Methods and Systems for Spatial Sound Field Optimization of Audio Signals in Marine Broadcasting Systems

Abstract

This invention provides a method and system for optimizing the spatial sound field of audio signals in a marine broadcasting system, relating to the field of signal processing technology. The method includes: acquiring actual sound field data; employing zoned active sound field control, deploying a ring-shaped anti-phase loudspeaker array and a pickup array to construct an acoustic isolation zone, and deploying independent directional loudspeaker arrays in each functional area; optimizing beam parameters in each area using a differentiated multi-objective reward function based on a deep reinforcement learning algorithm; and constructing a multi-channel adaptive filter bank based on ship navigation state parameters to output a compensation signal. This invention achieves differentiated sound field management in different functional areas of the ship, improving sound transmission quality.

Description

Technical Field

[0001] This invention relates to signal processing technology, and more particularly to a method and system for optimizing the spatial sound field of audio signals in marine broadcasting systems. Background Technology

[0002] Marine public address systems are a crucial component of shipboard communication and safety management, primarily used for information dissemination, emergency announcements, and routine broadcasts within the vessel. Traditional marine public address systems typically employ a uniformly distributed array of speakers, playing the same audio content throughout the ship. However, with the increasing size and multi-functionality of modern ships, the demands for audio quality and sound field characteristics in different functional areas (such as the bridge, passenger cabins, engine room, and entertainment areas) are becoming increasingly diverse. Simultaneously, noise interference during navigation, spatial reverberation, and sound interference between different areas also pose challenges to the sound quality and clarity of information transmission in public address systems. Existing marine public address systems still suffer from the following defects and shortcomings:

[0003] Existing marine broadcasting systems lack the ability to customize the sound field for different functional areas. They typically use a uniform audio processing scheme and speaker layout, which cannot meet the differentiated needs of different areas for sound field characteristics. For example, the bridge requires clear command transmission, while the rest area requires a more comfortable audio environment.

[0004] The lack of effective zonal isolation technology in the internal acoustic field management of ships leads to mutual interference between different functional areas, affecting the effective transmission of broadcast information and the comfort of the acoustic environment. In particular, it may affect the accurate transmission of critical information in emergency situations.

[0005] Existing marine broadcasting systems are not adaptable enough to changes in the ship's navigation status. They fail to fully consider the impact of dynamic factors such as sea state changes and engine noise on the sound field during navigation and lack a real-time adaptive compensation mechanism. As a result, the sound quality and information transmission effect of the broadcasting system are significantly reduced under adverse navigation conditions, affecting the safety of ship operation and the comfort of personnel. Summary of the Invention

[0006] This invention provides a method and system for optimizing the spatial sound field of audio signals in a marine broadcasting system, which can solve the problems in the prior art.

[0007] A first aspect of the present invention provides a method for optimizing the spatial sound field of audio signals in a marine broadcasting system, comprising:

[0008] Collect actual sound field data of the loudspeaker array of the ship's broadcasting system within the ship's interior space;

[0009] Based on the acoustic field requirements of different functional areas of the ship, a zoned active acoustic field control is adopted for acoustic field zoning management. An acoustic isolation zone is constructed by deploying a ring-shaped anti-phase loudspeaker array and a pickup array at the boundary of the zone. The ring-shaped anti-phase loudspeaker array generates a cancellation signal based on an adaptive filtering algorithm. Independent directional loudspeaker arrays are deployed in each functional area.

[0010] Based on the actual sound field data, a three-dimensional sound field distribution feature of the ship's internal space is constructed. According to the three-dimensional sound field distribution feature, a deep reinforcement learning algorithm is adopted, and different optimization objectives are set according to the needs of different functional areas. By setting differentiated multi-objective reward functions, the beam parameters of the loudspeaker array in each area are iteratively optimized.

[0011] A multi-channel adaptive filter bank is constructed based on the ship's navigation state parameters. The multi-channel adaptive filter bank contains multiple frequency band filters. Wavelet decomposition is performed on the navigation state parameters to obtain multi-frequency band signals. A hybrid adaptive filtering algorithm is used to adaptively filter the multi-frequency band signals to obtain a compensation signal. The compensation signal is superimposed with the optimized beam parameters and output to the directional loudspeaker array.

[0012] In one alternative implementation,

[0013] Based on the acoustic field requirements of different functional areas of the ship, a zoned active acoustic field control is adopted for acoustic field zoning management. Acoustic isolation zones are constructed by deploying a ring-shaped anti-phase loudspeaker array and a pickup array at the zone boundaries. The ring-shaped anti-phase loudspeaker array generates a cancellation signal based on an adaptive filtering algorithm. The step of deploying independent directional loudspeaker arrays in each functional area includes:

[0014] Based on the different requirements of sound pressure level distribution uniformity, reverberation time and signal-to-noise ratio in different functional areas, zoned active sound field control is adopted for sound field zoning management.

[0015] An acoustic isolation zone is constructed at the boundary between different functional areas. The acoustic isolation zone includes a ring-shaped anti-phase loudspeaker array and a pickup array. The radial positions of the array elements of the ring-shaped anti-phase loudspeaker array are parameterized by modulation coefficients and the number of cycles. The pickup array is used to collect the boundary sound pressure level.

[0016] Based on the boundary sound pressure level, acoustic amplitude, vibration location and propagation delay information are extracted, and an adaptive filtering algorithm is used to generate a cancellation signal. The parameters of the adaptive filtering algorithm are updated in real time based on the error signal collected by the microphone, and the cancellation signal is output to the ring anti-phase loudspeaker array.

[0017] A directional loudspeaker array is deployed in each functional area. The beamforming weight is calculated using the signal covariance matrix and the desired response vector. The main lobe direction of the directional loudspeaker array is dynamically compensated according to the partition coupling coefficient.

[0018] The sound field control effect is evaluated based on the mean square error, and the beamforming weights and the cancellation signal are iteratively optimized by using adaptive step size parameters.

[0019] In one alternative implementation,

[0020] Based on the actual sound field data, a three-dimensional sound field distribution feature of the ship's interior space is constructed. Then, based on this three-dimensional sound field distribution feature, a deep reinforcement learning algorithm is used to iteratively optimize the beam parameters of the loudspeaker array in each region by setting different optimization objectives according to the needs of different functional areas and by setting differentiated multi-objective reward functions.

[0021] Sound pressure and phase information are collected at spatial grid points. Spatial harmonic components are obtained by performing spatial Fourier transform on the sound pressure and phase information. The energy absorption ratio of the ship's bulkhead material to the sound wave is measured as the boundary sound absorption coefficient. The boundary sound absorption coefficient is used to correct the spatial harmonic components to obtain the three-dimensional sound field distribution characteristics of the ship's internal space.

[0022] A multi-objective reward function is constructed, which includes a sound field quality evaluation index, a regional coupling penalty value, and an energy efficiency index. The sound field quality evaluation index is calculated by speech intelligibility, sound pressure level attenuation, and sound field uniformity through regional functional differentiation weights. The regional coupling penalty value is calculated based on the inter-regional sound field coupling degree.

[0023] A deep reinforcement learning network is established, in which the state space includes the sound pressure distribution of each region, sound field gradient information and beam parameter matrix, and the action space includes beam direction angle, beam elevation angle and beam weight matrix.

[0024] The value of state actions is calculated, and the value is iterated based on the multi-objective reward function and discount factor. The balance between exploration and utilization is adjusted by temperature parameters. The network parameters are updated according to the gradient information of the value of state actions, and the target parameters of the deep learning network are optimized by experience replay mechanism. When the value change is less than the preset convergence threshold, the optimized results of beam direction angle, beam pitch angle and beam weight matrix are output.

[0025] In one alternative implementation,

[0026] The steps to construct a multi-objective reward function include:

[0027] The ship is divided into functional zones, and differentiated optimization objectives are set based on the acoustic requirements of each zone. These differentiated optimization objectives include speech intelligibility objectives, sound pressure level attenuation objectives, and sound field uniformity objectives. Based on the optimization objectives of each functional zone, zone-specific evaluation indicators are constructed, and dynamic weighting coefficients are set for each zone-specific evaluation indicator. These dynamic weighting coefficients are adjusted in real time according to the zone's usage status and task priority.

[0028] A regional beam response optimization strategy is constructed, and the sampling probability distribution of the beam parameter search space is adjusted according to the dynamic weight coefficient of the regional-specific evaluation index.

[0029] A regional differentiation compensation term is introduced into the value function calculation of reinforcement learning, which dynamically modulates the reward signal based on the regional weight coefficient;

[0030] Based on regional priority, samples in the experience replay pool are sampled in a stratified manner, and network parameters of high-priority regions are updated first.

