Intelligent identification model for internal solitary wave based on physical constraints and local deployment of submarine

Abstract

The application belongs to the technical field of marine observation and artificial intelligence, and in particular to an internal solitary wave intelligent identification model based on physical constraints and oriented to local deployment of a subsurface buoy, comprising a multi-modal data preprocessing module, a lightweight heterogeneous encoder, a flexible decoder cluster, a physical constraint embedding module, a model deployment and real-time inference module; the internal solitary wave can be autonomously identified and feature extracted locally on the subsurface buoy, and the dependence on the temperature-salinity chain and shore-based interpretation can be eliminated; the application has the advantages of strong real-time performance, low power consumption, anti-interference, physical credibility and the like, and is suitable for business monitoring and safety warning of internal solitary waves in key sea areas of China.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of marine observation and artificial intelligence, and in particular to an intelligent identification model for internal solitary waves based on physical constraints and oriented towards local deployment of underwater moorings. Background Technology

[0002] Internal solitary waves are strong nonlinear large-amplitude waves generated within the stratified ocean under the influence of tides, topography, and other factors. They are widely distributed in key sea areas such as the northern South China Sea, the East China Sea, the Yellow Sea, and the waters off Qingdao, with amplitudes ranging from tens to over two hundred meters. They are characterized by their sudden onset, rapid propagation, and significant destructive power. Internal solitary waves can cause underwater vehicles to experience severe depth drops, sonar signal distortion, and the formation of acoustic shadow zones. They are known as the "invisible killers" of underwater equipment, seriously threatening the safety of submarine navigation, marine engineering operations, and underwater communication and detection.

[0003] CN119622614B discloses a deep learning-based method for monitoring internal ocean waves, belonging to the field of internal ocean wave monitoring technology. This method involves collecting current velocity datasets from the observed sea area as the raw dataset, performing quality control on the raw dataset, distinguishing between normal and abnormal sequences, and dividing the data into training, validation, and test sets. An internal ocean wave anomaly monitoring model is constructed, anomaly scoring is optimized, and error parameters are reconstructed. The trained model is then used to monitor current velocity observation data in the sea area and identify internal ocean wave sequences. Internal wave dynamics feature analysis is performed on the sequences, and features are output. This existing technology, employing a deep learning model, can accurately capture the behavior patterns of normal current velocities and achieve anomaly detection of internal wave currents through error reconstruction, effectively distinguishing between normal current velocities and internal wave velocities, reducing misjudgments and missed judgments caused by traditional methods. However, it only uses current velocity time-series data and cannot detect the amplitude of internal solitary waves; it can only operate on shore-based systems and cannot be deployed locally on underwater mooring embedded platforms.

[0004] Currently, the identification and detection of internal solitary waves mainly rely on a combination of on-site observation, satellite remote sensing, and data post-processing. However, mainstream technologies and existing systems suffer from the following significant technical shortcomings:

[0005] First, temperature-salinity chains (TCS) and fixed-point CTDs are contact-based measurements that rely on inverting internal solitary wave amplitudes from TCS profiles. Their measurement range is limited, and they are highly susceptible to breakage and damage in strong ocean currents. Furthermore, equipment deployment and retrieval are costly, making continuous unattended observation year-round impossible. ADCP is only used for routine flow field statistics; the vertical disturbance information of the thermocline and acoustic scattering layer carried by its echo signals is not fully explored, and current velocity and internal wave amplitude cannot be acquired simultaneously, resulting in a waste of hardware observation resources.

[0006] Secondly, under the existing observation model, the massive amounts of raw acoustic and thermal current data collected by underwater gliders need to be transmitted back to the land-based center via underwater acoustic and satellite communications, and then rely on researchers for manual interpretation, signal extraction, and event determination. The entire process suffers from problems such as communication transmission delays, data packet loss, insufficient bandwidth, and long manual processing cycles; while the duration of a single transit of an internal solitary wave is usually less than 30 minutes, the delayed processing mode cannot meet the real-time early warning needs of sudden marine dynamic events.

[0007] Third, traditional identification methods in the industry mostly rely on fixed empirical thresholds for temperature variability and flow velocity shear to trigger judgments, failing to distinguish between interference signals such as mesoscale eddies, sub-mesoscale eddy filaments, ocean fronts, intra-diurnal tides, and bottom friction disturbances. This leads to frequent false alarms in nearshore waters with complex stratification and strong currents. Small-amplitude weak internal isolated waves are easily masked by background noise, resulting in a large number of missed alarms.

[0008] Fourth, although artificial intelligence technology has been gradually applied to the study of internal ocean waves, existing models are mainly geared towards tasks such as satellite SAR image recognition, internal wave parameter inversion, and data super-resolution reconstruction. They generally suffer from problems such as complex network structures, large parameter sets, high computational requirements, and inability to be embedded in edge computing platforms such as underwater moorings. Furthermore, existing AI models are mostly purely data-driven and do not deeply integrate prior knowledge such as the physical propagation laws of internal waves, frequency domain characteristics, and vertical structural features. This results in poor physical interpretability, weak generalization ability, and a tendency for non-physical distortions in the recognition results.

[0009] Fifth, as the core platform for long-term observation of internal solitary waves, underwater buoys face practical constraints such as limited computing power, energy shortages, unreliable communication links, and susceptibility to damage from typhoons and ships. Currently, there is no intelligent identification model for internal solitary waves that can be deployed locally on underwater buoys, perform real-time reasoning, and meet physical constraints. Furthermore, an integrated technical system of "in-situ sensing, local identification, and real-time early warning" has not yet been formed, which seriously restricts the development of internal solitary wave early warning systems towards long-term, intelligent, and operational applications.

