An adaptive deep neural network-based active noise control system and method thereof

Abstract

This invention discloses an active noise control system and method based on an adaptive deep neural network. The system includes a reference signal acquisition unit, an error signal acquisition unit, a causal deep neural network controller, a lightweight adaptive module, and a parameter update unit. The method comprises the following steps: pre-training the causal deep neural network controller using an active noise control database; fixing the trained controller parameters and inserting a lightweight adaptive module with initial parameters of zero into the network; inputting the reference signal into the controller to generate a control signal to drive a secondary sound source to generate anti-noise, and using the error signal to construct the loss and calculate the gradient; freezing the main controller parameters during the control process, and only updating the embedded lightweight adaptive module online. This invention can improve robustness and control stability under complex operating conditions, reduce the dependence of deep neural network-based active noise control methods on large-scale training data, and improve noise suppression performance.

Description

Technical Field

[0001] This invention belongs to the technical field of active noise control, specifically relating to an adaptive deep neural network active noise control system and method based on real-time parameter updates. Background Technology

[0002] Active Noise Control (ANC) technology generates an anti-noise signal with an amplitude equivalent to but opposite phase to the original noise from a secondary sound source, thereby achieving noise cancellation within the target area. Due to its significant advantages in low-frequency noise suppression, ANC technology has been widely used in cabin noise control scenarios for vehicles such as automobiles and aircraft. With the development trends of lightweighting and electrification in automobiles, the noise reduction capabilities of traditional passive sound insulation methods in the low-frequency range are gradually being limited. Active Road Noise Control (ARNC) has become an important technical approach to improve in-vehicle acoustic comfort.

[0003] Existing ARNC systems typically employ a feedforward control structure, utilizing accelerometers to collect vehicle structural vibration signals as reference signals and generating anti-noise within the vehicle through speakers to create a localized quiet zone near the occupants' ears. Control algorithms based on adaptive filtering theory, such as the Filtered-x Least Mean Square (FxLMS) algorithm and its improved versions, are widely used due to their simple structure and good stability. However, in real-world road environments, road noise is influenced by various factors such as road surface type, vehicle speed changes, and traffic conditions, exhibiting significant time-varying and non-stationary characteristics. Traditional adaptive algorithms often suffer from slow convergence speed, limited steady-state performance, and strong parameter dependence under such complex conditions. Furthermore, traditional algorithms are based on linear models, and their noise reduction performance is affected by speaker nonlinearity and other nonlinearities.

[0004] To improve the control performance of ARNC systems in complex environments, deep neural networks (DNNs) have been introduced into the field of active noise control in recent years. Related methods typically utilize the powerful nonlinear modeling capabilities of neural networks to directly or indirectly replace traditional control filters, and under certain conditions, can break through the performance limits of linear models. However, most existing DNN-based ANC methods rely on offline training, with network parameters remaining fixed after deployment. When the actual operating environment differs from the distribution of training data, model performance is prone to significant degradation, making it difficult to effectively address the distribution drift problem prevalent in real-world road noise.

[0005] Therefore, how to maintain the high modeling capability of deep neural networks while introducing online adaptive mechanisms to improve the tracking ability of time-varying road noise is a key challenge in the current development of ARNC technology. Furthermore, ARNC systems have strict requirements for real-time performance, causality, and system stability. Directly updating the parameters of large-scale neural networks online often leads to high computational overhead and potential instability risks, making it difficult to meet the engineering implementation requirements of practical vehicle-mounted systems. Therefore, there is an urgent need for an active road noise control technology that can adaptively suppress time-varying road noise while ensuring causal control and system stability. Summary of the Invention

[0006] To address the aforementioned technical issues, this invention proposes an active noise control system and method based on an adaptive deep neural network. This method can retain the high performance of causal neural networks while introducing an online adaptive mechanism in a real-time road noise control system under multi-channel complex scenarios, significantly alleviating the data distribution drift problem between training and actual operating environments, and improving noise control performance under complex scenarios.

