A sea area vertical section multi-dimensional parameter collaborative inversion system

Abstract

The application discloses a kind of sea area vertical section multi-dimensional parameter collaborative inversion system, and original observation data is compensated in situ by sensing data self-compensation module, and high-fidelity observation data is output;Profile adaptive sampling module is according to high-fidelity observation data and is divided into profile hierarchical type and is executed differentiation sampling, and output is simplified sampling data set;Multi-dimensional parameter collaborative inversion module utilizes the lightweight hybrid neural network model of marine fluid dynamics conservation equation constraint, and is coupled decoupled by adaptive weight distribution, and high-precision multi-dimensional parameter inversion profile data is obtained;Underwater weak communication optimization module includes data hierarchical coding, breakpoint continuation and communication state adaptive adjustment;Edge collaborative control module receives communication state and profile feature feedback signal, and uniformly dispatches sampling and transmission parameter.The application solves the problems of sensor drift, low sampling efficiency, physical inversion inconsistency and weak communication bottleneck in deep-sea long-term profile observation.

Description

Technical Field

[0001] This invention relates to the field of marine environmental monitoring technology, and in particular to a multi-dimensional parameter collaborative inversion system for vertical profiles of sea areas. Background Technology

[0002] Multidimensional parameters such as temperature, salinity, depth, turbidity, current velocity, and sound velocity in ocean vertical profiles are crucial foundational information for marine environmental sensing, resource surveys, marine engineering support, and research on marine dynamic processes. With the increasing demands for long-term deep-sea observation, unmanned platform operations, and multi-mission collaborative detection, marine vertical profile parameter acquisition systems are evolving towards long-term, refined, and intelligent capabilities. This places higher demands on the accuracy of observation data, the effectiveness of the sampling process, the reliability of parameter processing results, and the overall operational stability of the system.

[0003] However, due to factors such as the complex and changeable marine environment, the long duration of observation missions, limited platform resources, and special underwater operating conditions, there are still some shortcomings in the current process of acquiring and processing vertical profile parameters of marine areas.

[0004] First, under long-term continuous observation conditions, the stability and consistency of observation data are easily reduced due to factors such as environmental disturbances, changes in device status, and fluctuations in operating conditions, which in turn affects the long-term availability and reliability of profile data.

[0005] Secondly, ocean vertical profiles often exhibit significantly different variation characteristics at different depths, with some areas showing drastic parameter changes while others show relatively gentle changes. In current profile observation processes, the allocation of sampling resources is often difficult to fully match with the profile variation characteristics, easily leading to insufficient acquisition of key local information and some data redundancy, thus affecting observation efficiency, energy consumption control, and the burden of subsequent data processing.

[0006] Furthermore, multidimensional parameters are often strongly coupled and influenced by complex ocean processes. Existing parameter processing and inversion methods often face the challenge of balancing accuracy, stability, and adaptability, especially under complex sea conditions, limited computing power, and long-term operating conditions, where there is still room for improvement in the consistency and reliability of the results.

[0007] Furthermore, from the overall process of data acquisition, profile collection, parameter processing to result output, there is often a lack of close linkage and coordination between the various links. The system-level operation is not capable of responding to environmental changes, task status and resource conditions, which limits the further improvement of overall observation efficiency.

[0008] Meanwhile, data transmission conditions in marine, especially underwater, environments are complex, and communication links are easily affected by factors such as bandwidth, latency, and stability. This leads to problems such as limited efficiency, insufficient continuity, or difficulty in ensuring the integrity of observation data during transmission, which poses a greater challenge to long-term, continuous, and real-time marine vertical profile monitoring applications.

[0009] Therefore, how to improve the data reliability, sampling effectiveness, result accuracy, system coordination, and communication adaptability in the acquisition and processing of multidimensional parameters of vertical marine profiles in complex marine environments has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0010] To address the shortcomings of existing technologies, the present invention aims to provide a multi-dimensional parameter collaborative inversion system for marine vertical profiles, which improves the accuracy, sampling efficiency, system coordination, and data transmission reliability under weak communication conditions in complex marine environments.

[0011] To achieve the above objectives, the present invention provides the following technical solution: a multi-dimensional parameter collaborative inversion system for vertical profiles of a sea area, comprising: The sensor data self-compensation module is used to perform in-situ drift compensation on the raw observation data and output high-fidelity observation data. The profile adaptive sampling module is connected to the sensor data self-compensation module. It is used to calculate vertical gradient data based on the high-fidelity observation data and divide the profile layer type. Based on the closed-loop feedback control logic, it performs differentiated sampling for different layer types and outputs a simplified sampling dataset. The multidimensional parameter collaborative inversion module is connected to the profile adaptive sampling module. It is used to receive the simplified sampling dataset, drive a lightweight hybrid neural network model with physical constraint terms containing the ocean hydrodynamic conservation equation, and perform coupling and decoupling through adaptive weight allocation to obtain high-precision multidimensional parameter inversion profile data. An underwater weak communication optimization module, connected to the multi-dimensional parameter collaborative inversion module, includes: A hierarchical data encoding unit is used to encode the high-precision multidimensional parameter inversion profile data into core layer data and feature layer data according to the importance level of the high-precision multidimensional parameter inversion profile data. The breakpoint resume management unit is used to generate breakpoint logs and perform incremental resume transmission of incomplete data segments after communication is restored. The communication status adaptive unit is connected to the hierarchical data encoding unit and the breakpoint resume management unit, respectively, and is used to dynamically adjust the transmission and caching strategy of the core layer data and the feature layer data according to the underwater acoustic communication quality, and output the communication status feedback signal. The edge collaborative control module is connected to the sensor data self-compensation module, the profile adaptive sampling module, the multi-dimensional parameter collaborative inversion module, and the underwater weak communication optimization module, respectively. It is used to receive the communication status feedback signal and the profile feature feedback signal, and to uniformly schedule the sampling strategy parameters of the profile adaptive sampling module and the transmission parameters of the underwater weak communication optimization module.

[0012] Furthermore, the sensor data self-compensation module includes a drift modeling unit, which is used to construct a drift mathematical model that couples temperature parameters, pressure parameters and salinity parameters, and to perform online drift correction on the original observation data based on the drift mathematical model; The input variables of the drift mathematical model include temperature deviation, static pressure deviation, salinity deviation, and continuous working time. The output of the drift mathematical model is the sensor output drift. The coupling coefficient in the drift mathematical model is predetermined through offline calibration.

[0013] Furthermore, the sensor data self-compensation module includes a steady-state layer identification unit and a hysteresis correction unit; The steady-state layer identification unit is used to calculate the vertical gradient value of water parameters at each depth layer in the original observation data in real time, and identify the continuous depth interval where the vertical gradient value of water parameters is less than a preset gradient threshold as a uniform steady-state layer, and use the average value of the observation data in the uniform steady-state layer as the natural reference benchmark for the in-situ drift compensation. The hysteresis correction unit is used to perform dynamic hysteresis correction on the data after in-situ drift compensation using a second-order inertial element, so as to eliminate the profile depth offset distortion introduced by sensor response hysteresis.

[0014] Furthermore, the profile adaptive sampling module includes a gradient calculation unit, a layer determination unit, a dynamic control unit, and a data order preservation unit; The gradient calculation unit is used to calculate the vertical gradient value of the temperature, salinity and density parameters in the high-fidelity observation data layer by layer. The layer determination unit is connected to the gradient calculation unit and is used to compare the vertical gradient value of each depth layer with a preset first gradient threshold and a preset second gradient threshold respectively. When the vertical gradient value is greater than the preset first gradient threshold, it is determined to be a critical layer. When the vertical gradient value is less than the preset second gradient threshold, it is determined to be a stable layer. When the vertical gradient value is between the preset first gradient threshold and the preset second gradient threshold, it is determined to be a transition layer. When the vertical gradient value jumps between adjacent depth layers by more than a preset abrupt change threshold, it is determined to be an abnormal abrupt change layer. The dynamic control unit is connected to the hierarchical determination unit and is used to execute an encrypted sampling strategy for the critical layer, a sparse sampling strategy for the stable layer, and a fixed-point retesting strategy for the abnormal mutation layer based on the closed-loop feedback control logic. The data ordering unit is connected to the dynamic control unit and is used to extract feature points in each depth layer whose information density exceeds a preset density threshold and remove redundant sampling points, so as to output the simplified sampling dataset while preserving the complete morphological features of the profile.