[0031] By setting regional coupling constraints, when the coupling degree of adjacent regions exceeds the preset coupling degree threshold, the search direction of the beam parameters is adjusted through a penalty term to achieve collaborative optimization of the sound field between regions.

[0032] In one alternative implementation,

[0033] The steps to build a deep reinforcement learning network include:

[0034] A convolution operator based on sound field modes is constructed. The filter parameters of the convolution operator are generated by combining the spherical Hankel function and the spherical harmonic function. Sparse constraints are applied to the filter parameters and the weight coefficient matrix is ​​obtained through mode matching optimization. The weight coefficient matrix is ​​convolved with the partitioned sound pressure distribution to obtain the sound field feature vector.

[0035] A sound field gradient tensor is constructed based on the sound field feature vector, the eigenvalue distribution of the sound field gradient tensor is calculated, the density distribution of spatial sampling points is determined based on the singularity of the eigenvalue distribution, and the density distribution is mapped to the set of sampling point coordinates.

[0036] Sound pressure and sound field gradient information are obtained at the set of coordinates of the sampling points. The acoustic impedance and sound pressure jump value of the regional interface are calculated. A regional coupling constraint relationship containing the acoustic impedance of the regional interface and the sound pressure jump value is established. The regional coupling constraint relationship is used to characterize the sound field coupling relationship between adjacent regions. Based on the regional coupling constraint relationship, the Kolmogorov forward equation is constructed. The state prediction value is obtained by solving the Kolmogorov forward equation.

[0037] The control quantity is calculated based on the state prediction value. The beam direction angle, beam elevation angle and beam weight matrix are updated based on the control quantity. The updated beam parameters are then substituted into the deep reinforcement learning network for the next round of optimization.

[0038] In one alternative implementation,

[0039] The steps of constructing a multi-channel adaptive filter bank based on ship navigation state parameters, wherein the multi-channel adaptive filter bank contains multiple frequency band filters, performing wavelet decomposition on the navigation state parameters to obtain multi-frequency band signals, and using a hybrid adaptive filtering algorithm to adaptively filter the multi-frequency band signals to obtain a compensation signal include:

[0040] Ship navigation status parameters are collected, and the ship navigation status parameters are standardized to obtain standardized status parameters; wavelet decomposition of the standardized status parameters is performed using wavelet basis functions to obtain multiple frequency band components.

[0041] A hybrid filter combining recursive least squares and Kalman filtering algorithms is constructed. The corresponding filtering algorithm is selected based on the mean square error and abrupt change index of the frequency band components. The algorithm is smoothly switched through a hybrid gain matrix. The hybrid gain matrix is ​​adaptively updated based on the prediction error and noise variance. The hybrid filter is used to filter the frequency band components.

[0042] The output signal of the hybrid filter is weighted and combined using frequency band weights to obtain a reconstructed signal. The frequency band weights are obtained by minimizing the reconstruction error. The state deviation between the reconstructed signal and the standardized state parameters is calculated. A compensation gain matrix is ​​constructed based on the state deviation. The compensation gain matrix is ​​adaptively adjusted based on the changing trend of the performance index. The compensation gain matrix is ​​multiplied by the state deviation to obtain a compensation signal.

[0043] A second aspect of the present invention provides a spatial sound field optimization system for audio signals of a marine broadcasting system, comprising:

[0044] The first unit is used to collect actual sound field data of the loudspeaker array of the ship's broadcasting system in the ship's interior space.

[0045] The second unit is used to manage the sound field in zones by adopting zoned active sound field control according to the sound field requirements of different functional areas of the ship. It constructs an acoustic isolation zone by deploying a ring-shaped anti-phase loudspeaker array and a pickup array at the zone boundary. The ring-shaped anti-phase loudspeaker array generates a cancellation signal based on an adaptive filtering algorithm. Independent directional loudspeaker arrays are deployed in each functional area.

[0046] The third unit is used to construct the three-dimensional sound field distribution characteristics of the ship's internal space based on the actual sound field data. According to the three-dimensional sound field distribution characteristics, a deep reinforcement learning algorithm is used to set different optimization objectives according to the needs of different functional areas. By setting differentiated multi-objective reward functions, the beam parameters of the loudspeaker array in each area are iteratively optimized.

[0047] The fourth unit is used to construct a multi-channel adaptive filter bank based on the ship's navigation state parameters. The multi-channel adaptive filter bank includes multiple frequency band filters. The navigation state parameters are decomposed into wavelet signals to obtain multi-frequency signals. A hybrid adaptive filtering algorithm is used to adaptively filter the multi-frequency signals to obtain a compensation signal. The compensation signal is superimposed with the optimized beam parameters and output to the directional speaker array.

[0048] Collect actual sound field data of the loudspeaker array of the ship's broadcasting system within the ship's interior space;

[0049] Based on the acoustic field requirements of different functional areas of the ship, a zoned active acoustic field control is adopted for acoustic field zoning management. An acoustic isolation zone is constructed by deploying a ring-shaped anti-phase loudspeaker array and a pickup array at the boundary of the zone. The ring-shaped anti-phase loudspeaker array generates a cancellation signal based on an adaptive filtering algorithm. Independent directional loudspeaker arrays are deployed in each functional area.

[0050] Based on the actual sound field data, a three-dimensional sound field distribution feature of the ship's internal space is constructed. According to the three-dimensional sound field distribution feature, a deep reinforcement learning algorithm is adopted, and different optimization objectives are set according to the needs of different functional areas. By setting differentiated multi-objective reward functions, the beam parameters of the loudspeaker array in each area are iteratively optimized.

[0051] A multi-channel adaptive filter bank is constructed based on the ship's navigation state parameters. The multi-channel adaptive filter bank contains multiple frequency band filters. Wavelet decomposition is performed on the navigation state parameters to obtain multi-frequency band signals. A hybrid adaptive filtering algorithm is used to adaptively filter the multi-frequency band signals to obtain a compensation signal. The compensation signal is superimposed with the optimized beam parameters and output to the directional loudspeaker array.

[0052] Based on the acoustic field requirements of different functional areas of the ship, a zoned active acoustic field control is adopted for acoustic field zoning management. Acoustic isolation zones are constructed by deploying a ring-shaped anti-phase loudspeaker array and a pickup array at the zone boundaries. The ring-shaped anti-phase loudspeaker array generates a cancellation signal based on an adaptive filtering algorithm. The step of deploying independent directional loudspeaker arrays in each functional area includes:

[0053] Based on the different requirements of sound pressure level distribution uniformity, reverberation time and signal-to-noise ratio in different functional areas, zoned active sound field control is adopted for sound field zoning management.

[0054] An acoustic isolation zone is constructed at the boundary between different functional areas. The acoustic isolation zone includes a ring-shaped anti-phase loudspeaker array and a pickup array. The radial positions of the array elements of the ring-shaped anti-phase loudspeaker array are parameterized by modulation coefficients and the number of cycles. The pickup array is used to collect the boundary sound pressure level.

[0055] Based on the boundary sound pressure level, acoustic amplitude, vibration location and propagation delay information are extracted, and an adaptive filtering algorithm is used to generate a cancellation signal. The parameters of the adaptive filtering algorithm are updated in real time based on the error signal collected by the microphone, and the cancellation signal is output to the ring anti-phase loudspeaker array.

[0056] A directional loudspeaker array is deployed in each functional area. The beamforming weight is calculated using the signal covariance matrix and the desired response vector. The main lobe direction of the directional loudspeaker array is dynamically compensated according to the partition coupling coefficient.

[0057] The sound field control effect is evaluated based on the mean square error, and the beamforming weights and the cancellation signal are iteratively optimized by using adaptive step size parameters.

[0058] Based on the actual sound field data, a three-dimensional sound field distribution feature of the ship's interior space is constructed. Then, based on this three-dimensional sound field distribution feature, a deep reinforcement learning algorithm is used to iteratively optimize the beam parameters of the loudspeaker array in each region by setting different optimization objectives according to the needs of different functional areas and by setting differentiated multi-objective reward functions.

[0059] Sound pressure and phase information are collected at spatial grid points. Spatial harmonic components are obtained by performing spatial Fourier transform on the sound pressure and phase information. The energy absorption ratio of the ship's bulkhead material to the sound wave is measured as the boundary sound absorption coefficient. The boundary sound absorption coefficient is used to correct the spatial harmonic components to obtain the three-dimensional sound field distribution characteristics of the ship's internal space.

[0060] A multi-objective reward function is constructed, which includes a sound field quality evaluation index, a regional coupling penalty value, and an energy efficiency index. The sound field quality evaluation index is calculated by speech intelligibility, sound pressure level attenuation, and sound field uniformity through regional functional differentiation weights. The regional coupling penalty value is calculated based on the inter-regional sound field coupling degree.