[0010] In summary, existing internal solitary wave identification and detection technologies have significant shortcomings in terms of observation methods, identification modes, lightweight models, physical constraint fusion, and in-situ deployment of moorings, failing to meet the critical need for real-time, accurate, autonomous, and reliable identification of solitary waves in key coastal waters of my country. Therefore, developing an intelligent internal solitary wave identification model that is physically constrained, lightweight, locally deployable on mooring platforms, resistant to complex ocean process interference, and capable of low-power real-time inference has become a key technical problem urgently needing to be solved in this field. Summary of the Invention

[0011] To address the shortcomings of the existing technologies, a smart identification model for internal solitary waves based on physical constraints and designed for local deployment on underwater moorings is provided. This model can autonomously identify and extract features of internal solitary waves locally on the moorings, eliminating the dependence on temperature-salinity chains and shore-based interpretation. It has advantages such as strong real-time performance, low power consumption, anti-interference, and physical reliability, and is suitable for operational monitoring and safety early warning of internal solitary waves in key sea areas of China's coastal waters.

[0012] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is an intelligent identification model for internal solitary waves based on physical constraints and oriented to local deployment of underwater gliders, including a multimodal data preprocessing module, a lightweight heterogeneous encoder, a flexible decoder cluster, a physical constraint embedding module, and a model deployment and real-time inference module.

[0013] The multimodal data preprocessing module is used to perform super-resolution reconstruction, denoising and quality control on the acoustic scattering intensity, flow velocity and temperature-depth frequency conversion observation data of the underwater mooring ADCP, and to complete the quantitative physical conversion from the original ADCP echo intensity to the acoustic scattering intensity.

[0014] The lightweight heterogeneous encoder is used to differentially encode multimodal observation data and uniformly output a high-dimensional vector representation for reuse by downstream decoders.

[0015] The flexible decoder cluster is used to identify internal solitary wave occurrence events and extract internal solitary wave feature parameters;

[0016] The physical constraint embedding module uses a combination of hard and soft constraints to ensure that the model output conforms to the physical laws of ocean fluids.

[0017] The model deployment and real-time inference module is used to solidify the lightweight model onto the buoy embedded platform to achieve low-power monitoring and event-driven local real-time inference.

[0018] Data collected by the underwater glider ADCP and temperature-depth sensor is sent to the multimodal data preprocessing module to complete acoustic conversion and multi-level quality control. After processing, the data is input into the lightweight heterogeneous encoder to fuse features. The feature vector is sent to the flexible decoder cluster to output the probability and dynamic parameters of isolated wave events. The parameters are entered into the physical constraint embedding module to correct non-physical results. Finally, the model deployment and real-time inference module completes embedded solidification, frequency conversion sampling and local early warning.

[0019] The aforementioned intelligent identification model for internal solitary waves based on physical constraints and oriented towards local deployment of underwater gliders, in the multimodal data preprocessing module, uses the following formula to calculate the conversion of the original ADCP echo intensity into acoustic scattering intensity:

[0020] ;

[0021] All constants and parameters in the formula are taken from the original factory calibration manual of TRDI Workhorse Long Range 75kHz four-beam ADCP. The volumetric acoustic scattering intensity in water is expressed in dB. These are system calibration constants; The measured seawater temperature for the transducer is expressed in °C. D is the vertical distance from the probe to the scatterer. The standard beam angle for the equipment; This refers to the transducer's transmit power. For pulse distance loss, These are the factory-calibrated parameters for the equipment; The sound absorption coefficient of seawater; E is the beam conversion factor; E is the original echo count value of ADCP. This is the instrument's inherent noise floor count value.

[0022] The aforementioned intelligent identification model for internal solitary waves based on physical constraints and oriented towards local deployment of underwater buoys, wherein the multimodal data preprocessing module is equipped with a pruned and compressed lightweight generative adversarial network to achieve data denoising and super-resolution reconstruction, and is accompanied by an eight-step standardized quality control process: basic validity check, single-beam independent acoustic scattering calculation, beam median residual verification, opposing beam consistency verification, interlayer velocity mutation removal, spatiotemporal solitary point removal, missing data interpolation, and comprehensive quality control scoring;

[0023] The comprehensive quality control scoring formula is:

[0024] ;

[0025] in, For beam consistency scoring, Scoring the correlation coefficient To score attitude stability, For the completeness score of missing tests, Scoring is given for the rate of error.

[0026] when Data deemed high-quality can be used for reasoning throughout the entire process; For low-confidence available data, the inference results are downgraded; This is invalid data; invalid data will not be involved in event triggering.

[0027] The aforementioned intelligent identification model for internal solitary waves, based on physical constraints and designed for local deployment of underwater gliders, includes a lightweight heterogeneous encoder comprising three types of feature extraction units in parallel:

[0028] The temporal feature encoding unit is a lightweight LSTM network, with hidden unit values ​​ranging from 16 to 64;

[0029] The physical field feature encoding unit is a lightweight Transformer with 1 to 2 layers and 1 to 4 attention heads;

[0030] The spatial structure encoding unit is a lightweight depthwise separable CNN with 8 to 32 convolutional channels;

[0031] The output vectors of the three types of units are concatenated and fused into a unified 64-dimensional high-dimensional feature vector, which is then encoded once and used by two types of decoders.

[0032] The aforementioned intelligent identification model for internal solitary waves based on physical constraints and oriented towards local deployment of underwater gliders includes an internal solitary wave generation discrimination decoder and an internal solitary wave feature extraction decoder in the flexible decoder cluster.

[0033] The internal solitary wave occurrence discrimination decoder uses a lightweight SVM classifier to output the probability of internal solitary wave events. ;

[0034] The internal solitary wave feature extraction decoder is a bottleneck-lightweight MLP, outputting a parameter vector. A is the amplitude, C is the propagation speed, T is the duration, and L is the horizontal scale. The direction of propagation is determined by the output, which is then fed into a physical constraint verification layer for forced correction.