[0015] Furthermore, the multidimensional parameter collaborative inversion module includes a mechanism constraint unit and a lightweight network unit; The mechanism constraint unit is used to construct the ocean fluid continuity equation, temperature-salinity diffusion equation and water body radiation transfer equation as physical constraint terms, and to embed the physical constraint terms into the loss function of the lightweight hybrid neural network model in the form of regularization terms, so that the lightweight hybrid neural network model is guided by the physical constraint terms during training and inference, and outputs initial multidimensional inversion data that meets the requirements of physical conservation and profile self-consistency. The lightweight network unit is used to carry the lightweight hybrid neural network model, which adopts a hybrid structure of convolutional neural network and lightweight Transformer. The storage amount of model parameters is less than a preset storage threshold to support real-time inference on embedded microcontrollers.

[0016] Furthermore, the multidimensional parameter collaborative inversion module includes a coupling decoupling unit and a profile reconstruction unit; The coupling and decoupling unit is used to calculate the mutual information entropy value between each parameter in the initial multidimensional inversion data, and to assign corresponding inversion weights to each parameter according to the mutual information entropy value, so as to eliminate the strong coupling interference between temperature parameter, salinity parameter, turbidity parameter and flow velocity parameter, and obtain decoupled multidimensional parameter profile data. The profile reconstruction unit is connected to the coupling and decoupling unit and is used to perform full-depth interpolation and smoothing on the decoupled multidimensional parameter profile data to output high-precision multidimensional parameter inversion profile data without depth jumps or physical violations.

[0017] Furthermore, the edge collaborative control module includes a data scheduling unit and a closed-loop feedback unit; The data scheduling unit is used to uniformly manage the data flow and control flow between the sensor data self-compensation module, the profile adaptive sampling module, the multi-dimensional parameter collaborative inversion module, and the underwater weak communication optimization module, to ensure that the high-fidelity observation data, the simplified sampling dataset, the high-precision multi-dimensional parameter inversion profile data, and the core layer data and the feature layer data are connected between the modules according to a preset time sequence link. The closed-loop feedback unit is connected to the data scheduling unit and is used to receive the communication status feedback signal output by the underwater weak communication optimization module and the profile feature feedback signal output by the multi-dimensional parameter collaborative inversion module. Based on the underwater acoustic communication quality represented by the communication status feedback signal and the profile structure complexity represented by the profile feature feedback signal, it generates sampling strategy adjustment instructions and transmission strategy adjustment instructions.

[0018] Furthermore, the edge collaborative control module includes a low-power communication instruction unit and a local storage scheduling unit; The low-power communication instruction unit is used to send a communication mode switching instruction or a breakpoint resume trigger instruction to the underwater weak communication optimization module according to the underwater acoustic communication quality, so as to control the underwater weak communication optimization module to enter a data transmission state that matches the current underwater acoustic communication quality. The local storage scheduling unit is used to perform hierarchical storage management of the original observation data, the high-fidelity observation data, the simplified sampling dataset, and the high-precision multidimensional parameter inversion profile data. When the local storage space is detected to be lower than the preset storage threshold, the core layer data and the high-precision multidimensional parameter inversion profile data are retained first, and the non-core data are rolled over and stored in order of data timestamp.

[0019] Furthermore, the underwater weak communication optimization module also includes a redundancy fault-tolerant verification unit; The hierarchical data encoding unit is specifically used to divide the high-precision multi-dimensional parameter inversion profile data into the core layer data, the feature layer data, and the redundant layer data according to the profile layer type and the degree of drastic change of parameter gradient corresponding to each data block in the high-precision multi-dimensional parameter inversion profile data. The core layer data is encoded using a lossless compression algorithm, the feature layer data is encoded using a lossy compression algorithm, and the redundant layer data is discarded or periodically packaged. The redundancy fault-tolerant verification unit is connected to the hierarchical data encoding unit and is used to add cyclic redundancy check codes to the encoded core layer data and feature layer data, and to generate an additional XOR redundancy backup packet for the core layer data. When the receiving end detects that the core layer data has been lost, it recovers the lost data through the XOR redundancy backup packet. When the receiving end detects that the feature layer data has been lost, it triggers a retransmission request for the corresponding data block.

[0020] Furthermore, the breakpoint resume management unit includes a communication status monitoring subunit, a breakpoint recording subunit, and an incremental resume subunit; The communication status monitoring subunit is used to monitor the signal-to-noise ratio, bandwidth, and disconnection duration of the underwater acoustic communication link in real time. When the signal-to-noise ratio is lower than a preset signal-to-noise ratio threshold or the bandwidth is lower than a preset bandwidth threshold, a breakpoint recording operation is triggered. The breakpoint recording subunit is connected to the communication status monitoring subunit and is used to record the successfully transmitted data segment number, transmission timestamp and data segment check code in the local storage medium when a breakpoint recording trigger signal is received, forming a breakpoint log file. The incremental resume transmission subunit is connected to the breakpoint recording subunit. After detecting the restoration of the underwater acoustic communication link, it reads the breakpoint log file and, based on the transmitted data segment number recorded in the breakpoint log file, uploads only the remaining data segments that have not been transmitted to the receiving end to achieve incremental resume transmission. The communication state adaptive unit is connected to the incremental transmission subunit and is used to divide the current underwater acoustic communication link into a good communication state, a normal communication state, or a very poor communication state according to the signal-to-noise ratio value and the bandwidth value, and to execute corresponding transmission strategies for different communication states. Specifically, in the good communication state, the core layer data and the feature layer data are transmitted; in the normal communication state, only the core layer data is transmitted and the feature layer data is temporarily stored locally; and in the very poor communication state, active uploading is paused and the data to be transmitted is encrypted and stored locally.

[0021] The beneficial effects of this invention are: (1) The present invention provides a sensor data self-compensation module, which can perform in-situ drift compensation on the original observation data, thereby improving the stability and reliability of the observation data under long-term continuous observation conditions.

[0022] (2) The present invention sets up a profile adaptive sampling module, which can perform differentiated sampling on different profile regions according to the vertical gradient change, thereby reducing redundant sampling, improving sampling efficiency and reducing system energy consumption while ensuring the acquisition effect of key profile features.

[0023] (3) The present invention sets up a multi-dimensional parameter collaborative inversion module, which combines physical constraint terms with a lightweight hybrid neural network model to achieve multi-dimensional parameter collaborative inversion, thereby helping to improve the accuracy, physical rationality and robustness of the inversion results under complex sea conditions.

[0024] (4) By constructing a collaborative closed-loop control mechanism between sensor data compensation, profile sampling, parameter inversion and communication optimization, this invention can improve the overall linkage capability and autonomous operation capability of the system, thereby better meeting the application needs of long-term unattended marine observation.

[0025] (5) The present invention sets up an underwater weak communication optimization module and combines hierarchical data encoding, breakpoint resume and communication status adaptive adjustment mechanism to improve the reliability, continuity and integrity of data transmission in weak communication environment.

[0026] (6) The present invention has a clear overall structure, which is easy to implement by combining with existing sensors, embedded hardware platforms and communication links, and has good engineering application value and deployment adaptability. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the structure of the multi-dimensional parameter collaborative inversion system for vertical profiles of the sea area in this invention; Figure 2 This is a schematic diagram of the interruption point resume management unit of the present invention.

[0028] Reference numerals: 1. Sensor data self-compensation module; 11. Drift modeling unit; 12. Steady-state layer identification unit; 13. Hysteresis correction unit; 2. Profile adaptive sampling module; 21. Gradient calculation unit; 22. Layer determination unit; 23. Dynamic control unit; 24. Data order preservation unit; 3. Multi-dimensional parameter collaborative inversion module; 31. Mechanism constraint unit; 32. Lightweight network unit; 33. Coupling and decoupling unit; 34. Profile reconstruction unit; 4. Underwater weak communication optimization module; 41. Hierarchical data encoding unit; 42. Breakpoint resume management unit; 421. Communication status monitoring subunit; 422. Breakpoint recording subunit; 423. Incremental resume subunit; 43. Communication status adaptive unit; 44. Redundancy fault tolerance verification unit; 5. Edge collaborative control module; 51. Data scheduling unit; 52. Closed-loop feedback unit; 53. Low-power communication command unit; 54. Local storage scheduling unit. Detailed Implementation

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The same functional modules or functional units are denoted by the same reference numerals.

[0030] like Figure 1 As shown in Example 1, this example provides a multi-dimensional parameter collaborative inversion system for marine vertical profiles, which can be applied to deep-sea moorings, profiling buoys, unmanned underwater vehicles carrying profiling platforms, or near-shore long-term observation platforms. It is used for long-term continuous autonomous observation, differentiated sampling, collaborative inversion, and reliable data transmission under weak communication conditions of multi-dimensional parameters of marine vertical profiles in complex marine environments.

