An acoustic noise reduction module and a noise reduction control method for an enclosed area independent sound field
By combining a distributed acoustic sensor array with a deep reinforcement learning network, the sound field energy density distribution map is reconstructed and the listening zone is dynamically adjusted, solving the spatial overflow problem of traditional active noise cancellation technology and realizing the construction of a high-quality independent sound field and noise control.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- DONGGUAN ALLLIKE ELECTRONICS CO LTD
- Filing Date
- 2026-03-17
- Publication Date
- 2026-06-05
Smart Images
Figure CN122157634A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of acoustic signal processing technology, specifically to an audio noise reduction module and a noise reduction control method for an independent sound field in a closed area. Background Technology
[0002] As people's demands for quality of life increase, the creation of high-quality, independent sound fields in enclosed areas such as car cabins, home theaters, and office cubicles has become an urgent need. Traditional active noise cancellation (ANC) technology is usually based on adaptive filtering theory, achieving noise reduction by minimizing noise error at the error microphone. However, this method often suffers from the "spatial spillover effect," meaning that while eliminating noise at one point, it leads to an increase in noise energy at other locations in the space. Furthermore, existing deep learning-based noise reduction methods mostly focus on predicting time-domain waveforms, making it difficult to accurately control the distribution of sound field energy in three-dimensional space, and thus unable to create stable, independent quiet zones that move with the user in open spaces. Summary of the Invention
[0003] To address the shortcomings of existing technologies, this invention provides an audio noise reduction module and a noise reduction control method for an independent sound field in a closed area. By real-time reconstruction of sound field energy density and optimization control through deep reinforcement learning, it achieves efficient noise reduction and independent sound field construction in the target area, while suppressing noise spillover into non-target areas.
[0004] To achieve the above objectives, the present invention provides the following technical solution: an audio noise reduction module, comprising:
[0005] A distributed acoustic sensor array is used to collect sound pressure data in a closed area in real time.
[0006] An optical positioning unit is used to obtain the user's ear coordinates within an enclosed area;
[0007] The sound field reconstruction processing unit is used to calculate the sound field energy density distribution map based on the sound pressure data using a near-field acoustic holography algorithm.
[0008] A deep reinforcement learning control unit, comprising a pre-trained deep reinforcement learning network, is used to output control action vectors based on state vectors and update network parameters based on reward values.
[0009] The signal synthesis and driving unit is used to generate multi-channel driving signals based on the control action vector;
[0010] The distributed sound unit includes multiple speakers for playing sound according to multi-channel drive signals.
[0011] Preferably, the distributed acoustic sensor array consists of no fewer than 16 MEMS microphones, which are evenly distributed at the boundary or top of the enclosed area;
[0012] The distributed sound unit includes a high-frequency speaker embedded in the headrest and a low-frequency speaker embedded in the backrest.
[0013] Preferably, a noise reduction control method for an independent sound field in a closed area, using the aforementioned audio noise reduction module, includes the following steps:
[0014] S1: Real-time acquisition of sound pressure data within a closed area using a distributed acoustic sensor array, and reconstruction of the spatial sound field based on a near-field acoustic holography algorithm to obtain a sound field energy density distribution map;
[0015] S2: Based on the user's ear coordinates obtained by the optical positioning system, the target listening area and the non-target free area are divided in the sound field energy density distribution map, a state space for deep reinforcement learning is constructed, and a state vector in the state space is generated based on the measured data at the current moment.
[0016] S3: Input the state vector into a pre-trained deep reinforcement learning network, and output the control action vector for the distributed vocal unit through the policy network inference.
[0017] S4: Generate multi-channel driving signals based on the control action vector and play them through distributed sound units. At the same time, calculate the reward value at the current moment and update the parameters of the deep reinforcement learning network based on the reward value using the gradient descent method.
[0018] S5: When the user's head position is detected to have moved beyond the preset threshold, steps S2 to S4 are re-executed to achieve dynamic tracking of the acoustic potential energy trap.
[0019] Preferably, step S1 specifically includes:
[0020] Acquire the sound pressure signals collected by each microphone in the distributed acoustic sensor array at time t. Where k = 1, ..., K, and K is the number of microphones;
[0021] A sound field reconstruction model is established based on the equivalent source method, and the virtual equivalent source intensity vector is solved. Calculate the sound pressure at any point r in space. and normal particle velocity ;
[0022] According to the sound pressure and normal particle velocity Calculate the instantaneous sound field energy density at this location. ;
[0023] The closed region is discretized into a grid, based on each grid point. This constitutes the current sound field energy density distribution map. .