[0007] The technical solution adopted by the system of this invention is as follows: An active noise control system based on an adaptive deep neural network includes: The reference signal acquisition unit is used to acquire vibration or acoustic signals as reference input signals for the active noise control system. Error signal acquisition unit is used to acquire residual noise signals at the target location; A causal deep neural network controller, pre-trained using an active noise control database, is used to generate control signals based on the reference input signal under sample-by-sample causal constraints to drive a secondary sound source to generate anti-noise; all parameters of the causal deep neural network controller remain frozen during the control process. A lightweight adaptive module, embedded inside the causal deep neural network controller, is used to adjust the local parameters of the causal deep neural network controller online according to the updated parameters output by the parameter update unit during the control process, so as to compensate for the impact of environmental changes on control performance. The parameter update unit is used to calculate the loss function based on the reference input signal and the residual noise signal, and to update the parameters of the lightweight adaptive module online only based on the loss function.

[0008] Furthermore, the causal deep neural network controller adopts a time-domain structure and meets the real-time control requirements of sample-by-sample input and sample-by-sample output.

[0009] Furthermore, the lightweight adaptive modules are inserted into multiple network layers and connected to their respective network layers.

[0010] Furthermore, the lightweight adaptive module includes a serial adaptive module and / or a parallel adaptive module. The serial adaptive module is set serially before and / or after one or more network layers of the causal deep neural network controller to sequentially adjust the output features of the corresponding network layers, thereby realizing online correction of the local feature representation of the causal deep neural network controller. The parallel adaptive module is set in parallel with the corresponding network layer of the causal deep neural network controller and receives the same input as the corresponding network layer.

[0011] Furthermore, the serial adaptive module is a causal convolutional layer with a small convolutional kernel.

[0012] Furthermore, the parallel adaptive module is a bottleneck structure module, which includes a dimension-reducing convolutional layer, a nonlinear activation layer, and a dimension-upgrading convolutional layer connected sequentially along the channel dimension.

[0013] Furthermore, the active noise control system adopts a time-division processing structure, which includes a point-by-point control path and a frame-by-frame parameter update path; wherein, the control path is used to generate control signals in real time; and the parameter update path is used to update the parameters of the lightweight adaptive module online.

[0014] This invention also provides a control method for an active noise control system based on an adaptive deep neural network, comprising the following steps: S1, a causal deep neural network controller is pre-trained using an active noise control database; S2, fix the trained causal deep neural network controller parameters, and insert a lightweight adaptive module into the causal deep neural network, so that the lightweight adaptive module establishes a data connection relationship with the corresponding target network layer; initialize the lightweight adaptive module so that it does not affect the controller output in the initial state; freeze the main parameters of the causal deep neural network controller, and retain only the parameters of the lightweight adaptive module as updatable parameters; S3, the reference signal is input to the causal deep neural network controller to generate a control signal to drive the secondary sound source to generate anti-noise. After the control signal propagates through the secondary path, it is superimposed with the road noise to form a residual noise signal, and the residual noise signal is collected. S4 constructs an objective function based on the reference signal and residual noise signal, and updates the parameters of the lightweight adaptive module frame by frame while keeping all parameters of the causal deep neural network controller frozen, thereby achieving adaptive tracking of the time-varying characteristics of road noise.

[0015] Furthermore, in step S2, the lightweight adaptive module includes a serial adaptive module and / or a parallel adaptive module; the lightweight adaptive module adopts a zero initialization method during initialization, so that the initial output of the parallel adaptive module is zero, and the initial state of the serial adaptive module is equivalent to an identity mapping, thereby not affecting the overall output of the causal deep neural network controller at the initial moment.

[0016] Furthermore, the active noise control system adopts a time-division processing structure, which includes a point-by-point control path and a frame-by-frame parameter update path. In the control path, the causal deep neural network controller processes the input reference signal point by point, outputting a corresponding control signal sampling point for each input reference signal sampling point to meet the real-time control requirements of the active noise control system. In the parameter update path, the parameter update unit accumulates the error signals of multiple consecutive sampling points according to a preset frame length, and calculates the loss function and parameter update amount based on the error signal within the frame, updating the parameters of the lightweight adaptive module at an update frequency lower than that of the forward control path.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) By combining a causal deep neural network controller and a lightweight adaptive module, this invention introduces an online adaptive mechanism for the first time in an ARNC system based on a deep neural network. This mechanism enables the system to have both high-performance control capabilities and online adaptive capabilities.