[0061] A deep reinforcement learning network is established, in which the state space includes the sound pressure distribution of each region, sound field gradient information and beam parameter matrix, and the action space includes beam direction angle, beam elevation angle and beam weight matrix.

[0062] The value of state actions is calculated, and the value is iterated based on the multi-objective reward function and discount factor. The balance between exploration and utilization is adjusted by temperature parameters. The network parameters are updated according to the gradient information of the value of state actions, and the target parameters of the deep learning network are optimized by experience replay mechanism. When the value change is less than the preset convergence threshold, the optimized results of beam direction angle, beam pitch angle and beam weight matrix are output.

[0063] The steps to construct a multi-objective reward function include:

[0064] The ship is divided into functional zones, and differentiated optimization objectives are set based on the acoustic requirements of each zone. These differentiated optimization objectives include speech intelligibility objectives, sound pressure level attenuation objectives, and sound field uniformity objectives. Based on the optimization objectives of each functional zone, zone-specific evaluation indicators are constructed, and dynamic weighting coefficients are set for each zone-specific evaluation indicator. These dynamic weighting coefficients are adjusted in real time according to the zone's usage status and task priority.

[0065] A regional beam response optimization strategy is constructed, and the sampling probability distribution of the beam parameter search space is adjusted according to the dynamic weight coefficient of the regional-specific evaluation index.

[0066] A regional differentiation compensation term is introduced into the value function calculation of reinforcement learning, which dynamically modulates the reward signal based on the regional weight coefficient;

[0067] Based on regional priority, samples in the experience replay pool are sampled in a stratified manner, and network parameters of high-priority regions are updated first.

[0068] By setting regional coupling constraints, when the coupling degree of adjacent regions exceeds the preset coupling degree threshold, the search direction of the beam parameters is adjusted through a penalty term to achieve collaborative optimization of the sound field between regions.

[0069] The steps to build a deep reinforcement learning network include:

[0070] A convolution operator based on sound field modes is constructed. The filter parameters of the convolution operator are generated by combining the spherical Hankel function and the spherical harmonic function. Sparse constraints are applied to the filter parameters and the weight coefficient matrix is ​​obtained through mode matching optimization. The weight coefficient matrix is ​​convolved with the partitioned sound pressure distribution to obtain the sound field feature vector.

[0071] A sound field gradient tensor is constructed based on the sound field feature vector, the eigenvalue distribution of the sound field gradient tensor is calculated, the density distribution of spatial sampling points is determined based on the singularity of the eigenvalue distribution, and the density distribution is mapped to the set of sampling point coordinates.

[0072] Sound pressure and sound field gradient information are obtained at the set of coordinates of the sampling points. The acoustic impedance and sound pressure jump value of the regional interface are calculated. A regional coupling constraint relationship containing the acoustic impedance of the regional interface and the sound pressure jump value is established. The regional coupling constraint relationship is used to characterize the sound field coupling relationship between adjacent regions. Based on the regional coupling constraint relationship, the Kolmogorov forward equation is constructed. The state prediction value is obtained by solving the Kolmogorov forward equation.

[0073] The control quantity is calculated based on the state prediction value. The beam direction angle, beam elevation angle and beam weight matrix are updated based on the control quantity. The updated beam parameters are then substituted into the deep reinforcement learning network for the next round of optimization.

[0074] The steps of constructing a multi-channel adaptive filter bank based on ship navigation state parameters, wherein the multi-channel adaptive filter bank contains multiple frequency band filters, performing wavelet decomposition on the navigation state parameters to obtain multi-frequency band signals, and using a hybrid adaptive filtering algorithm to adaptively filter the multi-frequency band signals to obtain a compensation signal include:

[0075] Ship navigation status parameters are collected, and the ship navigation status parameters are standardized to obtain standardized status parameters; wavelet decomposition of the standardized status parameters is performed using wavelet basis functions to obtain multiple frequency band components.

[0076] A hybrid filter combining recursive least squares and Kalman filtering algorithms is constructed. The corresponding filtering algorithm is selected based on the mean square error and abrupt change index of the frequency band components. The algorithm is smoothly switched through a hybrid gain matrix. The hybrid gain matrix is ​​adaptively updated based on the prediction error and noise variance. The hybrid filter is used to filter the frequency band components.

[0077] The output signal of the hybrid filter is weighted and combined using frequency band weights to obtain a reconstructed signal. The frequency band weights are obtained by minimizing the reconstruction error. The state deviation between the reconstructed signal and the standardized state parameters is calculated. A compensation gain matrix is ​​constructed based on the state deviation. The compensation gain matrix is ​​adaptively adjusted based on the changing trend of the performance index. The compensation gain matrix is ​​multiplied by the state deviation to obtain a compensation signal.

[0078] A third aspect of the present invention provides an electronic device, comprising:

[0079] processor;

[0080] Memory used to store processor-executable instructions;

[0081] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0082] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0083] This invention collects actual sound field data of the loudspeaker array of a marine broadcasting system and combines it with zoned active sound field control technology to achieve precise sound field management in different functional areas. This effectively solves the sound field aliasing and interference problems of traditional marine broadcasting systems and improves the voice clarity and coverage uniformity of the broadcasting system.

[0084] This invention employs a deep reinforcement learning algorithm to model and optimize the three-dimensional sound field distribution inside a ship. Differentiated multi-objective reward functions are set for different functional areas, enabling intelligent optimization of the speaker array beam parameters. This makes the sound field distribution of each functional area more in line with actual usage requirements, enhancing the spatial adaptability and performance of the shipboard broadcasting system.

[0085] This invention constructs a multi-channel adaptive filter bank based on ship navigation state parameters. Through wavelet decomposition and hybrid adaptive filtering algorithm, it can compensate for the impact of changes in ship navigation state on the sound field in real time, improve the stability and reliability of the broadcasting system under complex sea conditions and navigation conditions, and ensure the accurate transmission of important information in various environments. Attached Figure Description

[0086] Figure 1 This is a flowchart illustrating the spatial sound field optimization method for audio signals in a marine broadcasting system according to an embodiment of the present invention.

[0087] Figure 2 A flowchart for optimizing the three-dimensional sound field distribution characteristics inside a ship based on deep reinforcement learning. Detailed Implementation

[0088] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0089] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0090] Figure 1 This is a flowchart illustrating the spatial sound field optimization method for audio signals in a marine broadcasting system according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0091] Collect actual sound field data of the loudspeaker array of the ship's broadcasting system within the ship's interior space;

[0092] Based on the acoustic field requirements of different functional areas of the ship, a zoned active acoustic field control is adopted for acoustic field zoning management. An acoustic isolation zone is constructed by deploying a ring-shaped anti-phase loudspeaker array and a pickup array at the boundary of the zone. The ring-shaped anti-phase loudspeaker array generates a cancellation signal based on an adaptive filtering algorithm. Independent directional loudspeaker arrays are deployed in each functional area.

[0093] Based on the actual sound field data, a three-dimensional sound field distribution feature of the ship's internal space is constructed. According to the three-dimensional sound field distribution feature, a deep reinforcement learning algorithm is adopted, and different optimization objectives are set according to the needs of different functional areas. By setting differentiated multi-objective reward functions, the beam parameters of the loudspeaker array in each area are iteratively optimized.

[0094] A multi-channel adaptive filter bank is constructed based on the ship's navigation state parameters. The multi-channel adaptive filter bank contains multiple frequency band filters. Wavelet decomposition is performed on the navigation state parameters to obtain multi-frequency band signals. A hybrid adaptive filtering algorithm is used to adaptively filter the multi-frequency band signals to obtain a compensation signal. The compensation signal is superimposed with the optimized beam parameters and output to the directional loudspeaker array.

[0095] In one optional implementation, based on the acoustic field requirements of different functional areas of the ship, a zoned active acoustic field control is adopted for acoustic field zoning management. This involves constructing acoustic isolation zones by deploying a ring-shaped anti-phase loudspeaker array and a pickup array at the zone boundaries. The ring-shaped anti-phase loudspeaker array generates a cancellation signal based on an adaptive filtering algorithm. The step of deploying independent directional loudspeaker arrays within each functional area includes:

[0096] Based on the different requirements of sound pressure level distribution uniformity, reverberation time and signal-to-noise ratio in different functional areas, zoned active sound field control is adopted for sound field zoning management.

[0097] An acoustic isolation zone is constructed at the boundary between different functional areas. The acoustic isolation zone includes a ring-shaped anti-phase loudspeaker array and a pickup array. The radial positions of the array elements of the ring-shaped anti-phase loudspeaker array are parameterized by modulation coefficients and the number of cycles. The pickup array is used to collect the boundary sound pressure level.