[0024] Preferably, the virtual equivalent source strength vector Solve using the following formula:
[0025] ;
[0026] in, Represents the identity matrix. Represents the regularization parameter. Indicates conjugate transpose;
[0027] Calculate the instantaneous sound field energy density at that moment. :
[0028] ;
[0029] in, Indicates air density, It indicates the speed of sound.
[0030] Preferably, the specific process of S2 includes:
[0031] Based on the user's ear coordinates obtained by the optical positioning system, the sound field energy density distribution map is divided into a target listening zone and a non-target free zone. A state space for deep reinforcement learning is constructed, and a state vector within this state space is generated based on the measured data at the current moment.
[0032] Using an optical camera to capture feature points on the user's head, the coordinates of the left ear are calculated. and right ear coordinates The target hearing zone is defined as a sphere with radius R centered on the ear coordinates. The remaining part of the closed region is defined as the non-target free zone. ;
[0033] Define the set of dimensions for the deep reinforcement learning agent to perceive the environment, and obtain the state space;
[0034] The state space consists of the following three physical characteristics: the sound field energy characteristics of the target listening area, the sound field energy characteristics of the non-target free area, and the spatial position characteristics of the ear in the closed area.
[0035] Extracting the target listening zone from the sound field energy density distribution map Non-target free zone Based on the energy statistics within the time frame and combined with user location information, a state vector at time t is constructed. :
[0036] ;
[0037] in, Indicates the average energy density of the target area. Indicates the standard deviation of energy in the target region. This represents the average energy density of the free region.
[0038] Preferably, the policy network receives the state vector. Output the motion vector for L loudspeakers. :
[0039] ;
[0040] in, This represents the gain coefficient of the l-th loudspeaker. This represents the phase offset of the l-th speaker.
[0041] Preferably, the policy network The specific process for outputting the optimal combination of gain coefficient and phase offset based on the current sound field state includes:
[0042] Normalization process: Receive the generated state vector The state vector is normalized.
[0043] Feature extraction and nonlinear transformation: The normalized state vector is input into the hidden layer of the policy network. The policy network adopts a multi-layer fully connected neural network structure. Through multi-layer linear transformation and nonlinear activation function, deep features in the sound field state are extracted, and the mapping relationship between the sound field energy distribution and the loudspeaker control parameters is established.
[0044] The calculation process for the l-th hidden layer is as follows:
[0045] ;
[0046] in, This indicates the output of the previous layer. This is the normalized state vector. and Let these represent the weight matrix and bias vector of the l-th layer, respectively. Represents the linear rectification activation function;
[0047] Output layer raw parameter calculation: After the last hidden layer, the network enters the output layer. The number of neurons in the output layer depends on the number of distributed vocal units L. The output layer outputs a raw parameter vector containing 2L elements.
[0048] The original vector can be divided into two parts: the first L elements correspond to the original values of the gain. The last L elements correspond to the original values of the phase. ;
[0049] Physical constraint mapping of motion parameters: The original gain values are mapped to the [0, 1] interval using the Sigmoid activation function;
[0050] Using the Sigmoid activation function in conjunction with a scaling factor, the original phase values are mapped to the interval [0, 2π].
[0051] Assembly and output of control action vectors: This involves mapping all gain coefficients... and phase offset The speakers are arranged in a staggered or grouped manner according to a preset speaker sequence, and assembled into the final control action vector. .
[0052] Preferably, the step of generating multi-channel drive signals based on control action vectors specifically includes:
[0053] Obtain the audio signal and reference noise signal that need to be played, and generate the final drive signal for the Lth speaker based on the control action vector;
[0054] ;
[0055] in, Indicates the final drive signal. This indicates the audio signal that needs to be played. Indicates based on reference noise signal The generated noise reduction reference signal.
[0056] Preferably, the reward value is calculated using a reward function, which is:
[0057] ;
[0058] in, This represents minimizing the energy of the target region. This represents minimizing the energy of the free region. This indicates maximizing the fidelity of the audio signal in the target area. This represents the feature vector of the audio signal actually acquired or reconstructed in the target listening area. This represents the feature vector of the original audio signal to be played. This indicates a power consumption limit.
