Active noise reduction method, system, terminal and medium based on deep neural network

Abstract

The application provides a deep neural network-based active noise reduction method, system, terminal and medium, and belongs to the active noise reduction technical field.The method comprises the following steps: collecting noise signals in real time through a reference microphone; obtaining a main channel estimation signal at a current moment; inputting the noise signal at the current moment and the main channel estimation signal after preprocessing into a trained deep neural network, obtaining a neural filter parameter through processing, generating a cancellation signal based on the neural filter parameter and the noise signal after preprocessing; and playing the cancellation signal through a loudspeaker to generate a reverse noise signal for offsetting a main channel noise signal, wherein the main channel noise signal is generated by the noise signal propagating in the main channel.The application simultaneously inputs the main channel estimation signal and the noise signal into the deep neural network for prediction, can effectively perceive the main channel transfer function, and further generates a cancellation signal highly matched with the noise information of the main channel.

Description

Technical Field

[0001] This invention relates to the field of active noise reduction technology, and more particularly to an active noise reduction method, system, terminal, and medium based on deep neural networks. Background Technology

[0002] Active noise cancellation (ANC) is a technology that cancels out noise by generating a sound wave that is out of phase with the noise. It is mainly used in headphones, car cabins, and to reduce the operating noise of home appliances. The system needs to complete the entire process of "noise acquisition → analysis and calculation → generation of the inverse sound wave" in a very short time to ensure that the inverse sound wave meets the noise precisely on time.

[0003] In existing technologies, deep learning-based active noise reduction methods have become an important research direction. The core idea of ​​these methods is to treat the active noise reduction process as a supervised learning task, that is, to use a deep neural network model to directly learn the end-to-end nonlinear mapping relationship from the noise signal to the inverse cancellation signal, thereby achieving active noise reduction. However, these methods typically use a reference signal directly as input and use a deep neural network to predict the inverse noise to achieve noise reduction. Since the input signal does not contain any information about the main channel transfer function, while the network needs to predict the inverse noise that includes the characteristics of the main channel transfer function, this mismatch between information and task requirements makes it difficult for the network to generate a cancellation signal that highly matches the noise information of the main channel.

[0004] Therefore, existing technologies have shortcomings and need to be improved and developed. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide an active noise reduction method, system, terminal and medium based on deep neural networks to address the above-mentioned deficiencies of the prior art, and to solve the problem that the prior art is unable to generate a cancellation signal that highly matches the noise information of the main channel.

[0006] The technical solution adopted by this invention to solve the technical problem is as follows:

[0007] In a first aspect, embodiments of the present invention provide an active noise reduction method based on a deep neural network, the method comprising:

[0008] Noise signals are acquired in real time using a reference microphone;

[0009] The most recent error signal recorded by the error microphone is acquired. Based on the most recent error signal and the historical reverse noise signal of the previous moment, the estimated signal of the main channel at the current moment is obtained.

[0010] The noise signal and the estimated main channel signal at the current moment are preprocessed and then input into the trained deep neural network. The neural filtering parameters are obtained after processing. Based on the neural filtering parameters and the preprocessed noise signal, a cancellation signal is generated.

[0011] The cancellation signal is played through a speaker to generate a reverse noise signal to cancel out the main channel noise signal, which is generated by the noise signal propagating in the main channel.

[0012] In one implementation, the main channel estimation signal for the current moment is obtained based on the most recent error signal and the historical reverse noise signal from the previous moment, including:

[0013] The main channel estimation signal at the current moment is obtained by subtracting the historical reverse noise signal from the most recent error signal.

[0014] In one implementation, the noise signal and the main channel estimation signal at the current moment are preprocessed and then input into a trained deep neural network. The deep neural network processes the signal to obtain neural filtering parameters. Based on the neural filtering parameters and the preprocessed noise signal, a cancellation signal is generated, including:

[0015] Perform a short-time Fourier transform on the noise signal and the main channel estimated signal at the current moment to obtain the time-frequency domain representation of the noise signal and the time-frequency domain representation of the main channel estimated signal at the current moment;

[0016] The time-frequency domain representation of the noise signal at the current moment and the time-frequency domain representation of the estimated signal of the main channel are input into the trained deep neural network, and the neural filtering parameters are obtained after processing by the deep neural network.

[0017] The noise signal is processed based on the neural filtering parameters and the time-frequency domain representation of the noise signal at the current moment to generate a cancellation signal.

[0018] In one embodiment, the deep neural network includes a first convolutional block, four sequentially connected state space blocks, and a second convolutional block; the time-frequency domain representation of the noise signal at the current moment and the time-frequency domain representation of the main channel estimated signal are input into the trained deep neural network, and the neural network processes them to obtain neural filtering parameters, including:

[0019] The time-frequency domain representation of the noise signal at the current moment and the time-frequency domain representation of the estimated signal of the main channel are input into the first convolutional block. After processing by the first convolutional block, the first fused time-frequency feature is obtained.

[0020] The first fused time-frequency feature is input into four sequentially connected state space blocks. After being processed sequentially by each of the state space blocks, the second fused time-frequency feature is obtained.

[0021] The second fused time-frequency feature is input into the second convolutional block, and the second convolutional block performs dimensionality reduction on the second fused time-frequency feature to obtain the neural filter parameters.