[0031] In this embodiment, the marine vertical profile multi-dimensional parameter collaborative inversion system includes: a sensor data self-compensation module 1, a profile adaptive sampling module 2, a multi-dimensional parameter collaborative inversion module 3, an underwater weak communication optimization module 4, and an edge collaborative control module 5.

[0032] The sensor data self-compensation module 1 is used to perform in-situ drift compensation on the original observation data and output high-fidelity observation data. The profile adaptive sampling module 2 is connected to the sensor data self-compensation module 1 and is used to calculate vertical gradient data and classify profile layers based on the high-fidelity observation data. It performs differentiated sampling for different layer types according to closed-loop feedback control logic and outputs a simplified sampling dataset. The multidimensional parameter collaborative inversion module 3 is connected to the profile adaptive sampling module 2 and is used to receive the simplified sampling dataset. It drives a lightweight hybrid neural network model using physical constraint terms containing the ocean hydrodynamic conservation equations and performs coupling and decoupling through adaptive weight allocation to obtain high-precision multidimensional parameter inversion profile data. The underwater weak communication optimization module 4 is connected to the multidimensional parameter collaborative inversion module 3 and includes a hierarchical data encoding unit 41, a breakpoint resume management unit 42, and a communication state adaptive unit 43. The system comprises a hierarchical data encoding unit 41, which encodes the profile data into core layer data and feature layer data based on the importance level of the high-precision multi-dimensional parameter inversion profile data; a breakpoint resumption management unit 42, which generates breakpoint logs and incrementally resumes incomplete data segments after communication is restored; a communication status adaptive unit 43, which is connected to the hierarchical data encoding unit 41 and the breakpoint resumption management unit 42, and dynamically adjusts the transmission and caching strategies of the core layer data and feature layer data according to the underwater acoustic communication quality, and outputs communication status feedback signals; and an edge collaborative control module 5, which is connected to the sensor data self-compensation module 1, the profile adaptive sampling module 2, the multi-dimensional parameter collaborative inversion module 3, and the underwater weak communication optimization module 4, and receives communication status feedback signals and profile feature feedback signals, and uniformly schedules the sampling strategy parameters of the profile adaptive sampling module 2 and the transmission parameters of the underwater weak communication optimization module 4.

[0033] I. Hardware Components and Supporting Environment; In this embodiment, the sensing part of the system includes a temperature, salinity, and depth (CTD) sensor, an optical turbidity sensor, an acoustic Doppler flow velocity sensor, and a MEMS pressure / attitude sensor, which are used to collect raw observation data such as temperature, salinity, depth, turbidity, flow velocity, and attitude.

[0034] The control and computing section employs an ARM-based embedded MCU or edge computing main control board with a main frequency of no less than 800MHz and a memory of no less than 512MB. It is used to run the sensor data self-compensation module 1, the profile adaptive sampling module 2, the multi-dimensional parameter collaborative inversion module 3, the underwater weak communication optimization module 4, and the edge collaborative control module 5. In practical applications, platforms based on NXPi.MX8M Mini, Rockchip RK3568, or equivalent ARM architecture processors can be selected.

[0035] The execution part includes a profiler lifting drive mechanism and a motor control circuit, which is used to perform vertical lifting sampling according to the sampling strategy output by the profile adaptive sampling module 2.

[0036] The communication component includes an underwater acoustic communication module and a BeiDou short message backup communication module. The underwater acoustic communication module supports bandwidth adaptation and link quality monitoring, while the BeiDou short message backup communication module is used for backhauling key summary information in extreme weak communication scenarios.

[0037] The storage section uses large-capacity FLASH storage units with a capacity of no less than 16GB to cache raw observation data, high-fidelity observation data, simplified sampling datasets, high-precision multi-dimensional parameter inversion profile data, and breakpoint logs.

[0038] The power supply section uses a low-power lithium battery and power management circuit to meet the requirements of long-term low-power operation.

[0039] In terms of software environment, the edge side can run embedded Linux or real-time operating system, the algorithm program is implemented in C / C++, and the lightweight hybrid neural network model can be deployed using TensorFlow Lite, ONNX Runtime or MNN to meet the real-time inference requirements of the edge.

[0040] II. Key parameter settings; To ensure that this embodiment can be clearly implemented and achieve stable results, the system adopts the following parameter settings: The temperature gradient threshold is 0.05℃ / m; the salinity gradient threshold is 0.02psu / m; the encrypted sampling step size is 10cm; the sparse sampling step size is 100cm; the model inference latency is less than 100ms; the sensor data self-compensation update cycle is 10 minutes / time; the profile stratification determination uses a 5m depth sliding window; the data block size is 128 bytes / block. The communication status judgment thresholds are as follows: when the signal-to-noise ratio is greater than or equal to 15dB, it is judged as good; when the signal-to-noise ratio is greater than or equal to 5dB and less than 15dB, it is judged as fair; when the signal-to-noise ratio is less than 5dB, it is judged as very poor; the breakpoint log storage interval is 5 minutes / time.

[0041] The above parameters can be adaptively adjusted according to different marine environments, platform sizes and mission requirements. However, in this embodiment, the above parameters can better balance sampling resolution, inversion real-time performance, energy consumption control and communication reliability.

[0042] III. Work Process; In this embodiment, the system's workflow is as follows: After the system is powered on, the edge collaborative control module 5 first loads drift compensation parameters, lightweight hybrid neural network model weights, sampling control parameters, and communication optimization parameters, and completes the initialization of each module.

[0043] Subsequently, the sensing section enters the vertical profile observation state, continuously collecting raw observation data such as temperature, salinity, depth, turbidity, flow velocity and attitude, and sending the raw observation data to the sensor data self-compensation module 1.

[0044] The sensor data self-compensation module 1 performs in-situ drift compensation on the original observation data, corrects zero drift, slow drift and dynamic response deviations that may occur during long-term operation, and outputs high-fidelity observation data.

[0045] The profile adaptive sampling module 2 receives high-fidelity observation data, calculates the temperature gradient and salinity gradient along the depth direction, and classifies the profile into layers based on the temperature gradient judgment threshold of 0.05℃ / m, the salinity gradient judgment threshold of 0.02psu / m, and a 5m depth sliding window. For feature layers with significant parameter changes, a dense sampling step size of 10cm is used, and for steady-state layers with gradual changes, a sparse sampling step size of 100cm is used, thereby outputting a simplified sampling dataset.

[0046] The multi-parameter collaborative inversion module 3 receives a simplified sampled dataset, drives a lightweight hybrid neural network model using physical constraints containing the ocean hydrodynamic conservation equations, and decouples the coupling relationship between multi-dimensional parameters through adaptive weight allocation, outputs high-precision multi-dimensional parameter inversion profile data, and extracts the corresponding profile feature feedback signals.

[0047] The underwater weak communication optimization module 4 receives high-precision multi-dimensional parameter inversion profile data. Among them, the hierarchical data encoding unit 41 encodes the data into core layer data and feature layer data according to the importance level of the data; the communication state adaptive unit 43 dynamically adjusts the transmission and caching strategy of core layer data and feature layer data according to the current underwater acoustic communication quality; the breakpoint resume management unit 42 generates breakpoint logs and performs incremental resume transmission of incomplete data segments after communication is restored.

[0048] The edge collaborative control module 5 receives communication status feedback signals and profile feature feedback signals, and coordinates the sampling strategy parameters of the profile adaptive sampling module 2 and the transmission parameters of the underwater weak communication optimization module 4 to form a closed-loop collaborative control of "observation-sampling-inversion-transmission".

[0049] When the system is running continuously, the above process is executed in a loop; when a communication interruption occurs, the system automatically enters the caching and breakpoint resume mode; when communication is restored, the system performs supplementary transmission of incomplete data segments based on the breakpoint log, thereby achieving long-term continuous autonomous observation.

[0050] IV. Working principle; The working principle of this embodiment is as follows: During vertical profile observations in the ocean, raw observation data is easily affected by deep-sea high pressure, low temperature, high salinity, and long-term operating conditions, resulting in measurement drift, dynamic response deviation, and local observation distortion. The sensor data self-compensation module 1 first performs in-situ drift compensation on the raw observation data, making the output high-fidelity observation data closer to real ocean environmental parameters, thus providing a reliable data foundation for subsequent profile stratification and multi-dimensional parameter inversion.

