A deep learning-based active noise reduction algorithm
By using deep neural networks with multi-resolution encoding and temporal dependency modeling, combined with physical acoustic path constraint optimization, the problems of insufficient feature extraction and training target deviation in existing active noise reduction algorithms under complex noise environments are solved, and more efficient noise reduction effect is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING UNIV
- Filing Date
- 2026-02-13
- Publication Date
- 2026-06-05
AI Technical Summary
Existing deep learning-based active noise reduction algorithms suffer from limitations in noise reduction performance when dealing with broadband complex noise, including insufficient diversity in frequency domain feature extraction, inadequate attention to key frequency bands, and deviation between the training target and the physical sound field.
A deep neural network employing multi-resolution encoding, temporal dependency modeling, and jump-connection fusion decoding is optimized in conjunction with physical acoustic path constraints. Through multi-scale feature extraction and adaptive frequency band weight allocation, the nonlinear mapping capability of deep learning is utilized, and the system identification results of physical acoustic paths are combined for training.
It significantly improves noise reduction performance in complex environments, better captures global and local noise characteristics, improves the suppression efficiency of key frequency bands, and makes the generated control signal more consistent with the actual physical sound field requirements, thus reducing performance degradation.
Smart Images
Figure CN122157633A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of active noise control technology, and in particular to a novel active noise reduction algorithm based on deep learning. Background Technology
[0002] Active Noise Control (ANC) technology utilizes the principle of acoustic wave interference to cancel noise by emitting a secondary, out-of-phase signal with the same amplitude as the primary noise signal through a loudspeaker. With the rapid development of electronic technology and digital signal processing algorithms, ANC technology has been widely applied in headphones, automotive cabin noise reduction, and industrial pipeline noise control. Traditional active noise cancellation systems mainly rely on adaptive filter algorithms, such as the Least Mean Square (LMS) algorithm and its improved variant, the X-Least Mean Square (FxLMS) algorithm. The FxLMS algorithm compensates for the influence of the secondary channel on the out-of-phase signal by estimating the transfer function of the secondary path, offering advantages such as low computational complexity and good stability. However, these linear adaptive algorithms are essentially time-domain linear filters, and their control strategies are often based on a uniform error minimization criterion across the entire frequency band, lacking the ability to finely model the noise characteristics and acoustic response differences of different frequency bands. When dealing with broadband noise with complex spectral structures or uneven energy distribution, single-scale filter weight updates cannot simultaneously ensure effective suppression of all key frequency bands, easily leading to insufficient noise reduction in some frequency bands. Furthermore, traditional algorithms typically decouple signal processing from acoustic paths, making it difficult to fully utilize the deep time-frequency correlations and physical sound field constraints in the data for global optimization, thus limiting the overall noise reduction potential of the system in complex acoustic environments.
[0003] In recent years, deep learning technology, represented by deep neural networks, has achieved remarkable results in speech enhancement and audio processing, and researchers have begun to explore its application in the field of Automatic Control (ANC). Most existing deep learning-based ANC schemes use simple convolutional neural networks (CNNs) or recurrent neural networks (RNNs) to directly fit control filter coefficients or generate inverted waveforms. Although these schemes possess stronger nonlinear mapping capabilities compared to traditional algorithms, they still have the following shortcomings:
[0004] 1. The diversity of feature extraction needs to be improved: Existing network structures mostly use convolutional kernels of fixed size for feature extraction. This single-scale feature extraction method has room for improvement in its ability to take into account both local and global features when dealing with different types of broadband noise, which may limit further noise reduction performance in certain specific frequency bands.
[0005] 2. Attention to Key Frequency Bands: The energy distribution of real-world environmental noise is often uneven and dynamically changing. Existing methods typically assign equal weight to the characteristics of all frequency bands, rarely introducing adaptive weighting mechanisms for key high-energy frequency bands. This, to some extent, affects the efficiency of targeted elimination of major noise components.
[0006] 3. Deviation between training objectives and physical reality: Many existing deep ANC algorithms use only mean squared error (MSE) as the loss function during training, focusing only on the mathematical fit between the network output and the ideal inverse signal, while ignoring the influence of secondary paths in the actual physical sound field on the final residual error. This results in high network performance after training, but the noise reduction (NR) in actual listening or physical tests is still limited.