[0059] This invention provides an audio noise reduction module and a noise reduction control method for an independent sound field in a closed area, which has the following beneficial effects:
[0060] This invention effectively solves the problems of "spatial spillover effect" in traditional active noise cancellation technology and the difficulty of accurately controlling the sound field distribution in three-dimensional space by existing deep learning methods through real-time reconstruction of sound field energy density and optimized control by deep reinforcement learning. The method utilizes a distributed acoustic sensor array combined with a near-field acoustic holography algorithm to reconstruct the sound field energy density distribution map, enabling precise perception of sound field details in space. It also uses an optical positioning unit to acquire the user's ear coordinates in real time, dynamically dividing the target listening zone and non-target free zone to construct a state space containing sound field energy features and positional features. Based on this, a pre-trained deep reinforcement learning network outputs a state space containing... The control vectors for gain weight and phase shift, through the design of a comprehensive reward function that includes minimizing energy in the target area, suppressing energy in the non-target area, maximizing audio fidelity, and limiting power consumption, effectively suppress noise spillover into non-target areas while eliminating noise in the target area, thus avoiding the phenomenon of increased noise in other locations in the space during the noise reduction process. In addition, the system can monitor the user's head movement in real time, and automatically trigger a repositioning process when the position change exceeds a threshold, realizing the dynamic tracking of the "sound potential energy trap". This creates a high-quality, highly stable, and spatially selective independent sound field for mobile users within a closed area, significantly improving the user's listening experience. Attached Figure Description
[0061] Figure 1 This is a schematic diagram of an audio noise reduction module provided in this embodiment;
[0062] Figure 2 This is a schematic diagram of the noise reduction control method for an independent sound field in a closed area according to the present invention.
[0063] Figure 3 This is a schematic diagram illustrating the process of the strategy network outputting the optimal combination of gain coefficient and phase offset according to the present invention. Detailed Implementation
[0064] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0065] like Figure 1 As shown, an embodiment of the present invention provides an audio noise reduction module, including:
[0066] A distributed acoustic sensor array is used to collect sound pressure data in a closed area in real time.
[0067] An optical positioning unit is used to obtain the user's ear coordinates within an enclosed area;
[0068] The sound field reconstruction processing unit is used to calculate the sound field energy density distribution map based on the sound pressure data using a near-field acoustic holography algorithm.
[0069] The deep reinforcement learning control unit includes a pre-trained deep reinforcement learning network for outputting control action vectors based on state vectors and updating deep reinforcement learning network parameters based on reward values.
[0070] The signal synthesis and driving unit is used to generate multi-channel driving signals based on the control action vector;
[0071] The distributed sound unit includes multiple speakers for playing sound according to multi-channel drive signals.
[0072] In this embodiment, the distributed acoustic sensor array consists of no fewer than 16 MEMS microphones, which are evenly distributed at the boundary or top of the enclosed area.
[0073] The distributed sound unit includes a high-frequency speaker embedded in the headrest and a low-frequency speaker embedded in the backrest.
[0074] like Figure 2 As shown, this embodiment also provides a noise reduction control method for an independent sound field in a closed area, which applies the above-mentioned audio noise reduction module and includes the following steps:
[0075] S1: Real-time acquisition of sound pressure data within a closed area using a distributed acoustic sensor array, and reconstruction of the spatial sound field based on a near-field acoustic holography algorithm to obtain a sound field energy density distribution map;
[0076] Specifically, K distributed acoustic sensor arrays are uniformly arranged at the boundary of the enclosed area (such as the top of the enclosed area, under the seat, etc.). The sound pressure signal collected by the k-th acoustic sensor at time t is... ;
[0077] Among them, the acoustic sensor array can be an array of microphones;
[0078] The sound field of a closed region is reconstructed using the equivalent source method;
[0079] Specifically, assuming there are N virtual equivalent sources within the closed region, their intensity vectors are... The sound pressure vector received by the acoustic sensor array It can be represented as: ;
[0080] in, To pass the matrix, elements Let Green's function represent the distance from the nth equivalent source to the kth acoustic sensor;
[0081] Solve for the equivalent source strength: ;
[0082] in, Represents the identity matrix. Represents the regularization parameter. This indicates the conjugate transpose.
[0083] Specifically, the sound pressure at any point r in space and normal particle velocity It can be calculated using the following formula:
[0084] ;
[0085] ;
[0086] in, and These are the sound pressure and vibration velocity transmission vectors from the equivalent source to position r, respectively.
[0087] Sound pressure at any point r in space and normal particle velocity Calculate the instantaneous sound field energy density at that moment. :
[0088] ;
[0089] in, Indicates air density, It indicates the speed of sound.