[0051] After acquiring high-fidelity observation data, the profile adaptive sampling module 2 identifies regions with abrupt parameter changes and regions with gradual changes in salinity in the vertical profile based on temperature and salinity gradients, and adopts different sampling step sizes for different stratification types. For regions with rapidly changing parameters, such as thermoclines, haloclines, and frontal transition zones, the system improves local resolution by increasing sampling density; for regions with gradual parameter changes, the system reduces redundant data and overall power consumption by using sparse sampling. Therefore, the system can balance key feature capture capabilities and sampling efficiency without significantly increasing computational and storage burdens.

[0052] After the simplified sampled dataset is input into the multidimensional parameter collaborative inversion module 3, the module does not simply rely on the data-driven model for inference. Instead, it combines physical constraints, including the conservation equations of ocean hydrodynamics, with a lightweight hybrid neural network model. Adaptive weight allocation is used to decouple the coupling relationships between multidimensional parameters such as temperature, salinity, current velocity, sound speed, and turbidity. In this way, the system can both leverage the correlation between multidimensional parameters to improve inversion accuracy and reduce mutual interference between different parameters, thereby enhancing the physical rationality and robustness of the inversion results under complex marine conditions.

[0053] After obtaining high-precision multi-dimensional parameter inversion profile data, the underwater weak communication optimization module 4 further implements hierarchical transmission based on the underwater acoustic link status. The hierarchical data encoding unit 41 divides the high-precision multi-dimensional parameter inversion profile data into core layer data and feature layer data to ensure priority transmission of critical data when communication conditions are poor. The communication status adaptive unit 43 classifies the link status into three categories: good, average, and extremely poor, based on the current signal-to-noise ratio, and dynamically adjusts the transmission and caching strategies of the core layer data and feature layer data for different communication statuses. The breakpoint resume management unit 42 records the sent and incomplete data segments through breakpoint logs and performs incremental resume transmission after the link is restored, thereby improving data integrity and transmission continuity in weak communication environments.

[0054] The edge collaborative control module 5 operates throughout the entire process. It receives communication status feedback signals and profile feature feedback signals, and accordingly coordinates sampling strategy parameters and transmission parameters, creating a real-time collaborative closed loop between sampling, inversion, and transmission. As a result, the system can dynamically adjust its operation mode based on changes in ocean profile features and communication status, enhancing its overall autonomous operation capability.

[0055] V. Technical Effects; Compared with the prior art, this embodiment has at least the following technical effects: First, the sensor data self-compensation module 1 can perform in-situ drift compensation on the original observation data, thereby improving the stability and reliability of the observation data under long-term continuous observation conditions.

[0056] Second, the profile adaptive sampling module 2 can perform differentiated sampling on different profile regions based on changes in vertical gradient, thereby reducing redundant sampling, improving sampling efficiency and reducing system energy consumption while ensuring the ability to acquire key profile features.

[0057] Third, the multi-dimensional parameter collaborative inversion module 3 combines physical constraint terms with a lightweight hybrid neural network model to achieve multi-dimensional parameter collaborative inversion, thereby helping to improve the accuracy, physical rationality and robustness of the inversion results under complex sea conditions.

[0058] Fourth, the underwater weak communication optimization module 4, through the coordinated action of the hierarchical data encoding unit 41, the breakpoint resume management unit 42, and the communication state adaptive unit 43, can improve the reliability, continuity, and integrity of data transmission in a weak communication environment.

[0059] Fifth, the edge collaborative control module 5 can uniformly schedule sampling strategy parameters and transmission parameters, improve the overall linkage capability and autonomous operation capability of the system, and thus better meet the application needs of long-term unattended marine observation.

[0060] VI. Experimental verification; To verify the application effect of this embodiment, a deployment test was conducted on a deep-sea observation mooring platform in the South China Sea for six consecutive months. During the test, the system continuously performed tasks such as vertical profile observation, differentiated sampling, multi-dimensional parameter collaborative inversion, and data transmission under weak communication conditions.

[0061] The test results show that: the long-term drift of the sensor is less than 0.03%; the amount of sampled data is reduced by 82%, and the equipment endurance is increased by 2.7 times; the mean square error of thermocline inversion is reduced by 43%, and no physically unreasonable profile is found; in the underwater weak communication scenario, the amount of transmitted data is further reduced by 65%, the success rate of breakpoint resume transmission reaches 99.2%, the data packet loss rate is reduced to 0.8%, and the long-term observation data integrity rate reaches 100%.

[0062] The above results demonstrate that this embodiment can achieve long-term stable observation, efficient sampling, reliable inversion, and complete transmission of multidimensional parameters of the vertical profile of the sea area under complex marine environments and weak communication conditions, and has good engineering application value.

[0063] In Example 2, based on Example 1, the sensor data self-compensation module 1 further includes a drift modeling unit 11, a steady-state layer identification unit 12, an online compensation unit, and a hysteresis correction unit 13; the profile adaptive sampling module 2 further includes a gradient calculation unit 21, a layer determination unit 22, a dynamic control unit 23, and a data order preservation unit 24; the multi-dimensional parameter collaborative inversion module 3 further includes a mechanism constraint unit 31, a lightweight network unit 32, a coupling / decoupling unit 33, and a profile reconstruction unit 34. The specific implementation methods of each unit are described below.

[0064] I. Further implementation of sensor data self-compensation module 1; 1. Drift modeling unit 11; The drift modeling unit 11 is used to construct a drift mathematical model that couples temperature, pressure and salinity parameters, and to perform online drift correction on the original observation data based on the drift mathematical model.

[0065] In this embodiment, the drift modeling unit 11 is implemented using a four-layer structure: "input normalization layer - coupling feature construction layer - nonlinear drift inference layer - drift output layer".

[0066] The input normalization layer converts temperature deviation, static pressure deviation, salinity deviation, and continuous operating time into dimensionless inputs; the coupling feature construction layer constructs interactive features between temperature and pressure, pressure and salinity, temperature and salinity, and temperature, pressure, and salinity; the nonlinear drift inference layer performs nonlinear compression and state gating on the multi-source coupled inputs using inverse hyperbolic sine and hyperbolic tangent functions; and the drift output layer outputs the sensor output drift at the current moment.

[0067] Preferably, the normalized input is defined as follows: ; in, This represents the normalized value of the temperature deviation. This represents the normalized value of static pressure deviation. This represents the normalized amount of salinity deviation. This represents the normalized value of continuous working duration; This indicates the temperature measurement value at the current moment. This represents the reference mean temperature of a uniform steady-state layer. Indicates the temperature normalization reference span; P represents the current static pressure measurement value. This represents the reference mean static pressure of a uniform steady-state layer. Indicates the hydrostatic normalized reference span; S represents the current salinity measurement value. This represents the reference mean salinity of a uniform steady-state layer. Indicates the salinity normalization reference span; t represents the continuous operating time. This represents the normalized benchmark for working hours.

[0068] Furthermore, the preferred drift mathematical model is: ; in, The sensor output drift is represented by M, where M represents the drift output scale coefficient; arsinh represents the inverse hyperbolic sine function; exp represents the natural exponential function; tanh represents the hyperbolic tangent function; a represents the linear coupling coefficient of temperature deviation; b represents the linear coupling coefficient of hydrostatic deviation; c represents the linear coupling coefficient of salinity deviation; d represents the second-order coupling coefficient of temperature deviation and hydrostatic deviation; e represents the second-order coupling coefficient of hydrostatic deviation and salinity deviation; f represents the second-order coupling coefficient of temperature deviation and salinity deviation; g represents the drift coefficient during continuous operation; h, ,j represent the gating coefficients for the squared terms of temperature deviation, static pressure deviation, and salinity deviation, respectively; k represents the third-order coupling gating coefficients for temperature deviation, static pressure deviation, and salinity deviation. The continuous working time gating coefficient is represented; l, m, and n represent the normalized suppression coefficients for temperature deviation amplitude, static pressure deviation amplitude, and salinity deviation amplitude, respectively; || represents absolute value operation.

[0069] In this embodiment, the coupling coefficients in the drift mathematical model are predetermined through offline calibration. During offline calibration, the temperature range is controlled to be 0°C to 25°C, the static pressure range to be 0.1MPa to 60MPa, and the salinity range to be 28psu to 40psu in a combined temperature, pressure, and salinity calibration environment. The continuous working time test nodes are set to 0h, 72h, 168h, 336h, and 720h. The deviation between the sensor output value and the standard reference value under each set of working conditions is collected. The parameter values ​​from a to p and M are obtained by fitting using the constrained nonlinear least squares method, and the parameters are written into the edge-side non-volatile memory.