[0007] Therefore, there is a need for an active noise reduction algorithm that can utilize multi-scale frequency domain information, rationally allocate frequency band attention, and be directly optimized based on physical acoustic path constraints, in order to improve noise reduction performance in complex environments. Summary of the Invention
[0008] This invention addresses the shortcomings of existing deep learning-based active noise reduction algorithms in broadband complex noise control, including insufficient diversity in frequency domain feature extraction, inadequate attention to key frequency bands, and deviations between training targets and the actual physical sound field. It provides a deep learning-based active noise reduction algorithm that improves the actual noise reduction performance of the system through multi-resolution encoding, temporal dependency modeling, and skip-connection fusion decoding, while also incorporating physical acoustic path constraints for optimization.
[0009] This invention proposes a deep learning-based active noise reduction algorithm, which mainly includes:
[0010] A deep learning-based active noise reduction algorithm, characterized by comprising a path modeling module, a frequency domain feature extraction module, a deep neural network noise reduction model training module, and a physical constraint optimization module; wherein:
[0011] The path modeling module identifies the main path P and secondary path S of the active noise reduction system and obtains the corresponding discrete impulse responses p(n) and s(n) for subsequent training of the deep neural network.
[0012] The frequency domain feature extraction module performs frame segmentation, windowing, and short-time Fourier transform (STFT) on the real-time acquired reference signal x(n), converting the one-dimensional time-domain signal into a complex spectral feature tensor organized according to the time-frequency-channel dimension. Where T is the number of time frames, F is the number of frequency points, and C is the number of channels (including real and imaginary features) to preserve complete amplitude and phase information;
[0013] The deep neural network denoising model training module constructs a frequency domain deep neural network that includes multi-resolution encoding, temporal dependency modeling, and skip-connection fusion decoding, and performs the following processing: 1) Multi-resolution encoding: Feature extraction and frequency domain dimensionality reduction of the input feature X are performed through M-level cascaded multi-resolution encoding blocks; the i-th level encoding block contains K parallel convolutional branches, each branch uses a frequency-dimensional convolutional kernel of different size to filter the input feature, obtaining multi-scale intermediate features; after concatenating the outputs of each branch, the concatenated features are adaptively weighted using a channel attention mechanism (Squeeze-and-Excitation, SE) to generate the i-th level encoded feature E. i 1) And retain this feature for subsequent skip connections; 2) Temporal dependency modeling: encode the last-level feature E M Flattened along the frequency and channel dimensions, the input is fed into a multi-layer long short-term memory (LSTM) network. The temporal recursive characteristics of the LSTM capture the long-term dependencies of the noise signal, outputting a bottleneck feature H containing the global temporal context. 3) Skip-connected fusion decoding: The bottleneck feature H is reconstructed in the frequency domain through M-level cascaded decoding blocks; the j-th level decoding block upsamples the feature using transposed convolution, and the upsampled result is combined with the corresponding level's encoded feature E in the encoder. M-j+1 Perform skip connection fusion; before fusion, prune the upsampled features in the frequency dimension to eliminate the dimensionality mismatch caused by the convolution stride and padding, and ensure strict alignment of the frequency dimension; finally, the decoded output is used to generate the frequency domain control tensor Y to cancel noise.
[0014] The physical constraint optimization module is used to construct the physical acoustic loss function L during the training phase, restore the network output to the time domain signal and calculate the residual error by combining the secondary path characteristics, and minimize the loss function through the backpropagation algorithm to update the network parameters.
[0015] Furthermore, in the path modeling module, the following features are defined: Modeling a primary path P, based on a reference signal x and an error signal e, using the LMS algorithm, where noise is emitted by a noise source, and the reference signal x received by the reference microphone and the error signal e received by the error microphone work together to update the primary path P; Modeling a secondary path S, based on the error signal e and the LMS algorithm, where the secondary sound source emitted by the secondary speaker and the error signal e received by the error microphone work together to update the secondary path S.
[0016] In the further frequency domain feature extraction module, the characteristic is that: the reference signal is sampled at a frequency f s According to the preset frame length L = B × f s Frame shift R = H × f sThe process involves framing the data, where B represents the time length of the STFT window and H represents the time length of the STFT frame shift. The reference signal for each frame is multiplied by a Hanning window function to reduce spectral leakage. An N-point FFT Fast Fourier Transform is performed on each windowed frame to obtain the complex frequency spectrum X(t, f), where t is the time frame index, f is the frequency index, and N is greater than or equal to the frame length. The real and imaginary parts of the complex spectrum X(t, f) for each frame are stacked as two separate channels. A feature tensor X of size T×F×2 (where T is the number of time frames, F is the number of frequency points, and 2 represents two channels, real and imaginary) is formed and used as the input to the deep neural network denoising model. The specific parameters are set as follows: sampling rate fs = 16000Hz; frame length L = 256, frame shift R = 128; number of STFT transform points Nfft = 256, generating F = Nfft / 2 + 1 = 129 frequency points; the size of the input feature tensor X is (375, 129, 2).
