the secondary channel.

[0035] The secondary channel response measurement and its inverse model training are carried out channel by channel.

[0036] Furthermore, the neural network controller adopts a multi-channel decoupled neural network (DecNet) structure, which specifically includes the following features:

[0037] The multi-channel decoupled neural network includes K sub-controllers, and the kth (k=1, 2, ..., K) sub-controller is denoted as NN k , used to generate a driving signal for the kth speaker;

[0038] Each sub-controller NN k The internal structure is composed of K three-layer deep neural networks, which process the output signal of the i-th (i=1,2,...,K) adaptive controller respectively, denoted as nn ik ;

[0039] Each nn ik It consists of an input layer, a hidden layer and an output layer, wherein the input layer contains D neurons, corresponding to the output signals of the i-th adaptive controller at the current moment and the previous D-1 moments; the hidden layer contains H neurons, and uses the Sigmoid activation function for nonlinear mapping; the output layer contains 1 neuron; the values ​​of D and H are determined by the secondary channel transfer function collected by the secondary channel measurement module;

[0040] The connection between each layer is a fully connected structure, that is, each neuron is connected to all neurons in the previous layer;

[0041] Each NN k All nn in the controller ik The output signals are linearly superimposed to generate driving signals for the corresponding speakers;

[0042] The secondary channel response measurement and its inverse model training are carried out channel by channel. First, the first secondary source and its corresponding secondary channel are measured and the inverse model training is carried out, and then the second secondary source and its corresponding secondary channel are measured and the inverse model training is carried out until the inverse models of all channels are completed.

[0043] Furthermore, the steps of training the neural network in the training module are as follows:

[0044] Step 1: Use a white noise signal with a mean of zero and a variance of 1 as the training input signal y wk (n), the input signal y wk (n) through the neural network model nn kk , and then through the secondary channel s between the kth loudspeaker and the kth error microphone kk , and obtain the secondary noise signal y at the error microphone k (n); introduce an appropriate time delay τ, set the secondary noise expected signal to the delayed signal; use the mean square error (MSE) between the secondary noise signal and the expected signal as the cost function, and use the Adam optimizer to train the inverse model nn kk , so that nn kk Approximately skk The inverse model of

[0045] Step 2: The inverse model nn kk The training results are used as the initial parameters, and the input signal is passed through nn kk Then it passes through the crossover secondary channel s between the kth loudspeaker and the ith error microphone ik Get the first output signal; pass the input signal through nn ki After passing through the secondary channel s ii , obtain the second output signal; superimpose the first output signal and the second output signal, and set the expected signal to 0; use the mean square error (MSE) between the superimposed signal and the expected signal as the cost function, and use the Adam optimizer to train the model nn ki ;

[0046] Step 3: Repeat step 2 so that i traverses K error microphones and trains the corresponding model nn respectively. ki ;

[0047] Step 4: Repeat steps 1 to 3, so that k traverses K speakers and trains the corresponding inverse model nn respectively. kk and model nn ki , complete the training of multi-channel decoupled neural network;

[0048] Step 5: All the inverse models nn that have been trained kk and model nn ki As the initial parameters, a multi-channel decoupled neural network is constructed; the performance of the trained neural network model is evaluated using the test set. If the model performance does not reach the preset threshold, the network structure or parameters are adjusted and retrained;

[0049] Step 6: deploying the trained neural network model to the neural network control module for real-time calculation and output of control signals to drive the speaker to achieve effective noise cancellation;

[0050] After the training is completed, the multi-channel decoupled neural network model can be approximately represented as a generalized inverse system of secondary channels with delays, and the signal passing through the network and then through the secondary channels can be regarded as passing through a time delay system.

[0051] Furthermore, an appropriate delay τ is introduced and the delay τ is compensated back to the adaptive controller; wherein the set delay τ is slightly larger than the acoustic delay of the secondary channel and much smaller than the acoustic delay of the primary channel.

[0052] Furthermore, the coefficient update method of the sub-controller is:

[0053] w k (n+1)=wk (n)-μe k (n)x w (n-τ),

[0054] Where μ is the controller step size, x w (n) is the reference signal vector received by the reference microphone, x w (n)=[x(n),...,x(n-L+1)] T , e k (n) is the residual signal received by the kth error output microphone, and L is the controller {w k (n)}, n is the length of the adaptive filter, n is the current iteration number, and τ is the compensation delay set by the initialization module.