[0090] The closed region is discretized into a grid, based on each grid point. This constitutes the current sound field energy density distribution map. .
[0091] S2: Based on the user's ear coordinates obtained by the optical positioning system, the target listening area and the non-target free area are divided in the sound field energy density distribution map, a state space for deep reinforcement learning is constructed, and a state vector in the state space is generated based on the measured data at the current moment.
[0092] Specifically, based on the user's ear coordinates obtained by the optical positioning system, the sound field energy density distribution map is divided into a target listening zone and a non-target free zone. A state space for deep reinforcement learning is constructed, and a state vector within this state space is generated based on the measured data at the current moment.
[0093] Using an optical camera to capture feature points on the user's head, the coordinates of the left ear are calculated. and right ear coordinates The target hearing zone is defined as a sphere with radius R centered on the ear coordinates. The remaining part of the closed region is defined as the non-target free zone. ;
[0094] Define the set of dimensions for the deep reinforcement learning agent to perceive the environment, and obtain the state space;
[0095] The state space consists of the following three physical characteristics: the sound field energy characteristics of the target listening area, the sound field energy characteristics of the non-target free area, and the spatial position characteristics of the ear in the closed area.
[0096] Understandably, the state space defines the range of environmental information required for an agent to make decisions, and determines the structure of the input layer of the subsequent deep reinforcement learning network.
[0097] Extracting the target listening zone from the sound field energy density distribution map Non-target free zone Based on the energy statistics within the time frame and combined with user location information, a state vector at time t is constructed. :
[0098] ;
[0099] in, Indicates the average energy density of the target area. This represents the standard deviation of energy in the target area (reflecting the stability of the sound field). This represents the average energy density of the free region.
[0100] It is understandable that the state vector It reflects the current sound field distribution and the user's location.
[0101] S3: Input the state vector into a pre-trained deep reinforcement learning network, and output the control action vector for the distributed vocal unit through the policy network inference.
[0102] The control motion vector includes the gain weight and phase offset of each sound unit;
[0103] In this embodiment, a deep reinforcement learning algorithm based on the Actor-Critic architecture is used;
[0104] Specifically, the policy network receives the state vector. Output the motion vector for L loudspeakers. :
[0105] ;
[0106] in, This represents the gain coefficient of the l-th loudspeaker. This represents the phase offset of the l-th speaker.
[0107] Policy Network Based on the current sound field state, output the optimal combination of gain coefficient and phase offset;
[0108] like Figure 3 As shown, in this embodiment, the policy network The specific process for outputting the optimal combination of gain coefficient and phase offset based on the current sound field state includes:
[0109] Normalization process: Receive the generated state vector The state vector is normalized.
[0110] Feature extraction and nonlinear transformation: The normalized state vector is input into the hidden layer of the policy network. The policy network adopts a multi-layer fully connected neural network structure. Through multi-layer linear transformation and nonlinear activation function, deep features in the sound field state are extracted, and the mapping relationship between the sound field energy distribution and the loudspeaker control parameters is established.
[0111] The calculation process for the l-th hidden layer is as follows:
[0112] ;
[0113] in, This indicates the output of the previous layer. This is the normalized state vector. and Let these represent the weight matrix and bias vector of the l-th layer, respectively. Represents the linear rectification activation function;
[0114] Output layer raw parameter calculation: After the last hidden layer, the network enters the output layer. The number of neurons in the output layer depends on the number of distributed vocal units L. The output layer outputs a raw parameter vector containing 2L elements.
[0115] The original vector can be divided into two parts: the first L elements correspond to the original values of the gain. The last L elements correspond to the original values of the phase. ;
[0116] Physical constraint mapping of motion parameters: The original gain values are mapped to the [0, 1] interval using the Sigmoid activation function;
[0117] Using the Sigmoid activation function in conjunction with a scaling factor, the original phase values are mapped to the interval [0, 2π].
[0118] Assembly and output of control action vectors: This involves mapping all gain coefficients... and phase offset The speakers are arranged in a staggered or grouped manner according to a preset speaker sequence, and assembled into the final control action vector. .
[0119] It should be noted that during the system's training phase, in order to explore better strategies, output actions will be... Superimposed Gaussian noise In this embodiment, the policy network is in exploitation mode and directly outputs the aforementioned determined... Random noise is no longer added to ensure the stability of noise reduction control.
