An adaptive cancellation method and system based on statistical characteristics of sea surface wind noise

Abstract

This application discloses an adaptive cancellation method and system based on the statistical characteristics of sea surface wind noise, relating to the field of data processing technology. The method includes: timestamping and fusing multimodal sensor data to generate a marine environmental state sequence; predicting sea surface wind noise based on the marine environmental state sequence using an air-sea interface wind noise generation model and a wave bubble field coupling model, outputting predicted statistical characteristics; constructing a reference noise signal based on the predicted spatial coherence coefficient matrix and the mixed acoustic signal in the predicted statistical characteristics; dividing the mixed acoustic signal into frequency bands according to the physical scale of sea surface wind noise; canceling noise in the scale components of each frequency band according to the reference noise signal, and then weighting and fusing the noise-cancelled scale components of each frequency band to generate a preliminary enhanced signal; and performing residual noise suppression and target signal fidelity restoration on the preliminary enhanced signal to obtain the target signal. This application can improve the accuracy of acoustic signals.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to an adaptive cancellation method and system based on the statistical characteristics of sea surface wind noise. Background Technology

[0002] With the rapid development of fields such as safeguarding maritime rights, exploring marine resources, monitoring underwater targets, and observing the marine environment, underwater acoustic communication, underwater acoustic detection, and acoustic sensing technologies for surface platforms have become key supports for the development of marine science and technology. Whether it's fixed seabed observation networks, autonomous underwater vehicles, or surface ships, buoy platforms, and unmanned aerial vehicles (UAVs) operating on the sea surface or near-shore, all rely heavily on acoustic sensors for underwater target detection, underwater communication, monitoring of marine environmental noise, and surface situational awareness. In these application scenarios, acoustic sensors are typically deployed on the surface or in shallow water, making them highly susceptible to strong noise interference from the sea surface environment. Among these, sea surface wind noise is one of the most common and powerful sources of interference.

[0003] The generation of sea surface wind noise has a unique physical mechanism: the sea surface, driven by wind, forms waves and turbulence. Simultaneously, the interaction between wind and the sea surface generates bubble layers, breaking waves, and turbulent boundary layers. These processes directly or indirectly produce strong sound radiation, forming a broadband underwater sound field interference. Furthermore, for aeroacoustic sensors deployed on surface buoys, ships, or UAV platforms, the direct impact of high-speed sea surface airflow on the sensors also generates strong wind noise interference. In the marine environment, wind noise often becomes a major factor limiting the performance of acoustic detection and communication systems, especially under medium to high sea states, where wind noise energy can completely drown out target signals, leading to underwater target detection failure, underwater acoustic communication interruption, or the unavailability of voice interaction on surface platforms. Summary of the Invention

[0004] In view of this, embodiments of this application provide an adaptive cancellation method and system based on the statistical characteristics of sea surface wind noise, so as to cancel wind noise and improve the accuracy of acoustic signals.

[0005] One aspect of this application provides an adaptive cancellation method based on the statistical characteristics of sea surface wind noise, the method comprising the following steps:

[0006] Acquire multimodal sensor data, perform timestamp alignment and data fusion on the multimodal sensor data, and generate a marine environmental state sequence that is time-aligned with the underwater acoustic signal;

[0007] Using a pre-established air-sea interface wind noise generation model and a wave bubble field coupling model, the sea surface wind noise within a set time window is predicted based on the marine environmental state sequence, and the predicted statistical characteristics of the sea surface wind noise are output.

[0008] An adaptive reference noise signal is constructed using a spatial filtering method based on the predicted spatial coherence coefficient matrix in the predicted statistical characteristics and the mixed acoustic signal in the multimodal sensor data.

[0009] The hybrid acoustic signal is divided into frequency bands according to the physical scale of sea surface wind noise to obtain the scale components of the hybrid acoustic signal in different frequency bands.

[0010] The scale components of each frequency band are noise-cancelled according to the adaptive reference noise signal, and then the scale components of each frequency band after noise cancellation are weighted and fused to generate a preliminary enhanced signal.

[0011] The preliminary enhanced signal is subjected to residual noise suppression and target signal fidelity restoration to obtain the final target signal.

[0012] In some embodiments, the noise cancellation of the scale component rows of each frequency band according to the adaptive reference noise signal includes the following steps:

[0013] Multidimensional features are extracted from the mixed acoustic signal, and the multidimensional features are fused to generate a comprehensive nonstationarity index;

[0014] The target strategy is selected from the preset strategy pool based on the comprehensive nonstationarity index.

[0015] The adaptive filter, driven by the target strategy, performs noise cancellation on the scale component rows of each frequency band based on the adaptive reference noise signal.

[0016] In some embodiments, the extraction of multidimensional features from the mixed acoustic signal includes the following steps:

[0017] The wave breaking event detection marker, temporal kurtosis coefficient, power spectrum fluctuation index, subband energy mutation rate, and sea state level normalization index are extracted from the hybrid acoustic signal as the multidimensional features.

[0018] The process of fusing the multidimensional features to generate a comprehensive nonstationary index includes the following steps:

[0019] The multidimensional features are adaptively weighted and fused to obtain the comprehensive nonstationarity index; wherein, when wave breaking events are continuously detected, the weight of the wave breaking event detection flag is increased.

[0020] In some embodiments, selecting a target strategy from a preset strategy pool based on the comprehensive nonstationarity index includes the following steps:

[0021] Construct and maintain a set of sea state conditions including calm, moderate, and severe sea states;

[0022] When the comprehensive nonstationarity index continuously exceeds a preset high threshold and the frequency of wave breaking events increases, the sea state is shifted to the severe sea state, and a strategy targeting wave breaking transient wind noise is selected from the preset strategy pool as the target strategy.

[0023] When the comprehensive nonstationarity index remains below a preset low threshold, the sea state is shifted to calm sea state, and a strategy targeting steady-state background wind noise is selected from the preset strategy pool as the target strategy.

[0024] In some embodiments, the method further includes the following steps:

[0025] Based on the current wind speed vector and significant wave height, the power spectrum template of the wind pressure transmission component is obtained from the air-sea interface wind noise generation model.

[0026] Based on the wave spectrum and wave steepness parameters, the power spectrum and transient pulse characteristics of the bubble radiation component are calculated using the wave-bubble field coupling model.

[0027] The change in wind noise propagation path is determined based on the power spectrum template of the wind pressure transmission component and the power spectrum and transient pulse characteristics of the bubble radiation component.

[0028] The platform's motion attitude data is used to perform noise spectrum attenuation compensation on the change in the wind noise propagation path in order to correct the predicted statistical characteristics.

[0029] In some embodiments, the step of constructing an adaptive reference noise signal based on the predicted spatial coherence coefficient matrix in the predicted statistical characteristics and the mixed acoustic signal in the multimodal sensor data using a spatial filtering method includes the following steps:

[0030] Using the dual-constraint minimum variance criterion, the predicted spatial coherence coefficient matrix is ​​used as the prior of wind noise spatial covariance, and the expected angle of arrival and array manifold of the target signal are used as the spatial features of the target signal to design spatial filtering coefficients.

[0031] The mixed acoustic signal is filtered using the spatial filtering coefficients so that the resulting adaptive reference noise signal satisfies the following conditions: its power spectrum is consistent with the predicted sea surface wind noise power spectrum, and its spatial coherence with the target signal is minimized.

[0032] In some embodiments, the method further includes the following steps:

[0033] During online operation, the system continuously caches the operation data within the most recent set time period. When the target signal distortion index is detected to be continuously deviating from the set conditions, the online fine-tuning mechanism is triggered.

[0034] The online fine-tuning mechanism replays the cached runtime data in small batches and introduces a loss function including a marine physical consistency regularization term to incrementally update the parameters of the mapping model used to map from marine environmental state to adaptive filter parameters.

[0035] Another aspect of this application embodiment provides an adaptive cancellation system based on the statistical characteristics of sea surface wind noise, the system comprising:

[0036] The data acquisition unit is used to acquire multimodal sensor data, perform timestamp alignment and data fusion on the multimodal sensor data, and generate a marine environmental state sequence that is time-aligned with the underwater acoustic signal.

[0037] The wind noise statistics unit is used to predict the sea surface wind noise within a set time window in the future based on the ocean environment state sequence using a pre-established air-sea interface wind noise generation model and a wave bubble field coupling model, and output the predicted statistical characteristics of the sea surface wind noise.

[0038] The reference noise construction unit is used to construct an adaptive reference noise signal based on the predicted spatial coherence coefficient matrix in the predicted statistical characteristics and the mixed acoustic signal in the multimodal sensor data using a spatial filtering method.

[0039] The signal segmentation unit is used to segment the mixed acoustic signal into frequency bands according to the physical scale of sea surface wind noise, so as to obtain the scale components of the mixed acoustic signal in different frequency bands.

[0040] The wind noise cancellation unit is used to cancel the noise of the scale components of each frequency band according to the adaptive reference noise signal, and then weight and fuse the scale components of each frequency band after noise cancellation to generate a preliminary enhanced signal.

[0041] The signal correction unit is used to suppress residual noise and restore the fidelity of the target signal in the preliminary enhanced signal to obtain the final target signal.

[0042] Another aspect of this application embodiment provides an electronic device, including a processor and a memory;

[0043] The memory is used to store programs;

[0044] The processor executes the program to implement any of the methods described above.

[0045] Another aspect of this application provides a computer-readable storage medium storing a program that is executed by a processor to implement the method described in any of the above embodiments.