[0055] Furthermore, the step size μ of the adaptive controller needs to satisfy:

[0056] 0<μ<μ max ,

[0057] Among them, μ max represents the maximum step size parameter of the controller, μ max =1 / trace(R xx ), R xx Represents the autocorrelation matrix of the noise source signal.

[0058] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0059] The present invention can effectively offset a wider range of noise, effectively reduce modeling errors and crosstalk interference, track the position changes of noise sources in real time, and improve the noise reduction performance in complex acoustic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is a schematic diagram of the layout of a decentralized multi-channel active noise control system based on a time-domain neural network according to an embodiment of the present invention.

[0061] Figure 2 It is a schematic diagram of the overall architecture of a decentralized multi-channel active noise control system based on a time-domain neural network according to an embodiment of the present invention.

[0062] Figure 3 The present invention is a schematic diagram of the module composition of a controller of a decentralized multi-channel active noise control system based on a time-domain neural network according to an embodiment of the present invention.

[0063] Figure 4 It is a structural schematic diagram of a neural network controller of a decentralized multi-channel active noise control system based on a time-domain neural network according to an embodiment of the present invention.

[0064] Figure 5 The present invention is a schematic diagram of a training process of a neural network of a decentralized multi-channel active noise control system based on a time-domain neural network according to an embodiment of the present invention.

[0065] Figure 6 It is a schematic diagram for comparing the noise reduction performance of a decentralized multi-channel active noise control system based on a time-domain neural network according to an embodiment of the present invention.

[0066] Figure 7 This is a noise tracking performance test result diagram of a decentralized multi-channel active noise control system based on a time-domain neural network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0067] The present invention is further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0068] Example:

[0069] Figure 1 1 is a schematic diagram of the layout of a decentralized multi-channel active noise control system based on a time domain neural network according to an embodiment of the present invention. In this embodiment, the system adopts a dual-channel configuration, that is, it includes two speakers as secondary sound sources, two error microphones, one reference microphone and a controller. Assume that the arrangement is carried out in an ordinary room with a length, width and height of 2.8 meters, 2.3 meters and 2.1 meters respectively, and a door is opened on one side wall of the room, so that a speaker is attached outside the door as the main source of external noise. Figure 1 A cross-section at 1.5 meters above the ground is shown.

[0070] The specific arrangement is as follows:

[0071] A reference microphone (marked as a) is installed near the door (i.e., the noise source), about 0.3 m from the door and 1.5 m from the ground, for real-time noise signal collection;

[0072] Two speakers (marked as b and c) are installed in the noise reduction area of ​​the room, with a horizontal interval of 0.6 meters between the two speakers and 1.5 meters from the ground to generate anti-noise signals to offset the noise;

[0073] Two error microphones (marked as d and e) are located directly behind the two speakers, 0.4 meters away from the speakers and 1.5 meters above the ground, to collect residual noise signals;

[0074] The controller is used to process the noise signal collected by the reference microphone and the residual noise signal detected by the error microphone in real time, and generate a control signal through a built-in control algorithm to drive the speaker to generate a signal with a phase opposite to the noise signal, thereby achieving noise cancellation.

[0075] Figure 2 This is a schematic diagram of the overall architecture of a decentralized multi-channel active noise control system based on a time domain neural network according to an embodiment of the present invention. In the figure, x(n) represents the reference signal received by the reference microphone at time n, and the signal is transmitted through two paths:

[0076] On the one hand, the transfer function p between the reference signal via the noise source to the kth error microphone is k After that, the noise signal d to be eliminated is formed at the kth error microphone. k (n);

[0077] On the other hand, the reference signal is first passed through the adaptive controller w k (n) filtering to obtain the output signal y of the adaptive controller wk (n). The output signal y wk (n) Then, a fixed parameter controller based on a neural network is used to generate a driving signal u k (n), and finally drives the kth (k = 1, 2, ..., K) speaker to emit secondary sound. The secondary sound passes through the linear acoustic channel s between the i-th speaker and the k-th error microphone. ki , a cancellation sound y is generated at the kth error microphone k (n).

[0078] After the primary noise and the cancellation sound are superimposed, the residual noise signal collected by the kth error microphone is e k (n). Each adaptive controller w k (n) According to the residual signal e received by the corresponding error microphone k (n) to update the coefficients and dynamically update its filter coefficients to minimize the residual noise energy.