[0070] In this embodiment, the drift modeling unit 11 performs online reasoning according to the following reasoning logic: First, it calculates the temperature deviation value, static pressure deviation value, and salinity deviation value based on the uniform steady-state layer reference mean value output by the steady-state layer identification unit 12; second, it obtains the value based on the preset normalization benchmark. , , and Then, the linear term, second-order cross term, and time slow drift term are input into the inverse hyperbolic sine function to form the main drift response, and the square term, third-order cross term, and continuous operating time term are input into the hyperbolic tangent gating expression to form the state gating factor; finally, the sensor output drift at the current moment is output. .

[0071] In this drift mathematical model, the inverse hyperbolic sine function is used to nonlinearly compress the multi-source coupled drift input, allowing the model to retain high sensitivity when the deviation is small and avoid drift estimation divergence when the deviation is large. The hyperbolic tangent function and the exponential gating term are used together to characterize the state saturation characteristics of the sensor under high pressure, low temperature, high salinity and long-term operation conditions, thereby improving the stability and accuracy of in-situ drift estimation under extreme deep-sea conditions. Thus, the drift modeling unit 11 can more accurately characterize the composite drift mechanism under the combined effects of temperature, pressure, salinity and continuous operating time, improving the accuracy of subsequent compensation.

[0072] 2. Steady-state layer identification unit 12; The steady-state layer identification unit 12 is used to calculate the vertical gradient values ​​of water parameters at each depth in the original observation data in real time, and to identify continuous depth intervals where the vertical gradient values ​​of water parameters are less than a preset gradient threshold as uniform steady-state layers, and to use the average value of the observation data within the uniform steady-state layer as a natural reference benchmark for in-situ drift compensation.

[0073] In this embodiment, a 5m depth sliding window is used for steady-state layer identification. Differential calculations are performed on data from adjacent depth layers within the window to obtain the vertical gradient values ​​of temperature, salinity, and density. Preferably, the temperature gradient threshold is 0.05℃ / m, the salinity gradient threshold is 0.02 psu / m, and the density gradient threshold is 0.03 (kg / m³). 3 ) / m.

[0074] When the absolute values ​​of the temperature gradient, salinity gradient, and density gradient are less than 0.05℃ / m, respectively, within a continuous depth range, the absolute values ​​of these gradients are also less than 0.02 psu / m and 0.03 kg / m³. 3 When the depth is ) / m, the continuous depth interval is identified as a uniform steady-state layer, and the average values ​​of temperature, hydrostatic pressure and salinity observations within the layer are calculated, which are used as the reference average values ​​for temperature, hydrostatic pressure and salinity of the uniform steady-state layer, respectively.

[0075] Using the mean of a uniform steady-state layer as a natural reference benchmark helps reduce the interference of drastically changing water layers on drift estimation, thereby improving the robustness of drift compensation.

[0076] 3. Online compensation unit; The online compensation unit is used to substitute real-time observations into the drift mathematical model, calculate the current drift, and correct the output in real time.

[0077] In this embodiment, the online compensation unit performs the correction in the following manner: ; in, This represents the observed output value after in-situ drift compensation has been completed. This represents the original observation output value. This represents the sensor output drift amount output by the drift modeling unit.

[0078] Preferably, the online compensation update cycle is 10 minutes per cycle. That is, every 10 minutes of operation, the system updates the drift estimate based on the latest identified uniform steady-state layer, the current continuous working duration, and real-time observation data, and compensates and corrects subsequent collected data.

[0079] 4. Lag correction unit 13; The hysteresis correction unit 13 is used to perform dynamic hysteresis correction on the data after in-situ drift compensation using a second-order inertial element, so as to eliminate the profile depth offset distortion introduced by the sensor response hysteresis.

[0080] In this embodiment, the dynamic response of the sensor is approximated as: ; Where G(s) represents the sensor dynamic response transfer function, and s represents the Laplace complex frequency domain variable. Represents the first inertial time constant. This represents the second inertial time constant. Preferably, Take 0.8s, Take 1.5s.

[0081] Furthermore, the lag-corrected observed output value is estimated using the following formula: ; in, This represents the observed output value after dynamic lag correction. express The first derivative with respect to time, express The second derivative with respect to time.

[0082] In the embedded implementation, the first and second derivatives are approximated by the difference between adjacent sampling points. This correction method can reduce the positional shift of the layer boundary and the local profile distortion caused by sensor response hysteresis.

[0083] II. Further implementation of the profile adaptive sampling module 2; 1. Gradient calculation unit 21; The gradient calculation unit 21 is used to calculate the vertical gradient value of temperature, salinity and density parameters in high-fidelity observation data layer by layer.

[0084] In this embodiment, the vertical gradient is calculated using the central difference method: ; in, Indicates that parameter X is in the depth layer The vertical gradient value at point X, where X represents the temperature parameter, salinity parameter, or density parameter. This represents the depth coordinates corresponding to the i-th depth layer. and These represent the depth coordinates of the next depth layer and the previous depth layer, respectively. This indicates that parameter X is in the depth coordinate. The value at that location, This indicates that parameter X is in the depth coordinate. The value at that location. The density parameter is calculated using the potential density anomaly.

[0085] 2. Layer determination unit 22; The layer determination unit 22 is connected to the gradient calculation unit 21 and is used to compare the vertical gradient value of each depth layer with the preset first gradient threshold and the preset second gradient threshold respectively. When the vertical gradient value is greater than the preset first gradient threshold, it is determined to be a critical layer. When the vertical gradient value is less than the preset second gradient threshold, it is determined to be a stable layer. When the vertical gradient value is between the preset first gradient threshold and the preset second gradient threshold, it is determined to be a transition layer. When the vertical gradient value jumps between adjacent depth layers by more than the preset abrupt change threshold, it is determined to be an abnormal abrupt change layer.

[0086] Preferably, the threshold values ​​for each parameter are set as follows: Temperature parameters: First gradient threshold is 0.05℃ / m, second gradient threshold is 0.02℃ / m, and abrupt change threshold is 0.08℃ / m; Salinity parameters: first gradient threshold is 0.02 psu / m, second gradient threshold is 0.008 psu / m, and mutation threshold is 0.03 psu / m; Density parameter: First gradient threshold is 0.03 (kg / m³). 3 The second gradient threshold is 0.01 (kg / m). 3 ) / m, mutation threshold is 0.05 (kg / m 3 ) / m.

[0087] In this embodiment, the current depth layer is determined as a critical layer as long as any one of the parameters, temperature, salinity, or density, meets the critical layer determination criteria; when all three parameters meet the stable layer determination criteria, it is determined as a stable layer; when at least one parameter is between the first gradient threshold and the second gradient threshold and no abnormal mutation determination is triggered, it is determined as a transition layer; when the gradient jump variable of any parameter between adjacent depth layers exceeds the corresponding mutation threshold, it is preferentially determined as an abnormal mutation layer.

[0088] 3. Dynamic control unit 23; The dynamic control unit 23 is connected to the hierarchical determination unit 22, and is used to execute a dense sampling strategy for the critical layer, a sparse sampling strategy for the stable layer, and a fixed-point retesting strategy for the abnormal mutation layer based on the closed-loop feedback control logic.

[0089] In this embodiment, the critical layer uses a 10cm dense sampling step size; the stable layer uses a 100cm sparse sampling step size; the transition layer uses a 50cm intermediate sampling step size; and the abnormal mutation layer performs fixed-point retesting near the current depth point, with the retesting range set at 0.5m above and below the current depth point, the retesting interval at 10cm, and the retesting number at 2 times.

[0090] 4. Data order preservation unit 24; The data ordering unit 24 is connected to the dynamic control unit 23 and is used to extract feature points in each depth layer whose information density exceeds a preset density threshold and remove redundant sampling points, so as to output a simplified sampling dataset while preserving the complete morphological features of the profile.

[0091] In this embodiment, the information density index is calculated as follows: ; in, This represents the information density index of the i-th depth layer. , , These represent temperature, salinity, and density at different depths. Vertical gradient value at that point This represents the local window variance centered on the current depth layer. , , , These represent the normalized temperature, salinity, and density at different depths. The vertical gradient value and local window variance at the location.

[0092] Preferably, the preset density threshold is 0.60. When When the value is greater than or equal to 0.60, the current sampling point is retained as a feature point; when... If the error is less than 0.60 and the local linear error formed by the current point and its adjacent points is less than the preset error limit, the current point will be discarded as a redundant sampling point.