[0046] This application includes at least the following beneficial effects:

[0047] This application can acquire multimodal sensor data, perform timestamp alignment and data fusion on the multimodal sensor data, and generate a marine environmental state sequence that is time-aligned with the underwater acoustic signal. Using a pre-established air-sea interface wind noise generation model and a wave-bubble field coupling model, it predicts sea surface wind noise within a future set time window based on the marine environmental state sequence, outputting the predicted statistical characteristics of sea surface wind noise. An adaptive reference noise signal is constructed based on the predicted spatial coherence coefficient matrix in the predicted statistical characteristics and the mixed acoustic signal in the multimodal sensor data using a spatial filtering method. The mixed acoustic signal is then segmented into frequency bands according to the physical scale of sea surface wind noise, obtaining the scale components of the mixed acoustic signal in different frequency bands. Noise cancellation is performed on the scale components of each frequency band according to the adaptive reference noise signal, and then the scale components after noise cancellation are weighted and fused to generate a preliminary enhanced signal. The preliminary enhanced signal is then subjected to residual noise suppression and target signal fidelity restoration to obtain the final target signal. Based on the physical generation mechanism of sea surface wind noise, this application constructs an adaptive cancellation scheme adapted to the special marine environment, which can accurately cancel wind noise signals in acoustic signals and improve the accuracy of acoustic signals. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 A flowchart illustrating an adaptive cancellation method based on the statistical characteristics of sea surface wind noise, provided for an embodiment of this application;

[0050] Figure 2 This is a structural block diagram of an adaptive cancellation system based on the statistical characteristics of sea surface wind noise, provided in an embodiment of this application. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0052] Before providing a detailed description of the embodiments of this application, some related technologies involved in the embodiments of this application will be described first, as follows:

[0053] As marine equipment develops towards unmanned operation, intelligence, and long endurance, higher demands are placed on the robustness of acoustic sensors in complex sea conditions. Traditional noise suppression methods are mostly designed for terrestrial environments, assuming that the noise source has relatively stable statistical characteristics or that a clean reference signal is readily available. However, wind noise in the marine environment has unique spatiotemporal evolution patterns. Its statistical characteristics are closely related to sea surface meteorological conditions, wave field distribution, sensor deployment depth, and platform motion, exhibiting far more complex non-stationarity, non-Gaussianity, and strong coupling than terrestrial wind noise. Therefore, how to construct an adaptive cancellation method adapted to the special marine environment based on the physical generation mechanism of sea surface wind noise has become a core technical challenge that urgently needs to be overcome in the fields of marine acoustic detection and underwater acoustic communication.

[0054] The unique physical characteristics of sea surface wind noise and the limitations of existing processing technologies are as follows:

[0055] The most fundamental difference between sea surface wind noise and land-based wind noise lies in its generation mechanism, which involves a complex coupling process at the air-sea interface. When sea surface wind speeds exceed approximately 3 m / s, wind energy is converted into underwater sound energy through two main mechanisms: first, the wind directly acts on the sea surface, generating pressure fluctuations that penetrate the water to form an underwater sound field; second, bubble clouds entrained during wave breaking radiate sound energy during their oscillation and collapse. These two mechanisms together determine the unique statistical characteristics of sea surface wind noise, making existing land-based wind noise cancellation technologies significantly unsuitable for marine applications.

[0056] (I) Multimodal coupling and spatiotemporal nonstationary characteristics of sea surface wind noise.

[0057] The energy distribution of sea surface wind noise is coupled with multiple factors, including sea surface wind speed, wind direction, wind zone length, wave height, wave direction, and atmospheric stability, exhibiting highly complex spatiotemporal nonstationarity. In the frequency domain, the main energy peak of sea surface wind noise typically shifts with wind speed within the range of several hundred hertz to several thousand hertz, and is superimposed on transient impulse noise generated by wave breaking, forming a complex spectral structure where continuous and discrete transient components coexist. In the time domain, sea surface wind noise exhibits significant intermittency and impulsivity, especially generating strong transient sound pulses of up to tens of milliseconds at the moment of wave breaking, with peak amplitudes exceeding the average background wind noise by more than 20 dB. This nonstationarity is much faster than the tracking capability of traditional adaptive filtering algorithms on a time scale, causing the algorithm to lag in response during strong wind noise transients and then over-adjust after the transient, resulting in speech distortion or signal distortion.

[0058] More importantly, the statistical characteristics of sea surface wind noise are strongly coupled with the sensor deployment depth, attitude, and sea surface wave field. Under the influence of waves, underwater sensors move vertically with the buoy or platform, and their relative distance to the sea surface changes in real time, leading to dynamic changes in the propagation path loss and angle of arrival of wind noise. The directivity of the surface sensors changes continuously during the platform's swaying, causing the relative spatial distribution of wind noise and target signals to change continuously. This coupling characteristic determines that the statistical characteristics of sea surface wind noise depend not only on meteorological conditions but also on the six degrees of freedom motion of the platform. However, existing technologies generally lack explicit modeling and utilization of this multi-factor coupling relationship.

[0059] (ii) Limitations of existing marine environmental noise suppression technologies.

[0060] Existing technologies for addressing wind noise interference in marine environments can be mainly categorized into the following types, each with significant limitations.

[0061] The first category is marine environmental noise estimation methods based on statistical models. These methods utilize empirical models of marine environmental noise, such as the Wenz curve or statistical results of noise spectral levels in specific sea areas, to estimate the wind noise power spectrum and then suppress it using spectral subtraction or Wiener filtering. Their core drawback is that empirical models reflect long-term statistical averages and cannot capture short-term transient changes and local anomalies in wind noise. When actual sea conditions deviate from model assumptions, the noise estimation error increases significantly, leading to undersuppression or oversuppression.

[0062] The second category involves the application of traditional adaptive noise cancellation methods on offshore platforms. Some marine buoys or shipboard systems employ a dual-sensor configuration: one sensor acquires the main signal, while the other is deployed at a depth or location less affected by wind noise as a reference. However, limited by the physical size and deployment conditions of the offshore platform, the reference sensor struggles to obtain a pure wind noise reference signal—the underwater reference channel is still affected by underwater target signal leakage, and the surface reference channel exhibits complex multipath propagation differences compared to the main channel. Furthermore, traditional LMS-type adaptive algorithms are slow to respond to the non-stationarity of sea surface wind noise and cannot handle the problem of severe overlap between wind noise and target signals in the time and frequency domains.

[0063] The third category comprises underwater noise suppression methods based on deep learning, which have been developed in recent years. With the introduction of deep neural networks into the field of underwater acoustic signal processing, some studies have attempted to use convolutional neural networks or recurrent neural networks for end-to-end suppression of sea surface wind noise. However, these methods face two prominent problems: First, acquiring training data for sea surface wind noise is extremely difficult; synchronous clean speech and wind noise data under real sea conditions are hard to collect, and there are significant differences between simulated and real data, severely limiting the model's generalization ability. Second, deep learning models have high computational complexity, making them difficult to deploy in real time on power-constrained platforms such as ocean buoys and underwater vehicles. More importantly, purely data-driven models lack utilization of the physical generation mechanism of sea surface wind noise, resulting in a sharp decline in performance under unseen marine environmental conditions.

[0064] The fourth category involves methods that combine physical acoustic models. Some studies attempt to construct forward models of wind noise generation based on physical models of wind noise (such as bubble oscillation models and turbulent pressure fluctuation models), and then design inverse filtering methods based on these models. However, these methods typically require precise knowledge of parameters that are difficult to obtain in real time, such as sea surface wave spectra and bubble distribution functions, and deviations between the forward model and the actual physical process can lead to the failure of inverse filtering. More importantly, existing methods fail to organically combine the physical model with the closed-loop optimization mechanism of the adaptive filter, lacking the ability to adaptively correct the physical model parameters based on real-time observation data.

[0065] In summary, existing technologies for handling sea surface wind noise share a common fundamental deficiency: the failure to deeply integrate prior knowledge of the physical coupling at the air-sea interface with adaptive signal processing mechanisms. Specifically, they lack the following three key capabilities: First, they lack the ability to quantify and predict the real-time coupling relationship between meteorology, waves, and the acoustic field during sea surface wind noise generation. Existing technologies either rely entirely on statistical estimation using acoustic signals, ignoring the value of physical observation data such as wind speed, wind direction, wave height, and platform motion; or they only use physical observations as external references, failing to embed them into the core optimization framework of adaptive cancellation. Second, they lack methods for estimating the statistical characteristics of wind noise by integrating multimodal ocean observation data (wind speed and direction instruments, wave sensors, inertial measurement units, and multi-channel acoustic arrays) with physical prior models, resulting in a slow system response to non-stationary transient sea surface wind noise. Third, they lack an adaptive mechanism that can dynamically adjust the structure and parameters of the adaptive filter based on the real-time statistical characteristics of sea surface wind noise, making it impossible to simultaneously ensure noise suppression depth and signal fidelity when sea states change rapidly.

[0066] In summary, acoustic systems for marine applications urgently require a novel wind noise cancellation technology that can deeply integrate prior physical models of the air-sea interface, multimodal marine observation data, and adaptive signal processing algorithms. This application addresses this need by proposing an adaptive cancellation method, system, and device based on the statistical characteristics of sea surface wind noise, aiming to overcome the fundamental limitations of existing technologies in the field of marine environmental noise suppression.

[0067] Disadvantages of existing technology:

[0068] I. Inherent defects in physical protection and acoustic structure design methods.

[0069] In acoustic detection and communication systems in marine environments, physical protection is the most intuitive and earliest adopted method for suppressing sea surface wind noise. These methods typically involve wrapping transducers or hydrophone arrays with waterproof and acoustically permeable membranes, elastic vibration-damping layers, or designing specific fairing structures to physically attenuate interference from wave breaking, bubble oscillation, and direct impact of turbulence on the sensors. However, these methods have insurmountable limitations in marine applications.

[0070] First, the complexity of the marine environment limits the effectiveness of physical isolation. Sea surface wind noise has an extremely wide energy spectrum, including not only low-frequency turbulent pressure fluctuations (typically below 100Hz) but also mid-to-high frequency components (up to several kHz) generated by wave breaking and bubble resonance. Traditional waterproof and acoustically permeable materials are extremely weak in suppressing low-frequency pressure fluctuations. To enhance low-frequency isolation, the material thickness or density needs to be increased, but this inevitably introduces significant acoustic impedance mismatch, causing severe reflection and attenuation of target underwater acoustic signals (such as marine biological acoustic signals, underwater communication signals, submarine radiated noise, etc.) when passing through the isolation layer. In deep-water or long-distance detection scenarios, the target signal itself is already extremely weak; the additional insertion loss introduced by physical isolation will directly cause the detection signal-to-noise ratio to drop to an unacceptable level.

[0071] Secondly, the dynamic nature of the marine environment renders static physical protection methods completely ineffective. The intensity of sea surface wind noise changes drastically in real time with wind speed, wave height, and sea state, while physical structures such as fairings and waterproof membranes remain fixed once installed and cannot adaptively adjust to dynamic changes in sea state. Under low sea state conditions, physical structures may cause unnecessary attenuation of weak target signals; under strong wind and wave conditions, their protective capabilities are inadequate. More importantly, physical protection structures face problems such as biofouling, seawater corrosion, and material aging during long-term marine service, and their acoustic performance gradually deteriorates over time, making it impossible to maintain a stable wind noise suppression effect.

[0072] Furthermore, for marine acoustic detection platforms such as towed arrays, moorings, and buoys, the volume, weight, and fluid resistance of physical protective structures are strictly limited. In towed arrays, excessively large fairings significantly increase towing drag, affecting platform maneuverability; in buoy systems, the weight of the protective structure consumes valuable buoyancy margins. This rigid constraint of physical space dictates that purely physical isolation methods can only serve as auxiliary means of suppressing sea surface wind noise and cannot fundamentally solve the problem of wind noise interference under strong sea states.