[0079] Figure 3 FIG. 1 is a schematic diagram of the module composition of a controller of a decentralized multi-channel active noise control system based on a time domain neural network according to an embodiment of the present invention. Figure 3 As shown, the controller includes a neural network controller and an adaptive controller, wherein the neural network controller part includes the following modules:

[0080] The secondary channel measurement module is used to collect the transfer function of the secondary channel through the system identification method to provide basic data for the neural network model training;

[0081] A training module, which uses the collected data to train a neural network to offline model a secondary channel inverse model with nonlinear mapping capabilities;

[0082] The neural network control module is used to deploy the trained neural network model in the system and output the control signal through real-time calculation to drive the speaker to achieve effective noise cancellation.

[0083] The adaptive controller part includes the following modules:

[0084] Initialization module, used to configure the initial parameters of the system, including adaptive step size, filter order L, and compensation delay, to ensure the convergence and stability of the system;

[0085] A local adaptive update module is used to update the coefficients of the adaptive controller according to real-time input data to adapt to environmental changes;

[0086] The filter output module is used to filter the input noise signal and output it to the neural network controller.

[0087] Figure 4 The structure diagram of a neural network controller of a decentralized multi-channel active noise control system based on a time domain neural network according to an embodiment of the present invention is shown in FIG. The network includes K controllers, and the kth (k=1, 2, ..., K) neural network controller is denoted as NN k , used to generate the driving signal for the kth speaker. Considering the crosstalk between the channels, the controller NN k The internal structure is composed of K three-layer deep neural networks, which process the output signal of the i-th (i=1,2,...,K) adaptive controller respectively, denoted as nn ik .like Figure 5 As shown in the upper right corner, each independent nn ik It consists of an input layer, a hidden layer, and an output layer. The input layer contains D neurons, the hidden layer contains H neurons, and uses the Sigmoid activation function for nonlinear mapping. The output layer contains 1 neuron. The connection between each layer is fully connected, that is, each neuron is connected to all neurons in the previous layer.

[0088] Since the proposed multi-channel decoupled neural network processes the signal in the time domain, nn ik The D neurons in the input layer correspond to the output signals of the i-th adaptive controller at the current moment and the previous D-1 moments. k All nn ik The output signal of the controller is then linearly superimposed to obtain the output of the controller as the driving signal of the corresponding speaker.

[0089] Figure 5 This is a schematic diagram of a training process of a neural network of a decentralized multi-channel active noise control system based on a time domain neural network according to an embodiment of the present invention. The process is as follows:

[0090] Step 1: Use a white noise signal with a mean of zero and a variance of 1 as the training input signal y wk (n), the input signal y wk (n) Through the neural network model nn kk , and then through the secondary channel s between the kth loudspeaker and the kth error microphone kk , and the secondary noise signal y at the error microphone is obtained k (n); introduce an appropriate time delay τ and set the secondary noise expected signal to the delayed signal; use the mean square error (MSE) between the secondary noise signal and the expected signal as the cost function and use the Adam optimizer to train the inverse model nn kk , so that nn kk Approximately s kk The inverse model of

[0091] Step 2: Inverse model nn kk The training results are used as the initial parameters, and the input signal is passed through nn kk Then it passes through the crossover secondary channel s between the kth loudspeaker and the ith error microphone ik Get the first output signal; pass the input signal through nn ki After passing through the secondary channel s ii , get the second output signal; superimpose the first output signal and the second output signal, and set the expected signal to 0; use the mean square error (MSE) between the superimposed signal and the expected signal as the cost function, and use the Adam optimizer to train the model nn ki ;

[0092] Step 3: Repeat step 2 so that i traverses K error microphones and trains the corresponding model nn respectively. ki ;

[0093] Step 4: Repeat steps 1 to 3, so that k traverses K speakers and trains the corresponding inverse model nn respectively. kk and model nn ki , complete the training of multi-channel decoupled neural network;

[0094] Step 5: All the inverse models nn that have been trained kk and model nn ki As the initial parameters, a multi-channel decoupled neural network is constructed; the performance of the trained neural network model is evaluated using the test set. If the model performance does not reach the preset threshold, the network structure or parameters are adjusted and retrained;

[0095] Step 6: Deploy the trained neural network model to the neural network control module for real-time calculation and output of control signals to drive the speaker to effectively cancel out the noise;

[0096] After training, the multi-channel decoupled neural network model can be approximately represented as a generalized inverse system of secondary channels with delays. The signal passing through the network and then through the secondary channels can be regarded as passing through a time delay system.