[0120] S4: Generate multi-channel driving signals based on the control action vector and play them through distributed sound units. At the same time, calculate the reward value at the current moment and update the parameters of the deep reinforcement learning network based on the reward value using the gradient descent method.
[0121] In this embodiment, the specific process of generating multi-channel drive signals based on control action vectors, playing them through distributed sound units, and simultaneously calculating the reward value at the current moment includes:
[0122] Obtain the audio signal and reference noise signal that need to be played, and generate the final drive signal for the Lth speaker based on the control action vector;
[0123] ;
[0124] in, Indicates the final drive signal. This indicates the audio signal that needs to be played. Indicates based on reference noise signal The generated noise reduction reference signal;
[0125] It should be noted that the noise reduction reference signal is usually the inverse of the reference noise or a filtered signal.
[0126] In this embodiment, the reward value is calculated using a reward function, which is:
[0127] ;
[0128] in, This means minimizing the energy of the target region to achieve noise reduction. This represents minimizing the energy of the free region to prevent noise from spilling out to other locations in space during the cancellation process. This means maximizing the fidelity of the audio signal in the target area to ensure that noise reduction does not affect listening experience. This represents the feature vector of the audio signal actually acquired or reconstructed in the target listening area. This represents the feature vector of the original audio signal to be played. This indicates a power consumption limit to prevent excessive gain.
[0129] It is understandable that using gradient descent to update the parameters of deep reinforcement learning networks is a current technique and will not be elaborated upon here.
[0130] S5: When the user's head position is detected to have moved beyond the preset threshold, steps S2 to S4 are re-executed to achieve dynamic tracking of the acoustic potential energy trap.
[0131] Specifically, the optical positioning system continuously monitors the user's ear coordinates and calculates the distance difference between the current position and the previous position. ;like If the value exceeds a preset threshold, it is determined that the user has moved significantly, and the relocation process is immediately triggered. The target listening zone and the non-target free zone are redefined, and reasoning and control are re-performed based on the new state vector, so that the formed "sound potential energy trap" (low energy zone) can closely follow the user's ear movement.
[0132] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A sound noise reduction module, characterized in that, include: A distributed acoustic sensor array is used to collect sound pressure data in a closed area in real time. An optical positioning unit is used to obtain the user's ear coordinates within an enclosed area; The sound field reconstruction processing unit is used to calculate the sound field energy density distribution map based on the sound pressure data using a near-field acoustic holography algorithm. A deep reinforcement learning control unit, comprising a pre-trained deep reinforcement learning network, is used to output control action vectors based on state vectors and update network parameters based on reward values. The signal synthesis and driving unit is used to generate multi-channel driving signals based on the control action vector; The distributed sound unit includes multiple speakers for playing sound according to multi-channel drive signals.
2. The audio noise reduction module according to claim 1, characterized in that, The distributed acoustic sensor array consists of no fewer than 16 MEMS microphones, which are evenly distributed at the boundary or top of the enclosed area. The distributed sound unit includes a high-frequency speaker embedded in the headrest and a low-frequency speaker embedded in the backrest.
3. A noise reduction control method for an independent sound field in a closed area, employing the audio noise reduction module described in claim 1 or 2, characterized in that, Includes the following steps: S1: Real-time acquisition of sound pressure data within a closed area using a distributed acoustic sensor array, and reconstruction of the spatial sound field based on a near-field acoustic holography algorithm to obtain a sound field energy density distribution map; S2: Based on the user's ear coordinates obtained by the optical positioning system, the target listening area and the non-target free area are divided in the sound field energy density distribution map, a state space for deep reinforcement learning is constructed, and a state vector in the state space is generated based on the measured data at the current moment. S3: Input the state vector into a pre-trained deep reinforcement learning network, and output the control action vector for the distributed vocal unit through the policy network inference. S4: Generate multi-channel driving signals based on the control action vector and play them through distributed sound units. At the same time, calculate the reward value at the current moment and update the parameters of the deep reinforcement learning network based on the reward value using the gradient descent method. S5: When the user's head position is detected to have moved beyond the preset threshold, steps S2 to S4 are re-executed to achieve dynamic tracking of the acoustic potential energy trap.