[0073] II. Algorithmic limitations of traditional adaptive noise cancellation methods.

[0074] Traditional noise cancellation techniques, which are based on adaptive filtering, have been successful in handling steady-state noise on land, but they have revealed profound algorithmic flaws when dealing with sea surface wind noise, a special type of interference.

[0075] First, the assumptions made by traditional adaptive algorithms regarding the statistical characteristics of noise are severely mismatched with the actual characteristics of sea surface wind noise. Classical adaptive algorithms such as LMS, NLMS, and RLS implicitly assume that the noise is stationary or at least slowly time-varying, and the convergence speed and steady-state error of the algorithms depend on the relative stability of the noise's statistical characteristics. However, the generation mechanism of sea surface wind noise is extremely complex—multiple physical processes, including wave breaking driven by sea surface wind speed, bubble field generation and collapse, and turbulent boundary layer pressure fluctuations, work together, and its statistical characteristics can change drastically on time scales of seconds or even milliseconds. This strong non-stationarity and non-Gaussianity means that traditional adaptive filters are always in a "catching up" state, making it difficult to converge to the optimal solution. Under strong wind and wave conditions, the filter often lags behind during bursts of wind noise energy, resulting in significant residual interference, and introduces additional distortion due to over-adjustment during periods of wind noise remission, severely damaging the performance of subsequent signal processing stages.

[0076] Second, the problem of acquiring the reference signal is amplified dramatically in sea surface wind noise scenarios. Traditional dual-channel or multi-channel adaptive noise cancellation structures require the reference channel to acquire wind noise interference signals as purely as possible while avoiding leakage of target signals (such as underwater target radiated noise and underwater acoustic communication signals). However, in marine acoustic detection, the target signal originates from underwater, while wind noise interference originates from the sea surface. Although there is some spatial separation between the two, in shallow sea environments or when towed arrays are close to the sea surface, the spatial aliasing of the target signal and wind noise on the receiving array is extremely severe. Even more challenging is the strong spatial coherence of sea surface wind noise—pressure fluctuations generated by the same turbulent field on the sea surface are highly correlated between adjacent array elements, while the spatial coherence of the target signal on the array exhibits a completely different structure due to the complex propagation path. Traditional methods cannot effectively utilize this difference in spatial coherence to construct a pure reference signal, inevitably leading to the mixing of target signal components into the reference channel, causing the "signal self-cancellation" problem, where the filter, while attempting to cancel wind noise, also cancels part of the target signal.

[0077] Third, traditional adaptive filters lack the ability to utilize the multi-scale and multi-physics coupling characteristics of sea surface wind noise. Sea surface wind noise is not a single noise source, but rather a superposition of physical processes at different scales: low-frequency pressure fluctuations generated by large-scale waves, pulsating pressure generated by mesoscale turbulence, and high-frequency transient impacts generated by small-scale bubble bursts. These interferences at different scales highly overlap in the time and frequency domains, and their statistical characteristics vary completely with sea state. Traditional adaptive filtering methods are based solely on a single error signal optimization criterion, lacking the ability to decompose and specifically process the multi-scale structure of wind noise. This often leads to the exacerbation of the influence of wind noise at other scales while suppressing wind noise at one scale.

[0078] III. Generalization and Deployment Bottlenecks of Data-Driven Wind Noise Suppression Methods

[0079] In recent years, with the development of deep learning technology, data-driven end-to-end wind noise suppression methods have begun to be explored and applied in the field of marine acoustics. These methods use architectures such as convolutional neural networks, recurrent neural networks, or Transformers to learn complex nonlinear mappings from noisy signals to clean target signals from large-scale labeled data. However, these methods have revealed multiple shortcomings in marine application scenarios.

[0080] First, the physical diversity of sea surface wind noise poses a fundamental challenge to the generalization ability of deep learning models. The statistical characteristics of sea surface wind noise depend on the complex coupling of various factors such as wind speed, wind direction, wave height, wave direction, seawater temperature, salinity, bubble field distribution, array depth, and towing speed. Its parameter space has extremely high dimensionality and is difficult to exhaustively enumerate. Even with large-scale data augmentation strategies, the training set cannot cover all possible sea state combinations in the real ocean environment. When the model faces unseen sea states that differ significantly from the training data distribution, its performance often drops sharply, either due to over-suppression leading to distortion of weak target signals or ineffective suppression resulting in residual wind noise. This lack of generalization ability leads to a dilemma for purely data-driven methods in marine acoustic engineering applications: "overfitting to data from specific sea areas and seasons" and "underfitting to real, complex sea states."

[0081] Secondly, the stringent requirements for real-time performance and power consumption in marine acoustic detection missions present a sharp contradiction with the computational complexity of deep learning methods. On platforms such as towed linear arrays, autonomous underwater vehicles, and buoys, signal processing systems are often limited by limited power supply and computing resources. Deep learning models typically involve a high number of parameters and computational complexity; even after quantization and pruning, their power consumption and latency remain significantly higher than traditional adaptive filtering methods. For marine acoustic systems requiring long-term continuous monitoring, continuously running deep learning models for wind noise suppression would drastically shorten mission endurance; for military applications requiring real-time target detection and tracking, the inference latency of complex neural networks may not meet real-time requirements.

[0082] Furthermore, the "black box" nature of deep learning methods makes it difficult to effectively integrate with prior physical knowledge in the field of marine acoustics. Purely data-driven models lack explicit utilization of the generation mechanism, propagation path, spatial coherence structure, and influence of marine environmental parameters on sea surface wind noise. The mapping relationships they learn are often statistical correlations rather than causal relationships. This mechanism leads to problems such as the destruction of weak features of the target signal and the loss of transient signals when the target signal and wind noise overlap significantly in the time and frequency domains (especially in the low-frequency band). At the same time, the decision-making process of deep learning models is difficult to interpret. In the field of marine acoustic detection, which has extremely high requirements for reliability and interpretability, abnormal behavior of the model output is difficult to trace and debug, bringing additional reliability risks to the equipment application of the system.

[0083] Looking at the three mainstream technical approaches mentioned above, we can find that they all share a common fundamental deficiency in addressing the problem of sea surface wind noise: they fail to systematically integrate the deep-seated marine physical statistical characteristics of sea surface wind noise into the core design of adaptive cancellation mechanisms.

[0084] Physical isolation methods completely ignore the dynamic changes and sea state dependence of sea surface wind noise statistical characteristics; although traditional adaptive filtering methods have adaptive capabilities, their optimization criteria are still based on the generalized stationarity assumption, failing to fully explore key marine physical statistical characteristics such as the multi-scale structure, non-Gaussian impulsivity, and spatial coherence of sea surface wind noise as a function of sea state; although deep learning methods can fit complex nonlinear relationships, their utilization of sea surface wind noise statistical characteristics is implicit and uncontrollable, lacking explicit modeling of marine physical laws.

[0085] Specifically, existing technologies generally lack the following three key capabilities: First, they lack a mechanism to integrate marine environmental parameters (wind speed, wave height, sea state level) and acoustic array data to estimate and predict the statistical characteristics of sea surface wind noise in real time, making it impossible to use prior sea state information as the basis for adaptive adjustment; second, they lack a method to explicitly utilize the difference between the spatial coherence structure of sea surface wind noise and the spatial coherence structure of the target signal in the adaptive cancellation architecture for reference signal optimization or filter constraints, resulting in signal leakage problems being particularly prominent in shallow sea and towed array scenarios; third, they lack a closed-loop adaptive framework that can dynamically adjust the cancellation strategy (such as multi-level filtering structure, frequency band segmentation processing, and hierarchical adaptation of update criteria) based on the multi-scale physical characteristics of sea surface wind noise (large-scale wave pressure, mesoscale turbulence, and small-scale bubble noise), making it impossible for the system to maintain optimal operating conditions in complex sea state environments.

[0086] It is precisely because of these fundamental deficiencies that current sea surface wind noise suppression technologies generally face a sharp contradiction between protecting weak target signals and suppressing strong wind noise in practical marine acoustic applications, making it difficult to meet the urgent needs of marine environmental monitoring, underwater communication support, and underwater acoustic detection equipment for high signal-to-noise ratio and high robustness signal acquisition.

[0087] Reference Figure 1 This application provides an adaptive cancellation method based on the statistical characteristics of sea surface wind noise, specifically including the following steps S100~S150:

[0088] S100: Acquire multimodal sensor data, perform timestamp alignment and data fusion on the multimodal sensor data, and generate a marine environmental state sequence that is time-aligned with the underwater acoustic signal;

[0089] S110: Using a pre-established air-sea interface wind noise generation model and a wave bubble field coupling model, based on the marine environmental state sequence, predict the sea surface wind noise within a future set time window, and output the predicted statistical characteristics of the sea surface wind noise.

[0090] S120: Construct an adaptive reference noise signal based on the predicted spatial coherence coefficient matrix in the predicted statistical characteristics and the mixed acoustic signal in the multimodal sensor data using a spatial filtering method;

[0091] S130: The hybrid acoustic signal is divided into frequency bands according to the physical scale of sea surface wind noise to obtain the scale components of the hybrid acoustic signal in different frequency bands;

[0092] S140: Based on the adaptive reference noise signal, noise cancellation is performed on the scale components of each frequency band, and then the scale components of each frequency band after noise cancellation are weighted and fused to generate a preliminary enhanced signal.

[0093] S150: The preliminary enhanced signal is subjected to residual noise suppression and target signal fidelity restoration to obtain the final target signal.

[0094] Optionally, the noise cancellation of the scale component rows of each frequency band based on the adaptive reference noise signal includes the following steps:

[0095] Multidimensional features are extracted from the mixed acoustic signal, and the multidimensional features are fused to generate a comprehensive nonstationarity index;

[0096] The target strategy is selected from the preset strategy pool based on the comprehensive nonstationarity index.

[0097] The adaptive filter, driven by the target strategy, performs noise cancellation on the scale component rows of each frequency band based on the adaptive reference noise signal.

[0098] Optionally, the extraction of multidimensional features from the mixed acoustic signal includes the following steps:

[0099] The wave breaking event detection marker, temporal kurtosis coefficient, power spectrum fluctuation index, subband energy mutation rate, and sea state level normalization index are extracted from the hybrid acoustic signal as the multidimensional features.

[0100] The process of fusing the multidimensional features to generate a comprehensive nonstationary index includes the following steps:

[0101] The multidimensional features are adaptively weighted and fused to obtain the comprehensive nonstationarity index; wherein, when wave breaking events are continuously detected, the weight of the wave breaking event detection flag is increased.