[0017] In the further multi-resolution coding step, the specific operation process of the i-th level multi-resolution coding block is as follows: Let the input feature of this level be F. in Two parallel 2D convolutional branches are set up. The kernel size of the first branch is 1 x k1, and the kernel size of the second branch is 1 x k2, where k1 is not equal to k2, and the stride of both is 1 x s. Let the output features of the two branches be O1 and O2, respectively, and concatenate them along the channel dimension to obtain O. cat =Concat(O1, O2); for O cat Batch normalization and exponential linear unit activation (ELU) processing are performed sequentially; channel attention weighting is constructed: the processed O cat Global average pooling is performed to obtain the channel descriptor z. A channel weight vector w is generated using a two-layer fully connected network and a sigmoid activation function. The weighted feature O is then calculated. se =O cat *w, where * denotes channel-wise multiplication; O is adjusted by a projective convolutional layer with a kernel size of 1x1. se The number of channels is used to obtain the output feature E of this stage. i .
[0018] Further specific implementation steps of the deep neural network denoising model training module include:
[0019] S1. Data preparation: Using the measured primary path impulse response p(n) and secondary path impulse response s(n), the original reference signal x(n) is processed to generate primary path noise d(n) = x(n) * p(n) and complex spectral features X for network input;
[0020] S2. Network Construction:
[0021] A) Multi-resolution coding: Initialize network input F in =X; Construct an M-level concatenated multi-resolution coding structure; For the i-th level code (where i = 1, 2, 3, ..., M), convert the current F... in The input is given to the multi-resolution coding block, and after parallel convolution and channel attention weighting, the output coding feature E is obtained. i E i As input to the next level of encoding (updating F) in =E i ), while retaining E i For use in the decoding stage;
[0022] B) Temporal dependency modeling: Receive the Mth level encoded output E M The frequency dimension is further compressed by a downsampling convolutional layer with a stride of 1x2 to obtain the compressed feature F. bn ; F bn Flatten along the frequency and channel dimensions to construct the time series vector S in ;S in The input is fed into an LSTM network containing two stacked layers, each with 1024 hidden units, and the output is a sequence S that captures long-term dependencies. out ;S out Reconstructing back to the three-dimensional tensor format yields the bottleneck feature H;
[0023] C) Jump-connection fusion decoding: Initialize decoding input D in =H; Construct an M-level cascaded decoding structure; For the j-th level decoding (where j = 1, 2, 3, ..., M), first, the current D... in The input is fed into a transposed convolutional layer for upsampling to obtain intermediate features D. up ; For D up Prune the frequency dimension so that its frequency point count matches the corresponding encoded feature E. M-j+1 Strict consistency yields the trimming feature D. crop ; Perform a skip connection: compute D out =D crop +E M-j+1 D out As input for the next stage of decoding (updating D) in =D out );
[0024] D) Output Generation: Decode the Mth level and output D M The input is fed into the output convolutional layer, and after transposed convolution upsampling and final frequency clipping, the frequency dimension is restored to 129 points, generating a final control tensor Y with 2 channels.
[0025] The further physical constraint optimization process includes: transforming the frequency domain tensor Y into a time domain signal frame using the iSTFT operator, and performing overlapping and addition to obtain a continuous time domain signal y(n); performing time domain convolution between y(n) and the secondary path impulse response s(n) to obtain the secondary sound field signal y. s [n] = s[n] * y[n]; correspondingly, the residual error energy P_e = Mean((d(n) + y_s(n)) is calculated. 2 The noise energy of the main road, P_d = Mean(d(n)). 2 Construct a physical acoustic loss function L = P_e / (P_d + epsilon), where epsilon is a numerical stability constant; and construct a noise reduction evaluation index based on the loss function. To measure the training effect, a gradient descent-based optimization algorithm (preferably the Adam algorithm) is used to iteratively update the weights and bias parameters of the convolutional layers, LSTM layers, and deconvolutional layers in the network, so that the loss function L gradually decreases until the preset convergence condition is met.