[0093] Preferably, the local linearity error limits are set as follows: temperature 0.01℃, salinity 0.005 psu, and density 0.005 kg / m³. 3 .

[0094] III. Further implementation of the multi-dimensional parameter collaborative inversion module 3; 1. Mechanism constraint unit 31; Mechanism constraint unit 31 is used to construct the ocean fluid continuity equation, temperature-salinity diffusion equation and water body radiation transfer equation as physical constraint terms, and embed the physical constraint terms into the loss function of the lightweight hybrid neural network model in the form of regularization terms, so that the lightweight hybrid neural network model is guided by the physical constraint terms during training and inference, and outputs initial multidimensional inversion data that meets the requirements of physical conservation and profile self-consistency.

[0095] In this embodiment, the total loss function is written as: ; in, Represents the total loss function. This indicates the monitoring loss of the observation data. Represents physical constraint loss. This represents the physical constraint weight coefficient, which is preferably set to 0.2.

[0096] Furthermore, the monitoring loss is expressed as: ; Where M represents the number of inversion parameter types, and N represents the number of depth layers. This represents the supervision weight of the m-th class parameter. This represents the reference true value of the m-th type of parameter at the i-th depth layer. This represents the model's output value.

[0097] Physical constraint loss is defined as: ; in, This represents the constraint loss term in the continuity equation. This represents the constraint loss term in the temperature-salinity diffusion equation. This represents the constraint loss term in the water body radiation transfer equation.

[0098] The continuity equation constraint terms are: ; in, Represents the gradient or divergence operator; This represents the inner product symbol in dot product or divergence operations. This represents the velocity vector field at the i-th depth layer; This represents the divergence of the velocity vector field at the i-th depth layer.

[0099] The constraint terms of the temperature-salinity diffusion equation are: ; Where ∂ represents the partial derivative operator; T represents temperature; S represents salinity; and t represents time. Represents the velocity vector field; ∇S represents the spatial gradient of temperature; ∇S represents the spatial gradient of salinity. Represents the Laplace operator; Indicates the temperature diffusivity; Indicates the salinity diffusion coefficient; Indicates the temperature source term; This indicates the salinity source term.

[0100] The constraint terms of the water body radiation transfer equation are: ; Where I represents the water body radiation intensity or optical response; z represents the depth coordinate; This represents the attenuation coefficient at the i-th depth layer; This represents the water radiation intensity or optical response at the i-th depth layer; This represents the scattering source term at the i-th depth layer.

[0101] Furthermore, during the inference phase, the mechanism constraint unit 31 performs a physical residual correction on the initial result output by the lightweight network unit 32: ; in, This represents the initial multidimensional inversion result. This represents the multidimensional inversion result after single-step correction of the physical residuals. This represents the physical correction step size during the inference phase, preferably 0.05. This represents the physical constraint loss with respect to the initial multidimensional inversion result. The partial derivatives are used to characterize the direction and magnitude of the current output deviating from the physical constraints.

[0102] 2. Lightweight network unit 32; The lightweight network unit 32 is used to carry a lightweight hybrid neural network model. This lightweight hybrid neural network model adopts a hybrid structure of convolutional neural network and lightweight Transformer, and the storage of model parameters is less than a preset storage threshold to support real-time inference on embedded microcontrollers.

[0103] Preferably, the preset storage threshold is 5MB. The model input is a profile data sequence after adaptive sampling. The input channels include temperature, salinity, pressure, turbidity, flow rate, depth index, sampling type identifier, and effective mask, for a total of 8 channels. The input depth sequence is uniformly resampled or padded to 256 layers.

[0104] In this embodiment, the model includes a local feature extraction layer, a global correlation modeling layer, and a multi-task output layer. The local feature extraction layer employs a three-layer one-dimensional convolutional structure; the global correlation modeling layer uses a two-layer lightweight Transformer encoder, each layer including four attention heads, with a hidden dimension of 128 and a feedforward network dimension of 256; the multi-task output layer sets four parameter heads: temperature branch, salinity branch, turbidity branch, and current velocity branch. The density profile is calculated from temperature, salinity, and pressure according to the seawater state equation, and the sound velocity profile is calculated from the reconstructed temperature, salinity, and depth parameters according to the empirical formula for sound velocity.

[0105] Preferably, the total number of model parameters is approximately 0.96M. After using 16-bit fixed-point quantization, the model parameter storage is approximately 1.92MB, which is less than the preset storage threshold of 5MB.

[0106] 3. Coupling / decoupling unit 33; The coupling / decoupling unit 33 is used to calculate the mutual information entropy value between each parameter in the initial multidimensional inversion data, and to assign corresponding inversion weights to each parameter according to the mutual information entropy value, so as to eliminate the strong coupling interference between temperature parameter, salinity parameter, turbidity parameter and flow velocity parameter, and obtain decoupled multidimensional parameter profile data.

[0107] In this embodiment, the mutual information between any two parameters is calculated using the following formula: ; in, Indicates parameters With parameters The mutual information entropy value between them Describes the joint probability distribution. and These represent the marginal probability distributions.

[0108] Furthermore, for a certain parameter Its total coupling strength Defined as: ; Independent branch weights Defined as: ; in, Indicates parameters The corresponding independent branch weights.

[0109] The decoupled parameter output is: ; in, This represents the output result after the parameter 𝑖 has been coupled and decoupled. This indicates the independent branch output of parameter 𝑖. This represents the parameter output generated from the shared backbone features.

[0110] 4. Section reconstruction unit 34; The profile reconstruction unit 34 is connected to the coupling and decoupling unit 33, which is used to perform full-depth interpolation and smoothing on the decoupled multi-dimensional parameter profile data, and output high-precision multi-dimensional parameter inversion profile data without depth jumps and physical violations.

[0111] In this embodiment, the profile reconstruction first uses piecewise cubic Hermite interpolation to recover continuous profile values ​​for unsampled depth points, and then uses constrained smoothing optimization for profile shaping. The optimization objective is: ; in, The final optimized profile sequence is represented by argmin; argmin represents the variable value that minimizes the objective function; Y represents the profile sequence to be optimized. This represents the interpolated profile sequence; represents the second-order difference term of the profile sequence; 𝜇 represents the smoothing coefficient, preferably 0.01.

[0112] Furthermore, within the stable layer, the density difference between adjacent depth layers satisfies: ; in, and These represent the density values ​​at adjacent depth layers.

[0113] IV. Working principle; The working principle of this embodiment is as follows: The sensor data self-compensation module 1 first obtains a uniform steady-state layer reference benchmark through the steady-state layer identification unit 12. The drift modeling unit 11 constructs a nonlinear coupled drift mathematical model based on temperature, pressure, salinity and continuous working time, and outputs the current sensor output drift amount. The online compensation unit performs real-time correction on the original observation data accordingly. The hysteresis correction unit 13 further eliminates the profile position offset caused by the hysteresis of the sensor dynamic response.

[0114] After obtaining high-fidelity observation data, the profile adaptive sampling module 2 determines the key layer, stable layer, transition layer and anomalous change layer based on the vertical gradient of temperature, salinity and density, and switches the sampling strategy according to different layer types; the data ordering unit 24 extracts high information density feature points and removes redundant points while maintaining the complete shape of the profile, and outputs a simplified sampling dataset.

[0115] After the simplified sampled dataset enters the multi-parameter collaborative inversion module 3, the mechanism constraint unit 31 introduces the ocean fluid continuity equation, temperature-salinity diffusion equation and water body radiation transfer equation into the loss function. The lightweight network unit 32 completes real-time inference at the edge. The coupling and decoupling unit 33 adjusts the inversion weights between parameters based on mutual information entropy. The profile reconstruction unit 34 completes full-depth interpolation and smoothing constraints, and finally outputs high-precision multi-parameter inversion profile data without depth jumps and physical violations.

[0116] V. Technical Effects; Compared with Embodiment 1, this embodiment further refines the sensor data self-compensation module 1, the profile adaptive sampling module 2, and the multi-dimensional parameter collaborative inversion module 3, and has at least the following technical effects: Firstly, by introducing a nonlinear coupled drift model of temperature-pressure-salinity-continuous working time, the accuracy and stability of in-situ drift compensation under long-term continuous deep-sea observation conditions are improved. Secondly, through fine-grained hierarchical determination of key layers, stable layers, transition layers and abnormal mutation layers, adaptive sampling that balances the ability to acquire key profile features and sampling efficiency is achieved. Third, through the synergistic effect of the mechanism constraint unit 31, lightweight network unit 32, coupling and decoupling unit 33 and profile reconstruction unit 34, the physical rationality, continuity and robustness of the multidimensional parameter inversion results under complex sea conditions are improved.