[0022] In one implementation, processing is performed based on the neural filtering parameters and the noise signal at the current moment to generate a cancellation signal, including:

[0023] The neural filtering parameters and the current noise signal's time-frequency domain representation are multiplied by a complex number at each time frequency point to obtain the time-frequency domain signal;

[0024] The time-frequency domain signal is subjected to an inverse discrete Fourier transform, and then the result of the inverse discrete Fourier transform is subjected to an inverse short-time Fourier transform based on preset synthesis window parameters to generate a time-domain frame signal.

[0025] The time-domain frame signals are summed, and the last sampling point is selected as the valid output from the summed time-domain signals of each frame.

[0026] All the valid outputs are concatenated in chronological order to form the cancellation signal for the current moment.

[0027] In one implementation, a training dataset is obtained, the training dataset comprising noise training signals at several time points, main channel estimation training signals, and main channel noise training signals;

[0028] For each time step, the noise training signal and the main channel estimation training signal at the current time step are used as input signals, and the main channel noise training signal at the M sampling points after the current time step step is used as the training target corresponding to the input signal.

[0029] During training, the input signal is input into the deep neural network to be trained to obtain the eliminated training signal. The time-domain mean square error of the superposition result of the eliminated training signal and the training target is calculated as the loss value. The backpropagation algorithm based on gradient descent is used to iteratively update the model parameters according to the loss value.

[0030] Repeat the model training steps until the preset termination condition is met to obtain the trained deep neural network model.

[0031] In one implementation, M is the sum of the secondary channel delay and the preset system algorithm delay.

[0032] Secondly, embodiments of the present invention also provide an active noise reduction system based on a deep neural network, the system comprising:

[0033] The first acquisition module is used to acquire noise signals in real time via a reference microphone;

[0034] The second acquisition module is used to acquire the most recent error signal recorded by the error microphone, and to obtain the current main channel estimation signal based on the most recent error signal and the historical reverse noise signal of the previous moment.

[0035] The cancellation signal generation module is used to input the noise signal and the main channel estimation signal at the current moment into the trained deep neural network after preprocessing, obtain neural filtering parameters after processing, and generate cancellation signal based on the neural filtering parameters and the preprocessed noise signal;

[0036] The reverse noise signal generation module plays the cancellation signal through a speaker to generate a reverse noise signal to cancel the main channel noise signal, which is generated by the noise signal propagating in the main channel.

[0037] Thirdly, embodiments of the present invention also provide a terminal, the terminal comprising: a memory, a processor, and an active noise reduction program based on a deep neural network stored in the memory and executable on the processor, wherein when the active noise reduction program based on a deep neural network is executed by the processor, the active noise reduction program based on a deep neural network implements the steps of the active noise reduction method based on a deep neural network as described above.

[0038] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing an active noise reduction program based on a deep neural network, the active noise reduction program based on a deep neural network being executable to implement the steps of the active noise reduction method based on a deep neural network as described above.

[0039] The beneficial effects of this invention are as follows: This invention acquires noise signals in real time using a reference microphone; obtains the estimated main channel signal at the current moment; inputs the noise signal and the estimated main channel signal at the current moment into a trained deep neural network after preprocessing, and obtains neural filtering parameters. Based on the neural filtering parameters and the preprocessed noise signal, a cancellation signal is generated; the cancellation signal is played through a speaker to generate an inverse noise signal to cancel the main channel noise signal, which is generated by the noise signal propagating in the main channel. This invention simultaneously inputs the estimated main channel signal and the noise signal into the deep neural network for prediction, which can effectively perceive the main channel transfer function and thus generate a cancellation signal that highly matches the noise information of the main channel. Attached Figure Description

[0040] Figure 1 This is a flowchart of a preferred embodiment of the active noise reduction method based on deep neural networks in this invention.

[0041] Figure 2 This is a schematic diagram of the architecture of an active noise reduction system based on deep neural networks.

[0042] Figure 3 This is a flowchart for generating the cancellation signal.

[0043] Figure 4 This is a schematic diagram of the algorithm delay of the present invention.

[0044] Figure 5 This is a schematic diagram illustrating how the present invention trains a deep neural network to predict M samples in advance.

[0045] Figure 6 This is a diagram illustrating the effect of active noise reduction for aliasing noise in this invention.

[0046] Figure 7 This is a diagram illustrating the effect of active noise reduction for engine noise in this invention.

[0047] Figure 8 This is a diagram illustrating the effect of active noise reduction in a factory using the present invention.

[0048] Figure 9 This is a schematic diagram of a preferred embodiment of the active noise reduction system based on deep neural networks in this invention.

[0049] Figure 10 This is a block diagram of the terminal principle of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail 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 are not intended to limit the invention.

[0051] Active noise cancellation (ANC) is a technology that cancels out noise by generating a sound wave that is out of phase with the noise. It is mainly used in headphones, car cabins, and to reduce the operating noise of home appliances. The system needs to complete the entire process of "noise acquisition → analysis and calculation → generation of the inverse sound wave" in a very short time to ensure that the inverse sound wave meets the noise precisely on time.