[0018] (2) The causal deep neural network controller used in this invention generates control signals based only on current and historical reference signals. Compared with the prior art, it can better meet the requirements of real-time and causality of active road noise control and can effectively improve the control performance in complex road noise environments.

[0019] (3) The lightweight adaptive module used in this invention allows only a small number of parameters to be updated online without adjusting the deep neural network model as a whole, thereby reducing the online computational complexity and deployment overhead and improving the feasibility of the system in actual vehicle platforms.

[0020] (4) The present invention can still maintain good noise reduction effect even with limited training data, and has fast convergence speed and strong anti-interference robustness, providing a feasible technical solution for the application of deep neural networks in practical active road noise control systems. Attached Figure Description

[0021] Figure 1 This is an overall structural block diagram of the method of the present invention.

[0022] Figure 2Here are the hardware configuration diagrams in this embodiment: (a) is a layout diagram of the acceleration sensor, and (b) is a physical diagram of the driver's control speaker and reference microphone.

[0023] Figure 3 This is a diagram of the WaveNet-VNN network structure used in the example.

[0024] Figure 4 These are the noise reduction curves at the error microphone in this embodiment. (a) is the noise reduction curve at the left ear microphone, and (b) is the noise reduction curve at the right ear microphone.

[0025] Figure 5 These are the spectrum diagrams at the error microphone in this embodiment. (a) is the spectrum diagram at the left ear microphone, and (b) is the spectrum diagram at the right ear microphone.

[0026] Figure 6 These are the noise reduction curves with unrelated interference signals added in this embodiment. (a) is the noise reduction curve at the left ear microphone, and (b) is the noise reduction curve at the right ear microphone. Detailed Implementation

[0027] The technical solution of the present invention will be fully and clearly described below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0028] This embodiment presents an active noise control system based on an adaptive deep neural network, such as... Figure 1 As shown, it includes: The reference signal acquisition unit is used to acquire vibration or acoustic signals as reference input signals for the active noise control system. The error signal acquisition unit is used to acquire the residual noise signal at the target location; the target location refers to the area where noise reduction is required. A causal deep neural network controller, pre-trained using an active noise control database, is used to generate control signals based on reference input signals under sample-by-sample causal constraints, thereby driving secondary sound sources to generate anti-noise. Lightweight adaptive modules, such as Figure 1As shown, an adaptive module is embedded within a causal deep neural network controller. During the control process, it adjusts the local parameters of the controller online based on the updated parameters output by the parameter update unit to compensate for the impact of environmental changes on control performance. The lightweight adaptive module includes a parallel adaptive module and / or a serial adaptive module. The parallel adaptive module is connected in parallel with the corresponding target network layer and receives the same input as the target network layer. Its output is fused with the output of the target network layer and then passed to subsequent network layers. The serial adaptive module is connected in series on the output path of the target network layer and processes its output. Simultaneously, the serial adaptive module also includes a skip connection branch, which directly transmits the input or output of the target network layer to the fusion node and fuses it with the processing result of the serial adaptive module before outputting it to subsequent network layers. The target network layer is one or more network layers in the causal deep neural network controller. The parallel adaptive module and the serial adaptive module can be selected individually or combined at different target network layers to form the lightweight adaptive module. The parameter update unit is used to calculate the loss function based on the reference input signal and the residual noise signal, and to update the parameters of the lightweight adaptive module online only based on the loss function.

[0029] Among them, the parameters of the causal deep neural network controller are frozen after all parameters are trained, which can support point-by-point operation. The lightweight adaptive module updates in real time frame by frame. This dual-rate design achieves adaptive suppression of time noise while ensuring no delay noise reduction.