[0035] The aforementioned intelligent identification model for internal solitary waves based on physical constraints and oriented towards local deployment of underwater gliders includes a physical constraint embedding module comprising hard constraints at the output end and soft constraints at the training end.

[0036] (1) The hard constraints of inference include six mandatory verification rules:

[0037] ①Amplitude upper limit H represents the deployment depth; if the model outputs A≤0, it is directly considered an invalid event.

[0038] ② Phase velocity of the first-order internal wave theory for two-layer fluids:

[0039] ;

[0040] Reduced gravity , For the theoretical first-order phase velocity, , These represent the thicknesses of the upper and lower seawater layers, respectively. , The density of the upper and lower layers, For reference density, The model output wave velocity must satisfy the following conditions: gravitational acceleration; ;

[0041] ③ Amplitude inversion from scattering layer , The reference amplitude of the internal solitary wave is directly calculated from the vertical displacement of the acoustic scattering layer; constraint Weighted adjustment when not satisfied Weight Take a value of 0.5~0.8;

[0042] ④ Kinematically constrained residuals , The relative residuals between the theoretical horizontal scale and the model prediction scale of the internal wave; residuals Use directly Coverage correction level scale ;

[0043] ⑤ Polarity consistency constraint: the sinking of the scattering layer corresponds to the sinking type of internal solitary wave, and the upward layer corresponds to the upward type. If the amplitude sign and the flow velocity rotation structure do not match, the confidence of the event will be reduced or it will be directly eliminated.

[0044] ⑥ Fluid incompressibility continuity constraint: Constructing velocity divergence residuals If the residual exceeds three times the standard deviation of the background residual, the prediction result is determined to have non-physical bias and the confidence level of the event is reduced.

[0045] (2) The weights of the total loss function with soft constraints are globally fixed and used uniformly for training across the entire sea area; they cannot be adjusted.

[0046] The soft constraint introduces a physical guiding term into the loss function during the model training phase. The total loss function is:

[0047] ;

[0048] in, The event binary cross-entropy classification loss is used. For amplitude and wave velocity regression Huber loss, For amplitude consistency loss, The loss is due to the phase velocity range constraint. For time-series smoothing loss, For the incompressible residual loss of flow velocity, This is for beam consistency loss.

[0049] The aforementioned intelligent identification model for internal solitary waves based on physical constraints and oriented towards local deployment of underwater gliders consists of a model deployment and real-time inference module divided into five stages: offline training, model lightweighting, model solidification, local loading of underwater gliders, and event-driven inference.

[0050] Lightweight model design techniques include network structure lightweighting, 20%~60% structured pruning, knowledge distillation, INT8 fixed-point quantization, and operator fusion. Pruning adaptive rules: fixed 40% pruning for deep water scenarios ≥500m; fixed 60% pruning for shallow sea scenarios with strong tidal interference <100m; and linear adjustment within the 40%~60% range for medium-depth sea areas 100~500m, balancing accuracy and storage usage.

[0051] Floating-point weighted INT8 quantization formula:

[0052] ;

[0053] In the formula, For floating-point weights, Quantize integer weights, For quantitative scale, Zero-point offset;

[0054] Quantization-based perceptual training ensures stable embedded inference; operator fusion is used to derive general-purpose embedded models for ONNX, TFLiteMicro, and TVM.

[0055] When the model is solidified, the model weights, standardized parameters, complete set of acoustic conversion parameters, and physical constraint thresholds are stored in the buoy's non-volatile memory. The device automatically performs CRC verification upon startup, and automatically switches to the backup model if the verification fails.

[0056] The aforementioned intelligent identification model for internal solitary waves, based on physical constraints and designed for local deployment of underwater gliders, employs a multi-physical quantity joint discrimination mechanism for event triggering, calculating the background disturbance scale using the median absolute deviation (MAD).

[0057] ;

[0058] In the formula, The standard deviation of real-time background disturbance. This is a long-term background observation sequence. The data represents the current time-series observations, with MAD representing the median absolute deviation; the background window duration is 6–24 h, and the event judgment window is 20–40 min.

[0059] Fixed event activation threshold Hysteresis exit threshold ;

[0060] Sampling is divided into three working modes: low-frequency monitoring 10 min / time, candidate confirmation 2-5 min / time, and high-frequency inference 1-2 min / time; the longest duration of a single high-frequency sampling is 3 hours in deep water and 2 hours in shallow water.

[0061] Switching high-frequency sampling criteria: Comprehensive score meets the standard. At least two of the four physical anomalies must meet any one of the following conditions, and the quality control score must be [missing information]. ;

[0062] High-risk warning trigger condition: Model predicted amplitude or maximum flow velocity disturbance or If any condition is met, a high-priority early warning will be immediately sent via underwater acoustic / satellite communication.

[0063] The beneficial effects of the intelligent identification model for internal solitary waves based on physical constraints and oriented towards local deployment of underwater moorings in this invention are that it adopts the standard ADCP acoustic scattering quantitative conversion model, and can simultaneously obtain the current velocity and internal solitary wave amplitude using only conventional acoustic profilers, without the need to deploy and maintain expensive and easily damaged temperature-salinity chains and CTDs; the acoustic non-contact measurement has a large range and is resistant to strong current impacts, making it suitable for long-term unattended observation in deep water, and significantly reducing the investment in marine observation engineering.

[0064] Data preprocessing, feature encoding, event discrimination, and parameter verification are all completed in the embedded hardware of the underwater buoy, eliminating the need to transmit massive amounts of raw observation data back, avoiding problems such as insufficient underwater communication bandwidth, data packet loss, and transmission lag. Identification and early warning can be completed within an isolated 30-minute transit window, meeting real-time business needs.