[0098] Based on the boundary sound pressure level, acoustic amplitude, vibration location and propagation delay information are extracted, and an adaptive filtering algorithm is used to generate a cancellation signal. The parameters of the adaptive filtering algorithm are updated in real time based on the error signal collected by the microphone, and the cancellation signal is output to the ring anti-phase loudspeaker array.

[0099] A directional loudspeaker array is deployed in each functional area. The beamforming weight is calculated using the signal covariance matrix and the desired response vector. The main lobe direction of the directional loudspeaker array is dynamically compensated according to the partition coupling coefficient.

[0100] The sound field control effect is evaluated based on the mean square error, and the beamforming weights and the cancellation signal are iteratively optimized by using adaptive step size parameters.

[0101] For example, in a shipboard environment, zoned active acoustic field control technology is implemented based on the acoustic requirements of different functional areas. Taking the command and control room and rest area of ​​a ship as an example, these two adjacent areas have different acoustic requirements. The command and control room requires a low-noise environment (sound pressure level controlled below 45 dB(A)) and clear communication (speech intelligibility index greater than 0.75), while the rest area requires a comfortable acoustic environment (sound pressure level controlled below 55 dB(A)) and appropriate privacy protection.

[0102] When implementing zoned active sound field control in these two regions, an acoustic isolation zone is constructed at the boundary between the two regions. This isolation zone is 1.2 meters wide and consists of a ring array of 16 anti-phase loudspeakers and a monitoring array of 8 microphones. Each element of the anti-phase loudspeaker array is an 8 cm diameter full-range loudspeaker, with an element spacing of 15 cm, forming a ring structure with a diameter of 80 cm. The radial positions of the elements are parameterized using a modulation coefficient of 0.85 and a period number of 4, resulting in a micro-ripple distribution to enhance the spatial coverage of sound field cancellation. The microphone array uses high-sensitivity condenser microphones with a sampling rate of 48 kHz and a sensitivity of -38 dB, distributed on both the inner and outer sides of the isolation zone to acquire boundary sound pressure level information in real time.

[0103] The operation of the acoustic isolation zone involves processing boundary sound pressure level information and generating a cancellation signal. Boundary noise is collected using a microphone array, and time-frequency analysis is performed to extract acoustic features, including noise amplitude (average 58 dB at the command room boundary and 62 dB at the rest area boundary), vibration location information (the main noise source is determined to be located at the port side mechanical equipment using a sound source localization algorithm), and propagation delay information (the average delay time from the noise source to the boundary is 78 milliseconds). Based on this information, an adaptive filtering algorithm (filter order 256, initial step size parameter 0.001) is used to generate a cancellation signal. Boundary error signals are collected at 20-millisecond intervals, and the filter coefficients are updated based on the measured error. The optimized cancellation signal is played through an anti-phase loudspeaker array, forming an acoustic "barrier" at the boundary, achieving acoustic isolation between the two areas. The isolation effect can reach over 15 dB in the mid-frequency range (500 Hz-2 kHz).

[0104] Within each functional area, directional loudspeaker arrays are deployed for precise sound field control. A 5×5 planar array is arranged in the command and control room, with each element being a 3-inch full-range loudspeaker spaced 25 cm apart; a 3×4 linear array is arranged in the rest area, with elements spaced 30 cm apart. Beamforming weights are calculated based on sound field requirements to ensure that the sound field uniformity in the command and control room is controlled within ±3dB, the reverberation time is controlled within 0.45 seconds, and the signal-to-noise ratio is maintained above 12dB; the sound field uniformity in the rest area is controlled within ±5dB, the reverberation time is 0.65 seconds, and the signal-to-noise ratio is maintained above 8dB.

[0105] During beamforming of the directional loudspeaker array, the acoustic environment characteristics of each region are acquired to construct a signal covariance matrix. For the command and control room, a 16×16 covariance matrix is ​​constructed, and the desired response vector is set to point towards the key communication location; for the rest area, a 12×12 covariance matrix is ​​constructed, and the desired response vector covers the entire rest space. When calculating the beamforming weights, the partition coupling coefficient between the two regions (measured value of 0.37) is considered, and the main lobe direction of the loudspeaker array is dynamically compensated. In the command room, the main lobe angle is deflected inward by 6 degrees, and in the rest area, the main lobe angle is deflected inward by 4 degrees to reduce sound energy leakage to adjacent areas.

[0106] Mean squared error (MSE) was used as the evaluation metric to monitor the control effect in real time. The initial MSE was -8.5 dB, which stabilized at -22.3 dB through iterative optimization, indicating effective suppression of acoustic interference between regions. During optimization, the adaptive step size parameter was automatically adjusted based on the error change rate. A larger step size (0.005) was used in the initial stage to accelerate convergence, and then reduced to 0.0008 in the stable stage to improve accuracy. In this way, initial convergence was completed within 10 seconds, achieving the goals of improving the speech intelligibility index in the command and control room to 0.82 and the privacy protection index in the rest area to 0.75.

[0107] This invention employs zoned active sound field control to achieve precise sound field zoning management. It constructs a highly efficient acoustic isolation zone through an innovative ring-shaped anti-phase loudspeaker array and pickup array, solving the problem of traditional methods being unable to cope with the complex spatial sound field interference on ships. In particular, beamforming technology based on radial parameterized configuration and dynamic compensation of array elements significantly improves boundary sound insulation and sound field uniformity within the region.

[0108] In one optional implementation, the steps of constructing a three-dimensional sound field distribution feature of the ship's interior space based on the actual sound field data, employing a deep reinforcement learning algorithm based on the three-dimensional sound field distribution feature, setting different optimization objectives according to the needs of different functional areas, and iteratively optimizing the beam parameters of the loudspeaker array in each area by setting differentiated multi-objective reward functions include:

[0109] Sound pressure and phase information are collected at spatial grid points. Spatial harmonic components are obtained by performing spatial Fourier transform on the sound pressure and phase information. The energy absorption ratio of the ship's bulkhead material to the sound wave is measured as the boundary sound absorption coefficient. The boundary sound absorption coefficient is used to correct the spatial harmonic components to obtain the three-dimensional sound field distribution characteristics of the ship's internal space.

[0110] A multi-objective reward function is constructed, which includes a sound field quality evaluation index, a regional coupling penalty value, and an energy efficiency index. The sound field quality evaluation index is calculated by speech intelligibility, sound pressure level attenuation, and sound field uniformity through regional functional differentiation weights. The regional coupling penalty value is calculated based on the inter-regional sound field coupling degree.

[0111] A deep reinforcement learning network is established, in which the state space includes the sound pressure distribution of each region, sound field gradient information and beam parameter matrix, and the action space includes beam direction angle, beam elevation angle and beam weight matrix.

[0112] The value of state actions is calculated, and the value is iterated based on the multi-objective reward function and discount factor. The balance between exploration and utilization is adjusted by temperature parameters. The network parameters are updated according to the gradient information of the value of state actions, and the target parameters of the deep learning network are optimized by experience replay mechanism. When the value change is less than the preset convergence threshold, the optimized results of beam direction angle, beam pitch angle and beam weight matrix are output.

[0113] Combination Figure 2 This document describes a flowchart for optimizing the 3D sound field distribution characteristics inside a ship's interior space based on deep reinforcement learning. For example, a sound pressure sensor array is deployed inside the ship's interior space, with a 10cm × 10cm × 10cm 3D grid defined for each functional area. Sound pressure level (dB) and phase information (rad) are measured at each grid node. For instance, in the bridge area, the sound pressure level at grid point (3.2m, 4.5m, 2.1m) is 72.3dB, and the phase is 1.25rad; in the rest area, the sound pressure level at grid point (12.5m, 5.8m, 1.8m) is 65.7dB, and the phase is 0.87rad. The collected sound pressure and phase information are then subjected to a spatial Fourier transform, converting the sound field data in the physical space into spatial harmonic components. These spatial harmonic components contain amplitude and phase information, reflecting the distribution characteristics of the sound field at different spatial frequencies. For a ship's cockpit measuring 10m × 8m × 3m, the spatial harmonic components are decomposed into 50 spatial harmonics, yielding the amplitude and phase of each harmonic. For example, the first harmonic has an amplitude of 0.85 and a phase of 0.32 rad; the second harmonic has an amplitude of 0.63 and a phase of 1.45 rad.

[0114] The proportion of sound wave energy absorbed by the ship's bulkhead materials is measured and used as the boundary absorption coefficient. For example, the absorption coefficient of metal panels is 0.05, that of glass is 0.03, and that of areas treated with sound-absorbing materials is 0.65. Based on these boundary absorption coefficients, spatial harmonic components are corrected to generate a more accurate three-dimensional sound field distribution characteristic of the ship's interior space. The corrected harmonic components satisfy the energy absorption condition at the boundaries, with a first-order harmonic correction coefficient of 0.92 and a second-order harmonic correction coefficient of 0.88.