[0097] After training is completed, the multi-channel decoupled neural network model can be approximately represented as a generalized inverse system of secondary channels with delays. The signal passing through the network and then through the secondary channel can be regarded as passing through a time delay system.

[0098] Furthermore, the system introduces an appropriate delay τ to ensure the causality of the neural network controller and compensates the delay τ back to the adaptive controller. The delay τ should be slightly larger than the acoustic delay of the secondary channel but much smaller than the acoustic delay of the primary channel to meet the overall causality constraint of the system.

[0099] Furthermore, the local adaptive update module of the adaptive controller uses a decentralized structure, that is, a sub-controller is set corresponding to each error microphone, and each sub-controller independently updates the coefficient only according to the residual signal received by the corresponding error microphone, and realizes the coefficient update through one of the following algorithms:

[0100] Minimum mean square error algorithm;

[0101] Normalized minimum mean square error algorithm;

[0102] The algorithm is selected based on the statistical characteristics of the input noise and the real-time requirements of the system to ensure the convergence and noise reduction performance of the adaptive controller.

[0103] Furthermore, the coefficient update method of the sub-controller is:

[0104] w k (n+1)=w k (n)-μe k (n)x w (n-τ),

[0105] Where μ is the controller step size, x w (n) is the reference signal vector received by the reference microphone, x w (n)=[x(n),...,x(n-L+1)] T , e k (n) is the residual signal received by the kth error output microphone, and L is the controller {w k (n)}, n is the current iteration number, and τ is the compensation delay set by the initialization module.

[0106] Furthermore, the step size μ of the adaptive controller needs to satisfy:

[0107] 0<μ<μ max ,

[0108] Among them, μ max represents the maximum step size parameter of the controller, μ max =1 / trace(R xx ), R xx Represents the autocorrelation matrix of the noise source signal.

[0109] The present invention also tests the above dual-channel active noise control system and compares the performance of the controller using different algorithms with white noise and interior noise as reference signals. The adaptive controller length of all test methods is set to L=160. Figure 6 FIG. 1 is a schematic diagram comparing the noise reduction performance of a distributed multi-channel active noise control system based on a time domain neural network according to an embodiment of the present invention. Figure 6 For the white noise input in (a), compared with the adaptive controllers using the centralized filter xLMS algorithm (CFxLMS) and the decentralized filter xLMS algorithm (DCFxLMS), the controller using the DecNet-LMS algorithm proposed in the present invention can make the system converge to the steady-state excess mean square error value at the fastest speed, while the Deep MANC method using only the neural network controller achieves a lower modeling error (lower by 3dB). When the input signal becomes the interior noise, Figure 6 The results in (b) show similar performance.

[0110] Figure 7 This is a noise tracking performance test result diagram of a decentralized multi-channel active noise control system based on a time domain neural network according to an embodiment of the present invention. The position of the primary noise is changed at the 10th second to test the tracking performance of the algorithm. Figure 7 The results in show that the system based on the Deep MANC algorithm cannot track the changes of the primary path, while the systems based on the traditional CFxNLMS and DCFxNLMS algorithms are able to converge and slowly track the changes. In contrast, the DecNet-LMS system combining the neural network with the adaptive controller of the present invention maintains satisfactory noise reduction performance in both static and non-static situations.

[0111] In addition, under the setting of this example, the computational amount of each iteration of the main algorithms of the above different types of controllers is shown in Table 1. Where T is the number of time frames after the short-time Fourier transform. Since the Deep MANC method cannot achieve real-time processing, in order to evaluate its computational amount, the average computational amount of a single iteration is estimated by dividing the total computational amount by the total number of iterations N.

[0112] Table 1 Comparison of the computational effort per iteration of the main algorithms of different controllers

[0113]

[0114] The computational complexity of the control algorithm combining the adaptive controller and the neural network in this example is related to the complexity of the acoustic environment and the optimization performance. Taking into account the more complex acoustic scene learning, a larger number of neurons are currently set, which is an increase compared to the traditional ANC adaptive algorithm, but it is still lower than the average computational complexity of each iteration of the Deep MANC algorithm based on the neural network.