[0030] Preferably, the causal deep neural network controller structure used in this embodiment is as follows: Figure 3As shown, a time-domain structure is adopted, and the real-time control requirements of sample-by-sample input and output are met. Specifically, the controller includes one WaveNet, two VNN (Volterra Neural Network) modules, five Conv1d modules, and three Tanh activation function modules. Among them, WaveNet (A. Van Den Oord, S. Dieleman, H. Zen, K. Simonyan, O. Vinyals, A. Graves, N. Kalchbrenner, A. Senior, K. Kavukcuoglu, et al., WaveNet: Agenerative model for raw audio, 2016, arXiv preprint arXiv:1609.03499 12.) is used as the main body due to its powerful acoustic modeling capabilities. VNN (S. Roheda, H. Krim, and B. Jiang, “Volterra neural networks (VNNs),” J. Mach. Learn. Res. 25(182), 1–29(2024.) enhances the nonlinear modeling capabilities of the model by introducing an intersection term. The Conv1d module is used to match the number of input channels within the network, and the Tanh activation function introduces nonlinearity. The causal neural network controller is used to learn the nonlinear mapping relationship from the multi-channel reference signal to the multi-channel control signal, and achieves a noise reduction effect that surpasses that of control filters based on linear models.

[0031] In this embodiment, the lightweight adaptive module is a lightweight neural network module, whose parameter size is significantly smaller than the main parameter size of the causal deep neural network controller. For example... Figure 3 As shown, the lightweight adaptive module includes two forms: a serial adaptive module (S-Adapter) and a parallel adaptive module (P-Adapter), which are set in the causal deep neural network controller according to their different insertion methods. The serial adaptive module consists of causal convolutional layers with small kernels, which are sequentially placed before and / or after one or more network layers of the causal deep neural network controller. It is used to sequentially adjust the output features of the corresponding network layers, thereby achieving online correction of the controller's local feature representation.

[0032] The parallel adaptive module is the bottleneck structure module, comprising a dimensionality-reducing convolutional layer, a nonlinear activation layer, and an updimensional convolutional layer connected sequentially along the channel dimension. The parallel adaptive module is connected in parallel with the corresponding target network layer of the causal deep neural network controller and receives the same input as the target network layer. The correction output of the parallel adaptive module is fused with the original output of the target network layer and used as the input to subsequent network layers to perform online compensation of the controller output while maintaining the backbone network structure. Further, the output of the parallel adaptive module can be expressed as follows: the input signal of the target network layer is first subjected to channel dimensionality reduction by the first causal convolutional layer, then processed by a nonlinear activation function, subsequently restored to the output channel dimension by the second causal convolutional layer, and finally output as the correction component. To adjust the correction intensity of each channel, the output of the parallel adaptive module is multiplied by a learnable scaling parameter.

[0033] To avoid disturbing the original performance of the causal deep neural network controller during the initial operation of the system, the lightweight adaptive module adopts a zero initialization method during initialization, so that the initial output of the parallel adaptive module is zero, and the output of the serial adaptive module is equal to the input, which is equivalent to an identity mapping, so as not to affect the overall output of the causal deep neural network controller at the initial moment.

[0034] In this embodiment, the active noise control system adopts a time-division processing structure, such as... Figure 1 As shown, the time-division processing structure includes a point-by-point control path and a frame-by-frame parameter update path. The control path generates control signals in real time. The causal deep neural network controller processes the input reference signals point-by-point, outputting a corresponding control signal sample point for each input reference signal sample point to meet the real-time control requirements of the active noise control system. The parameter update path updates the parameters of the lightweight adaptive module online. The parameter update unit accumulates the error signals of multiple consecutive sample points according to a preset frame length, calculates the loss function and parameter update amount based on the error signals within that frame, and updates the parameters of the lightweight adaptive module at a lower update frequency than the forward control path. Thus, the control signal generation process and parameter update process of the causal deep neural network controller are separated in time, achieving real-time operation of the control output and low-frequency execution of parameter updates.