[0052] In existing technologies, deep learning-based active noise reduction methods have become an important research direction. The core idea of ​​these methods is to treat the active noise reduction process as a supervised learning task, that is, to directly learn the end-to-end nonlinear mapping relationship from the noise signal to the inverse cancellation signal using a deep neural network model, thereby achieving active noise reduction. However, these methods typically use a reference signal directly as input and use a deep neural network to predict inverse noise to achieve noise reduction. Since the input signal does not contain any information about the main channel transfer function, while the network needs to predict inverse noise that includes the characteristics of the main channel transfer function, this mismatch between information and task requirements makes it difficult for the network to generate a cancellation signal that highly matches the noise information of the main channel.

[0053] To address the aforementioned deficiencies in existing technologies, this invention provides an active noise reduction method, system, terminal, and medium based on a deep neural network. The method includes: acquiring noise signals in real-time using a reference microphone; obtaining a current-time main channel estimation signal; inputting the current-time noise signal and the main channel estimation signal, after preprocessing, into a trained deep neural network to obtain neural filtering parameters; generating a cancellation signal based on the neural filtering parameters and the preprocessed noise signal; and playing the cancellation signal through a speaker to generate an inverse noise signal to cancel the main channel noise signal, wherein the main channel noise signal is generated by the noise signal propagating in the main channel. This invention simultaneously inputs the main channel estimation signal and the noise signal into the deep neural network for prediction, effectively sensing the main channel transfer function and thereby generating a cancellation signal that highly matches the noise information of the main channel.

[0054] Please see Figure 1 The active noise reduction method based on deep neural networks described in this embodiment of the invention includes the following steps:

[0055] Step S100: Acquire noise signals in real time using a reference microphone.

[0056] Specifically, the architecture diagram of the active noise reduction system based on deep neural networks is as follows: Figure 2 As shown. The reference microphone is positioned near the noise source. In active noise cancellation scenarios, the reference microphone can be positioned on the outside of the headphones. When the noise source emits noise, the reference microphone collects the noise signal in real time. This noise signal can also be called the reference signal. The noise signal passes through the main channel transfer function. After processing, it will become a main channel noise signal. The noise signal in this main channel is the target signal to be eliminated.

[0057] Please see Figure 1 The active noise reduction method based on deep neural networks described in this embodiment of the invention further includes the following steps:

[0058] Step S200: Obtain the most recent error signal recorded by the error microphone, and obtain the current main channel estimation signal based on the most recent error signal and the historical reverse noise signal of the previous moment.

[0059] Specifically, error microphones are typically positioned near the target noise reduction area to monitor the residual sound waves after interference between the main channel noise signal and the inverse noise signal in the target noise reduction area, thus obtaining the error signal. In active noise cancellation scenarios using headphones, the error microphone can be placed inside the ear canal. The historical inverse noise signal from the previous moment is also considered. It is an active noise reduction system based on deep neural networks at the previous time step (i.e. (Time) Output cancellation signal Then, the reverse noise signal is generated after propagation through the secondary channel. Since the main channel noise signal cannot be directly measured in practical applications and needs to be estimated, the most recent error signal recorded by the error microphone is used. Subtract the historical reverse noise signal from the previous moment. The estimated signal of the main channel at the current moment can be obtained. The estimated main channel signal is then used as input to the deep neural network along with the noise signal at the current moment. This input includes the main channel transfer function, allowing the trained deep neural network to dynamically identify the current main channel characteristics and make the generated canceled signal more closely match the noise signal.

[0060] Please see Figure 1 The active noise reduction method based on deep neural networks described in this embodiment of the invention further includes the following steps:

[0061] Step S300: The noise signal and the estimated signal of the main channel at the current moment are preprocessed and then input into the trained deep neural network. The neural filtering parameters are obtained after processing. Based on the neural filtering parameters and the preprocessed noise signal, a cancellation signal is generated.

[0062] Specifically, the flowchart for generating the cancellation signal is as follows: Figure 3As shown, the noise signal and the estimated main channel signal at the current moment are subjected to a Short-Time Fourier Transform (STFT) to obtain the time-frequency domain representations of the noise signal and the main channel signal at the current moment. These time-frequency domain representations are then input into a trained deep neural network, which processes them to obtain neural filtering parameters. Based on these neural filtering parameters and the time-frequency domain representation of the noise signal at the current moment, a cancellation signal is generated. The deep neural network includes a first convolutional block, four sequentially connected state space blocks (S4D blocks), and a second convolutional block. Finally, the cancellation signal is generated based on the neural filtering parameters and the noise signal at the current moment. In generating the cancellation signal, the neural filtering parameters and the time-frequency domain representation of the noise signal are first multiplied by a complex number at each time frequency point. Then, the result is sequentially subjected to an Inverse Discrete Fourier Transform (Inverse Short-Time Fourier Transform) and an Inverse Short-Time Fourier Transform (ISTFT) to generate a time-domain frame signal.

[0063] In one implementation, the fused time-frequency features are input into a trained deep neural network, and the neural network processes them to obtain neural filter parameters, including:

[0064] The fused time-frequency features are input into the first convolutional block, and after processing by the first convolutional block, the first fused time-frequency features are obtained.