[0026] In summary, this invention proposes a deep learning-based active noise reduction method. This method constructs a multi-scale feature encoding and bottleneck layer temporal modeling network, and trains it using a physical loss function. Its significant advantages are:
[0027] 1. This invention, through multi-resolution convolution branches, can simultaneously capture the global overview and local fine spectral structure of broadband noise, thus solving the problem of insufficient feature extraction in complex noise environments caused by single-scale convolution.
[0028] 2. This invention introduces an SE channel attention mechanism, which enables the network to automatically adjust the weights of different frequency band features according to the characteristics of the input noise, thereby improving noise reduction efficiency.
[0029] 3. This invention combines the local feature extraction capability of CNN with the long temporal memory capability of LSTM, and improves the real-time performance of the system by performing temporal modeling on the bottleneck layer after frequency domain compression.
[0030] 4. This invention, through a physical constraint optimization module, directly incorporates the system identification results of the acoustic path into the training process via a loss function, achieving a deep integration of the "signal processing algorithm" and the "physical acoustic environment." This training method makes the generated control signal more consistent with the destructive interference requirements of the actual physical sound field, significantly reducing performance degradation during actual deployment. Attached Figure Description
[0031] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.
[0032] Figure 2This is a schematic diagram of a deep neural network model architecture.
[0033] Figure 3 This is a schematic diagram of the multi-resolution coding block structure.
[0034] Figure 4 A schematic diagram of the module structure for modeling time-series dependencies.
[0035] Figure 5 This is a schematic diagram of the jump connection fusion and decoding module structure.
[0036] Figure 6 A schematic diagram illustrating the principle of physical constraint optimization training.
[0037] Figure 7 The graph shows a comparison of the noise reduction performance of the algorithm of this invention and the traditional FxLMS algorithm: (a) time domain and (b) frequency domain. Detailed Implementation
[0038] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings. However, the following embodiments are merely descriptive and not restrictive.
[0039] This embodiment discloses an active noise reduction method based on deep learning. For example... Figure 1 As shown, the method as a whole includes: path modeling preparatory steps, frequency domain feature extraction steps, deep neural network denoising model training, and physical constraint optimization steps; its specific implementation process includes the following detailed steps:
[0040] Step 1: Model the path of the actual system using a reference microphone and an error microphone. First, based on the reference signal x and the error signal e, the primary path p(n) is updated by the noise source emitting noise, the reference signal x received by the reference microphone, and the error signal e received by the error microphone, together with the LMS algorithm. Then, based on the error signal e and the LMS algorithm, the secondary sound source emitted by the secondary speaker and the error signal e received by the error microphone work together to update the secondary path s(n). The order of p(n) and s(n) is 192.
[0041] Step 2: As Figure 1To construct the signal feature tensor, the reference microphone signal x(n) and the error microphone signal e(n) in the active noise cancellation system are first obtained. The system sampling rate is set to fs = 16000Hz. The acquired time-domain signal is then processed by framing and windowing. In this embodiment, the frame length is set to 256 points (corresponding to 16ms), the frame shift is set to 128 points (corresponding to 8ms), and the Hann window function is used to reduce spectral leakage. The windowed signal is then subjected to a short-time Fourier transform (STFT) to obtain a complex spectrum with F = 129 frequency points. The real and imaginary parts of the reference and error signals are concatenated along the channel dimension to construct an input feature tensor Min with dimensions (T, 129, 2) or (T, 129, 4), where T is the number of time frames. In this embodiment, a 3-second audio segment is extracted during the training phase, corresponding to approximately 375 STFT frames T.