[0115] Constructing a multi-objective reward function is the core of deep reinforcement learning algorithms. The speech intelligibility metric uses a weighted average speech transmission index, with regional averages of 0.75 (cabin) and 0.65 (rest area). The sound pressure level attenuation metric measures the attenuation of sound pressure level with distance, with an average attenuation rate of 3.2 dB / m. The sound field uniformity metric measures the consistency of sound pressure distribution within a region, with a standard deviation not exceeding 3 dB. Differentiated weights are set according to the region's function; for example, in the driver's cab, speech intelligibility has a weight of 0.6, sound pressure level attenuation has a weight of 0.3, and sound field uniformity has a weight of 0.1; in the rest area, speech intelligibility has a weight of 0.4, sound pressure level attenuation has a weight of 0.2, and sound field uniformity has a weight of 0.4.

[0116] The regional coupling penalty is calculated based on the inter-regional sound field coupling degree, which is defined as the reciprocal of the sound pressure level difference at the boundary of adjacent regions. For example, the average sound pressure level difference at the boundary between the driver's cab and the rest area is 15 dB, corresponding to a coupling degree of 0.067. The energy efficiency index measures the sound field quality per unit input power and is defined as the ratio of the sound field quality evaluation index to the total power consumption, expressed in decibels per watt (dB / W). Before optimization, the energy efficiency was 1.2 dB / W; after optimization, it increased to 1.8 dB / W.

[0117] The multi-objective reward function integrates the sound field quality evaluation index, regional coupling penalty value, and energy efficiency index into a single reward value through weighted summation. The weighting coefficients are α, β, and γ, respectively, satisfying α + β + γ = 1. For example, in the command area where voice communication quality needs to be prioritized, α is set to 0.6, β to 0.3, and γ to 0.1; while in the rest area where comfort requirements are higher, α is set to 0.3, β to 0.5, and γ to 0.2, reflecting the different priorities under different application scenarios.

[0118] A deep reinforcement learning network is established. The state space includes the sound pressure distribution of each region (matrix size is the number of grid points in the region × 1), sound field gradient information (matrix size is the number of grid points in the region × 3, representing the gradients in the x, y, and z directions), and beam parameter matrix (matrix size is the number of loudspeakers × 3, representing the azimuth angle, pitch angle, and weight of each loudspeaker). The action space includes beam azimuth angle adjustment (range -5° to 5°), beam pitch angle adjustment (range -3° to 3°), and beam weight adjustment coefficients (range 0.9 to 1.1).

[0119] The network architecture employs a dual-network structure, comprising an evaluation network and a target network. The evaluation network consists of a four-layer fully connected neural network. The number of neurons in the input layer is equal to the state space dimension, hidden layer 1 contains 256 neurons, hidden layer 2 contains 128 neurons, and the number of neurons in the output layer is equal to the action space dimension multiplied by the number of discretized samples. The target network has the same structure as the evaluation network, but its parameters are updated at a lower frequency, once every 100 iterations.

[0120] When calculating state-action value, the Q-learning algorithm is used with a discount factor of 0.95. A temperature parameter is used to balance exploration and utilization; the initial temperature parameter is set to 5.0, linearly decreasing to 0.5 during training. In the early stages of training, a higher temperature parameter encourages the algorithm to explore more new beam parameter combinations; in the later stages, a lower temperature parameter makes the algorithm more inclined to utilize known high-value actions. Network parameters are updated based on the gradient information of state-action value using the Adam optimizer with a learning rate of 0.001 and a batch size of 64 samples. An experience replay mechanism is used to optimize the deep learning network, with an experience replay buffer size of 10000, from which training samples are randomly sampled. During optimization, if the value change is less than a preset convergence threshold of 0.01 for 50 consecutive iterations, the optimization is considered converged, and the optimal beam parameters are output.

[0121] This invention introduces an innovative multi-objective reward function to achieve differentiated sound field control for different functional regions. This method overcomes the limitations of traditional sound field optimization, simultaneously considering multiple objectives such as speech intelligibility, sound pressure level attenuation, and sound field uniformity. Through iterative optimization using a deep reinforcement learning network, beam parameters can be adaptively adjusted to achieve precise sound field control in complex ship environments.

[0122] In one alternative implementation, the steps of constructing the multi-objective reward function include:

[0123] The ship is divided into functional zones, and differentiated optimization objectives are set based on the acoustic requirements of each zone. These differentiated optimization objectives include speech intelligibility objectives, sound pressure level attenuation objectives, and sound field uniformity objectives. Based on the optimization objectives of each functional zone, zone-specific evaluation indicators are constructed, and dynamic weighting coefficients are set for each zone-specific evaluation indicator. These dynamic weighting coefficients are adjusted in real time according to the zone's usage status and task priority.

[0124] A regional beam response optimization strategy is constructed, and the sampling probability distribution of the beam parameter search space is adjusted according to the dynamic weight coefficient of the regional-specific evaluation index.

[0125] A regional differentiation compensation term is introduced into the value function calculation of reinforcement learning, which dynamically modulates the reward signal based on the regional weight coefficient;

[0126] Based on regional priority, samples in the experience replay pool are sampled in a stratified manner, and network parameters of high-priority regions are updated first.

[0127] By setting regional coupling constraints, when the coupling degree of adjacent regions exceeds the preset coupling degree threshold, the search direction of the beam parameters is adjusted through a penalty term to achieve collaborative optimization of the sound field between regions.

[0128] For example, the ship is divided into multiple functional areas, including command area, living area, and work area, and each area is set with different acoustic optimization goals according to its functional characteristics.

[0129] For the command area, the primary optimization objective is speech intelligibility, using the Speech Transmission Index (STI) as the evaluation metric. The target value for speech intelligibility is set at 0.75. When the actual STI value is below 0.6, the weighting coefficient for that area is increased. In practical applications, the STI value of the command area is calculated by measuring the energy ratio of direct speech to reflected speech. For example, when the direct speech energy is 85 dB and the reflected speech energy is 70 dB, the calculated STI value is approximately 0.68.

[0130] The optimization target for the living area was sound pressure level attenuation, using A-weighted sound pressure level (SPL) as the evaluation metric. The target SPL was set at 45 dB(A), and weighting coefficient adjustments were triggered when it exceeded 55 dB(A). Measurements showed that by optimizing beam parameters, the SPL in the living area could be reduced from the original 58 dB(A) to 47 dB(A), approaching the target.

[0131] The primary objective for the work area is sound field uniformity, using the spatial sound pressure level standard deviation (SPLSD) as the evaluation metric. The target uniformity is set to a standard deviation not exceeding 3 dB; when the standard deviation exceeds 5 dB, the weight of that area is increased. Before optimization, the sound pressure levels at the six measuring points in the work area were [72, 65, 78, 68, 75, 70] dB with a standard deviation of 4.8 dB; after optimization, the sound pressure levels at the measuring points were [71, 69, 73, 70, 72, 69] dB, and the standard deviation decreased to 1.6 dB.

[0132] The dynamic weight coefficient adjustment mechanism is based on the area usage status and task priority. A basic weight matrix [0.4, 0.3, 0.3] is set, corresponding to the command area, living area, and work area, respectively. When the command area is in an important task status, its weight can be dynamically increased to 0.6, and the weights of other areas are reduced accordingly. The weight adjustment formula is achieved by multiplying the area status index and the task priority index, with the status index ranging from [0.8, 1.5] and the task priority index ranging from [0.9, 1.3].

[0133] The beam response optimization strategy employs an adaptive sampling method in the parameter space. The beamformer parameters include the azimuth angle θ, the aperture angle φ, and the gain g. The initial search spaces are set as θ∈[-60°, 60°], φ∈[30°, 120°], and g∈[0.5, 2.0]. The parameter sampling probability distribution is adjusted according to the weight coefficients of each region, with increased sampling probability near target parameters in high-weight regions. For example, when the command area weight is 0.6, the sampling probability within ±10° of its optimal azimuth angle increases by 50%, and the actual sampling point density increases from 2 points per degree to 3 points per degree.

[0134] The reinforcement learning value function calculation introduces a regional differentiation compensation term Ci, which is calculated based on the difference between the performance of each region and the target value. Taking the command area as an example, when the STI value is 0.65 and the target value is 0.75, the difference is 0.1, and the corresponding compensation value is 0.2. The compensation value increases non-linearly with the increase of the difference; when the difference exceeds 0.2, the compensation value growth rate increases by 50%. This mechanism ensures that regions that deviate far from the target are prioritized for optimization.