[0065] The first fused time-frequency feature is input into four sequentially connected state space blocks. After being processed sequentially by each of the state space blocks, the second fused time-frequency feature is obtained.

[0066] The second fused time-frequency feature is input into the second convolutional block, and the second convolutional block performs dimensionality reduction on the second fused time-frequency feature to obtain the neural filter parameters.

[0067] Specifically, the State Space Block (S4D block) captures long-term dependencies and models complex noise characteristics. Each State Space Block models the long-term dependencies of the input features and outputs features processed by activation functions and normalization.

[0068] In one implementation, processing is performed based on the neural filtering parameters and the noise signal at the current moment to generate a cancellation signal, including:

[0069] The neural filtering parameters and the current noise signal's time-frequency domain representation are multiplied by a complex number at each time frequency point to obtain the time-frequency domain signal;

[0070] The time-frequency domain signal is subjected to an inverse discrete Fourier transform, and then the result of the inverse discrete Fourier transform is subjected to an inverse short-time Fourier transform based on preset synthesis window parameters to generate a time-domain frame signal.

[0071] The time-domain frame signals are summed, and the last sampling point is selected as the valid output from the summed time-domain signals of each frame.

[0072] All the valid outputs are concatenated in chronological order to form the cancellation signal for the current moment.

[0073] Specifically, the algorithm delay diagram of the present invention is as follows: Figure 4 As shown. The initial input signals are the noise signal and the main channel estimation signal at the current moment. These original time-domain signals are first divided into multiple frames of time-domain input signals, from frame t to frame (t+1 to t+K-1). Some frames undergo zero-padding adaptation and subsequent transformation, forming the original data basis for time-frequency domain processing. K is the number of points in the Discrete Fourier Transform. Then, an analysis window (using an analysis window g with a window length of 64 and a window shift of 1) is applied to each frame, and then a Short-Time Fourier Transform is performed through Discrete Fourier Transform to convert the time-domain frame signal into a time-frequency domain signal, obtaining the time-frequency domain representation of the noise signal and the time-frequency domain representation of the main channel estimation signal at the current moment. After the Discrete Fourier Transform, the frequency dimension of the signal is... The time dimension is kept consistent with the time dimension of the time-domain information. Then, the time-frequency domain representation of the noise signal at the current moment and the time-frequency domain representation of the main channel estimated signal are input into the trained deep neural network, and the neural filtering parameters are obtained after processing. The neural filtering parameters and the time-frequency domain representation of the noise signal at the current moment are then multiplied by complex numbers at each time frequency point to obtain the time-frequency domain signal.

[0074] Perform an inverse discrete Fourier transform on the time-frequency domain signal to obtain the inverse discrete Fourier transform result (i.e. Figure 4 The predicted frame t, the predicted frame t+1, ..., the predicted frame t+K-1 are then used to calculate the synthesis window using the following formula. , Where N is the window length of the synthesis window, N=2, and k represents the index of the last two sampling points. The synthesis window parameters are window length 2 and window shift 1. Then, based on the preset synthesis window parameters, windowing processing is performed on the predicted t-th frame signal, the predicted t+1-th frame signal, ..., the predicted t+K-1-th frame signal. After windowing processing, the overlapping and summed signal at predicted frame t and the predicted frame t are obtained. The overlapping and summed signals at position 1... and the overlapping and summed signals at position t-K+1 of the predicted frame, etc., are intermediate signals. Finally, these windowed frames are summed, and the last sample point from each frame is selected as the valid output.

[0075] By employing the aforementioned strategy, the inherent latency of the algorithm is reduced to just one sampling point. This design, combined with the analysis window (window length 64, window shift 1), minimizes the total system latency while ensuring frequency domain processing accuracy, thus meeting the stringent real-time requirements of active noise cancellation systems.

[0076] In one implementation, the training steps of the deep neural network include:

[0077] Obtain a training dataset, which includes noise training signals at several time points, main channel estimation training signals, and main channel noise training signals;

[0078] For each time step, the noise training signal and the main channel estimation training signal at the current time step are used as input signals, and the main channel noise training signal at the M sampling points after the current time step step is used as the training target corresponding to the input signal.

[0079] During training, the input signal is input into the deep neural network to be trained to obtain the eliminated training signal. The time-domain mean square error of the superposition result of the eliminated training signal and the training target is calculated as the loss value. The backpropagation algorithm based on gradient descent is used to iteratively update the model parameters according to the loss value.

[0080] Repeat the model training steps until the preset termination condition is met to obtain the trained deep neural network model.

[0081] Specifically, a total of 10,000 noise signals were collected from preset noise websites as noise training signals, with a total duration of 126 hours. This was based on a preset main channel transfer function. The noise training signal is converted into a main channel noise training signal to simulate the main channel noise signal generated during propagation in a real-world scenario. An initial cancellation training signal is generated using existing filtering methods, such as the Filtered-x Least Mean Square (FxLMS) algorithm. Then, the Scaled Error Function (SEF) is used to simulate the nonlinear distortion of the initial cancellation training signal played by the loudspeaker. The formulas related to nonlinear distortion are as follows:

[0082] ;

[0083] in, This is the speaker input, i.e., the initial cancellation training signal. The nonlinear distortion of the simulated loudspeaker input is represented by z, where z is the integral variable. Defined as nonlinear intensity, when As it approaches infinity, the proportional error function becomes linear, while when... When the value approaches zero, the proportional error function becomes a hard limiter. It is a natural constant.