[0042] Step 3: Multi-resolution feature encoding. Input the feature tensor obtained in Step 1 into the CRN network (e.g., ...). Figure 2 The encoder section (shown) contains M=4 cascaded multi-resolution encoding modules. Figure 3 As shown, the specific processing steps of each encoding module are as follows: Parallel multi-scale extraction: Input features are simultaneously fed into two convolutional branches. The first branch uses a 1×3 convolutional kernel for local fine feature extraction; the second branch uses a 1×5 convolutional kernel for broadband contextual feature extraction. Both branches use a stride of (1, 2) for downsampling in the frequency dimension. Feature fusion and activation: The output features of the two branches are concatenated in the channel dimension, and then passed through a batch normalization layer and an ELU activation function layer. Channel attention enhancement: The fused features enter the compression-activation module. First, the (T, Fi, C) features are compressed into (T, 1, C) channel descriptors through global average pooling; then, they pass through the first fully connected layer (activation function is ReLU, dimensionality reduction ratio is 8) and the second fully connected layer (activation function is Sigmoid) to generate channel weights; finally, the weights are multiplied back into the original features to adaptively suppress background noise bands and enhance speech or pass through frequency bands. Specific parameter examples: The first-stage encoder has 16 output channels and the frequency dimension is reduced from 129 to about 65; the fourth-stage encoder has 128 output channels and the frequency dimension is reduced to about 9.
[0043] Step 4: Temporal Dependency Modeling To capture the long-term dependencies of noise signals over time, a bottleneck layer is set at the encoder output. For example... Figure 4As shown: Feature compression and flattening: The output features of the fourth-level encoder (with dimensions of approximately T×9×128) are first further compressed and reshaped through a 1×3 convolutional layer, flattening the frequency and channel dimensions to form a time-series vector. LSTM recursive processing: The flattened sequence is input into a two-layer stacked LSTM. In this embodiment, each LSTM layer contains 1024 hidden units to provide sufficient capacity to remember long-term acoustic environment changes. Feature reconstruction: The output sequence of the LSTM is projected through a fully connected layer and reshaped into a three-dimensional tensor format, serving as the initial input to the decoder.
[0044] Step 5: The skip-connect decoder with frequency clipping also contains a 4-level structure, designed to progressively restore low-dimensional features to the original spectral dimension. For example... Figure 5 As shown, for the j-th level decoding (j = 1, 2, 3, 4): Transposed convolutional upsampling: The input features are upsampled in the frequency dimension using a transposed convolutional layer of size 1×3 and stride (1, 2). Frequency clipping: Due to the padding strategy in the encoding process, the number of frequency points after direct upsampling may be 1-2 more than the corresponding encoding layer. This embodiment introduces an explicit clipping layer to remove redundant frequency points at the end of the frequency axis of the upsampled features, making them consistent with the corresponding encoder features E. M-j+1 Strict dimensional alignment. Skip connection fusion: The pruned decoded features are matched with the corresponding encoded features E. M-j+1 Element-wise addition is performed to fuse high-level semantic information with low-level spectral details. Final output: After the output of the last-stage decoder is pruned, a complex spectral prediction value Y with dimensions (T, 129, 2) is obtained. pred (Including real and imaginary parts).
[0045] Step 6: Parameter optimization under the physical loop. This invention does not use the traditional mean square error (MSE) as the loss function, but instead constructs a differentiable physical acoustic loop for end-to-end optimization. For example... Figure 6As shown: Differentiable time-domain reconstruction: Using the inverse short-time Fourier transform (iSTFT) operator and the overlap-add method, the frequency domain tensor Ypred output by the network is transformed into a continuous time-domain noise-resistant signal y(n). Secondary path filtering: y(n) is convolved in the time domain with the pre-identified secondary path impulse response s(n) to simulate the physical process of sound waves propagating from the loudspeaker to the error microphone, resulting in the secondary sound field signal ys(n) = y(n) * s(n). Physical error calculation: The physical residual error e(n) = d(n) + ys(n) is calculated, where d(n) is the main path noise. Energy ratio loss function: The ratio of residual error energy Pe to main path noise energy Pd is used as the loss function L = P_e / (P_d + epsilon). Using the Adam optimizer (learning rate set to 0.001), gradient descent is applied to all convolutional kernel weights, LSTM unit parameters, and bias terms in the network based on this loss function L until the loss function converges, thereby directly maximizing the actual physical noise reduction. For example... Figure 7 In the noise reduction experiment, this embodiment achieved a noise reduction of 16.63 dB, which is a significant improvement compared to the 10.49 dB noise reduction of the traditional FxLMS algorithm. As can be seen from the comparison, the method of the present invention has achieved significantly better noise suppression effect in a wide frequency band. This fully demonstrates that the multi-resolution feature extraction and physical constraint optimization mechanism proposed in this invention can effectively break through the performance bottleneck of traditional linear algorithms and achieve deeper active elimination of environmental noise.
[0046] It should be noted that the above embodiments are not intended to limit the scope of protection of the present invention. Equivalent transformations or substitutions made on the basis of the above technical solutions, as well as several improvements such as extensions based on them, all fall under the protection of the claims of the present invention.