[0135] The stratified sampling strategy of the experience replay pool divides samples into three priority layers: high, medium, and low. Samples in the high-priority region have a 50% probability of being selected during replay, medium-priority samples have a 30% probability, and low-priority samples have a 20% probability.

[0136] The regional coupling degree constraint is measured by the correlation coefficient ρ of the acoustic field characteristics of adjacent regions. When ρ exceeds a preset threshold of 0.7, a coupling penalty term is introduced, which is proportional to the coupling degree. For example, the coupling degree between the command area and the living area is 0.8, which exceeds the threshold of 0.7. The calculated penalty value is 0.15, and this value is incorporated into the beam parameter optimization process to correct the search direction.

[0137] This invention innovatively constructs a differentiated multi-objective reward function framework, achieving a high degree of customization in sound field control by precisely dividing ship functional areas and setting targeted optimization objectives. The introduced dynamic weight coefficients can be adjusted in real time according to the area's usage status and task priority, giving it strong adaptability to environmental changes. The design of regional beam response optimization strategies and regional differentiated compensation terms significantly improves the algorithm's convergence speed and optimization effect. In particular, through a hierarchical sampling mechanism based on regional priorities and regional coupling constraints, it is possible to achieve sound field collaborative optimization between adjacent areas while ensuring the sound field quality of high-priority areas, solving the technical challenge of balancing the sound field requirements of multiple areas in traditional methods.

[0138] In one alternative implementation, the steps of building a deep reinforcement learning network include:

[0139] A convolution operator based on sound field modes is constructed. The filter parameters of the convolution operator are generated by combining the spherical Hankel function and the spherical harmonic function. Sparse constraints are applied to the filter parameters and the weight coefficient matrix is ​​obtained through mode matching optimization. The weight coefficient matrix is ​​convolved with the partitioned sound pressure distribution to obtain the sound field feature vector.

[0140] A sound field gradient tensor is constructed based on the sound field feature vector, the eigenvalue distribution of the sound field gradient tensor is calculated, the density distribution of spatial sampling points is determined based on the singularity of the eigenvalue distribution, and the density distribution is mapped to the set of sampling point coordinates.

[0141] Sound pressure and sound field gradient information are obtained at the set of coordinates of the sampling points. The acoustic impedance and sound pressure jump value of the regional interface are calculated. A regional coupling constraint relationship containing the acoustic impedance of the regional interface and the sound pressure jump value is established. The regional coupling constraint relationship is used to characterize the sound field coupling relationship between adjacent regions. Based on the regional coupling constraint relationship, the Kolmogorov forward equation is constructed. The state prediction value is obtained by solving the Kolmogorov forward equation.

[0142] The control quantity is calculated based on the state prediction value. The beam direction angle, beam elevation angle and beam weight matrix are updated based on the control quantity. The updated beam parameters are then substituted into the deep reinforcement learning network for the next round of optimization.

[0143] For example, the filter parameters of the sound field mode-based convolution operator are generated by combining spherical Hankel functions and spherical harmonic functions. The spherical Hankel function characterizes the radial characteristics of the sound field, while the spherical harmonic function characterizes the angular characteristics. In practical applications, an N-order spherical Hankel function and an M-order spherical harmonic function can be selected to generate N×M basic filter parameters. To improve the sparsity of the filter, an L1 norm constraint is applied to the filter parameters, with the value controlled between 0.01 and 0.1. The modal matching optimization method is used, employing a gradient descent algorithm for iterative optimization. Iteration stops when the change in the objective function is less than a preset threshold of 0.001, ultimately yielding the weight coefficient matrix W. The weight coefficient matrix W is then convolved with the partitioned sound pressure distribution P to obtain the sound field feature vector F. Specifically, a 5×5 filter window is slid across the partitioned sound pressure distribution with a step size of 2. Corresponding elements are multiplied and summed to obtain the components of the feature vector F.

[0144] A sound field gradient tensor G is constructed based on the sound field feature vector F. The sound field gradient tensor G contains information about the rate of change of the sound field in three spatial directions, obtained by calculating the differences between the sound field feature vector F in the x, y, and z directions. The eigenvalue distribution λ of the sound field gradient tensor G is calculated using singular value decomposition (SVD) to obtain eigenvalues ​​λ1, λ2, and λ3. The density distribution D of spatial sampling points is determined based on the singularity of the eigenvalue distribution. The singularity criterion is: when the ratio of the largest eigenvalue to the smallest eigenvalue is greater than a preset threshold of 10, singularity is considered to exist. The density distribution D is proportional to the difference in eigenvalues, which can be expressed as D = α·(λmax / λmin), where α is a proportionality coefficient with a value of 0.05. The density distribution D is mapped to the set of sampling point coordinates S. An importance sampling strategy is adopted, allocating more sampling points in high-density regions. Specifically, the space is divided into a 10×10×10 grid, and the number of sampling points is allocated according to the density value of each grid, with the total number of sampling points controlled within 500.

[0145] Acquire sound pressure p and sound field gradient g at the set of sampling point coordinates S. Obtain the sound pressure and sound field gradient values ​​at each sampling point using sound field simulation software or actual measurements. Calculate the regional interface acoustic impedance Z and sound pressure jump value Δp. The regional interface acoustic impedance Z is the ratio of sound pressure to normal particle velocity, typically ranging from 400 to 1600 Pa·s / m. The sound pressure jump value Δp is the difference in sound pressure on both sides of the interface, typically between 5 and 20 Pa based on actual measurement data. Establish a regional coupling constraint relationship C that includes the regional interface acoustic impedance Z and the sound pressure jump value Δp. This constraint relationship characterizes the sound field coupling relationship between adjacent regions and can be described as a linear combination of the sound pressure gradient and the sound pressure jump value between adjacent regions. Construct the Kolmogorov forward equation based on the regional coupling constraint relationship C, discretize the equation using the finite difference method with a spatial step size of 0.05 m and a time step size of 0.001 s, and obtain the state prediction value X through iterative calculation. The state prediction value X includes two parts: the sound field distribution state and the control parameter state.

[0146] The control quantity U is calculated based on the predicted state value X. The control quantity U is obtained through a policy network π, which is a four-layer fully connected neural network with 128, 256, and 128 hidden layer nodes, and the activation function is ReLU. The beam azimuth angle θ, beam elevation angle φ, and beam weight matrix W are updated based on the control quantity U. The update range for the beam azimuth angle θ is 0° to 360°, with an adjustment step size of 2°; the update range for the beam elevation angle φ is -90° to 90°, with an adjustment step size of 1°; the update of the beam weight matrix W uses the gradient ascent method with a learning rate of 0.01. The updated beam parameters are then substituted into a deep reinforcement learning network for the next round of optimization. During optimization, a reward function R is used to evaluate the control effect. The reward function R is defined as the weighted sum of the sound pressure gain in the target area and the sound pressure suppression in the interference area, with weight coefficients of 0.7 and 0.3, respectively. The optimization process terminates when the change in the reward function R is less than a preset threshold of 0.005 or when the maximum number of iterations (500) is reached.

[0147] This invention effectively extracts sound field features through a convolution operator based on sound field modes, significantly improving the characterization capability of complex sound field environments. The introduction of the sound field gradient tensor and eigenvalue distribution greatly enhances computational efficiency and optimization accuracy. The calculation methods for regional interface acoustic impedance and sound pressure jump values ​​provide a precise mathematical description of the sound field coupling relationship between adjacent regions, while the state prediction mechanism based on the Kolmogorov forward equation provides theoretical support for beam parameter optimization. This structural design enables efficient handling of the nonlinear characteristics and spatial variations of the sound field, achieving intelligent optimization and control of the ship's sound field.

[0148] In one optional implementation, the steps of constructing a multi-channel adaptive filter bank based on ship navigation state parameters, wherein the multi-channel adaptive filter bank includes multiple frequency band filters, performing wavelet decomposition on the navigation state parameters to obtain multi-frequency band signals, and using a hybrid adaptive filtering algorithm to adaptively filter the multi-frequency band signals to obtain a compensation signal include:

[0149] Ship navigation status parameters are collected, and the ship navigation status parameters are standardized to obtain standardized status parameters; wavelet decomposition of the standardized status parameters is performed using wavelet basis functions to obtain multiple frequency band components.

[0150] A hybrid filter combining recursive least squares and Kalman filtering algorithms is constructed. The corresponding filtering algorithm is selected based on the mean square error and abrupt change index of the frequency band components. The algorithm is smoothly switched through a hybrid gain matrix. The hybrid gain matrix is ​​adaptively updated based on the prediction error and noise variance. The hybrid filter is used to filter the frequency band components.