[0035] The parameter update unit is used to update the parameters of the lightweight adaptive module online based on the error signal. In this embodiment, the parameter update unit includes a loss calculation module and a parameter optimization module. The loss calculation module calculates a loss function value based on the error signal, and the loss function can be an A-weighted sound pressure level loss function. The parameter optimization module updates the parameters of the lightweight adaptive module based on the loss function value, and the parameter optimization module can be implemented using the Adam optimization algorithm. Furthermore, the parameter update unit performs parameter updates at the frame level, that is, it accumulates the error signals corresponding to multiple consecutive sampling points and then performs a unified parameter update, thereby not affecting the sample-by-sample control process of the causal deep neural network controller.

[0036] The active noise control system of this invention is applicable to multi-channel active road noise control scenarios consisting of multiple reference signals, multiple control signals, and multiple error signals. In the presence of unrelated interference signals, this system can maintain the stability of the control process and reduce the impact of interference on the adaptive process by limiting the range of online parameter updates.

[0037] This embodiment also provides an adaptive deep neural network active noise control method for real-time parameter updates, including the following steps: Step 1: Collect multi-channel road noise data of the vehicle under different road conditions and driving conditions. The data includes vehicle structural vibration signals as reference signals and corresponding noise signals at the human ear inside the vehicle. The collected data is preprocessed and the corresponding secondary paths are estimated to build a training database for active road noise control. Step 2: Based on the training database, construct a deep neural network controller that satisfies the sample-by-sample causal constraint. With multi-channel reference signals as input and control signals as output, train the neural network offline to learn the nonlinear mapping relationship between the reference signals and the control signals. Step 3: Set up lightweight adaptive modules at one or more target network layers in the causal deep neural network controller that has completed offline training, and establish data connection relationships between the lightweight adaptive modules and the corresponding target network layers to form an adaptive neural network model; initialize the lightweight adaptive modules so that they do not affect the controller output in the initial state; freeze the main parameters of the causal deep neural network controller, and retain only the parameters of the lightweight adaptive modules as updatable parameters; Step 4: In the active road noise control process, the real-time collected reference signal is input into the adaptive neural network model to generate a control signal to drive the secondary sound source; after the control signal propagates through the secondary path, it is superimposed with the road noise to form a residual noise signal, and the residual noise signal is collected. Step 5: Construct an objective function based on the reference signal and residual noise signal, and while keeping the main parameters of the causal deep neural network controller frozen, update the parameters of the lightweight adaptive module frame by frame to achieve adaptive tracking of the time-varying characteristics of road noise.

[0038] The following is a detailed introduction using a 42×2×2 multi-channel ARNC system as an example: 1. Database Acquisition In this example, a 30-hour database of vehicle road noise was recorded. A classic five-seat small electric car was selected for the measurement. Eight capacitive microphones were placed near the ears on the headrests of the driver, front passenger, and left and right rear seats to collect main noise. Eight speakers were placed on the shoulder pads of the four seats to generate control signals. Ten triaxial accelerometers and twelve single-axis accelerometers were distributed in the undercarriage, doors, wheel wells, and hood, with specific locations as follows: Figure 2 (a) The recorded vibration signal was used as a reference signal. Acquisition was performed using a 128-channel NI chassis, with a sampling rate set to 8 kHz during acquisition. Data was recorded by driving the car at constant speeds of 50 km / h, 60 km / h, 80 km / h, and 100 km / h, including 9 hours of data at 50 km / h, 9 hours at 60 km / h, 9 hours at 80 km / h, and 3 hours at 100 km / h. Data at 50 km / h and 80 km / h were collected over three days, the 60 km / h dataset over two days, and the 100 km / h dataset within a single day.

[0039] 2. Training the backbone network This example uses two microphones and two speakers from the driver's side; a physical image is shown below. Figure 2 (b) and all accelerometers were simulated, forming a 42×2×2 multichannel system. 1.5 hours of data were randomly selected from a 30-hour database as the test set, and the remaining data as the training set. The backbone network used was the previously proposed WaveNet-VNN, as it is a temporal causal network capable of sample-by-sample inference and has been proven to outperform various advanced traditional algorithms in ANC tasks. The loss function used was A-weighted sound pressure level. The system was trained 30 times using the Adam optimizer.