[0084] Then, using the preset secondary channel transfer function The nonlinearly distorted signal is processed to obtain the inverse noise training signal. An error training signal is calculated based on the main channel noise training signal and the inverse noise training signal. The inverse noise training signal and the corresponding error training signal are then added to obtain the main channel noise estimation training signal. This main channel noise training signal is used to simulate the main channel noise estimation signal in practical applications.

[0085] A schematic diagram of the present invention for training a deep neural network to predict M samples in advance is shown below. Figure 5 As shown. During the training of the deep neural network, the estimated main channel noise signal is... It is also used as one of the inputs, however, to estimate the signal. and real signals They are not aligned in timing. This is because the actual timing is not aligned. It will cancel out the inverse noise in real time. And the estimated signal... It is made by eliminating noise signals After nonlinear distortion and the secondary channel transfer function, the transmission of the secondary channel will have an inherent delay. Therefore, in order to obtain a usable [transmission function] at the current moment... To enable the network to predict the future, a method was designed whereby, for each time step, the noise training signal and the estimated main channel training signal at the current time step are used as input signals, and the main channel noise training signal at M sampling points after the current time step is used as the corresponding training target. This method allows the deep neural network to predict M sampling points backward, where M is the sum of the secondary channel delay and the preset system algorithm delay. Through this look-ahead prediction, the deep neural network outputs a cancellation signal aligned with the future at the current time step, thereby compensating for the secondary channel delay and improving system stability. This delay compensation training strategy ensures strict synchronization between the output signal and the main channel noise, avoiding the cancellation failure problem caused by delay in traditional methods, and enabling the system to maintain stable noise reduction performance under different acoustic paths and application scenarios.

[0086] The loss function of a deep neural network is expressed as:

[0087] ;

[0088] Where T is the number of sampling points, and n represents the number of sampling points. The error signal is obtained by superimposing the training signal and the training target into the output of the deep neural network during the training process.

[0089] The Adam optimizer is used during training, with an initial learning rate of 0.001. If the validation loss does not improve for two consecutive epochs, the learning rate is halved. If no performance improvement is observed for eight consecutive training epochs, an early termination mechanism is enabled.

[0090] Please see Figure 1 The active noise reduction method based on deep neural networks described in this embodiment of the invention further includes the following steps:

[0091] Step S400: Play the cancellation signal through a speaker to generate an inverse noise signal to cancel the main channel noise signal.

[0092] Specifically, such as Figure 2 As shown, the cancellation signal generated by the system The signal is sent to a speaker for playback. During playback, the speaker introduces nonlinear distortion, an effect that can be expressed as... The cancellation signal emitted by the loudspeaker, as it propagates to the target noise reduction area (such as near the ear), is affected by the secondary channel transfer function. The noise signal, after passing through the secondary channel, ultimately generates a reverse noise signal. The generated reverse noise signal In the target noise reduction area and the main channel noise signal They meet. According to the principle of sound wave superposition, the two waves cancel each other out due to their opposite phase, thus significantly reducing the noise level in the area. The residual sound wave after cancellation is the error signal. Their relationship is The error signal is recorded in real time by the error microphone and immediately fed back to the system as part of the calculation of the reverse noise signal at the next moment. The key input is used to form an adaptive closed-loop control circuit.

[0093] To verify the technical effectiveness of this invention, test data was used to compare it with the Filtered-x Least Mean Square (FxLMS) algorithm and the Adaptive Recursive Network (ARN). Three typical noise types from the NoiseX-92 dataset were used as test data to evaluate the generalization ability of the method under different noise characteristics. Specifically, these included aliasing noise (Babble), engine noise, and factory noise. Different nonlinear conditions were set to test the robustness of the algorithm in real-world scenarios. ,in, It refers to nonlinear intensity. The performance is uniformly evaluated using the Normalized Mean Square Error (NMSE) as the evaluation metric. This metric calculates the error signal after noise reduction. With main channel noise signal The mean square error relative to The ratio of self-energy. The lower the normalized mean square error, the less residual noise after denoising, and the better the denoising performance of the algorithm. Table 1 shows a performance comparison between this invention and other methods.

[0094] Table 1

[0095]

[0096] As shown in Table 1, the method proposed in this invention achieves better active noise reduction performance under different nonlinear conditions.

[0097] The effect diagrams of active noise reduction of aliasing noise, engine noise, and factory noise using the method of this invention are shown below. Figure 6 , Figure 7 and Figure 8 As shown. Figure 6 (a) in the figure is the spectrum of the main channel noise corresponding to the aliasing noise. Figure 6 (b) in the figure is the spectrum of the error signal corresponding to the aliasing noise. Figure 7 (a) in the diagram is the spectrum of the main channel noise corresponding to the engine noise. Figure 7 (b) in the figure is the spectrum of the error signal corresponding to the engine noise. Figure 8 (a) in the diagram is the spectrum of the main channel noise corresponding to the factory noise. Figure 8 (b) in the diagram is the spectrum of the error signal corresponding to the factory noise. (Through...) Figure 6 , Figure 7 and Figure 8 It can be seen that the method of the present invention significantly reduces the noise amplitude under aliasing noise, engine noise and factory noise types. In the comparison segment, there is almost no perceptible noise in the quiet area, which further proves the effectiveness of the active noise reduction method in actual time sequence.