[0065] The model training phase introduces a fluid dynamics loss term, and the inference phase forces the verification of amplitude, wave velocity, kinematics, and continuous physical boundaries. This effectively suppresses interference signals such as internal tides, mesoscale eddies, fish schools, and bubbles, solving the pain points of non-physical distortion and weak generalization ability in pure data-driven model outputs, and significantly reducing the false alarm rate of nearshore environment.

[0066] Through network simplification, channel pruning, knowledge distillation, and INT8 quantization compression, the number of model parameters and the computational power required for inference are significantly reduced.

[0067] The design incorporates a complete eight-step quality control process for beam, flow velocity, and spatiotemporal noise, with tiered data confidence levels. Low-quality data is excluded from model inference, reducing the probability of misidentification caused by equipment noise and ocean interference from the data source.

[0068] It is natively compatible with TRDI75kHz ADCP and mainstream underwater mooring systems. By simply adjusting the upper limit of amplitude corresponding to water depth, internal wave phase velocity constraint, and sampling period, it can be adapted to different sea areas such as the deep water of the South China Sea and the shallow water of the Yellow Sea. It can be widely used in various scenarios such as submarine protection, offshore platforms, underwater communication, and transparent ocean monitoring. Attached Figure Description

[0069] Figure 1 This is a flowchart of the process of the present invention;

[0070] Figure 2This is a time series diagram of acoustic backscattering intensity depth-time profile obtained from long-term in-situ observations of the nearshore mooring ADCP; the vertical axis represents seawater depth (unit: m), the horizontal axis represents the consecutive observation dates, and the color scale on the right represents the acoustic backscattering intensity (unit: dB).

[0071] Figure 3 Comparison of acoustic scattering-current velocity dual-mode profiles and automatic internal wave identification markers from nearshore mooring ADCP synchronous observations; (a) Acoustic scattering intensity Sv detected by ADCP in the northern continental slope of the South China Sea on April 13, 2024. (b) Horizontal current velocity observed by ADCP. The black lines in the figure represent isotherms at 6°C, 10°C, and 19°C, and the red dots represent identified internal waves. The results show that the disturbance of the acoustic scattering layer is consistent with the isotherm fluctuations, indicating that the internal wave amplitude inversion based on the acoustic scattering layer fluctuations from ADCP is reliable. Detailed Implementation

[0072] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0073] like Figure 1-3 As shown, the intelligent identification model for internal solitary waves based on physical constraints and oriented towards local deployment of underwater gliders includes a multimodal data preprocessing module, a lightweight heterogeneous encoder, a flexible decoder cluster, a physical constraint embedding module, and a model deployment and real-time inference module.

[0074] The multimodal data preprocessing module is used to perform super-resolution reconstruction, denoising and quality control on the acoustic scattering intensity, flow velocity and temperature-depth frequency conversion observation data of the underwater mooring ADCP, and to complete the quantitative physical conversion from the original ADCP echo intensity to the acoustic scattering intensity.

[0075] This module is equipped with a lightweight generative adversarial network that has been optimized through structured pruning and quantized compression. It completes noise reduction and time-depth dimension super-resolution reconstruction of the original ADCP observation data. The core implementation content is the quantitative conversion of the original ADCP echo intensity to the acoustic scattering intensity of the water body, and simultaneously performs eight standardized multi-level data quality control steps.

[0076] The formula for converting the original echo intensity to acoustic scattering intensity using ADCP is as follows:

[0077] ;

[0078] in, The original echo intensity, The intensity of acoustic scattering; It is a system constant. It refers to the temperature at the transducer probe. It is the distance from the transducer probe to the scatterer along the emitted wavenumber, and D is the perpendicular distance from the transducer probe to the scatterer. It is the beam angle of the ADCP. The transducer's transmit power. Calculated from the distance of the transmitted pulse, Provided by the equipment manufacturer. It is the sound absorption coefficient of water. It is the beam scaling factor. It is the background signal reception strength.

[0079] It uniformly corrects four types of systematic errors: temperature, spherical diffusion, water absorption, and equipment background noise, and outputs a physically meaningful acoustic scattering field for inverting thermocline displacement and calculating internal solitary wave amplitude. The equipment compatible model is TRDIWorkhorseLongRange75kHzADCP.

[0080] The eight-step quality control parameters are divided into two categories: fixed thresholds applicable to the entire sea area and adjustable thresholds for different sea areas. Fixed thresholds: error velocity 0.05m / s, current velocity change 0.3m / s, beam residual 6dB, instrument tilt angle 15°; Adjustable thresholds: 30 times the standard deviation of background current velocity, maximum number of interpolation layers, event window duration. The deep and shallow sea implementation examples can be configured separately.

[0081] Eight-step standardized quality control implementation process:

[0082] (1) Basic validity point-by-point verification: Time-by-time, depth layer, and beam screening; any case of low echo, insufficient correlation coefficient, effective sampling ratio <50%, tilt angle >15°, or exceeding range is marked as invalid; layer error rate judgment conditions This satisfies the condition of directly removing the sampling point.

[0083] (2) Calculation of independent acoustic scattering of a single beam: Calculate separately (b beam number, i depth, t time), the original echo is not averaged in advance, and beam differences are retained for consistency verification.

[0084] (3) Beam center residual verification: Calculate the median value of acoustic scattering of all beams at the same depth. Single-beam residual The residual exceeding the threshold marks the abnormal beam.

[0085] (4) Judgment of the consistency of the opposing beams: the difference between the two sets of opposing beams of the four beams , If the difference is greater than 6dB and is accompanied by attitude / flow rate anomalies, all data at that depth will be deleted.