[0040] 3. Construct an adaptive deep neural network controller Two different lightweight adaptive modules (lightweight adapters) are inserted into the trained backbone network in parallel and series configurations, such as... Figure 3As shown. The cascaded insertion adapter (S-Adapter) is a convolutional layer with a small kernel, while the parallel insertion adapter (P-Adapter) employs a bottleneck structure design, consisting of two causal convolutional layers and a Tanh activation function. All adapter parameters are initialized to zero, so that they have no or near-zero effect on the controller output in the initial state. The number of network parameters and computational cost added by both the parallel and cascade insertion adapters are 5% of the original network.

[0041] 4. Real-time control and parameter updates Simulations were performed using data from the test set. A reference signal was input into the adaptive neural network model to generate a control signal. After propagation through a secondary path, the control signal was superimposed on road noise to form a residual noise signal. An objective function was constructed based on the reference signal and the residual noise signal. While keeping the main parameters of the causal deep neural network controller frozen, the parameters of the lightweight adaptive module were updated frame by frame to achieve adaptive tracking of the time-varying characteristics of road noise.

[0042] 5. Noise reduction performance Figure 4 This paper presents the normalized noise reduction curves of the proposed scheme and other methods under the condition of cross-date distribution data at a speed of 50 km / h in this example. Among them, MIMO-Wiener is the multi-channel Wiener solution, which is the theoretically optimal solution; TD-FxLMS is the most basic adaptive filtering algorithm; ODW-FDFeLMS is a recently proposed traditional adaptive filtering algorithm optimized by reference whitening secondary path decoupling (S. Lian, T. Li, J. Gu, Y. Hu, C. Zhu, S. Wang, and J. Lu, “An online decoupling-whitening frequency domain filtered-errorleast mean square algorithm for active road noise control,” J. Acoust. Soc. Am. 156(2),1413–1424 (2024)), whose performance is significantly better than other adaptive filtering algorithms, and whose convergence effect is close to the Wiener solution; WaveNet-VNN (1-Day-Adaptive) is the test result of the model trained on date one on date three. The proposed scheme converges faster than the state-of-the-art ODW-FDFeLMS algorithm and has a higher upper limit for noise reduction performance, even surpassing the ideal Wiener solution. It effectively solves the problem of poor generalization performance of the ANC scheme in fixed-parameter deep neural networks. Figure 5The corresponding left and right ear spectrograms are shown, and the proposed scheme is significantly better than various traditional algorithms in the mid-to-high frequency band.

[0043] 6. Convergence speed and anti-interference performance Figure 6 This example demonstrates the normalized noise reduction curves of various schemes at a speed of 50 km / h, with irrelevant interference added to the error signal. At t = 15 s, 1 second of Gaussian white noise with zero mean (6 dB SNR) is introduced to simulate sudden disturbances. At t = 25 s and t = 50 s, a 10-second 0 dB SNR speech signal and a 6-second classic iPhone ringtone (marimba sound effect, -6 dB SNR) are introduced, respectively, to simulate continuous interference of varying intensity and type. Figure 6 The results demonstrate the performance of the ODW-FDFeLMS algorithm under three different step sizes. Smaller step sizes result in slower convergence, while larger step sizes lead to faster convergence, but both are prone to divergence due to interference. The recently proposed C-DWFDFeLMS algorithm, designed to address irrelevant noise interference, significantly improves this problem. Furthermore, the proposed scheme exhibits significantly better anti-interference performance than the recently proposed C-DWFDFeLMS algorithm, and its convergence speed is also superior to a series of traditional adaptive algorithms.

Claims

1. An active noise control system based on an adaptive deep neural network, characterized in that, include: The reference signal acquisition unit is used to acquire vibration or acoustic signals as reference input signals for the active noise control system. Error signal acquisition unit is used to acquire residual noise signals at the target location; A causal deep neural network controller, pre-trained using an active noise control database, is used to generate control signals based on the reference input signal under sample-by-sample causal constraints to drive a secondary sound source to generate anti-noise; all parameters of the causal deep neural network controller remain frozen during the control process. A lightweight adaptive module, embedded inside the causal deep neural network controller, is used to adjust the local parameters of the causal deep neural network controller online according to the updated parameters output by the parameter update unit during the control process, so as to compensate for the impact of environmental changes on control performance. The parameter update unit is used to calculate the loss function based on the reference input signal and the residual noise signal, and to update the parameters of the lightweight adaptive module online only based on the loss function.