[0098] Furthermore, the performance of this invention is compared with that of a non-filtered neural network. The performance metrics and nonlinearity settings used in this comparison are the same as those in Table 1, and the specific results are shown in Table 2.

[0099] Table 2

[0100]

[0101] As can be seen from Table 2, under the condition of strong nonlinearity ( =0.1), non-filtered neural networks exhibit relatively high active noise reduction performance. However, as the degree of nonlinearity decreases, the performance advantage of the method of this invention gradually increases, and it performs better under near-linear conditions. Overall, the method of this invention demonstrates better overall performance and robustness under different nonlinear scenarios.

[0102] In existing technologies, besides deep learning-based active noise reduction methods, there are also adaptive filtering methods and hybrid methods combining traditional filtering and neural networks. The core of adaptive filtering methods is to collect noise signals through a reference microphone and dynamically generate an "anti-noise wave" with opposite phase and similar amplitude to the original noise using algorithms such as Least Mean Square (LMS) or Filtered-x Least Mean Square (FxLMS) that considers acoustic path delay. Noise cancellation is achieved through acoustic wave interference. The advantage of this method is its adaptability to changing noise environments (such as variations in the frequency of road noise inside a vehicle), and its relatively flexible noise reduction effect. However, its disadvantages include relatively high computational resource consumption and the need for precise control of algorithm parameters to balance noise reduction performance and system stability. It may also experience transient misalignment or convergence speed issues when noise changes abruptly. The core idea of ​​hybrid methods combining traditional filtering and neural networks is to combine the efficient adaptive capability of traditional adaptive filters with the powerful nonlinear modeling capability of deep neural networks to overcome the limitations of a single method. These methods typically employ a cascaded architecture: adaptive filters (such as FxLMS) are responsible for quickly tracking and processing the linear components of noise and handling abrupt changes in the acoustic path, while deep neural networks focus on modeling the intractable nonlinear parts of the system (such as the nonlinear distortion of loudspeakers) or accurately estimating the inverse model of complex acoustic paths. The advantage of this division of labor is that it retains the relatively low computational cost and real-time adaptability of traditional adaptive filters, while significantly improving the noise reduction depth for nonlinear noise and the system's generalization potential in unforeseen scenarios through neural networks. However, its disadvantages include increased system complexity, requiring careful trade-offs in the collaborative work of the two algorithms and resource allocation, and making the overall stability analysis of the hybrid system more challenging.

[0103] The model input of this invention includes noise signals and main channel estimation signals, enabling the deep neural network to perceive the main channel transfer function during learning, effectively learn channel characteristics, and generate cancellation signals that highly match the noise information of the main channel. This significantly improves the model's generalization ability and robustness under different acoustic environments. This invention uses a unique signal reconstruction measurement to reduce algorithm latency to one sampling point. Extremely low algorithm latency means faster system response and faster tracking of noise changes, thereby reducing the transient misalignment problem of adaptive filters during sudden noise changes and improving stability. This invention adopts a simple and efficient end-to-end deep learning architecture, avoiding the complexity of resource allocation. All computations are completed within a unified network framework, making it easier to optimize and deploy, achieving controllable computational complexity while ensuring performance.

[0104] In one embodiment, such as Figure 9 As shown, based on the above-mentioned active noise reduction method based on deep neural networks, the present invention also provides an active noise reduction system based on deep neural networks, the system comprising:

[0105] The first acquisition module 100 is used to acquire noise signals in real time via a reference microphone;

[0106] The second acquisition module 200 is used to acquire the most recent error signal recorded by the error microphone, and to obtain the current main channel estimation signal based on the most recent error signal and the historical reverse noise signal of the previous moment.

[0107] The cancellation signal generation module 300 is used to input the noise signal and the main channel estimation signal at the current moment into a trained deep neural network after preprocessing, obtain neural filtering parameters after processing, and generate cancellation signal based on the neural filtering parameters and the preprocessed noise signal.

[0108] The reverse noise signal generation module 400 plays the cancellation signal through a speaker to generate a reverse noise signal to cancel the main channel noise signal, which is generated by the noise signal propagating in the main channel.

[0109] In one embodiment, the system further includes:

[0110] The calculation unit is used to subtract the historical reverse noise signal from the previous moment from the most recent error signal to obtain the current moment's main channel estimation signal.

[0111] In one embodiment, the signal cancellation generation module includes:

[0112] The short-time Fourier transform unit is used to perform short-time Fourier transform on the noise signal and the main channel estimated signal at the current time to obtain the time-frequency domain representation of the noise signal and the time-frequency domain representation of the main channel estimated signal at the current time.

[0113] The network processing unit is used to input the time-frequency domain representation of the noise signal at the current moment and the time-frequency domain representation of the estimated signal of the main channel into the trained deep neural network, and obtain the neural filtering parameters through the deep neural network processing.