[0115] Therefore, the system proposed in the present invention effectively solves the coupling and nonlinear problems in the secondary channel by cascading the adaptive controller with the neural network controller. The algorithm can not only respond to the dynamic changes of the primary channel in real time, but also significantly reduce the computational complexity, and has higher practicality and scalability than other neural network-based methods.

[0116] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0117] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A decentralized multi-channel active noise control system based on time domain neural network, characterized in that: The system comprises: a reference microphone, a loudspeaker, an error microphone, and a controller; wherein: The reference microphone is arranged at one end close to the noise source; The loudspeaker is arranged near the area to be noise-reduced, and is used to output a control signal to offset the noise, and each loudspeaker is equipped with the error microphone and the channel controller; The controller includes an adaptive controller and a neural network controller, which are used to process the signals of the reference microphone and the error microphone and generate the control signal to drive the speaker; The reference microphone and the error microphone are connected to the input of the adaptive controller; The adaptive controller is connected in series with the neural network controller, and the output end of the neural network controller is connected to the speaker.

2. The decentralized multi-channel active noise control system based on time domain neural network according to claim 1, characterized in that: The neural network controller is used for offline modeling of the inverse model of the secondary channel and suppressing the crosstalk between channels; the adaptive controller is used for real-time estimation of the transfer function of the primary channel.

3. The decentralized multi-channel active noise control system based on time domain neural network according to claim 2 is characterized in that: The neural network controller includes the following modules: The secondary channel measurement module is used to collect the transfer function of the secondary channel through the system identification method to provide basic data for the neural network model training; The training module uses the collected data to train the neural network and offline model the secondary channel inverse model with nonlinear mapping capabilities; The neural network control module is used to deploy the trained neural network in the system and calculate the output control signal in real time to drive the speaker to achieve effective noise cancellation.

4. The decentralized multi-channel active noise control system based on time domain neural network according to claim 2 is characterized in that: The adaptive controller includes the following modules: Initialization module, used to configure the initial parameters of the system, including adaptive step size μ, filter order L, and compensation delay τ, to ensure the convergence and stability of the system; A local adaptive updating module, used to update the coefficients of the adaptive controller according to real-time input data to adapt to environmental changes; The filter output module is used to filter the input noise signal and output it to the neural network controller.

5. The control system according to claim 4, characterized in that: The local adaptive update module of the adaptive controller adopts a distributed structure. A sub-controller is set corresponding to each error microphone. Each sub-controller independently updates the coefficient according to the residual signal received by the corresponding error microphone, and the coefficient update is realized by one of the following methods: minimum mean square error algorithm; normalized minimum mean square error algorithm; wherein the selection of the algorithm is optimized based on the statistical characteristics of the input noise and the real-time requirements of the system.

6. A decentralized multi-channel active noise control system based on a time domain neural network according to any one of claims 1 to 5, characterized in that: The input and output configurations of the adaptive controller and the neural network controller are as follows: The input of the adaptive controller includes: each controller receives a noise signal from a path corresponding to the reference microphone and a residual signal from a path corresponding to the error microphone; The output of the adaptive controller is a signal after being filtered by the controller and serves as one of the inputs of the neural network controller; The input of the neural network controller includes: the filtered output signal from the adaptive controller, the secondary channel transfer function data collected by the secondary channel measurement module, and the residual signal fed back by the error microphone; The output of the neural network controller is a control signal for driving the speaker, which is used to generate a reverse sound wave with a phase opposite to that of the noise signal to achieve effective noise cancellation.

7. The decentralized multi-channel active noise control system based on time domain neural network according to claim 6 is characterized in that: The secondary channel measurement module collects the transfer function of the secondary channel through the following steps: Outputting a preset test signal through the speaker, and collecting a response signal of the secondary channel using the error microphone; Calculate the transfer function of the secondary channel based on the system identification method; The calculated transfer function data is stored in the secondary channel measurement module as basic data for neural network training and used for offline modeling of the inverse model of the secondary channel; The secondary channel response measurement and its inverse model training are carried out channel by channel.