[0102] Optionally, selecting a target strategy from a preset strategy pool based on the comprehensive nonstationarity index includes the following steps:

[0103] Construct and maintain a set of sea state conditions including calm, moderate, and severe sea states;

[0104] When the comprehensive nonstationarity index continuously exceeds a preset high threshold and the frequency of wave breaking events increases, the sea state is shifted to the severe sea state, and a strategy targeting wave breaking transient wind noise is selected from the preset strategy pool as the target strategy.

[0105] When the comprehensive nonstationarity index remains below a preset low threshold, the sea state is shifted to calm sea state, and a strategy targeting steady-state background wind noise is selected from the preset strategy pool as the target strategy.

[0106] Optionally, the method further includes the following steps:

[0107] Based on the current wind speed vector and significant wave height, the power spectrum template of the wind pressure transmission component is obtained from the air-sea interface wind noise generation model.

[0108] Based on the wave spectrum and wave steepness parameters, the power spectrum and transient pulse characteristics of the bubble radiation component are calculated using the wave-bubble field coupling model.

[0109] The change in wind noise propagation path is determined based on the power spectrum template of the wind pressure transmission component and the power spectrum and transient pulse characteristics of the bubble radiation component.

[0110] The platform's motion attitude data is used to perform noise spectrum attenuation compensation on the change in the wind noise propagation path in order to correct the predicted statistical characteristics.

[0111] Optionally, the step of constructing an adaptive reference noise signal based on the predicted spatial coherence coefficient matrix in the predicted statistical characteristics and the mixed acoustic signal in the multimodal sensor data using a spatial filtering method includes the following steps:

[0112] Using the dual-constraint minimum variance criterion, the predicted spatial coherence coefficient matrix is ​​used as the prior of wind noise spatial covariance, and the expected angle of arrival and array manifold of the target signal are used as the spatial features of the target signal to design spatial filtering coefficients.

[0113] The mixed acoustic signal is filtered using the spatial filtering coefficients so that the resulting adaptive reference noise signal satisfies the following conditions: its power spectrum is consistent with the predicted sea surface wind noise power spectrum, and its spatial coherence with the target signal is minimized.

[0114] Optionally, the method further includes the following steps:

[0115] During online operation, the system continuously caches the operation data within the most recent set time period. When the target signal distortion index is detected to be continuously deviating from the set conditions, the online fine-tuning mechanism is triggered.

[0116] The online fine-tuning mechanism replays the cached runtime data in small batches and introduces a loss function including a marine physical consistency regularization term to incrementally update the parameters of the mapping model used to map from marine environmental state to adaptive filter parameters.

[0117] The following section will provide a detailed introduction and explanation of the solutions in the embodiments of this application, using specific application examples.

[0118] I. Overall Technical Solution Framework.

[0119] This embodiment proposes an adaptive cancellation scheme based on the statistical characteristics of sea surface wind noise, aiming to solve the core technical challenges in the field of marine acoustic detection and communication, where existing technologies cannot effectively address the strong non-stationarity of sea surface wind noise, the coupling characteristics of multiple physics fields, and the difficulty in simultaneously achieving noise suppression and target signal fidelity due to dynamic changes in the marine environment. The core innovation of this embodiment lies in constructing a full-link technical architecture of "marine physics prior guidance—multimodal ocean perception—quantification of sea surface wind noise statistical characteristics—adaptive filtering strategy coordination—closed-loop self-evolution—heterogeneous platform collaborative processing." The key difference between this architecture and existing technologies lies in the deep integration of the physical mechanisms of sea surface wind noise generation (air-sea interface coupling, wave breaking, bubble field evolution, turbulent boundary layer pressure pulsation) with the motion characteristics of marine platforms, forming an intelligent closed-loop system capable of real-time perception of sea conditions, dynamic prediction of sea surface wind noise statistical characteristics, adaptive adjustment of cancellation strategies, and continuous self-optimization.

[0120] The overall technical solution framework is divided into seven core layers: The first layer is the multimodal ocean perception and physical prior layer, responsible for collecting sea surface meteorological parameters, ocean wave parameters, platform motion attitude, and multi-channel acoustic signals; the second layer is the sea surface wind noise statistical characteristic estimation and prediction layer, which accurately estimates and predicts the time-varying statistical characteristics of sea surface wind noise based on ocean hydrodynamic prior models and real-time perception data; the third layer is the adaptive cancellation core processing layer, which includes an adaptive reference signal construction module, a non-stationarity quantization evaluation module, and a dynamic switching module for filtering strategies to achieve adaptive and accurate cancellation of sea surface wind noise; the fourth layer is the multi-scale wind noise decomposition and hierarchical cancellation layer, which decomposes and processes the multi-scale physical structure of sea surface wind noise; the fifth layer is the closed-loop self-evolution and optimization layer, which achieves continuous optimization and adaptive evolution of system parameters through an online learning mechanism guided by a physical model; the sixth layer is the heterogeneous platform collaboration and task offloading layer, which realizes dynamic allocation and collaborative processing of computing tasks in the scenario of multi-platform networking in the ocean; and the seventh layer is the target signal output and post-processing layer, which completes the synthesis, enhancement, and output of underwater acoustic target signals or speech signals.

[0121] The following section provides a detailed explanation of the composition, function, and specific implementation process of each module.

[0122] II. Composition and Functions of Each Module

[0123] (a) Multimodal ocean perception and physical prior layer.

[0124] This layer consists of four types of sensor units: sea surface meteorological sensing units, ocean wave sensing units, inertial measurement units, and multi-channel underwater acoustic / acoustic array units. The sea surface meteorological sensing units employ shipborne or buoy-borne anemometers, barometers, and temperature and humidity sensors to collect real-time sea surface wind speed vectors (including instantaneous wind speed, gust intensity, and wind direction), atmospheric pressure, and atmospheric stability parameters. The ocean wave sensing units use wave radar, accelerometer-type wave buoys, or pressure-type wave meters to collect real-time information on effective wave height, wave direction, wave period, and wave spectrum. The inertial measurement units include three-axis accelerometers, three-axis gyroscopes, and magnetometers to collect six-degree-of-freedom motion attitude, vibration state, and spatial orientation changes of offshore platforms (buoys, ships, towed bodies, etc.). The multi-channel underwater acoustic / acoustic array unit consists of at least two hydrophones or microphones, with an optimized layout based on the platform form. It is used to collect mixed acoustic signals, including underwater target signals (such as submarine radiated noise, underwater acoustic communication signals, and marine biological acoustic signals) and sea surface wind noise interference signals.

[0125] The core functions of this layer are twofold: firstly, to provide prior information about the marine physical environment for subsequent estimation of sea surface wind noise statistical characteristics; and secondly, to compensate for the inadequacy of a single underwater acoustic signal in sensing sea surface wind noise through the spatiotemporal alignment and fusion of multimodal data. Sea surface meteorological data, wave data, and inertial measurement data are fused using extended Kalman filtering to eliminate sensor noise and measurement delay, forming a smooth marine environmental state sequence that is time-aligned with the underwater acoustic signal.

[0126] (II) Statistical characteristics estimation and prediction of sea surface wind noise.

[0127] This layer is one of the core innovations that distinguishes this embodiment from existing technologies. Existing methods only estimate noise characteristics from underwater acoustic signals, lacking utilization of the physical generation mechanism of sea surface wind noise. This layer innovatively combines a priori ocean hydrodynamic models with real-time ocean sensor data to construct a statistical feature estimator for sea surface wind noise.

[0128] This layer comprises three sub-modules: the air-sea interface wind noise generation model module, the wave-bubble field coupling model module, and the statistical feature prediction module. The air-sea interface wind noise generation model, based on wind-generated noise theory in marine acoustics, describes the spectral characteristics of pressure fluctuations generated by wind acting on the sea surface and their transmission coefficients propagating underwater under different wind speeds, directions, and atmospheric stability conditions. This model is pre-established through a combination of marine acoustic simulation and on-sea experimental calibration, forming a parameterizable prior model library. The wave-bubble field coupling model module, based on wave breaking dynamics and bubble oscillation acoustics theory, describes the physical processes of bubble cloud generation, evolution, oscillation, and noise radiation during wave breaking, outputting the power spectral density and transient impulse characteristics of bubble noise. The statistical feature prediction module uses the current sea surface meteorological data, wave data, and platform motion attitude data, combined with the two physical models mentioned above, to output the predicted statistical characteristics of sea surface wind noise within a short time window (e.g., 50 milliseconds to 200 milliseconds), including the predicted power spectrum (subdivided into wind pressure transmission components and bubble radiation components), the predicted spatial coherence coefficient matrix, and the predicted nonstationarity index (including the probability prediction of wave breaking events).

[0129] (iii) Adaptive cancellation core processing layer.

[0130] This layer is the core of the algorithm execution in this embodiment, and contains three tightly coupled modules, among which the adaptive reference signal construction module and the dynamic switching module for filtering strategies are specially designed for sea surface wind noise characteristics.

[0131] The adaptive reference signal construction module is responsible for constructing the optimal reference noise signal from the multi-channel underwater acoustic / acoustic array. Traditional methods simply use auxiliary sensor signals as a reference, resulting in target signal leakage. This embodiment innovatively introduces an adaptive reference signal construction method based on the spatial coherence spectrum analysis of sea surface wind noise and the spatial coherence spectrum analysis of the target signal. This module uses the statistical characteristics of sea surface wind noise to estimate the predicted spatial coherence coefficient matrix output (this matrix describes the spatial coherence structure of wind noise on the array and is closely related to sea state and array geometry), combined with real-time multi-channel acoustic signals, and extracts the signal component dominated by sea surface wind noise from each channel signal through spatial filtering methods, while suppressing the leakage of underwater target signals. Crucially, this module also utilizes the prior spatial coherence characteristics of the target signal—the underwater target signal presents a far-field plane wave or a spatial structure with a specific beam direction on the array, which is fundamentally different from the near-field, strongly coherent structure of wind noise. The module designs a dual-constraint spatial filter to maximize the energy proportion of wind noise in the output reference signal while minimizing the energy proportion of the target signal. This process dynamically adjusts with changes in sea state. When the predicted probability of wave breaking events increases, the filter parameters are adaptively adjusted to enhance the response to bubble pulse noise.

[0132] The nonstationarity quantification assessment module is responsible for real-time quantification of the time-varying intensity of sea surface wind noise, with particular attention to the transient pulse characteristics caused by wave breaking events. Traditional methods only judge the noise magnitude with coarse granularity based on short-time energy or signal-to-noise ratio, making it difficult to finely characterize the nonstationar characteristics of sea surface wind noise. This embodiment innovatively constructs a multi-dimensional quantitative index system for the nonstationarity of sea surface wind noise, including: wave breaking event detection indicators (based on the detection of sudden increases in bubble noise energy), time-domain kurtosis coefficient (reflecting the intensity of impulse), power spectrum fluctuation index (reflecting abrupt changes in spectral structure), sub-band energy mutation rate (focusing on frequency bands where wind noise energy is concentrated), and sea state normalization index (directly provided by meteorological wave sensors). These indicators describe the degree of nonstationarity of sea surface wind noise in the time domain, frequency domain, and statistical distribution from different dimensions. The module integrates the above indicators into a comprehensive nonstationarity index, which changes continuously over time, reflecting the intensity of the "suddenness" and "dynamics" of sea surface wind noise at each moment, and has a particularly high sensitivity response to wave breaking events.