[0151] The output signal of the hybrid filter is weighted and combined using frequency band weights to obtain a reconstructed signal. The frequency band weights are obtained by minimizing the reconstruction error. The state deviation between the reconstructed signal and the standardized state parameters is calculated. A compensation gain matrix is ​​constructed based on the state deviation. The compensation gain matrix is ​​adaptively adjusted based on the changing trend of the performance index. The compensation gain matrix is ​​multiplied by the state deviation to obtain a compensation signal.

[0152] For example, ship navigation state parameters are collected, including but not limited to speed, heading, attitude angles, and acceleration. Taking a specific example, assume the collected raw speed data is 12.5 knots, heading is 78.3 degrees, roll angle is 3.2 degrees, pitch angle is 2.1 degrees, and bow angle is 1.5 degrees. These parameters are standardized, mapping each parameter to the interval [-1, 1]. The standardization process uses a maximum-minimum normalization method. Specifically, for the speed parameter, if its historical maximum value is 25 knots and its minimum value is 0 knots, then the standardized speed value is (12.5-0) / (25-0)×2-1=0. Similarly, other parameters are standardized using the same method.

[0153] After standardization, wavelet decomposition is performed on the standardized state parameters using wavelet basis functions. The db4 wavelet basis function is selected for a three-level decomposition, dividing the signal into components of different frequency bands. Taking the standardized heading signal as an example, wavelet decomposition yields a low-frequency approximate component A3 and three high-frequency detail components D1, D2, and D3. A3 reflects the main trend of the signal, with a frequency range of 0-0.125Hz; D3 reflects mid-frequency variations, with a frequency range of 0.125-0.25Hz; D2 reflects higher-frequency variations, with a frequency range of 0.25-0.5Hz; and D1 reflects high-frequency variations, with a frequency range of 0.5-1Hz.

[0154] A hybrid filter combining recursive least squares and Kalman filtering algorithms is constructed. For different frequency band components, a suitable filtering algorithm is selected based on their signal characteristics. Specifically, the mean square error (MSE) and abrupt change index (ADI) of each frequency band component are calculated. The MSE represents the degree of signal fluctuation, and the ADI is obtained by calculating the rate of change between adjacent sampling points. Taking the frequency bands of the heading signal as an example, if the MSE of component A3 is 0.02, the ADI is 0.005; the MSE of component D3 is 0.05, the ADI is 0.01; the MSE of component D2 is 0.08, the ADI is 0.03; and the MSE of component D1 is 0.15, the ADI is 0.06.

[0155] When the mean square error of the frequency band component is less than 0.05 and the abrupt change index is less than 0.01, the Kalman filter algorithm is selected; when the mean square error is greater than 0.1 or the abrupt change index is greater than 0.05, the recursive least squares algorithm is selected; in the intermediate state, a smooth switching between the two algorithms is achieved through a hybrid gain matrix. The hybrid gain matrix is ​​dynamically adjusted based on the prediction error and noise variance. The prediction error is calculated by the difference between the actual measured value and the predicted value, and the noise variance is obtained through statistical analysis of historical data. Specifically, if the current prediction error is 0.03, the noise variance is 0.01, and the measurement noise variance is 0.02, then the hybrid gain coefficient is set to 0.6, indicating that the weight of the Kalman filter is 0.6 and the weight of the recursive least squares algorithm is 0.4.

[0156] After processing by the hybrid filter, the filtered output signals for each frequency band are obtained. To reconstruct the complete signal, the outputs of each frequency band need to be weighted and combined. The frequency band weights are obtained by minimizing the reconstruction error and are iteratively optimized using the gradient descent method. Taking the heading signal as an example, after 10 iterations of optimization, the frequency band weights are: A3 component weight 0.5, D3 component weight 0.3, D2 component weight 0.15, and D1 component weight 0.05. Multiplying each frequency band output by its corresponding weight and summing the results yields the reconstructed signal.

[0157] There is a state deviation between the reconstructed signal and the standardized state parameters. The difference between the two is calculated to obtain the deviation vector. Taking heading as an example, assuming the reconstructed signal value is 0.02 and the standardized state parameter value is 0, the state deviation is -0.02. A compensation gain matrix is ​​constructed based on this state deviation. This matrix is ​​adaptively adjusted according to the changing trends of performance indicators. Performance indicators include steady-state error, overshoot, and response speed. When the steady-state error increases, the corresponding gain coefficient is increased; when the overshoot is too large, the corresponding gain coefficient is decreased; when the response speed is slow, the gain coefficient is appropriately increased.

[0158] Taking heading control as an example, if the current steady-state error is 0.015, which is 0.005 greater than the previous time step, the corresponding gain coefficient is adjusted from 0.8 to 0.85. If the overshoot is 5%, which is larger than the expected value of 3%, the gain coefficient is adjusted from 0.85 to 0.82. The resulting compensation gain matrix is ​​a diagonal matrix, with the diagonal elements corresponding to the compensation gain coefficients for different state parameters. Multiplying the compensation gain matrix by the state deviation vector yields the compensation signal. For a heading deviation of -0.02, if the corresponding gain coefficient is 0.82, the compensation signal is -0.02 × 0.82 = -0.0164.

[0159] The compensation signal is converted into beam parameter adjustment values, and the optimized beam azimuth angle, pitch angle, and weight matrix are dynamically corrected. For example, when a compensation signal of -0.0164 is detected due to a change in heading, it is mapped to a fine adjustment of -1.23° in the beam azimuth angle. The beam parameters of the speaker array are then adjusted in real time by a digital signal processor to ensure that the optimal sound field distribution is maintained even when the ship's navigation status changes.

[0160] The hybrid filter design of this invention, combining recursive least squares and Kalman filtering algorithms, can intelligently switch filtering strategies based on signal characteristics, significantly improving the processing capability for both abrupt and stationary signals. The smooth switching of algorithms achieved through a hybrid gain matrix solves the discontinuity problem encountered in traditional methods during algorithm switching, while the adaptive update mechanism based on prediction error and noise variance ensures robustness. Optimal combination of frequency band weights and adaptive adjustment of the compensation gain matrix enable precise handling of various complex navigation conditions, providing stable and reliable sound field control for shipboard broadcasting.

[0161] A second aspect of the present invention provides a spatial sound field optimization system for audio signals of a marine broadcasting system, comprising:

[0162] The first unit is used to collect actual sound field data of the loudspeaker array of the ship's broadcasting system in the ship's interior space.

[0163] The second unit is used to manage the sound field in zones by adopting zoned active sound field control according to the sound field requirements of different functional areas of the ship. It constructs an acoustic isolation zone by deploying a ring-shaped anti-phase loudspeaker array and a pickup array at the zone boundary. The ring-shaped anti-phase loudspeaker array generates a cancellation signal based on an adaptive filtering algorithm. Independent directional loudspeaker arrays are deployed in each functional area.

[0164] The third unit is used to construct the three-dimensional sound field distribution characteristics of the ship's internal space based on the actual sound field data. According to the three-dimensional sound field distribution characteristics, a deep reinforcement learning algorithm is used to set different optimization objectives according to the needs of different functional areas. By setting differentiated multi-objective reward functions, the beam parameters of the loudspeaker array in each area are iteratively optimized.

[0165] The fourth unit is used to construct a multi-channel adaptive filter bank based on the ship's navigation state parameters. The multi-channel adaptive filter bank includes multiple frequency band filters. The navigation state parameters are decomposed into wavelet signals to obtain multi-frequency signals. A hybrid adaptive filtering algorithm is used to adaptively filter the multi-frequency signals to obtain a compensation signal. The compensation signal is superimposed with the optimized beam parameters and output to the directional speaker array.