2. The active noise control system based on an adaptive deep neural network according to claim 1, characterized in that, The causal deep neural network controller adopts a time-domain structure and meets the real-time control requirements of sample-by-sample input and sample-by-sample output.

3. The active noise control system based on an adaptive deep neural network according to claim 1, characterized in that, Multiple lightweight adaptive modules are inserted into multiple network layers and connected to their respective network layers.

4. The active noise control system based on an adaptive deep neural network according to claim 3, characterized in that, The lightweight adaptive module includes a serial adaptive module and / or a parallel adaptive module. The serial adaptive module is set serially before and / or after one or more network layers of the causal deep neural network controller to adjust the output features of the corresponding network layers sequentially, thereby realizing online correction of the local feature representation of the causal deep neural network controller. The parallel adaptive module is set in parallel with the corresponding network layer of the causal deep neural network controller and receives the same input as the corresponding network layer.

5. The active noise control system based on an adaptive deep neural network according to claim 4, characterized in that, The serial adaptive module is a causal convolutional layer with small convolutional kernels.

6. The active noise control system based on an adaptive deep neural network according to claim 4, characterized in that, The parallel adaptive module is a bottleneck structure module, which includes a dimension-reducing convolutional layer, a nonlinear activation layer and a dimension-upgrading convolutional layer connected sequentially along the channel dimension.

7. The active noise control system based on an adaptive deep neural network according to claim 1, characterized in that, The active noise control system adopts a time-division processing structure, which includes a point-by-point control path and a frame-by-frame parameter update path. The control path is used to generate control signals in real time, and the parameter update path is used to update the parameters of the lightweight adaptive module online.

8. The control method for an active noise control system based on an adaptive deep neural network as described in claim 1, characterized in that, Includes the following steps: S1, a causal deep neural network controller is pre-trained using an active noise control database; S2, fix the trained causal deep neural network controller parameters, and insert a lightweight adaptive module into the causal deep neural network, so that the lightweight adaptive module establishes a data connection relationship with the corresponding target network layer; initialize the lightweight adaptive module so that it does not affect the controller output in the initial state; freeze the main parameters of the causal deep neural network controller, and retain only the parameters of the lightweight adaptive module as updatable parameters; S3, the reference signal is input to the causal deep neural network controller to generate a control signal to drive the secondary sound source to generate anti-noise. After the control signal propagates through the secondary path, it is superimposed with the road noise to form a residual noise signal, and the residual noise signal is collected. S4 constructs an objective function based on the reference signal and residual noise signal, and updates the parameters of the lightweight adaptive module frame by frame while keeping all parameters of the causal deep neural network controller frozen, thereby achieving adaptive tracking of the time-varying characteristics of road noise.

9. The control method according to claim 8, characterized in that, In step S2, the lightweight adaptive module includes a serial adaptive module and / or a parallel adaptive module; the lightweight adaptive module adopts a zero initialization method during initialization, so that the initial output of the parallel adaptive module is zero, and the initial state of the serial adaptive module is equivalent to an identity mapping, so as not to affect the overall output of the causal deep neural network controller at the initial moment.

10. The control method according to claim 8, characterized in that, The active noise control system adopts a time-division processing structure, which includes a point-by-point control path and a frame-by-frame parameter update path. In the control path, the causal deep neural network controller processes the input reference signal point by point, outputting a corresponding control signal sampling point for each input reference signal sampling point to meet the real-time control requirements of the active noise control system. In the parameter update path, the parameter update unit accumulates the error signals of multiple consecutive sampling points according to a preset frame length, calculates the loss function and parameter update amount based on the error signals within the frame, and updates the parameters of the lightweight adaptive module at an update frequency lower than that of the forward control path.

Images