[0133] The dynamic switching module for filtering strategies selects and switches the update strategy and core parameters of the adaptive filter in real time based on the comprehensive non-stationarity index output by the non-stationarity quantification evaluation module. This embodiment innovatively designs a strategy pool for sea surface wind noise, containing multiple adaptive filtering strategies with different convergence characteristics and steady-state performance. Strategy 1 is a wave-tracking strategy, employing an update mechanism based on higher-order statistics to prioritize the filter's ability to quickly track sudden changes in wind noise, targeting strong transient pulses caused by wave breaking. Strategy 2 is a steady-state background strategy, using a recursive least squares algorithm combined with an adaptive forgetting factor to prioritize steady-state accuracy and target signal fidelity for the background continuous spectrum components of sea surface wind noise. Strategy 3 is an impact-resistant protection strategy, introducing nonlinear limiting and filter coefficient preservation mechanisms to prevent filter divergence in the face of strong impact noise under extremely harsh sea conditions. The switching logic adopts a sea state machine design, combining the predicted probability of wave breaking events and the comprehensive non-stationarity index for state transitions. During the switching process, a filter state weighted transition mechanism is used to achieve smooth output connection.

[0134] (iv) Multi-scale wind noise decomposition and layered cancellation layer.

[0135] This layer is another core innovation of this embodiment. Addressing the characteristic that sea surface wind noise is formed by the superposition of different physical processes, it achieves multi-scale decomposition and layered cancellation. Sea surface wind noise can be decomposed into three physical scales: a large-scale component (low-frequency pressure fluctuations generated by long waves at the sea surface, typically below 50Hz), a medium-scale component (generated by pressure pulsations in the turbulent boundary layer, with a frequency range of 50-500Hz), and a small-scale component (generated by wave breaking, bubble cloud oscillations, and radiation, with a frequency range of 500Hz-10kHz, including strong transient pulses). The statistical characteristics, spatial coherence, and masking effects on target signals of these three scale components are drastically different, making it difficult to achieve optimal performance using a uniform filter.

[0136] This layer comprises three parallel sub-modules: a low-frequency fluctuation cancellation module, a turbulence noise cancellation module, and a bubble pulse cancellation module. The low-frequency fluctuation cancellation module employs a low-order adaptive filter, using a long window and small step size filtering strategy to address the low-frequency, slowly varying characteristics of large-scale wind noise components, avoiding over-modulation of the low-frequency components of the target signal. The turbulence noise cancellation module uses a frequency band segmentation adaptive filtering structure, dividing the mid-frequency band into several sub-bands. Each sub-band undergoes independent adaptive noise cancellation, and the filter parameters between sub-bands are independently optimized to adapt to the non-uniform spectral characteristics of turbulence noise. The bubble pulse cancellation module employs an event detection-based adaptive transient suppression structure. First, a bubble pulse detector identifies the time window of wave breaking events. Then, within this window, a dedicated transient pulse suppression filter is activated. This filter uses an adaptive algorithm based on high-order statistics, enabling rapid tracking of the pulse's start and end times and amplitude changes. Outside the window, background processing mode is maintained.

[0137] The outputs of the three modules are weighted and fused to form a preliminary enhanced signal. The fusion weights are dynamically adjusted according to the energy proportion of wind noise at each scale under the current sea state. When the sea state is dominated by wave breaking, the output weight of the bubble pulse cancellation module is automatically increased.

[0138] (v) Closed-loop self-evolution and optimization layer.

[0139] This layer constructs a closed-loop self-evolving system guided by a physical model, which solves the problem in existing technologies where system parameters are fixed and cannot self-optimize in response to long-term changes in the marine environment.

[0140] The offline training phase utilizes large-scale marine environment simulation data and actual sea experimental data to establish a mapping model from multimodal marine sensor inputs to optimal adaptive filter parameters. Unlike purely data-driven black-box models, this embodiment introduces marine physical constraints during model training—using the air-sea interface wind noise generation model and the wave-bubble field coupling model as physical consistency regularization terms to constrain the output space of the mapping model. This ensures that the parameter relationships learned by the model conform to the laws of marine acoustic physics, avoiding the learning of spurious correlations. This design enables the model to provide reasonable outputs based on physical priors, even under rare sea state conditions with insufficient training data coverage.

[0141] The online fine-tuning process runs continuously during the actual deployment of the marine platform. The system records key state variables for each adaptive cancellation process, including the marine sensor input sequence, filter parameter adjustment trajectory, residual noise energy, and target signal distortion index (obtained by comparing changes in the target signal's harmonic structure). When the system detects that the target signal distortion index is consistently high or the noise suppression performance is lower than expected within a certain time period, it triggers the online fine-tuning mechanism. This mechanism replays the recorded operational data in small batches and, combined with the current marine physical model output, incrementally updates some parameters of the mapped model. Specifically, the online fine-tuning mechanism is designed with slow parameter tracking capabilities to address the long-term time-varying characteristics of the marine environment (such as seasonal sea state changes and long-term meteorological trends). By introducing cumulative gradient updates with a forgetting factor, the model can smoothly adapt to the long-term evolution of the marine environment.

[0142] (vi) Heterogeneous platform collaboration and task offloading layer.

[0143] In collaborative observation network scenarios composed of multiple heterogeneous marine platforms, such as a three-dimensional marine observation network consisting of surface unmanned vessels, underwater autonomous vehicles, moored buoys, and shore-based base stations, this layer realizes the dynamic allocation and collaborative processing of computational tasks. This layer includes a platform capability discovery module, a task segmentation module, and a result fusion module.

[0144] The platform capability discovery module utilizes underwater acoustic communication or satellite communication links to monitor the computing resource status, energy level, communication bandwidth, and current task load of each platform within the network in real time. The task segmentation module dynamically divides the adaptive cancellation task into several sub-tasks based on the severity of sea surface wind noise and the system's processing complexity. These sub-tasks include multi-channel signal acquisition, sea surface wind noise statistical characteristic estimation, multi-scale decomposition, adaptive filtering calculation, and post-processing. Surface platforms with sufficient energy and strong computing power undertake the computationally intensive multi-scale decomposition and adaptive filtering tasks, while energy-constrained underwater autonomous vehicles (AUVs) only undertake signal acquisition and preliminary preprocessing tasks. The decision-making basis for task offloading includes not only the platform status but also the estimation of sea surface wind noise statistical characteristics—when the sea surface wind noise non-stationarity index is high and wave breaking events are present, the system prioritizes scheduling tasks to surface platforms, utilizing their stronger computing power and richer external sensor data.

[0145] The result fusion module collects the intermediate results processed by each platform and generates the final output signal through a weighted fusion strategy. This fusion strategy dynamically adjusts the fusion weights based on the signal-to-noise ratio improvement of each platform's processing results and the degree of matching with the ocean physical model. At the same time, it uses the channel state information of the underwater acoustic communication link to correct the fusion weights, ensuring that robust fusion results can still be obtained even when the communication quality is poor.

[0146] (vii) Target signal output and post-processing layer.

[0147] This layer receives the initial enhanced signals from the adaptive cancellation core processing layer and the multi-scale decomposition layer, and performs residual noise suppression and target signal fidelity restoration. The post-processing module employs a lightweight processing method tailored to the statistical characteristics of sea surface wind noise, compensating for subtle spectral distortions that may be introduced during multi-scale cancellation. For underwater target signals (such as submarine radiated noise and underwater acoustic communication signals), the post-processing module features a specially designed line spectrum enhancement function, utilizing the prior harmonic structure of the target signal to enhance weak line spectra, further improving detection performance. This layer ultimately outputs a clear, high-fidelity target signal for subsequent underwater acoustic target detection, identification, tracking, or underwater acoustic communication demodulation.

[0148] III. Specific implementation process of each module.

[0149] (I) Implementation process of multimodal ocean sensing and physical prior layer.

[0150] After the offshore platform is deployed, the system initializes all sensors and establishes a unified time reference. The sea surface meteorological sensor unit continuously collects wind speed vector, air pressure, and temperature and humidity data at a sampling rate of no less than 10Hz. After low-pass filtering, it outputs instantaneous wind speed, gust intensity (based on the peak-to-mean ratio of wind speed), wind direction, and atmospheric stability parameters (calculated using Richardson's number). The ocean wave sensor unit outputs significant wave height, dominant wave period, wave direction, and wave spectrum data, with a sampling rate synchronized with the meteorological unit. The inertial measurement unit collects acceleration and angular velocity data at a sampling rate of 100Hz or higher, and calculates the platform's roll, pitch, bow angles, and vertical displacement in real time using attitude calculation algorithms. The multi-channel underwater acoustic / acoustic array synchronously collects acoustic signals at a sampling rate of no less than twice the highest frequency of the target signal, typically 32kHz or higher.

[0151] The system timestamps and aligns the four data streams before inputting them into the data fusion module. The fusion module employs an unscented Kalman filter, using inertial measurement data as the equation of motion and meteorological and wave data as the observation equation to estimate the platform's true motion state in the marine environment, eliminating sensor measurement errors caused by platform motion. Specifically, the fusion module also outputs real-time depth or height information of the platform relative to the sea surface, which is crucial for subsequent calculations of wind noise propagation path loss. The fused marine environment state sequence is output to the sea surface wind noise statistical characteristic estimation and prediction layer at fixed frame lengths (e.g., 20 milliseconds).

[0152] (II) Implementation process of sea surface wind noise statistical characteristic estimation and prediction layer.

[0153] The system was pre-calibrated using marine acoustic simulation software and on-sea experiments to establish air-sea interface wind noise generation models under different sea states. Different combinations of wind speed, direction, wave height, and wave period were set in the simulation, and the wind noise power spectral density at the hydrophone was recorded to establish a three-dimensional mapping table of wind speed-wave height-noise spectrum. Simultaneously, based on the wave breaking dynamics model, a model was established relating the probability of wave breaking events to wave steepness and wind speed, as well as a time-domain waveform template library for bubble cloud radiated noise.