[0114] The cancellation signal generation unit is used to process the neural filtering parameters and the time-frequency domain representation of the noise signal at the current moment to generate a cancellation signal.

[0115] In one embodiment, the deep neural network includes a first convolutional block, four sequentially connected state space blocks, and a second convolutional block; the system further includes:

[0116] The first feature processing unit is used to input the fused time-frequency features into the first convolutional block, and after processing by the first convolutional block, obtain the first fused time-frequency features;

[0117] The second feature processing unit is used to input the first fused time-frequency feature into four sequentially connected state space blocks, and after sequential processing by each of the state space blocks, obtain the second fused time-frequency feature.

[0118] The third feature processing unit is used to input the second fused time-frequency feature into the second convolutional block, and then perform dimensionality reduction processing on the second fused time-frequency feature through the second convolutional block to obtain neural filter parameters.

[0119] In one embodiment, the system further includes:

[0120] The dot product unit is used to perform a complex multiplication of the neural filter parameters and the time-frequency domain representation of the noise signal at the current moment on a time-frequency basis to obtain the time-frequency domain signal;

[0121] The inverse short-time Fourier transform unit is used to perform inverse discrete Fourier transform on the time-frequency domain signal, and then perform inverse short-time Fourier transform on the inverse discrete Fourier transform result based on preset synthesis window parameters to generate a time-domain frame signal.

[0122] The filtering unit is used to sum the time-domain frame signals and select the last sampling point as the valid output from the summed time-domain signals of each frame.

[0123] The effective output splicing unit is used to splice all the effective outputs in chronological order to synthesize the cancellation signal at the current moment.

[0124] In one embodiment, the system further includes:

[0125] The training data acquisition unit is used to acquire a training dataset, which includes noise training signals at several times, main channel estimation training signals, and main channel noise training signals.

[0126] The training data integration unit is used to take the noise training signal and the main channel estimated training signal at the current time as input signals for each time step, and take the main channel noise training signal at the M sampling points after the current time step as the training target corresponding to the input signal.

[0127] The training unit is used to input the input signal into the deep neural network to be trained during the training process, calculate the time-domain mean square error of the result of eliminating the superposition of the training signal and the training target as the loss value, and use a backpropagation algorithm based on gradient descent to iteratively update the model parameters according to the loss value.

[0128] The training termination unit is used to repeat the model training steps until a preset termination condition is reached, thereby obtaining the trained deep neural network model.

[0129] Based on the above embodiments, the present invention also provides a terminal, the structural schematic diagram of which is as follows: Figure 10 As shown, the terminal includes a processor, memory, network interface, and display screen connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and a deep neural network-based active noise reduction program. The internal memory provides an environment for the operation of the operating system and the deep neural network-based active noise reduction program in the non-volatile storage medium. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the deep neural network-based active noise reduction program implements the steps of any of the aforementioned deep neural network-based active noise reduction methods. The display screen can be a liquid crystal display (LCD) or an e-ink display.

[0130] Those skilled in the art will understand that Figure 10 The structural schematic diagram shown is only a schematic diagram of a part of the structure related to the present invention solution, and does not constitute a limitation on the terminal on which the present invention solution is applied. The specific terminal may include more or fewer components than shown in the figure, or combine some components, or have different component arrangements.

[0131] In one embodiment, a terminal is provided, the terminal including a memory, a processor, and an active noise reduction program based on a deep neural network stored in the memory and executable on the processor. When the active noise reduction program based on a deep neural network is executed by the processor, it implements the steps of any active noise reduction method based on a deep neural network provided in the embodiments of the present invention.

[0132] This invention also provides a computer-readable storage medium storing an active noise reduction program based on a deep neural network. When executed by a processor, the active noise reduction program based on a deep neural network implements the steps of any active noise reduction method based on a deep neural network provided in this invention.

[0133] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0134] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the above system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this invention. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0135] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0136] Those skilled in the art will recognize that the units and algorithm steps of the various examples 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 implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0137] In the embodiments provided by this invention, it should be understood that the disclosed system / terminal device and method can be implemented in other ways. For example, the system / terminal device embodiments described above are merely illustrative. For instance, the division of modules or units described above is only a logical functional division, and in actual implementation, it can be divided in other ways. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0138] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not mean that the essence of the corresponding technical solutions deviates from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. An active noise reduction method based on deep neural networks, characterized in that, The method includes: Noise signals are acquired in real time using a reference microphone; The most recent error signal recorded by the error microphone is acquired. Based on the most recent error signal and the historical reverse noise signal of the previous moment, the estimated signal of the main channel at the current moment is obtained. The noise signal and the estimated main channel signal at the current moment are preprocessed and then input into the trained deep neural network. The neural filtering parameters are obtained after processing. Based on the neural filtering parameters and the preprocessed noise signal, a cancellation signal is generated. The deep neural network includes a first convolutional block, four state space blocks connected in sequence, and a second convolutional block. The cancellation signal is played through a speaker to generate a reverse noise signal to cancel out the main channel noise signal, which is generated by the noise signal propagating in the main channel. Based on the most recent error signal and the historical reverse noise signal from the previous moment, the current main channel estimation signal is obtained, including: Subtract the historical reverse noise signal from the previous moment from the most recent error signal to obtain the current main channel estimation signal; The noise signal and the estimated main channel signal at the current moment are preprocessed and then input into the trained deep neural network. After processing, the neural filtering parameters are obtained, including: Perform a short-time Fourier transform on the noise signal and the main channel estimated signal at the current moment to obtain the time-frequency domain representation of the noise signal and the time-frequency domain representation of the main channel estimated signal at the current moment; The time-frequency domain representation of the noise signal at the current moment and the time-frequency domain representation of the estimated signal of the main channel are input into the first convolutional block. After processing by the first convolutional block, the first fused time-frequency feature is obtained. The first fused time-frequency feature is input into four sequentially connected state space blocks. After being processed sequentially by each of the state space blocks, the second fused time-frequency feature is obtained. The second fused time-frequency feature is input into the second convolutional block, and the second convolutional block performs dimensionality reduction on the second fused time-frequency feature to obtain the neural filter parameters.