[0117] In Example 3, based on Examples 1 and 2, the edge collaborative control module 5 further includes a data scheduling unit 51, a closed-loop feedback unit 52, a low-power communication command unit 53, and a local storage scheduling unit 54; the underwater weak communication optimization module 4 further includes a redundancy fault-tolerant verification unit 44; as shown Figure 2 As shown, the breakpoint resume management unit 42 includes a communication status monitoring subunit 421, a breakpoint recording subunit 422, and an incremental resume subunit 423.

[0118] The core of this embodiment is: taking weak communication state perception as the trigger, the communication state feedback signal is introduced into the edge collaborative control module 5, and further applied to the result output link of the profile adaptive sampling module 2, the multi-dimensional parameter collaborative inversion module 3, and the transmission link of the underwater weak communication optimization module 4, thereby forming a closed-loop collaborative autonomous architecture that runs through the entire link of sensing-sampling-inversion-transmission.

[0119] I. Further implementation methods of edge collaborative control module 5; 1. Data scheduling unit 51; The data scheduling unit 51 is used to uniformly manage the data flow and control flow between the sensor data self-compensation module 1, the profile adaptive sampling module 2, the multi-dimensional parameter collaborative inversion module 3, and the underwater weak communication optimization module 4, so as to ensure that high-fidelity observation data, simplified sampling dataset, high-precision multi-dimensional parameter inversion profile data, as well as core layer data and feature layer data are connected between the modules according to the preset time sequence link.

[0120] In this embodiment, the data scheduling unit 51 is implemented using a combination of "timestamp queue + priority cache + event-triggered scheduling table".

[0121] The timestamp queue is used to attach unified timestamps and depth indexes to high-fidelity observation data, simplified sampling datasets, high-precision multi-dimensional parameter inversion profile data, and hierarchical coded data; the priority cache is used to queue data in layers; and the event-triggered scheduling table is used to trigger sampling strategy adjustments, inversion result output adjustments, and transmission strategy switching based on communication status feedback signals and profile feature feedback signals.

[0122] Preferably, the data scheduling priority is set as follows: The first priority is the core layer data; the second priority is the key layer results and profile feature summary in the high-precision multi-dimensional parameter inversion profile data; the third priority is the simplified sampled dataset; the fourth priority is the high-fidelity observation data; the fifth priority is the feature layer data; and the sixth priority is the redundant layer data.

[0123] Furthermore, the data scheduling unit 51 adjusts the output method of the multi-dimensional parameter collaborative inversion module 3 under different communication states: Under good communication conditions, schedule complete, high-precision, multi-dimensional parameter inversion profile data and feature summaries to enter the transmission link; Under normal communication conditions, full inversion is maintained at the edge, but only key layer results and core feature summaries are prioritized for entry into the real-time transmission link; In the case of extremely poor communication, maintain local full inversion, only schedule the core layer results to enter the transmission link, and transfer the feature layer results to the local cache.

[0124] 2. Closed-loop feedback unit 52; The closed-loop feedback unit 52 is connected to the data scheduling unit 51 and is used to receive the communication status feedback signal output by the underwater weak communication optimization module 4 and the profile feature feedback signal output by the multi-dimensional parameter collaborative inversion module 3. Based on the underwater acoustic communication quality represented by the communication status feedback signal and the profile structure complexity represented by the profile feature feedback signal, it generates sampling strategy adjustment instructions and transmission strategy adjustment instructions.

[0125] In this embodiment, the communication status feedback signal includes at least the current signal-to-noise ratio, current bandwidth, disconnection duration, link recovery flag, and current transmission mode; the profile feature feedback signal includes at least the proportion of key layers, the number of abnormal mutation layers, the maximum normalized gradient value, the location of the jump layer depth, the profile complexity label, and the proportion of core layer data.

[0126] Preferably, the complexity of the cross-sectional structure is classified according to the following threshold: When the cumulative depth of the critical layer and the abnormal mutation layer is less than 15% and the maximum normalized gradient value is less than 1.0, it is judged as a low-complexity profile. When the cumulative depth of the critical layer and the abnormal mutation layer is greater than or equal to 15% and less than 35%, or the maximum normalized gradient value is greater than or equal to 1.0 and less than 2.0, it is judged as a medium complexity profile. When the cumulative depth of the critical layer and the abnormal mutation layer is greater than or equal to 35%, or the maximum normalized gradient value is greater than or equal to 2.0, it is judged as a high-complexity profile.

[0127] The maximum normalized gradient value is taken as the maximum value after normalization according to the temperature, salinity and density gradient thresholds in Example 2.

[0128] In this embodiment, the closed-loop feedback unit 52 generates control commands according to the following causal response rules: When the communication status is good and the current profile is a high-complexity profile, output the adjustment command "key layer encrypted sampling and preservation + full core layer data and feature layer data upload"; When the communication status is normal and the current profile is a high-complexity profile, output the adjustment command "key layer encryption sampling maintenance + stable layer sampling period extension + only upload core layer data"; When the communication status is extremely poor, the following adjustment instructions are output: "Prioritize the retention of samples from critical layers and abnormal mutation layers + extend the sampling period of stable layers + suspend active uploading + transfer feature layer data and redundant layer data to local cache". When the communication status recovers from a poor communication state to a normal or good communication state, the adjustment command "Resume breakpoint resume + Gradually release buffered feature layer data + Restore normal stable layer sampling period" is output.

[0129] Preferably, the start-up period of the profile containing the stable layer under the poor communication condition is extended from 30 minutes to 60 minutes, and then restored to 30 minutes after the communication condition is restored to good; the critical layer and the abnormal mutation layer still maintain the sampling step size recorded in Example 1 and Example 2 within a single profile.

[0130] 3. Low-power communication instruction unit 53; The low-power communication instruction unit 53 is used to send a communication mode switching instruction or a breakpoint resume trigger instruction to the underwater weak communication optimization module 4 according to the underwater acoustic communication quality, so as to control the underwater weak communication optimization module 4 to enter a data transmission state that matches the current underwater acoustic communication quality.

[0131] In this embodiment, the low-power communication instruction unit 53 supports the following three communication modes: Mode A is the full-priority transmission mode, corresponding to a good communication status; Mode B is the core-priority transmission mode, corresponding to the general communication state; Mode C is the local cache waiting mode, which corresponds to a very poor communication state.

[0132] Preferably, the control word output by the low-power communication instruction unit 53 includes at least a mode field, a continuation trigger field, a priority field, and a timestamp field.

[0133] Among them, the mode field is used to identify the current communication mode, the resume trigger field is used to identify whether to start the breakpoint resume, the priority field is used to identify the data level that is currently sent first, and the timestamp field is used to ensure the consistency of the execution order of control commands.

[0134] Preferably, in normal communication state and poor communication state, the low-power communication instruction unit 53 adopts the working mode of "batch wake-up, centralized transmission, and fast sleep"; when the pending buffer reaches 16 data blocks, a centralized transmission attempt is triggered. If the current link does not meet the transmission conditions, it re-enters the low-power waiting state.

[0135] 4. Local storage scheduling unit 54; The local storage scheduling unit 54 is used to perform hierarchical storage management of raw observation data, high-fidelity observation data, simplified sampling datasets and high-precision multidimensional parameter inversion profile data. When the local storage space is detected to be lower than the preset storage threshold, the core layer data and high-precision multidimensional parameter inversion profile data are retained first, and non-core data are rolled over and stored in order of data timestamp.

[0136] In this embodiment, the local storage scheduling unit 54 divides the local storage space into a raw observation data partition, a high-fidelity observation data partition, a simplified sampled dataset partition, a high-precision multidimensional parameter inversion profile data partition, and a breakpoint log and pending transmission cache partition.

[0137] Preferably, the preset storage threshold is triggered when the remaining local available space is less than 15% of the total storage space or less than 2GB.

[0138] After the storage threshold is triggered, the following management rules will be executed: Core layer data and high-precision multidimensional parameter inversion profile data are retained at the highest level and are not included in the coverage. The feature layer data is the second highest retention level; The original observation data, high-fidelity observation data, and redundant layer data are rolled over in a time-stamp manner from oldest to newest. The breakpoint log file will not be overwritten until the corresponding data segment has been transmitted and verified by the receiving end.

[0139] Preferably, the minimum granularity of rolling coverage is one data block, and the data block size is 128 bytes / block.