[0154] During real-time operation, the system acquires the fused marine environmental state sequence from the multimodal perception layer. First, based on the current wind speed vector and significant wave height, the power spectrum template of the wind pressure transmission component is retrieved or interpolated from the three-dimensional mapping table. Then, based on the wave spectrum and wave steepness parameters, the power spectrum and transient pulse characteristics of the bubble radiation component are calculated using a wave-bubble field coupling model, including the expected amplitude, duration, and probability of the pulse. Next, the propagation path changes caused by the platform's vertical motion are corrected using platform motion attitude data—as the platform dives or surfaces, the propagation attenuation of wind noise changes accordingly, and the system compensates for the noise spectrum attenuation based on the real-time depth. Finally, the statistical feature prediction module predicts the power spectrum evolution trend of wind noise, the probability of bubble pulse occurrence, and the coherence coefficient change trend within a 100-millisecond time window based on the wind speed change trend (predicted through differencing and autoregression of the wind speed time series) and the wave field evolution model (based on the wave propagation speed and the relative velocity of the platform). The prediction results are output as parameter vectors to the adaptive cancellation core processing layer and the multi-scale decomposition layer.

[0155] (III) Implementation process of adaptive reference signal construction module.

[0156] The system acquires the time-domain signals of the multi-channel underwater acoustic array in real time and performs short-time Fourier transform on a frame-by-frame basis to convert them to the time-frequency domain. For each time-frequency unit, the module obtains the predicted spatial coherence coefficient matrix output by the sea surface wind noise statistical characteristic estimation layer. This matrix describes the phase and amplitude correlation of sea surface wind noise between channels, and its structure is related to sea state, array geometry, and platform depth.

[0157] The core of reference signal construction is designing a set of spatial filter coefficients that ensure the filtered reference signal satisfies two conditions: its power spectrum is consistent with the predicted sea surface wind noise power spectrum (including wind pressure transmission components and bubble radiation components); and its correlation with the spatial coherence structure of the target signal is minimized. The module employs a dual-constraint minimum variance criterion, using the predicted wind noise coherence coefficient matrix as prior information for the wind noise spatial covariance matrix, and simultaneously using the expected angle of arrival and array manifold of the target signal as spatial features to construct the spatial filter. The optimization objective of this filter is to maximize the energy proportion of wind noise in the output reference signal while constraining the energy of the target signal in the output to be below a preset threshold. Since the coherence coefficient matrix of sea surface wind noise changes dynamically over time, the coefficients of the spatial filter are also adaptively updated frame by frame.

[0158] The filtered reference signal is returned to the time domain via inverse Fourier transform and used as the reference input for the adaptive filter. Compared to traditional methods that directly use the original signal from the auxiliary hydrophone, the reference signal constructed in this module has higher wind noise purity and lower target signal leakage, fundamentally improving the input conditions for adaptive noise cancellation.

[0159] (iv) The joint implementation process of the nonstationarity quantification evaluation module and the dynamic switching module of filtering strategy.

[0160] The system extracts multidimensional non-stationary features from each frame of the underwater acoustic signal. Wave breaking event detection indicators are obtained by monitoring energy spikes and waveform characteristics within the bubble radiation band (typically 1-10kHz). A wave breaking event is identified when the energy exceeds an adaptive threshold and the waveform exhibits a typical pulse shape. The temporal kurtosis coefficient is obtained by calculating the ratio of the fourth moment to the square of the second moment of the signal amplitude within a short time window, reflecting the intensity of the pulse-like abrupt changes in the signal; the kurtosis coefficient increases significantly during wave breaking. The power spectrum fluctuation index is obtained by comparing the Barthel distance between the current frame's power spectrum and the previous frame's power spectrum, reflecting the degree of abrupt changes in the spectral structure. The subband energy mutation rate focuses on the low- to mid-frequency subbands where wind noise energy is concentrated, calculating the statistical distribution characteristics of the inter-frame energy change rate of this subband. The sea state normalization index is directly provided by the wave sensor and, after normalization, serves as a macroscopic reference for sea surface wind noise non-stationarity.

[0161] The system adaptively weights and fuses the above five indicators to output a comprehensive nonstationarity index. The weight coefficients are adaptively adjusted by the system. When wave breaking events occur consecutively, the weight of the breaking event detection marker is automatically increased, enabling the comprehensive index to more sensitively reflect the strong nonstationarity caused by wave breaking.

[0162] The dynamic switching module for filtering strategies has three preset strategy modes. Mode 1 is a wave-tracking strategy, employing a normalized LMS algorithm combined with a step-size adjustment mechanism based on higher-order statistics. The step-size factor is positively correlated with the comprehensive non-stationarity index and the probability of wave breaking events; the stronger the non-stationarity, the larger the step-size, ensuring the filter responds quickly to sudden changes in wind noise. Mode 2 is a steady-state background strategy, employing a recursive least squares algorithm combined with an exponential forgetting factor. The forgetting factor is related to the sea state stability index, prioritizing the steady-state accuracy of the filter. Mode 3 is an anti-shock protection strategy, introducing a nonlinear amplitude limiting mechanism and a filter coefficient freezing mechanism for extremely severe sea conditions to prevent the filter from diverging due to excessively large input signal amplitudes.

[0163] The switching logic employs a state machine design based on sea state. The system maintains sea state variables, including three states: "calm sea state," "moderate sea state," and "severe sea state." When the comprehensive nonstationarity index continuously exceeds a high threshold and the frequency of wave breaking events increases, the state transitions to "severe sea state," switching to either a shock-resistant protection strategy or a wave-tracking strategy. When the index remains below a low threshold, it returns to "calm sea state," switching to a steady-state background strategy. During the switching process, the system performs a weighted transition of the filter states under the old and new strategies, ensuring a smooth connection of the output signals.

[0164] (V) Implementation process of multi-scale wind noise decomposition and hierarchical cancellation layer.

[0165] The system receives multi-channel underwater acoustic signals and prediction information output from the sea surface wind noise statistical characteristic estimation layer. First, signal frequency band segmentation is performed, dividing the frequency band boundaries according to the three physical scales of sea surface wind noise: a low-frequency boundary of 50Hz, a mid-frequency boundary of 500Hz, forming three frequency band signals. For each frequency band, the signal is sent to the corresponding cancellation module.

[0166] The low-frequency fluctuation cancellation module (<50Hz) employs second- to fourth-order adaptive filters. Targeting the low-frequency, slowly varying characteristics of large-scale wind noise components, the filters have low order and small step size to protect any low-frequency line spectrum components that may exist in the target signal (such as the line spectrum components of submarine radiated noise). The reference signal for this module is provided by the adaptive reference signal construction module, but bandwidth optimization has been performed for low-frequency characteristics.

[0167] The turbulence noise cancellation module (50-500Hz) employs a frequency band segmentation adaptive filtering structure, dividing the frequency band into 8 or 16 sub-bands. Each sub-band undergoes independent adaptive noise cancellation using a normalized LMS algorithm. The sub-band segmentation is non-uniform, with narrower bandwidths in low-frequency sub-bands to protect fine structures and wider bandwidths in high-frequency sub-bands to improve computational efficiency. The step size of each sub-band filter is independently adjusted based on the non-stationarity index within the sub-band.

[0168] The bubble pulse cancellation module (>500Hz) employs an event-detection-based adaptive transient suppression structure. The system continuously monitors the high-frequency signal and identifies bubble pulse events by matching a short-time energy detector with a waveform template. When a pulse event is detected, a transient suppression filter is activated. This filter uses an adaptive algorithm based on the minimum mean absolute deviation criterion to quickly track and suppress the pulse. The suppression window length is adaptively adjusted according to the pulse duration (typically 5-20 milliseconds). Outside the window, the system maintains a background processing mode and employs a steady-state filtering strategy to protect the high-frequency details of the target signal.

[0169] The outputs of the three modules are weighted and fused in the time domain. The fusion weights are dynamically adjusted based on the energy proportion of each scale in the output of the estimation layer according to the statistical characteristics of sea surface wind noise. At the same time, a smoothing mechanism is introduced to avoid abrupt weight changes that may introduce auditory artifacts or affect processing.

[0170] (vi) Implementation process of closed-loop self-evolution and optimization layer.

[0171] During the offline training phase, the system constructs a training dataset containing tens of thousands of marine environment scene-optimal parameter pairs. Each dataset includes an input feature vector (wind speed and direction sequence, wave parameter sequence, inertial measurement sequence, acoustic features) and a label (optimal adaptive filter parameters obtained through multi-objective optimization search, with the optimization objective being to minimize the weighted sum of residual noise energy and target signal distortion index). During training, the model employs a structure combining convolutional neural networks and gated recurrent units, outputting recommended filter parameters after inputting temporal features. Unlike traditional training, this embodiment adds a marine physics consistency regularization term to the loss function. This regularization term calculates the difference between theoretical and measured residual noise under the output parameters by substituting the parameters output by the model into the air-sea interface wind noise generation model and the wave-bubble field coupling model, while simultaneously checking whether the output parameters meet physical boundary conditions (such as the relationship between filter order and wind noise coherence length), thereby penalizing outputs that violate physical laws.

[0172] During the online fine-tuning phase, the device continuously caches the operational data from the most recent 60 seconds. The system calculates performance metrics every 10 seconds, including average residual noise energy and target signal distortion metrics (obtained by comparing harmonic structure changes and line spectrum preservation within the target signal's active detection segment). When the target signal distortion metric exceeds a set threshold for 30 consecutive seconds, or when residual noise energy abnormally increases under stable sea conditions, the system determines that the current model parameters are not optimal for the current marine environment and triggers fine-tuning. The fine-tuning process employs mini-batch stochastic gradient descent, using the cached data from the most recent 60 seconds as training samples, and constraining the update direction with a marine physics consistency regularization term, updating the parameters of the last few layers of the model in a limited number of steps. Specifically, to address long-term changes in the marine environment (such as changes in sea state distribution due to seasonal changes), the system designs a slow-varying tracking mechanism. By introducing a cumulative gradient smoothing term, the model can smoothly adapt to environmental changes over long time scales, avoiding drastic parameter fluctuations caused by short-term anomalies.

[0173] (vii) Implementation process of heterogeneous platform collaboration and task unloading layer.

[0174] In a multi-platform marine networking scenario, the system first discovers neighboring platforms via underwater acoustic communication or satellite communication links, exchanging platform capability information (computing resources, energy levels, sensor configurations, current task load) and real-time marine environmental perception data. The platform capability discovery module then aggregates this information to form a network resource view.