8. The decentralized multi-channel active noise control system based on time domain neural network according to claim 7 is characterized in that: The neural network controller adopts a multi-channel decoupled neural network (DecNet) structure, which specifically includes the following features: The multi-channel decoupled neural network includes K sub-controllers, and the kth (k=1, 2, ..., K) sub-controller is denoted as NN k , used to generate a driving signal for the kth speaker; Each sub-controller NN k The internal structure is composed of K three-layer deep neural networks, which process the output signal of the i-th (i=1,2,...,K) adaptive controller respectively, denoted as nn ik ; Each nn ik It consists of an input layer, a hidden layer and an output layer, wherein the input layer contains D neurons, corresponding to the output signals of the i-th adaptive controller at the current moment and the previous D-1 moments; the hidden layer contains H neurons, and uses the Sigmoid activation function for nonlinear mapping; the output layer contains 1 neuron; the values ​​of D and H are determined by the secondary channel transfer function collected by the secondary channel measurement module; The connection between each layer is a fully connected structure, that is, each neuron is connected to all neurons in the previous layer; Each NN k All nn in the controller ik The output signals are linearly superimposed to generate driving signals for the corresponding speakers; The secondary channel response measurement and its inverse model training are carried out channel by channel. First, the first secondary source and its corresponding secondary channel are measured and the inverse model training is carried out, and then the second secondary source and its corresponding secondary channel are measured and the inverse model training is carried out until the inverse models of all channels are completed.

9. The control method of the system according to any one of claims 1 to 8, characterized in that: The steps of training the neural network in the training module are as follows: Step 1: Use a white noise signal with a mean of zero and a variance of 1 as the training input signal y wk (n), the input signal y wk (n) through the neural network model nn kk , and then through the secondary channel s between the kth loudspeaker and the kth error microphone kk , and obtain the secondary noise signal y at the error microphone k (n); introduce an appropriate time delay τ, set the secondary noise expected signal to the delayed signal; use the mean square error (MSE) between the secondary noise signal and the expected signal as the cost function, and use the Adam optimizer to train the inverse model nn kk , so that nn kk Approximately s kk The inverse model of Step 2: The inverse model nn kk The training results are used as the initial parameters, and the input signal is passed through nn kk Then it passes through the crossover secondary channel s between the kth loudspeaker and the ith error microphone ik Get the first output signal; pass the input signal through nn ki After passing through the secondary channel s ii , obtain the second output signal; superimpose the first output signal and the second output signal, and set the expected signal to 0; use the mean square error (MSE) between the superimposed signal and the expected signal as the cost function, and use the Adam optimizer to train the model nn ki ; Step 3: Repeat step 2 so that i traverses K error microphones and trains the corresponding model nn respectively. ki ; Step 4: Repeat steps 1 to 3, so that k traverses K speakers and trains the corresponding inverse model nn respectively. kk and model nn ki , complete the training of multi-channel decoupled neural network; Step 5: All the inverse models nn that have been trained kk and model nn ki As the initial parameters, a multi-channel decoupled neural network is constructed; the performance of the trained neural network model is evaluated using the test set. If the model performance does not reach the preset threshold, the network structure or parameters are adjusted and retrained; Step 6: deploying the trained neural network model to the neural network control module for real-time calculation and output of control signals to drive the speaker to achieve effective noise cancellation; After the training is completed, the multi-channel decoupled neural network model can be approximately represented as a generalized inverse system of secondary channels with delays, and the signal passing through the network and then through the secondary channels can be regarded as passing through a time delay system.

10. The control method according to claim 9, characterized in that: Introducing an appropriate delay τ and compensating the delay τ back to the adaptive controller; wherein the set delay τ is slightly larger than the acoustic delay of the secondary channel and much smaller than the acoustic delay of the primary channel; The coefficient update method of the sub-controller is: w k (n+1)=w k (n)-μe k (n)x w (n-τ), Where μ is the controller step size, x w (n) is the reference signal vector received by the reference microphone, x w (n)=[x(n),...,x(n-L+1)] T , e k (n) is the residual signal received by the kth error output microphone, and L is the controller {w k (n)}, n is the current iteration number, and τ is the compensation delay set by the initialization module; The step size μ of the adaptive controller must satisfy: 0<μ<μ max , Among them, μ max represents the maximum step size parameter of the controller, μ max =1 / trace(R xx ), R xx Represents the autocorrelation matrix of the noise source signal.

Images