[0086] (5) Elimination of abrupt changes in interlayer velocity: velocity difference between adjacent depths ,satisfy Furthermore, if only a single layer anomaly is detected, isolated noise is determined and deleted.

[0087] (6) Spatiotemporal isolated point filtering: Construct a two-dimensional spatiotemporal matrix of acoustic scattering, flow velocity, temperature and depth, and directly remove single points with no continuous anomalies in the 3×3 neighborhood.

[0088] (7) Data interpolation for missing data: For continuous missing data of ≤2 sampling / depth units, spatiotemporal bilinear interpolation and median filtering are used to fill the missing data; for long missing data, only low quality data is marked, and interpolation to generate false internal wave signals is prohibited.

[0089] (8) Calculation of comprehensive quality control score:

[0090] ;

[0091] Beam score Correlation coefficient score Flow rate error score Posture rating Missing test completeness score; graded execution standard: Q≥0.8 high-quality data is used throughout the entire process; 0.6≤Q<0.8 low-confidence data inference results are downgraded; Q<0.6 invalid data does not participate in event triggering and model calculation.

[0092] The lightweight heterogeneous encoder is used to differentially encode multimodal observation data and uniformly output a high-dimensional vector representation for reuse by downstream decoders.

[0093] The encoder is divided into three types of parallel feature extraction units, and the output features are fused into a unified high-dimensional vector, enabling one-time encoding to be used for two types of decoders: discrimination and parameter extraction, thereby reducing the power consumption of repeated calculations.

[0094] 1. Temporal feature encoding unit: Lightweight LSTM, hidden unit configuration range 16~64, extracting temporal fluctuation features of flow velocity and acoustic scattering;

[0095] 2. Physical field feature encoding unit: Lightweight Transformer, 1~2 layers, 1~4 attention heads, introduce physical attention to automatically focus on the layered regions of the thermocline, improve the physical interpretability of the model;

[0096] 3. Spatial structure encoding unit: Lightweight depthwise separable CNN with 8~32 convolutional channels to extract vertical spatial correlation features of flow velocity at different depths.

[0097] The output vectors of the three types of units are concatenated and fused before being sent to the downstream flexible decoder cluster.

[0098] The flexible decoder cluster is used to identify internal solitary wave occurrence events and extract internal solitary wave feature parameters;

[0099] The implementation of a flexible decoder cluster includes:

[0100] The cluster contains two lightweight decoding branches that simultaneously perform event discrimination and internal wave parameter regression:

[0101] 1. Internal solitary wave generation discriminant decoder: Lightweight linear SVM classifier, input fused features, output event probability. ;

[0102] 2. Internal solitary wave feature extraction decoder: Lightweight MLP with bottleneck structure, output parameter vector

[0103] A. Amplitude, C. Propagation speed, T. Duration, L. Horizontal scale The propagation direction is determined, and the output is then connected to the physical constraint verification layer to force correction of non-physical prediction values.

[0104] The physical constraint embedding module uses a combination of hard and soft constraints to ensure that the model output conforms to the physical laws of ocean fluids.

[0105] The implementation of the physical constraint embedding module includes:

[0106] By combining hard constraints at the inference end and soft constraints at the training end, the problem of pure data-driven model output violating fluid dynamics is eliminated.

[0107] (1) Hard constraints at the output end (during the inference phase, the verification layer does not participate in network training)

[0108] Amplitude constraint: Upper limit of amplitude H represents the depth at which the buoy is deployed; predictions of A≤0 are directly deemed invalid. Revised to And lower the confidence level.

[0109] Propagation speed constraint: Phase velocity of the first-order theory of internal waves The predicted wave velocity must meet the following conditions. Approximate formula for two-layer fluid:

[0110] , , , The thickness of the upper and lower layers; , Density of the upper and lower layers of seawater; Reference density, Gravitational acceleration.

[0111] Amplitude matching constraint: Amplitude inversion in the scattering layer , must meet The weighted adjustment is not satisfied. , Take a value of 0.5 to 0.8.

[0112] Kinematic constraints: theory residual Use directly Coverage at horizontal scale.

[0113] Polarity constraints: a sinking scattering layer corresponds to a concave inner wave, and an upward scattering layer corresponds to a convex inner wave. If the amplitude sign and flow velocity rotation structure do not match, the confidence level is reduced or the event is eliminated.

[0114] Incompressible constraints: Constructing velocity divergence residuals If the value exceeds three times the standard deviation of the background, it is considered a non-physical output, and the event level is reduced.

[0115] (2) Soft constraints at the training end (the total loss function is embedded with a physical penalty term)

[0116] Total loss calculation formula:

[0117] ;

[0118] Meaning of implementation in sections:

[0119] Classification loss , Manually labeled with authentic tags;

[0120] Regression loss This helps to suppress large errors in amplitude and wave velocity regression.

[0121] Amplitude Consistency Loss ;

[0122] Phase velocity loss Punishment for exceeding the range of wave speed;

[0123] Temporal smoothing loss Suppress parameter timing jumps;

[0124] Incompressible residual loss ;

[0125] Beam consistency loss This reduces beam noise interference.

[0126] The model deployment and real-time inference module is used to solidify the lightweight model into the buoy embedded platform to achieve low-power monitoring and event-driven local real-time inference.

[0127] The implementation of the model deployment and real-time inference module includes:

[0128] The system is divided into five execution stages: shore-based offline training, network lightweighting, model solidification, buoy loading, and event-driven inference. The hardware adopts a time-sharing architecture with a low-power MCU and an NPU.

[0129] 1. Implementation of Offline Shore-Based Training

[0130] A dataset was constructed by collecting ADCP acoustic, 3D flow velocity, temperature depth, and attitude time-series data, combined with manually labeled events; the original data standardization formula is as follows:

[0131] ;

[0132] X represents the original observation value. training set mean The standard deviation and standardized parameters are packaged and stored in the buoy memory along with the model.