[0175] The task partitioning module determines the task allocation method based on the current sea surface wind noise non-stationarity index, wave breaking event probability, and system processing frame rate requirements. When the sea surface wind noise non-stationarity index is low and the sea state is stable, a centralized processing mode is adopted, with the main platform (such as an unmanned surface vessel or shore-based base station) completing all calculations, and only the raw underwater acoustic signals being uploaded from the platform (such as an autonomous underwater vehicle). When the sea surface wind noise non-stationarity index is high and wave breaking events are present, a distributed processing mode is adopted. The multi-scale decomposition and adaptive filtering tasks with the highest computational complexity are assigned to the platform with the strongest computing power and the most abundant energy; the sea surface wind noise statistical characteristic estimation task is assigned to the surface platform equipped with meteorological wave sensors; and the reference signal construction task is assigned to the platform with the most underwater acoustic array channels. The task partitioning module encapsulates the computational tasks into independently executable operator units, distributes them to each platform through a reliable underwater acoustic communication protocol, and sets task timeout and retransmission mechanisms to cope with the instability of underwater acoustic communication.

[0176] After each platform completes its assigned computational task, it returns the processing results and metadata (such as computation time, intermediate states, and local signal-to-noise ratio improvement) to the main platform. The result fusion module performs a quality assessment on the returned multiple results, using the weighted reciprocal of the residual noise energy and the target signal distortion index corresponding to each result as the fusion weight. Simultaneously, a communication link quality factor (based on signal-to-noise ratio and bit error rate estimation) is introduced to correct the weights, performing weighted superposition in the time domain or time-frequency domain to generate the final enhanced target signal. When some platforms are unable to return results due to communication interruption, the system automatically degrades to a fusion mode with usable results, ensuring the system's robustness in complex marine environments.

[0177] (viii) Implementation process of target signal output and post-processing layer.

[0178] The post-processing layer receives the initial enhanced signal output from the adaptive cancellation core processing layer and the multi-scale decomposition layer. The system performs a short-time Fourier transform on this signal to analyze the harmonic structure and continuity of each time-frequency unit. For frequency bands with significant residual wind noise, a spectral post-filtering technique based on an auditory masking threshold is employed to further attenuate residual noise while preserving the harmonic structure of the target signal. For underwater target detection scenarios, the post-processing module features a specially designed line spectrum enhancement submodule. By tracking the frequency stability of the target signal line spectrum, it adaptively boosts the gain of the detected stable line spectrum, further improving the detection signal-to-noise ratio.

[0179] After post-processing, the system outputs the final target signal. This output signal can be used simultaneously on multiple channels: one channel outputs to the underwater acoustic communication demodulator for communication reception; another outputs to the target detection and recognition module for underwater target classification and identification; one channel is saved to local storage for post-processing analysis; and another is transmitted back to the shore-based or mother ship monitoring center via a communication link. The system also outputs key performance indicators of this processing (such as wind noise suppression, target signal fidelity, current sea state, and the operating status of each module) to the platform's health management system for reference by operators or the automated decision-making system.

[0180] This embodiment systematically addresses the core pain points of existing sea surface wind noise cancellation technologies in marine acoustic detection and communication applications by constructing a full-link technology system guided by prior marine physics, including multimodal sensing, multi-scale decomposition and statistical characteristic quantification of sea surface wind noise, dynamic switching of adaptive filtering strategies, closed-loop self-evolutionary learning guided by physical models, and collaborative processing on heterogeneous platforms. The tightly coupled and collaborative modules form a complete, implementable, and significantly innovative technical solution, providing a novel technical path for acquiring high signal-to-noise ratio acoustic signals in marine environments.

[0181] In summary, this embodiment includes the following key technical solutions:

[0182] An adaptive cancellation method based on wind noise statistical characteristics includes: acquiring multimodal sensor data, which includes at least wind speed and direction sensing data, inertial measurement data, and acoustic signals from a multi-microphone array; estimating and predicting the time-varying statistical characteristics of wind noise in real time based on a prior model of wind field hydrodynamics and the multimodal sensor data, whereby the time-varying statistical characteristics include at least a predicted power spectrum, a predicted spatial coherence coefficient matrix, and a predicted nonstationarity index; performing spatial filtering on the multi-microphone array signal according to the predicted spatial coherence coefficient matrix to construct an adaptive reference signal, wherein the optimization objective of the spatial filtering is to minimize the leakage energy of the target speech in the reference signal; and dynamically selecting and switching the update strategy and core parameters of the adaptive filter according to the predicted nonstationarity index to perform adaptive wind noise cancellation.

[0183] Furthermore, based on the wind field hydrodynamic prior model and multimodal sensor data, the time-varying statistical characteristics of wind noise are estimated and predicted in real time. Specifically, this includes: establishing the power spectral density function of wind noise at the microphone port and the coherence coefficient matrix between multiple microphones under different wind speeds, turbulence intensities, and wind direction angles through hydrodynamic simulation in advance, forming a wind field-wind noise prior model library; acquiring the fused wind speed vector and equipment attitude data in real time, and retrieving or interpolating the wind noise power spectral density template and coherence coefficient template under the current wind field conditions from the prior model library; using inertial measurement data to correct the influence of relative wind direction changes caused by equipment movement on the coherence coefficient template; and predicting the power spectral density evolution trend and coherence coefficient change trend of wind noise within future time windows based on the wind speed change trend and turbulence inertial model.

[0184] Furthermore, an adaptive reference signal is constructed by spatial filtering the multi-microphone array signal based on the predicted spatial coherence coefficient matrix. Specifically, this includes: converting the multi-microphone array signal to the time-frequency domain; for each time-frequency unit, using the predicted spatial coherence coefficient matrix as prior information for the wind noise spatial covariance matrix, constructing a minimum variance distortion-free response spatial filter; the optimization objective of the spatial filter is to minimize the leakage energy of the target speech component in the output reference signal, with the constraint of maintaining the response to the wind noise component; and the output of the spatial filter is inversely transformed and used as the reference input of the adaptive filter.

[0185] Furthermore, the update strategy and core parameters of the adaptive filter are dynamically selected and switched based on the predicted nonstationarity index. Specifically, this includes: constructing a multi-dimensional wind noise nonstationarity quantification index system, which includes time-domain kurtosis coefficient, power spectrum fluctuation index, subband energy mutation rate, and turbulence intensity normalization index; integrating the multi-dimensional indexes into a comprehensive nonstationarity index; pre-setting a strategy pool containing a fast tracking strategy, a steady-state low-distortion strategy, and an anti-impact strategy; switching to a fast tracking strategy when the comprehensive nonstationarity index exceeds a first threshold, using an adaptive variable step size mechanism positively correlated with the comprehensive nonstationarity index; switching to a steady-state low-distortion strategy when the comprehensive nonstationarity index is below a second threshold, using a recursive least squares algorithm combined with an exponential forgetting factor; temporarily switching to an anti-impact strategy when the comprehensive nonstationarity index exceeds a third threshold higher than the first threshold, introducing a nonlinear limiting and filter coefficient preservation mechanism; and using a state machine and hysteresis comparison logic during the switching process, and introducing a filter state weighted transition mechanism to achieve smooth output connection.

[0186] Furthermore, it also includes closed-loop self-evolution and optimization steps: In the offline stage, a mapping model from the wind field scene to the optimal adaptive filter parameters is constructed. During the model training process, a prior model of wind field hydrodynamics is introduced as a physical consistency regularization term to constrain the output space of the mapping model to conform to the physical laws of wind noise generation and propagation. In the online stage, running data is cached and speech distortion index and residual noise energy are monitored. When the speech distortion index continuously exceeds the threshold, online fine-tuning is triggered. The cached data is replayed in small batches, and some parameters of the mapping model are incrementally updated under the constraint of the physical consistency regularization term.

[0187] Furthermore, the physical consistency regularization term calculates the difference between the theoretical residual noise and the measured residual noise under the filter parameters by substituting the filter parameters output by the mapping model into the wind field-wind noise prior model. This difference is then added as a regularization term to the loss function of the model training to penalize parameter outputs that violate physical laws.

[0188] Furthermore, it includes: a multimodal perception and physical prior layer, comprising a wind speed and direction sensing unit, an inertial measurement unit, and a multi-microphone array unit, used to collect environmental wind speed vectors, equipment motion attitude, and multi-channel acoustic signals; a wind noise statistical characteristic estimation and prediction layer, used to estimate and predict the time-varying statistical characteristics of wind noise in real time based on the wind field hydrodynamic prior model and the data output by the multimodal perception layer; an adaptive cancellation core processing layer, comprising an adaptive reference signal construction module, a non-stationarity quantization evaluation module, and a filtering strategy dynamic switching module, used to construct the optimal reference signal based on the predicted statistical characteristics of wind noise and dynamically switch the adaptive filtering strategy; a closed-loop self-evolution and optimization layer, used to continuously optimize system parameters through offline training and online fine-tuning mechanisms guided by physical models; and a multi-device collaboration and task offloading layer, used to dynamically allocate adaptive cancellation calculation tasks based on the computing power of each device and the wind noise non-stationarity index in device networking scenarios.

[0189] Furthermore, the multi-device collaboration and task offloading layer includes: a device capability discovery module, used to monitor the computing resource status, power level, and communication bandwidth of each device in the network; a task segmentation module, used to dynamically divide the adaptive cancellation task into acquisition subtasks, statistical characteristic estimation subtasks, adaptive filtering calculation subtasks, and post-processing subtasks according to the current wind noise non-stationarity index and system processing frame rate requirements, and to allocate the computationally intensive adaptive filtering subtasks to the devices with the strongest computing capabilities; and a result fusion module, used to collect the processing results and their metadata returned by each device, and to perform weighted superposition using the weighted reciprocal of the residual noise energy and speech distortion index corresponding to each result as the fusion weight to generate the final output signal.

[0190] Furthermore, the nonstationarity quantification evaluation module is configured to: extract the temporal kurtosis coefficient, power spectrum fluctuation index, subband energy mutation rate, and turbulence intensity normalization index of each frame of acoustic signal; wherein, the temporal kurtosis coefficient reflects the intensity of the impulsive mutation component of the signal, the power spectrum fluctuation index is obtained by comparing the power spectrum of the current frame with that of the previous frame by the Bach distance, the subband energy mutation rate is calculated to calculate the inter-frame change rate of the low-frequency subband energy in which wind noise energy is concentrated, and the turbulence intensity normalization index is directly provided by the wind speed and wind direction sensing unit; the above four indicators are adaptively weighted and fused to output a comprehensive nonstationarity index, and when any indicator changes drastically in a continuous time period, its fusion weight is automatically increased.