2. The active noise reduction method based on deep neural networks according to claim 1, characterized in that, Based on the neural filtering parameters and the preprocessed noise signal, a canceled signal is generated, including: The noise signal is processed based on the neural filtering parameters and the time-frequency domain representation of the noise signal at the current moment to generate a cancellation signal.

3. The active noise reduction method based on deep neural networks according to claim 2, characterized in that, Based on the neural filtering parameters and the noise signal at the current moment, a cancellation signal is generated, including: The neural filtering parameters and the current noise signal's time-frequency domain representation are multiplied by a complex number at each time frequency point to obtain the time-frequency domain signal; The time-frequency domain signal is subjected to an inverse discrete Fourier transform, and then the result of the inverse discrete Fourier transform is subjected to an inverse short-time Fourier transform based on preset synthesis window parameters to generate a time-domain frame signal. The time-domain frame signals are summed, and the last sampling point is selected as the valid output from the summed time-domain signals of each frame. All the valid outputs are concatenated in chronological order to form the cancellation signal for the current moment.

4. The active noise reduction method based on deep neural networks according to claim 1, characterized in that, The training steps of the deep neural network include: Obtain a training dataset, which includes noise training signals at several time points, main channel estimation training signals, and main channel noise training signals; For each time step, the noise training signal and the main channel estimation training signal at the current time step are used as input signals, and the main channel noise training signal at the M sampling points after the current time step step is used as the training target corresponding to the input signal. During training, the input signal is input into the deep neural network to be trained to obtain the eliminated training signal. The time-domain mean square error of the superposition result of the eliminated training signal and the training target is calculated as the loss value. The backpropagation algorithm based on gradient descent is used to iteratively update the model parameters according to the loss value. Repeat the model training steps until the preset termination condition is met to obtain the trained deep neural network model.

5. The active noise reduction method based on deep neural networks according to claim 4, characterized in that, M is the sum of the secondary channel delay and the preset system algorithm delay.

6. An active noise reduction system based on a deep neural network, characterized in that, include: The first acquisition module is used to acquire noise signals in real time via a reference microphone; The second acquisition module is used to acquire the most recent error signal recorded by the error microphone, and to obtain the current main channel estimation signal based on the most recent error signal and the historical reverse noise signal of the previous moment. The second acquisition module includes: The calculation unit is used to subtract the historical reverse noise signal from the previous moment from the most recent error signal to obtain the main channel estimation signal at the current moment. The signal cancellation generation module is used to input the noise signal and the main channel estimation signal at the current moment into the trained deep neural network after preprocessing. The neural filtering parameters are obtained after processing. Based on the neural filtering parameters and the preprocessed noise signal, the cancellation signal is generated. The deep neural network includes a first convolutional block, four state space blocks connected in sequence, and a second convolutional block. The signal cancellation generation module includes: The short-time Fourier transform unit is used to perform short-time Fourier transform on the noise signal and the main channel estimated signal at the current time to obtain the time-frequency domain representation of the noise signal and the time-frequency domain representation of the main channel estimated signal at the current time. The first feature processing unit is used to input the time-frequency domain representation of the noise signal at the current moment and the time-frequency domain representation of the main channel estimated signal into the first convolutional block, and obtain the first fused time-frequency feature after processing by the first convolutional block; The second feature processing unit is used to input the first fused time-frequency feature into four sequentially connected state space blocks, and after sequential processing by each of the state space blocks, obtain the second fused time-frequency feature. The third feature processing unit is used to input the second fused time-frequency feature into the second convolutional block, and perform dimensionality reduction processing on the second fused time-frequency feature through the second convolutional block to obtain neural filtering parameters; A reverse noise signal generation module is used to play the cancellation signal through a speaker to generate a reverse noise signal to cancel the main channel noise signal, which is generated by the noise signal propagating in the main channel.

7. A terminal, characterized in that, The terminal includes: a memory, a processor, and an active noise reduction program based on a deep neural network stored in the memory and executable on the processor. When the active noise reduction program based on the deep neural network is executed by the processor, it implements the steps of the active noise reduction method based on a deep neural network as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an active noise reduction program based on a deep neural network. When the active noise reduction program based on a deep neural network is executed by a processor, it implements the steps of the active noise reduction method based on a deep neural network as described in any one of claims 1-5.

Images