[0133] 2. Model Lightweight Implementation Process

[0134] Network simplification: Limit the size of LSTM, Transformer, CNN, and MLP, and use bottleneck structures and depthwise separable convolutions;

[0135] Structured pruning: Pruning ratio of 20%~60%, deleting low-contribution channels and neurons, and fine-tuning the recovery accuracy after pruning;

[0136] Knowledge distillation: High-precision teacher model outputs soft tags, distillation temperature Training lightweight student networks;

[0137] INT8 fixed-point quantization, quantization formula:

[0138] ;

[0139] Floating-point weights Quantize integers, Quantitative scale Zero-point offset; quantitative perception training is used to ensure stable underwater inference.

[0140] Operator fusion: convolution, BN, activation merging, outputting embedded general-purpose models such as TFLiteMicro and ONNX.

[0141] 3. Model solidification operation

[0142] The network structure, quantization weights, acoustic scattering conversion constants, quality control thresholds, all physical constraint parameters, and CRC checksums are written into the non-volatile flash memory of the submersible; the device automatically verifies upon power-up, and automatically switches to the backup model if the verification fails.

[0143] 4. Submersible buoy frequency conversion monitoring and event reasoning mechanism

[0144] (1) Implementation of three sampling modes:

[0145] Low-frequency monitoring: Once every 10 minutes, only the MCU runs, the NPU sleeps, and only basic quality control and background updates are performed;

[0146] Candidate confirmation: If the pre-score meets the standard, the time should be shortened to 2-5 minutes, and the disturbance should be observed for 2-3 consecutive cycles;

[0147] High-frequency reasoning: The sampling interval is switched to 1-2 minutes, and the NPU is woken up to execute the complete model inference.

[0148] (2) Hysteresis threshold control: enable threshold , =0.45, unchanged across the entire sea area; exit hysteresis threshold The frequency can only be switched back to low frequency after several consecutive cycles below the threshold and the disturbance disappears for 20 to 30 minutes, in order to avoid frequent start-stop and power consumption.

[0149] (3) High-frequency duration limit: The maximum duration for a single high-frequency event is 2-4 hours. If the duration exceeds this limit and the confidence level is <0.7, the event will enter the buffer recovery phase. Amplitude >50m, flow velocity disturbance >0.5m / s, High-risk events will be continuously alerted without time limits.

[0150] 5. Implementation triggered by multiple physical quantity events

[0151] Robust estimation formula for background disturbance:

[0152] ;

[0153] MAD is the median absolute deviation, with a background window of 6-24 hours and an event judgment window preferably of 30 minutes; it combines five indicators: acoustic scattering, scattering layer displacement, flow velocity, flow velocity acceleration, and vertical coherence; switching high-frequency judgment conditions: the comprehensive score meets the standard or Or at least two physical anomalies must be met simultaneously, and the quality control Q ≥ 0.6.

[0154] Example 1

[0155] Case Study of Deep-Water Mooring Implementation on the Dongsha Slope in the Northern South China Sea

[0156] 1. Deployment environment and hardware configuration

[0157] Deployment depth 1120m (water depth > 500m, pruning ratio fixed at 40%); 75kHz ADCP layered at 5m intervals, maximum range 1000m; 200Ah deep-sea lithium battery, designed to be self-sufficient for 12 months; equipped with underwater acoustic and Iridium satellite dual communication, real-time transmission of high-risk events back to the shore-based early warning platform.

[0158] 2. Customized parameters for shore-based training

[0159] Dataset: 3-year time-series data from the South China Sea, with 1267 annotated isolated wave samples; encoder: LSTM with 32 hidden units, Transformer single-layer 2-head, CNN with 8 channels; pruning ratio: 40%, quantized model size: 1.2MB; sea area constraint parameters: Theoretical phase velocity The wave velocity range is 0.375~1.5m / s. .

[0160] 3. Submersible buoy operation configuration

[0161] Low frequency 10min, candidate 3min, high frequency 2min, single high frequency up to 3h; automatic satellite alarm for amplitude > 50m.

[0162] 4. On-site implementation results

[0163] The sea area is constantly disturbed by mesoscale eddies and deep-water internal tides. The false alarm rate of the traditional threshold method is 32.7%; the false alarm rate of this model is reduced to 4.1%; after 11 months of continuous and stable operation, the remaining power is sufficient; the acoustic inversion amplitude and the measured error of the temperature-salt chain are within ±4m, which meets the operational monitoring standards.

[0164] Example 2: Implementation Case of Shallow Shelf Mooring in the Central Yellow Sea

[0165] Deployment environment and hardware configuration

[0166] Deployed at a water depth of 86m (water depth < 100m, pruning ratio fixed at 60%); ADCP layer spacing of 2m; 100Ah lithium battery, self-sustaining observation for 6 months; equipped only with local underwater acoustic storage, no satellite communication, event data will be analyzed after equipment recovery.

[0167] Custom parameters for shore-based training

[0168] Yellow Sea tidal and inland tide dataset, annotated with 583 events; Lightweight upgrade: LSTM 16 units, CNN 4 channels, pruning ratio increased to 60%, quantized model 0.68MB; Sea area constraints: Theoretical phase velocity The range is 0.16~0.64 m / s. Relax the quality control threshold for sudden flow velocity changes to adapt to strong current noise.

[0169] Submarine buoy operation configuration

[0170] Low-frequency events are marked for 10 minutes, candidate events for 5 minutes, and high-frequency events for 2 minutes, with a maximum duration of 2 hours for a single high-frequency event; high-risk events are marked locally only, with no remote push notifications.