4. The noise reduction control method for an independent sound field in a closed area according to claim 3, characterized in that, The specific steps in S1 include: Acquire the sound pressure signals collected by each microphone in the distributed acoustic sensor array at time t. Where k = 1, ..., K, and K is the number of microphones; A sound field reconstruction model is established based on the equivalent source method, and the virtual equivalent source intensity vector is solved. Calculate the sound pressure at any point r in space. and normal particle velocity ; According to the sound pressure and normal particle velocity Calculate the instantaneous sound field energy density at this location. ; The closed region is discretized into a grid, based on each grid point. This constitutes the current sound field energy density distribution map. .
5. The audio noise reduction module according to claim 4, characterized in that, The virtual equivalent source intensity vector Solve using the following formula: ； in, Represents the identity matrix. Represents the regularization parameter. Indicates conjugate transpose; Calculate the instantaneous sound field energy density at that moment. : ； in, Indicates air density, It indicates the speed of sound.
6. The audio noise reduction module according to claim 3, characterized in that, The specific process of S2 includes: Based on the user's ear coordinates obtained by the optical positioning system, the sound field energy density distribution map is divided into a target listening zone and a non-target free zone. A state space for deep reinforcement learning is constructed, and a state vector within this state space is generated based on the measured data at the current moment. Using an optical camera to capture feature points on the user's head, the coordinates of the left ear are calculated. and right ear coordinates The target hearing zone is defined as a sphere with radius R centered on the ear coordinates. The remaining part of the closed region is defined as the non-target free zone. ; Define the set of dimensions for the deep reinforcement learning agent to perceive the environment, and obtain the state space; The state space consists of the following three physical characteristics: the sound field energy characteristics of the target listening area, the sound field energy characteristics of the non-target free area, and the spatial position characteristics of the ear in the closed area. Extracting the target listening zone from the sound field energy density distribution map Non-target free zone Based on the energy statistics within the time frame and combined with user location information, a state vector at time t is constructed. : ； in, Indicates the average energy density of the target area. Indicates the standard deviation of energy in the target region. This represents the average energy density of the free region.
7. The noise reduction control method for an independent sound field in a closed area according to claim 3, characterized in that, The policy network receives the state vector Output the motion vector for L speakers. : ； in, This represents the gain coefficient of the l-th loudspeaker. This represents the phase offset of the l-th speaker.
8. The noise reduction control method for an independent sound field in a closed area according to claim 7, characterized in that, The policy network The specific process for outputting the optimal combination of gain coefficient and phase offset based on the current sound field state includes: Normalization process: Receive the generated state vector The state vector is normalized. Feature extraction and nonlinear transformation: The normalized state vector is input into the hidden layer of the policy network. The policy network adopts a multi-layer fully connected neural network structure. Through multi-layer linear transformation and nonlinear activation function, deep features in the sound field state are extracted, and the mapping relationship between the sound field energy distribution and the loudspeaker control parameters is established. The calculation process for the l-th hidden layer is as follows: ； in, This indicates the output of the previous layer. This is the normalized state vector. and Let these represent the weight matrix and bias vector of the l-th layer, respectively. Represents the linear rectification activation function; Output layer raw parameter calculation: After the last hidden layer, the network enters the output layer. The number of neurons in the output layer depends on the number of distributed vocal units L. The output layer outputs a raw parameter vector containing 2L elements. The original vector can be divided into two parts: the first L elements correspond to the original values of the gain. The last L elements correspond to the original values of the phase. ; Physical constraint mapping of motion parameters: The original gain values are mapped to the [0, 1] interval using the Sigmoid activation function; Using the Sigmoid activation function in conjunction with a scaling factor, the original phase values are mapped to the interval [0, 2π]. Assembly and output of control action vectors: This involves mapping all gain coefficients... and phase offset The speakers are arranged in a staggered or grouped manner according to a preset speaker sequence, and assembled into the final control action vector. .
9. The noise reduction control method for an independent sound field in a closed area according to claim 3, characterized in that, The process of generating multi-channel drive signals based on control action vectors specifically includes: Obtain the audio signal and reference noise signal that need to be played, and generate the final drive signal for the Lth speaker based on the control action vector; ； in, Indicates the final drive signal. This indicates the audio signal that needs to be played. Indicates based on reference noise signal The generated noise reduction reference signal.
10. The noise reduction control method for an independent sound field in a closed area according to claim 3, characterized in that, The reward value is calculated using a reward function, which is: ； in, This represents minimizing the energy of the target region. This represents minimizing the energy of the free region. This indicates maximizing the fidelity of the audio signal in the target area. This represents the feature vector of the audio signal actually acquired or reconstructed in the target listening area. This represents the feature vector of the original audio signal to be played. This indicates a power consumption limit.