Claims
1. A deep learning-based active noise reduction algorithm, characterized in that, It includes a path modeling module, a frequency domain feature extraction module, a deep neural network noise reduction model training module, and a physical constraint optimization module; among which: The path modeling module identifies the main path P and secondary path S of the active noise reduction system and obtains the corresponding discrete impulse responses p(n) and s(n) for subsequent training of the deep neural network. The frequency domain feature extraction module performs frame segmentation, windowing, and short-time Fourier transform (STFT) on the real-time acquired reference signal x(n), converting the one-dimensional time-domain signal into a complex spectral feature tensor organized according to the time-frequency-channel dimension. Where T is the number of time frames, F is the number of frequency points, and C is the number of channels (including real and imaginary features) to preserve complete amplitude and phase information; The deep neural network denoising model training module constructs a frequency domain deep neural network that includes multi-resolution encoding, temporal dependency modeling, and skip-connection fusion decoding, and performs the following processing: 1) Multi-resolution encoding: Feature extraction and frequency domain dimensionality reduction of the input feature X are performed through M-level cascaded multi-resolution encoding blocks; the i-th level encoding block contains K parallel convolutional branches, each branch uses a frequency-dimensional convolutional kernel of different size to filter the input feature, obtaining multi-scale intermediate features; after concatenating the outputs of each branch, the concatenated features are adaptively weighted using a channel attention mechanism (Squeeze-and-Excitation, SE) to generate the i-th level encoded feature E. i 1) And retain this feature for subsequent skip connections; 2) Temporal dependency modeling: encode the last-level feature E M Flattened along the frequency and channel dimensions, the input is fed into a multi-layer long short-term memory (LSTM) network. The temporal recursive characteristics of the LSTM capture the long-term dependencies of the noise signal, outputting a bottleneck feature H containing the global temporal context. 3) Skip-connected fusion decoding: The bottleneck feature H is reconstructed in the frequency domain through M-level cascaded decoding blocks; the j-th level decoding block upsamples the feature using transposed convolution, and the upsampled result is combined with the corresponding level's encoded feature E in the encoder. M-j+1 Perform skip connection fusion; before fusion, prune the upsampled features in the frequency dimension to eliminate the dimensionality mismatch caused by the convolution stride and padding, and ensure strict alignment of the frequency dimension; finally, the decoded output is used to generate the frequency domain control tensor Y to cancel noise. The physical constraint optimization module is used to construct the physical acoustic loss function L during the training phase, restore the network output to the time domain signal and calculate the residual error by combining the secondary path characteristics, and minimize the loss function through the backpropagation algorithm to update the network parameters.
2. The active noise reduction algorithm based on deep learning according to claim 1, wherein the path modeling module of the system software algorithm is characterized in that: Model the primary path P. Based on the reference signal x and the error signal e, and according to the LMS algorithm, the noise source emits noise, and the reference signal x received by the reference microphone and the error signal e received by the error microphone work together to update the primary path P. Model the secondary path S. Based on the error signal e and the LMS algorithm, the secondary sound source emitted by the secondary speaker and the error signal e received by the error microphone work together to update the secondary path S.
3. The active noise reduction algorithm based on deep learning according to claim 1, wherein the frequency domain feature extraction module is characterized in that: For the reference signal at a sampling frequency f s According to the preset frame length L = B × f s Frame shift R = H × f s Frame segmentation is performed, where B represents the time length of the STFT window and H represents the time length of the STFT frame shift. The reference signal of each frame is multiplied by a Hanning window function to reduce spectral leakage. For each windowed frame, an N-point FFT fast Fourier transform is performed to obtain the frequency domain complex spectrum X(t, f) of that frame, where t is the time frame index, f is the frequency index, and N is greater than or equal to the frame length. The real and imaginary parts of the complex spectrum X(t,f) of each frame of STFT are stacked as two channels to form a feature tensor X of size T×F×2 (where T is the number of time frames, F is the number of frequency points, and 2 represents two channels for the real and imaginary parts), which is then used as the input to the deep neural network denoising model.
4. The active noise reduction algorithm based on deep learning according to claim 1, characterized in that, The specific parameters of the frequency domain feature extraction module are set as follows: the input audio is 3 seconds long, the sampling rate is fs = 16000Hz; the frame length is L = 256, the frame shift is R = 128; the number of STFT transformation points is Nfft = 256, generating F = Nfft / 2 + 1 = 129 frequency points; the size of the input feature tensor X is (375, 129, 2).