[0171] On-site implementation effect

[0172] Tidal currents, schools of fish, and bubbles cause significant interference, and traditional methods have an identification efficiency of less than 60%. This model has an identification accuracy of 93.5%, has been running continuously for 5.8 months, and has an inversion error of ±3m for amplitudes of 5~45m in shallow seas. It does not require a temperature-salinity chain and significantly reduces the cost of array observation deployment.

[0173] A comparative summary of the two implementation examples

[0174] Structured pruning ratio 40% 60% Shallow waters experience high tidal noise, leading to increased compression and reduced computing power; deep water stratification clearly preserves accuracy. Model storage volume 1.2MB 0.68MB Suitable for shallow-sea small-capacity flash memory underwater moorings Maximum allowable amplitude 250m 68.8m <![CDATA[Follow A max =min(0.8H, 250m) water depth constraint]]> Theoretical wave speed range 0.375~1.5m / s 0.16~0.64m / s <![CDATA[There are differences in the calculation of C0 using the two-layer fluid formula due to the different thicknesses and density differences of deep and shallow seawater layers]]> Maximum duration of a single high-frequency event 3h 2h Solitary waves in shallow seas have shorter durations, reducing power consumption. Communication mode Real-time early warning from satellite and underwater acoustics Local storage recycling and reading Deep water requires remote early warning, while nearshore areas can utilize equipment recovery and analysis. Flow rate mutation quality control threshold 0.3m (General Fixed) 0.2m (adjustable for sea areas) The Yellow Sea has a high tidal current velocity, so relaxing the threshold can help avoid accidentally deleting valid signals. Self-sustaining observation duration 12 months 6 months Differences in lithium battery capacity configuration

[0175] General Complete Operation Flow: The underwater glider, equipped with an ADCP and a temperature and depth sensor, acquires raw observation data via frequency conversion. A preprocessing module then performs acoustic scattering conversion and eight-step full-dimensional quality control. The data is then fed into a lightweight heterogeneous encoder to extract unified fusion features. These features are input into a flexible decoder, which outputs event probabilities and a complete set of internal wave parameters. A physical constraint module then corrects non-physical prediction results through both hardware and software verification. A low-power MCU triggers the data based on a fixed trigger threshold. Determine whether to wake up the NPU, and transmit the identified high-risk events back to shore or local storage for archiving via the communication link.

[0176] The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent identification model for internal solitary waves based on physical constraints and oriented towards local deployment of underwater gliders, characterized by: It includes a multimodal data preprocessing module, a lightweight heterogeneous encoder, a flexible decoder cluster, a physical constraint embedding module, and a model deployment and real-time inference module; The multimodal data preprocessing module is used to perform super-resolution reconstruction, denoising and quality control on the acoustic scattering intensity, flow velocity and temperature-depth frequency conversion observation data of the underwater mooring ADCP, and to complete the quantitative physical conversion from the original ADCP echo intensity to the acoustic scattering intensity. The lightweight heterogeneous encoder is used to differentially encode multimodal observation data and uniformly output a high-dimensional vector representation for reuse by downstream decoders. The flexible decoder cluster is used to identify internal solitary wave occurrence events and extract internal solitary wave feature parameters; The physical constraint embedding module uses a combination of hard and soft constraints to ensure that the model output conforms to the physical laws of ocean fluids. The model deployment and real-time inference module is used to solidify the lightweight model onto the buoy embedded platform to achieve low-power monitoring and event-driven local real-time inference. Data collected by the underwater glider ADCP and temperature-depth sensor is sent to the multimodal data preprocessing module to complete acoustic conversion and multi-level quality control. After processing, the data is input into the lightweight heterogeneous encoder to fuse features. The feature vector is sent to the flexible decoder cluster to output the probability and dynamic parameters of isolated wave events. The parameters are entered into the physical constraint embedding module to correct non-physical results. Finally, the model deployment and real-time inference module completes embedded solidification, frequency conversion sampling and local early warning.

2. The intelligent identification model for internal solitary waves based on physical constraints and oriented towards local deployment of underwater gliders as described in claim 1, characterized in that: In the multimodal data preprocessing module, the formula for converting the original ADCP echo intensity into acoustic scattering intensity is as follows: ； All constants and parameters in the formula are taken from the original factory calibration manual of TRDI Workhorse Long Range 75kHz four-beam ADCP. The volumetric acoustic scattering intensity in water is expressed in dB. These are system calibration constants; The measured seawater temperature for the transducer is expressed in °C. D is the vertical distance from the probe to the scatterer. The standard beam angle for the equipment; This refers to the transducer's transmit power. For pulse distance loss, These are the factory-calibrated parameters for the equipment; The sound absorption coefficient of seawater; E is the beam conversion factor; E is the original echo count value of ADCP. This is the instrument's inherent noise floor count value.

3. The intelligent identification model for internal solitary waves based on physical constraints and oriented towards local deployment of underwater gliders as described in claim 2, is characterized in that... The multimodal data preprocessing module is equipped with a pruned and compressed lightweight generative adversarial network to achieve data denoising and super-resolution reconstruction, and is equipped with an eight-step standardized quality control process: basic validity check, single beam independent acoustic scattering calculation, beam median residual verification, opposing beam consistency verification, interlayer velocity mutation removal, spatiotemporal isolated point removal, missing data interpolation, and comprehensive quality control scoring. The comprehensive quality control scoring formula is: ； in, For beam consistency scoring, Scoring the correlation coefficient, To score attitude stability, For the completeness score of missing tests, Score the error rate; when Data deemed high-quality can be used for reasoning throughout the entire process; For low-confidence available data, the inference results are downgraded; This is invalid data; invalid data will not be involved in event triggering.