[0166] A third aspect of the present invention provides an electronic device, comprising:

[0167] processor;

[0168] Memory used to store processor-executable instructions;

[0169] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0170] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0171] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0172] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; 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 or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing the spatial sound field of audio signals in a marine broadcasting system, characterized in that, include: Collect actual sound field data of the loudspeaker array of the ship's broadcasting system within the ship's interior space; Based on the acoustic field requirements of different functional areas of the ship, a zoned active acoustic field control is adopted for acoustic field zoning management. An acoustic isolation zone is constructed by deploying a ring-shaped anti-phase loudspeaker array and a pickup array at the boundary of the zone. The ring-shaped anti-phase loudspeaker array generates a cancellation signal based on an adaptive filtering algorithm. Independent directional loudspeaker arrays are deployed in each functional area. Based on the actual sound field data, a three-dimensional sound field distribution feature of the ship's internal space is constructed. According to the three-dimensional sound field distribution feature, a deep reinforcement learning algorithm is adopted, and different optimization objectives are set according to the needs of different functional areas. By setting differentiated multi-objective reward functions, the beam parameters of the loudspeaker array in each area are iteratively optimized, including: collecting sound pressure values ​​and phase information at spatial grid points, performing spatial Fourier transform on the sound pressure values ​​and phase information to obtain spatial harmonic components, measuring the energy absorption ratio of the ship's bulkhead material to the sound waves as the boundary absorption coefficient, and performing boundary correction on the spatial harmonic components according to the boundary absorption coefficient to obtain the three-dimensional sound field distribution feature of the ship's internal space. A multi-objective reward function is constructed, comprising a sound field quality evaluation index, a regional coupling penalty value, and an energy efficiency index. The sound field quality evaluation index is calculated by weighting speech intelligibility, sound pressure level attenuation, and sound field uniformity using regional functional differentiation. The regional coupling penalty value is calculated based on the inter-regional sound field coupling degree. A deep reinforcement learning network is established, wherein the state space includes the sound pressure distribution of each region, sound field gradient information, and beam parameter matrix, and the action space includes beam azimuth angle, beam pitch angle, and beam weight matrix. The state-action value is calculated by iterating the value based on the multi-objective reward function and a discount factor, and adjusting the balance between exploration and utilization using a temperature parameter. The network parameters are updated according to the gradient information of the state-action value, and the objective parameters of the deep learning network are optimized through an experience replay mechanism. When the value change is less than a preset convergence threshold, the optimized results of the beam azimuth angle, beam pitch angle, and beam weight matrix are output. The steps of establishing the deep reinforcement learning network include: constructing a convolution operator based on sound field modes, wherein the filter parameters of the convolution operator are generated by combining spherical Hankel functions and spherical harmonic functions. The filter parameters are subjected to sparse constraints and optimized by modal matching to obtain a weight coefficient matrix. The weight coefficient matrix is ​​convolved with the partitioned sound pressure distribution to obtain a sound field feature vector. A sound field gradient tensor is constructed based on the sound field feature vector, and the eigenvalue distribution of the sound field gradient tensor is calculated. The density distribution of spatial sampling points is determined according to the singularity of the eigenvalue distribution, and the density distribution is mapped to a set of sampling point coordinates. Sound pressure and sound field gradient information are obtained at the set of sampling point coordinates. The acoustic impedance and sound pressure jump value of the regional interface are calculated, and a regional coupling constraint relationship containing the acoustic impedance of the regional interface and the sound pressure jump value is established. The regional coupling constraint relationship is used to characterize the sound field coupling relationship between adjacent regions. A Kolmogorov forward equation is constructed based on the regional coupling constraint relationship, and the Kolmogorov forward equation is solved to obtain a state prediction value. A control quantity is calculated based on the state prediction value, and the beam direction angle, beam elevation angle, and beam weight matrix are updated based on the control quantity. The updated beam parameters are substituted into the deep reinforcement learning network for the next round of optimization. A multi-channel adaptive filter bank is constructed based on the ship's navigation state parameters. The multi-channel adaptive filter bank contains multiple frequency band filters. Wavelet decomposition is performed on the navigation state parameters to obtain multi-frequency band signals. A hybrid adaptive filtering algorithm is used to adaptively filter the multi-frequency band signals to obtain a compensation signal. The compensation signal is superimposed with the optimized beam parameters and output to the directional loudspeaker array.

2. The method according to claim 1, characterized in that, Based on the acoustic field requirements of different functional areas of the ship, a zoned active acoustic field control is adopted for acoustic field zoning management. An acoustic isolation zone is constructed by deploying a ring-shaped anti-phase loudspeaker array and a pickup array at the zone boundary. The ring-shaped anti-phase loudspeaker array generates a cancellation signal based on an adaptive filtering algorithm. The steps for deploying independent directional loudspeaker arrays in each functional area include: Based on the different requirements of sound pressure level distribution uniformity, reverberation time and signal-to-noise ratio in different functional areas, zoned active sound field control is adopted for sound field zoning management. An acoustic isolation zone is constructed at the boundary between different functional areas. The acoustic isolation zone includes a ring-shaped anti-phase loudspeaker array and a pickup array. The radial positions of the array elements of the ring-shaped anti-phase loudspeaker array are parameterized by modulation coefficients and the number of cycles. The pickup array is used to collect the boundary sound pressure level. Based on the boundary sound pressure level, acoustic amplitude, vibration location and propagation delay information are extracted, and an adaptive filtering algorithm is used to generate a cancellation signal. The parameters of the adaptive filtering algorithm are updated in real time based on the error signal collected by the microphone, and the cancellation signal is output to the ring anti-phase loudspeaker array. A directional loudspeaker array is deployed in each functional area. The beamforming weight is calculated using the signal covariance matrix and the desired response vector. The main lobe direction of the directional loudspeaker array is dynamically compensated according to the partition coupling coefficient. The sound field control effect is evaluated based on the mean square error, and the beamforming weights and the cancellation signal are iteratively optimized by using adaptive step size parameters.

3. The method according to claim 1, characterized in that, The steps to construct a multi-objective reward function include: The ship is divided into functional zones, and differentiated optimization objectives are set based on the acoustic requirements of each zone. These differentiated optimization objectives include speech intelligibility objectives, sound pressure level attenuation objectives, and sound field uniformity objectives. Based on the optimization objectives of each functional zone, zone-specific evaluation indicators are constructed, and dynamic weighting coefficients are set for each zone-specific evaluation indicator. These dynamic weighting coefficients are adjusted in real time according to the zone's usage status and task priority. A regional beam response optimization strategy is constructed, and the sampling probability distribution of the beam parameter search space is adjusted according to the dynamic weight coefficient of the regional-specific evaluation index. A regional differentiation compensation term is introduced into the value function calculation of reinforcement learning, which dynamically modulates the reward signal based on the regional weight coefficient; Based on regional priority, samples in the experience replay pool are sampled in a stratified manner, and network parameters of high-priority regions are updated first. By setting regional coupling constraints, when the coupling degree of adjacent regions exceeds the preset coupling degree threshold, the search direction of the beam parameters is adjusted through a penalty term to achieve collaborative optimization of the sound field between regions.

4. The method according to claim 1, characterized in that, The steps of constructing a multi-channel adaptive filter bank based on ship navigation state parameters, wherein the multi-channel adaptive filter bank contains multiple frequency band filters, performing wavelet decomposition on the navigation state parameters to obtain multi-frequency band signals, and using a hybrid adaptive filtering algorithm to adaptively filter the multi-frequency band signals to obtain a compensation signal include: Ship navigation status parameters are collected, and the ship navigation status parameters are standardized to obtain standardized status parameters; wavelet decomposition of the standardized status parameters is performed using wavelet basis functions to obtain multiple frequency band components. A hybrid filter combining recursive least squares and Kalman filtering algorithms is constructed. The corresponding filtering algorithm is selected based on the mean square error and abrupt change index of the frequency band components. The algorithm is smoothly switched through a hybrid gain matrix. The hybrid gain matrix is ​​adaptively updated based on the prediction error and noise variance. The hybrid filter is used to filter the frequency band components. The output signal of the hybrid filter is weighted and combined using frequency band weights to obtain a reconstructed signal. The frequency band weights are obtained by minimizing the reconstruction error. The state deviation between the reconstructed signal and the standardized state parameters is calculated. A compensation gain matrix is ​​constructed based on the state deviation. The compensation gain matrix is ​​adaptively adjusted based on the changing trend of the performance index. The compensation gain matrix is ​​multiplied by the state deviation to obtain a compensation signal.

5. A spatial sound field optimization system for audio signals of a marine broadcasting system, used to implement the method of any one of claims 1-4, characterized in that, include: The first unit is used to collect actual sound field data of the loudspeaker array of the ship's broadcasting system in the ship's interior space. The second unit is used to manage the sound field in zones by adopting zoned active sound field control according to the sound field requirements of different functional areas of the ship. It constructs an acoustic isolation zone by deploying a ring-shaped anti-phase loudspeaker array and a pickup array at the zone boundary. The ring-shaped anti-phase loudspeaker array generates a cancellation signal based on an adaptive filtering algorithm. Independent directional loudspeaker arrays are deployed in each functional area. The third unit is used to construct the three-dimensional sound field distribution characteristics of the ship's internal space based on the actual sound field data. According to the three-dimensional sound field distribution characteristics, a deep reinforcement learning algorithm is used to set different optimization objectives according to the needs of different functional areas. By setting differentiated multi-objective reward functions, the beam parameters of the loudspeaker array in each area are iteratively optimized. The fourth unit is used to construct a multi-channel adaptive filter bank based on the ship's navigation state parameters. The multi-channel adaptive filter bank includes multiple frequency band filters. The navigation state parameters are decomposed into wavelet signals to obtain multi-frequency signals. A hybrid adaptive filtering algorithm is used to adaptively filter the multi-frequency signals to obtain a compensation signal. The compensation signal is superimposed with the optimized beam parameters and output to the directional speaker array.

6. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 4.

Images