[0191] Furthermore, the device includes: at least two microphones forming a multi-microphone array; a wind speed and direction sensor for acquiring environmental wind speed vectors; an inertial measurement unit for acquiring the device's motion attitude; a processor; and a memory storing executable instructions. When the processor executes the executable instructions, it performs the following steps: acquiring wind speed and direction sensing data, inertial measurement data, and acoustic signals from the multi-microphone array; estimating and predicting the time-varying statistical characteristics of wind noise in real time based on a priori model and data of wind field hydrodynamics, wherein the time-varying statistical characteristics include at least a predicted spatial coherence coefficient matrix and a predicted non-stationarity index; constructing an adaptive reference signal by spatially filtering the multi-microphone array signal according to the predicted spatial coherence coefficient matrix, wherein the optimization objective of the spatial filtering is to minimize the leakage energy of the target speech in the reference signal; dynamically selecting and switching the update strategy and core parameters of the adaptive filter from a strategy pool containing fast tracking strategies, steady-state low distortion strategies, and impact-resistant strategies according to the predicted non-stationarity index, and performing adaptive wind noise cancellation; and outputting the canceled target speech signal.

[0192] Furthermore, when the processor executes executable instructions, it also performs the following steps: when the device is in the same network as at least one other device, it monitors the computing resource status and power level of each device in the network; based on the current wind noise nonstationarity index, it offloads at least some computing subtasks in the adaptive cancellation task to other devices with sufficient computing resources and power levels higher than a preset threshold for execution; it receives the processing results returned by other devices, performs weighted fusion, and outputs the final voice signal; wherein, the decision criteria for task offloading include the level of the wind noise nonstationarity index, and when the nonstationarity index is higher than the threshold, computing-intensive tasks are preferentially offloaded.

[0193] Furthermore, the device is a true wireless stereo headset, a noise-canceling headset, a hearing aid, an in-vehicle voice interaction terminal, or a smart speaker; the wind speed and direction sensor and the inertial measurement unit are integrated inside the device cavity or near the microphone port; the spacing of each microphone in the multi-microphone array is optimized according to the device form so that the spatial coherence coefficient matrix of wind noise on the multi-microphone array meets the distinguishability condition.

[0194] Reference Figure 2 This application provides an adaptive cancellation system based on the statistical characteristics of sea surface wind noise, comprising:

[0195] The data acquisition unit is used to acquire multimodal sensor data, perform timestamp alignment and data fusion on the multimodal sensor data, and generate a marine environmental state sequence that is time-aligned with the underwater acoustic signal.

[0196] The wind noise statistics unit is used to predict the sea surface wind noise within a set time window in the future based on the ocean environment state sequence using a pre-established air-sea interface wind noise generation model and a wave bubble field coupling model, and output the predicted statistical characteristics of the sea surface wind noise.

[0197] The reference noise construction unit is used to construct an adaptive reference noise signal based on the predicted spatial coherence coefficient matrix in the predicted statistical characteristics and the mixed acoustic signal in the multimodal sensor data using a spatial filtering method.

[0198] The signal segmentation unit is used to segment the mixed acoustic signal into frequency bands according to the physical scale of sea surface wind noise, so as to obtain the scale components of the mixed acoustic signal in different frequency bands.

[0199] The wind noise cancellation unit is used to cancel the noise of the scale components of each frequency band according to the adaptive reference noise signal, and then weight and fuse the scale components of each frequency band after noise cancellation to generate a preliminary enhanced signal.

[0200] The signal correction unit is used to suppress residual noise and restore the fidelity of the target signal in the preliminary enhanced signal to obtain the final target signal.

[0201] It is understood that the content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0202] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this application are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0203] Furthermore, although this application is described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the functions and / or features may be integrated into a single physical system and / or software module, or one or more functions and / or features may be implemented in a separate physical system or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding this application. Rather, given the properties, functions, and internal relationships of the various functional modules in the system disclosed herein, the actual implementation of the module will be understood within the scope of ordinary skill of an engineer. Therefore, those skilled in the art can implement the application set forth in the claims using ordinary skill without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of this application, which is determined by the full scope of the appended claims and their equivalents.

[0204] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0205] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, system, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, system, or device). For the purposes of this specification, "computer-readable medium" can mean any system that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, system, or device.

[0206] More specific examples (a non-exhaustive list) of computer-readable media include: electrical connections with one or more wires (electronic systems), portable computer disk drives (magnetic systems), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic systems, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0207] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0208] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0209] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

[0210] The above is a detailed description of the preferred embodiments of this application, but this application is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. An adaptive cancellation method based on the statistical characteristics of sea surface wind noise, characterized in that, The method includes the following steps: Acquire multimodal sensor data, perform timestamp alignment and data fusion on the multimodal sensor data, and generate a marine environmental state sequence that is time-aligned with the underwater acoustic signal; Using a pre-established air-sea interface wind noise generation model and a wave bubble field coupling model, the sea surface wind noise within a set time window is predicted based on the marine environmental state sequence, and the predicted statistical characteristics of the sea surface wind noise are output. An adaptive reference noise signal is constructed using a spatial filtering method based on the predicted spatial coherence coefficient matrix in the predicted statistical characteristics and the mixed acoustic signal in the multimodal sensor data. The hybrid acoustic signal is divided into frequency bands according to the physical scale of sea surface wind noise to obtain the scale components of the hybrid acoustic signal in different frequency bands. The scale components of each frequency band are noise-cancelled according to the adaptive reference noise signal, and then the scale components of each frequency band after noise cancellation are weighted and fused to generate a preliminary enhanced signal. The preliminary enhanced signal is subjected to residual noise suppression and target signal fidelity restoration to obtain the final target signal.

2. The adaptive cancellation method based on the statistical characteristics of sea surface wind noise according to claim 1, characterized in that, The noise cancellation of the scale component rows of each frequency band based on the adaptive reference noise signal includes the following steps: Multidimensional features are extracted from the mixed acoustic signal, and the multidimensional features are fused to generate a comprehensive nonstationarity index; The target strategy is selected from the preset strategy pool based on the comprehensive nonstationarity index. The adaptive filter, driven by the target strategy, performs noise cancellation on the scale component rows of each frequency band based on the adaptive reference noise signal.

3. The adaptive cancellation method based on the statistical characteristics of sea surface wind noise according to claim 2, characterized in that, The extraction of multidimensional features from the mixed acoustic signal includes the following steps: The wave breaking event detection marker, temporal kurtosis coefficient, power spectrum fluctuation index, subband energy mutation rate, and sea state level normalization index are extracted from the hybrid acoustic signal as the multidimensional features. The process of fusing the multidimensional features to generate a comprehensive nonstationary index includes the following steps: The multidimensional features are adaptively weighted and fused to obtain the comprehensive nonstationarity index; wherein, when wave breaking events are continuously detected, the weight of the wave breaking event detection flag is increased.

4. The adaptive cancellation method based on the statistical characteristics of sea surface wind noise according to claim 2, characterized in that, The step of selecting a target strategy from a preset strategy pool based on the comprehensive non-stationarity index includes the following steps: Construct and maintain a set of sea state conditions including calm, moderate, and severe sea states; When the comprehensive nonstationarity index continuously exceeds a preset high threshold and the frequency of wave breaking events increases, the sea state is shifted to the severe sea state, and a strategy targeting wave breaking transient wind noise is selected from the preset strategy pool as the target strategy. When the comprehensive nonstationarity index remains below a preset low threshold, the sea state is shifted to calm sea state, and a strategy targeting steady-state background wind noise is selected from the preset strategy pool as the target strategy.

5. The adaptive cancellation method based on the statistical characteristics of sea surface wind noise according to claim 1, characterized in that, The method further includes the following steps: Based on the current wind speed vector and significant wave height, the power spectrum template of the wind pressure transmission component is obtained from the air-sea interface wind noise generation model. Based on the wave spectrum and wave steepness parameters, the power spectrum and transient pulse characteristics of the bubble radiation component are calculated using the wave-bubble field coupling model. The change in wind noise propagation path is determined based on the power spectrum template of the wind pressure transmission component and the power spectrum and transient pulse characteristics of the bubble radiation component. The platform's motion attitude data is used to perform noise spectrum attenuation compensation on the change in the wind noise propagation path in order to correct the predicted statistical characteristics.

6. The adaptive cancellation method based on the statistical characteristics of sea surface wind noise according to claim 1, characterized in that, The step of constructing an adaptive reference noise signal using a spatial filtering method based on the predicted spatial coherence coefficient matrix in the predicted statistical characteristics and the mixed acoustic signal in the multimodal sensor data includes the following steps: Using the dual-constraint minimum variance criterion, the predicted spatial coherence coefficient matrix is ​​used as the prior of wind noise spatial covariance, and the expected angle of arrival and array manifold of the target signal are used as the spatial features of the target signal to design spatial filtering coefficients. The mixed acoustic signal is filtered using the spatial filtering coefficients so that the resulting adaptive reference noise signal satisfies the following conditions: its power spectrum is consistent with the predicted sea surface wind noise power spectrum, and its spatial coherence with the target signal is minimized.

7. An adaptive cancellation method based on the statistical characteristics of sea surface wind noise according to any one of claims 1 to 6, characterized in that, The method further includes the following steps: During online operation, the system continuously caches the operation data within the most recent set time period. When the target signal distortion index is detected to be continuously deviating from the set conditions, the online fine-tuning mechanism is triggered. The online fine-tuning mechanism replays the cached runtime data in small batches and introduces a loss function including a marine physical consistency regularization term to incrementally update the parameters of the mapping model used to map from marine environmental state to adaptive filter parameters.

8. An adaptive cancellation system based on the statistical characteristics of sea surface wind noise, characterized in that, The system includes: The data acquisition unit is used to acquire multimodal sensor data, perform timestamp alignment and data fusion on the multimodal sensor data, and generate a marine environmental state sequence that is time-aligned with the underwater acoustic signal. The wind noise statistics unit is used to predict the sea surface wind noise within a set time window in the future based on the ocean environment state sequence using a pre-established air-sea interface wind noise generation model and a wave bubble field coupling model, and output the predicted statistical characteristics of the sea surface wind noise. The reference noise construction unit is used to construct an adaptive reference noise signal based on the predicted spatial coherence coefficient matrix in the predicted statistical characteristics and the mixed acoustic signal in the multimodal sensor data using a spatial filtering method. The signal segmentation unit is used to segment the mixed acoustic signal into frequency bands according to the physical scale of sea surface wind noise, so as to obtain the scale components of the mixed acoustic signal in different frequency bands. The wind noise cancellation unit is used to cancel the noise of the scale components of each frequency band according to the adaptive reference noise signal, and then weight and fuse the scale components of each frequency band after noise cancellation to generate a preliminary enhanced signal. The signal correction unit is used to suppress residual noise and restore the fidelity of the target signal in the preliminary enhanced signal to obtain the final target signal.

9. An electronic device, characterized in that, The electronic device includes a processor and a memory; The memory is used to store programs; The processor executes the program to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a program that is executed by a processor to implement the method as described in any one of claims 1 to 7.

Images