4. The intelligent identification model for internal solitary waves based on physical constraints and oriented towards local deployment of underwater gliders as described in claim 3, is characterized in that, The lightweight heterogeneous encoder comprises three types of feature extraction units in parallel: The temporal feature encoding unit is a lightweight LSTM network, with hidden unit values ​​ranging from 16 to 64; The physical field feature encoding unit is a lightweight Transformer with 1 to 2 layers and 1 to 4 attention heads; The spatial structure encoding unit is a lightweight depthwise separable CNN with 8 to 32 convolutional channels; The output vectors of the three types of units are concatenated and fused into a unified 64-dimensional high-dimensional feature vector, which is then encoded once and used by two types of decoders.

5. The intelligent identification model for internal solitary waves based on physical constraints and oriented towards local deployment of underwater gliders as described in claim 4, characterized in that, The flexible decoder cluster includes an internal solitary wave generation discrimination decoder and an internal solitary wave feature extraction decoder. The internal solitary wave occurrence discrimination decoder uses a lightweight SVM classifier to output the probability of internal solitary wave events. ; The internal solitary wave feature extraction decoder is a bottleneck-lightweight MLP, outputting a parameter vector. A is the amplitude, C is the propagation speed, T is the duration, and L is the horizontal scale. The direction of propagation is determined by the output, which is then fed into a physical constraint verification layer for forced correction.

6. The intelligent identification model for internal solitary waves based on physical constraints and oriented towards local deployment of underwater gliders as described in claim 4, is characterized in that, The physical constraint embedding module includes hard constraints at the output end and soft constraints at the training end. (1) The hard constraints of inference include six mandatory verification rules: ①Amplitude upper limit H represents the deployment depth, and the model output A≤0 directly determines the event as invalid; ② Phase velocity of the first-order internal wave theory for two-layer fluids: ； Reduced gravity , For the theoretical first-order phase velocity, , These represent the thicknesses of the upper and lower seawater layers, respectively. , The density of the upper and lower layers, For reference density, It is the acceleration due to gravity; The model output wave velocity must meet the following requirements. ; ③ Amplitude inversion from scattering layer , The reference amplitude of the internal solitary wave is directly calculated from the vertical displacement of the acoustic scattering layer; constraint Weighted adjustment when not satisfied Weight Take a value of 0.5~0.8; ④ Kinematically constrained residuals , The relative residuals between the theoretical horizontal scale and the model prediction scale of the internal wave; residuals Use directly Coverage correction level scale ; ⑤ Polarity consistency constraint: the sinking of the scattering layer corresponds to the sinking type of internal solitary wave, and the upward layer corresponds to the upward type. If the amplitude sign and the flow velocity rotation structure do not match, the confidence of the event will be reduced or it will be directly eliminated. ⑥ Fluid incompressibility continuity constraint: Constructing velocity divergence residuals If the residual exceeds three times the standard deviation of the background residual, the prediction result is determined to have non-physical bias and the confidence level of the event is reduced. (2) The weights of the total loss function with soft constraints are globally fixed and used uniformly for training across the entire sea area; they cannot be adjusted. The soft constraint introduces a physical guiding term into the loss function during the model training phase. The total loss function is: ； in, The event binary cross-entropy classification loss is used. For amplitude and wave velocity regression Huber loss, For amplitude consistency loss, The loss is due to the phase velocity range constraint. For time-series smoothing loss, For the incompressible residual loss of the flow rate, This is for beam consistency loss.

7. The intelligent identification model for internal solitary waves based on physical constraints and oriented towards local deployment of underwater gliders as described in claim 6, is characterized in that, The model deployment and real-time inference module is divided into five stages: offline training, model lightweighting, model solidification, local loading of latent markers, and event-driven inference. Lightweight model design techniques include network structure lightweighting, 20%~60% structured pruning, knowledge distillation, INT8 fixed-point quantization, and operator fusion. Pruning adaptive rules: fixed 40% pruning for deep water scenarios ≥500m; fixed 60% pruning for shallow sea scenarios with strong tidal interference <100m; and linear adjustment within the 40%~60% range for medium-depth sea areas 100~500m, balancing accuracy and storage usage. Floating-point weighted INT8 quantization formula: ； In the formula, For floating-point weights, Quantize integer weights, As a quantitative scale, Zero-point offset; Quantization-based perceptual training ensures stable embedded inference; operator fusion is used to derive general-purpose embedded models for ONNX, TFLiteMicro, and TVM. When the model is solidified, the model weights, standardized parameters, complete set of acoustic conversion parameters, and physical constraint thresholds are stored in the buoy's non-volatile memory. The device automatically performs CRC verification upon startup, and automatically switches to the backup model if the verification fails.

8. The intelligent identification model for internal solitary waves based on physical constraints and oriented towards local deployment of underwater gliders as described in claim 7, is characterized in that... The event triggering employs a multi-physical quantity joint discrimination mechanism, calculating the background perturbation scale using the median absolute deviation (MAD). ； In the formula, The standard deviation of real-time background disturbance. This is a long-term background observation sequence. The data represents the current time-series observations, with MAD representing the median absolute deviation; the background window duration is 6–24 h, and the event judgment window is 20–40 min. Fixed event activation threshold Hysteresis exit threshold ; Sampling is divided into three working modes: low-frequency monitoring 10 min / time, candidate confirmation 2-5 min / time, and high-frequency inference 1-2 min / time; the longest duration of a single high-frequency sampling is 3 hours in deep water and 2 hours in shallow water. Switching high-frequency sampling criteria: Comprehensive score meets the standard. At least two of the four physical anomalies must meet any one of the following conditions, and the quality control score must be [missing information]. ; High-risk warning trigger condition: Model predicted amplitude or maximum flow velocity disturbance or If any condition is met, a high-priority early warning will be immediately sent via underwater acoustic / satellite communication.

Images