5. The active noise reduction algorithm based on deep learning according to claim 1, characterized in that, In the multi-resolution coding step, the specific operation process of the i-th level multi-resolution coding block is as follows: Let the input feature of this level be F. in Two parallel 2D convolutional branches are set up. The kernel size of the first branch is 1 x k1, and the kernel size of the second branch is 1 x k2, where k1 is not equal to k2, and the stride of both is 1 x s. Let the output features of the two branches be O1 and O2, respectively, and concatenate them along the channel dimension to obtain O. cat =Concat(O1, O2); for O cat Batch normalization and exponential linear unit activation (ELU) processing are performed sequentially; channel attention weighting is constructed: the processed O cat Global average pooling is performed to obtain the channel descriptor z. A channel weight vector w is generated using a two-layer fully connected network and a sigmoid activation function. The weighted feature O is then calculated. se =O cat *w, where * denotes channel-wise multiplication; O is adjusted by a projective convolutional layer with a kernel size of 1x1. se The number of channels is used to obtain the output feature E of this stage. i .
6. The active noise reduction algorithm based on deep learning according to claim 1, characterized in that, The specific implementation steps of the deep neural network noise reduction model training module include: S1. Data preparation: Using the measured primary path impulse response p(n) and secondary path impulse response s(n), the original reference signal x(n) is processed to generate primary path noise d(n) = x(n) * p(n) and complex spectral features X for network input; S2. Network Construction: A) Multi-resolution coding: Initialize network input F in =X; Construct an M-level concatenated multi-resolution coding structure; For the i-th level code (where i = 1, 2, 3, ..., M), convert the current F... in The input is given to the multi-resolution coding block, and after parallel convolution and channel attention weighting, the output coding feature E is obtained. i E i As input to the next level of encoding (updating F) in =E i ), while retaining E i For use in the decoding stage; B) Temporal dependency modeling: Receive the Mth level encoded output E M The frequency dimension is further compressed by a downsampling convolutional layer with a stride of 1x2, resulting in the compressed feature F. bn ; F bn Flatten along the frequency and channel dimensions to construct the time series vector S in ;S in The input is fed into an LSTM network containing two stacked layers, each with 1024 hidden units, and the output is a sequence S that captures long-term dependencies. out ;S out Reconstructing back to the three-dimensional tensor format yields the bottleneck feature H; C) Jump-connection fusion decoding: Initialize decoding input D in =H; Construct an M-level cascaded decoding structure; For the j-th level decoding (where j = 1, 2, 3, ..., M), first, the current D... in The input is fed into a transposed convolutional layer for upsampling to obtain intermediate features D. up ; For D up Prune the frequency dimension so that its frequency point count matches the corresponding encoded feature E. M-j+1 Strict consistency yields the clipping feature D. crop ; Perform a skip connection: compute D ou t = D crop +E M-j+1 D out As input for the next stage of decoding (updating D) in =D out ); D) Output Generation: Decode the Mth level and output D M The input is fed into the output convolutional layer, and after transposed convolution upsampling and final frequency clipping, the frequency dimension is restored to 129 points, generating a final control tensor Y with 2 channels.
7. The active noise reduction algorithm based on deep learning according to claim 1, characterized in that, The specific implementation process of the physical constraint optimization includes: transforming the frequency domain tensor Y into a time domain signal frame using the inverse short-time Fourier transform (iSTFT) operator, and performing overlapping and addition to obtain a continuous time domain signal y(n); performing time domain convolution between y(n) and the secondary path impulse response s(n) to obtain the secondary sound field signal y. s [n] = s[n] * y[n]; correspondingly, the residual error energy P_e = Mean((d(n) + y_s(n)) is calculated. 2 The noise energy of the main road, P_d = Mean(d(n)). 2 Construct a physical acoustic loss function L = P_e / (P_d + epsilon), where epsilon is a numerical stability constant; and construct a noise reduction evaluation index based on the loss function. To measure the training effect, a gradient descent-based optimization algorithm (preferably the Adam algorithm) is used to iteratively update the weights and bias parameters of the convolutional layers, LSTM layers, and deconvolutional layers in the network, so that the loss function L gradually decreases until the preset convergence condition is met.