[0140] II. Further implementation of underwater weak communication optimization module 4; 1. Hierarchical data encoding unit 41; The hierarchical data encoding unit 41 is used to divide the high-precision multi-dimensional parameter inversion profile data into core layer data, feature layer data and redundant layer data according to the profile layer type and the degree of drastic change of parameter gradient corresponding to each data block in the high-precision multi-dimensional parameter inversion profile data, and to adopt different encoding strategies for each.

[0141] In this embodiment, the hierarchical rules are as follows: When a data block corresponds to a critical layer or an anomalous mutation layer, or when the normalized gradient strength is greater than or equal to 1.0, it is classified as core layer data. When a data block corresponds to a transition layer, or when the normalized gradient strength is greater than or equal to 0.4 and less than 1.0, it is classified as feature layer data. When a data block corresponds to a stable layer and the normalized gradient strength is less than 0.4, it is classified as redundant layer data.

[0142] The normalized gradient intensity is taken as the maximum value among the three normalized gradient values ​​of temperature, salinity, and density.

[0143] Furthermore, the core layer data is encoded using a lossless compression algorithm, preferably the improved LZ77 encoding.

[0144] Preferably, before entering LZ77 encoding, differential encoding is performed on the timestamp sequence, depth index sequence, and block number sequence, and then input together with the core parameter byte stream into the LZ77 encoder; the sliding window size is set to 4KB, the lookahead buffer size is set to 256B, and the minimum matching length is set to 4B; the encoded output adopts the "offset-match length-successor byte" triple structure.

[0145] The feature layer data is encoded using a lightweight lossy compression algorithm.

[0146] In this embodiment, linear prediction is first performed on the continuous depth layer data, then fixed-step quantization is performed on the prediction residuals, and finally run-length encoding is performed on the quantization results.

[0147] Preferably, the quantization step size is set as follows: temperature 0.01℃, salinity 0.005psu, flow rate 0.01m / s, turbidity 0.05NTU, and sound velocity 0.05m / s.

[0148] Redundant layer data is either discarded or periodically packaged.

[0149] Preferably, only the first and last sampling points are retained for every 5m depth interval, and the remaining duplicate points are discarded; if it is necessary to retain the redundant layer summary, a redundant layer statistical package is generated every 30 minutes.

[0150] 2. Redundant fault-tolerant verification unit 44; The redundancy fault-tolerant verification unit 44 is connected to the hierarchical data encoding unit 41 and is used to add cyclic redundancy check codes to the encoded core layer data and feature layer data, and to generate an additional XOR redundancy backup package for the core layer data.

[0151] In this embodiment, both the core layer data and the feature layer data are divided into blocks of 128 bytes each.

[0152] The core layer data is given a CRC-32 cyclic redundancy check code, and the feature layer data is given a CRC-16-CCITT cyclic redundancy check code.

[0153] Preferably, the CRC-32 generator polynomial is 0x04C11DB7, and the CRC-16-CCITT generator polynomial is 0x1021.

[0154] Furthermore, for the core layer data, the redundancy fault-tolerant verification unit 44 generates an XOR redundant backup package according to a group of four core layer data blocks. If the four core layer data blocks within the same group are respectively denoted as... , , , Then the XOR redundancy backup package P satisfies: ; Where P represents the XOR redundancy backup package, and ⊕ represents the bitwise XOR operation.

[0155] When the receiving end detects that only one core layer data block is lost in the same redundancy group, it can be recovered as follows: If If lost, then there is ; When the number of lost blocks in the same redundancy group is greater than 1, XOR recovery is not performed, and the receiving end initiates a retransmission request for the missing blocks in that group.

[0156] For feature layer data, when CRC check fails or data block is lost, a retransmission request for the corresponding data block is triggered directly, without using XOR recovery.

[0157] 3. Resumable download management unit 42; The breakpoint resume management unit 42 includes a communication status monitoring subunit 421, a breakpoint recording subunit 422, and an incremental resume subunit 423.

[0158] (1) Communication status monitoring subunit 421; The communication status monitoring subunit 421 is used to monitor the signal-to-noise ratio, bandwidth and disconnection duration of the underwater acoustic communication link in real time. When the signal-to-noise ratio is lower than the preset signal-to-noise ratio threshold or the bandwidth is lower than the preset bandwidth threshold, the breakpoint recording operation is triggered.

[0159] In this embodiment, the communication status monitoring subunit 421 samples the link status every 2 seconds and uses a 10-second sliding window to calculate the average signal-to-noise ratio and average bandwidth.

[0160] Preferably, when the signal-to-noise ratio (SNR) is greater than or equal to 15 dB and the bandwidth is greater than or equal to 2.0 kbps, the communication is considered to be in a good state; when the SNR is greater than or equal to 5 dB and less than 15 dB, or the bandwidth is greater than or equal to 0.5 kbps and less than 2.0 kbps, the communication is considered to be in a normal state; when the SNR is less than 5 dB, or the bandwidth is less than 0.5 kbps, the communication is considered to be in a very poor state.

[0161] The preset signal-to-noise ratio threshold is 5dB, and the preset bandwidth threshold is 0.5kbps.

[0162] If the continuous disconnection lasts for more than 30 seconds, the breakpoint recording operation will be triggered directly.

[0163] (2) Breakpoint recording subunit 422; The breakpoint recording subunit 422 is connected to the communication status monitoring subunit 421. When a breakpoint recording trigger signal is received, it records the successfully transmitted data segment number, transmission timestamp, and data segment checksum in the local storage medium to form a breakpoint log file.

[0164] In this embodiment, the breakpoint log file adopts a structure of "file header + segment index area + block status area + checksum area".

[0165] The file header should include at least the task number, profile number, log version number, block size field, and log generation timestamp; the segment index area should include at least the data segment number, data level identifier, start block number, end block number, and redundancy group number; the block status area should include at least the transmission status flag, reception acknowledgment flag, retransmission count, and transmission timestamp for each data block; and the checksum area should include at least the CRC checksum of the corresponding data segment, the XOR redundancy backup packet index, and the final acknowledgment status.

[0166] The sending status flags include at least four types of statuses: "not sent, sent but not acknowledged, acknowledged, and awaiting retransmission".

[0167] (3) Incremental transmission subunit 423; The incremental resume transmission subunit 423 is connected to the breakpoint recording subunit 422. After detecting the restoration of the underwater acoustic communication link, it reads the breakpoint log file and uploads only the remaining data segments that have not been transmitted to the receiving end according to the transmitted data segment number recorded in the breakpoint log file, so as to realize incremental resume transmission.

[0168] In this embodiment, the link recovery conditions are set as follows: the signal-to-noise ratio value recovers to greater than or equal to 5dB and the bandwidth value recovers to greater than or equal to 0.5kbps, and remains there for 20 seconds.

[0169] Once the recovery conditions are met, the incremental resume transmission subunit 423 organizes the data to be resumed in the order of "core layer data priority, old timestamp priority, and small remaining amount priority".

[0170] If some blocks in a data segment have been successfully transmitted and confirmed by the receiving end, only the incomplete blocks will be retransmitted; if a single block is lost in a core layer redundancy group and the receiving end can recover it through XOR redundancy backup packets, the lost block will not be retransmitted.

[0171] 4. Communication state adaptive unit 43; The communication state adaptive unit 43 is connected to the incremental transmission subunit 423, and is used to divide the current underwater acoustic communication link into a good communication state, a normal communication state, or a very poor communication state according to the signal-to-noise ratio and bandwidth value, and execute the corresponding transmission strategy for different communication states.

[0172] In this embodiment, core layer data and feature layer data are transmitted when communication is good; only core layer data is transmitted when communication is normal, and feature layer data is temporarily stored locally; when communication is extremely poor, active uploading is suspended, and the data to be transmitted is encrypted and stored locally.

[0173] Preferably, the local cached data is written to the storage medium after being encrypted using AES-128 symmetric encryption.

[0174] Furthermore, the communication state adaptive unit 43 feeds back the current communication state to the edge collaborative control module 5, enabling the data scheduling unit 51 and the closed-loop feedback unit 52 to further adjust the sampling strategy and the output link of the inversion results.

[0175] III. Working Principle; The working principle of this embodiment is as follows: using the underwater acoustic link status as the trigger source, the communication status feedback signal is introduced into the edge collaborative control module 5, and the closed-loop feedback unit 52 transforms it into a joint adjustment of the sampling strategy, the output mode of the inversion result, and the transmission strategy, thereby constructing a closed-loop collaborative autonomous link that runs through the entire process of sensing, sampling, inversion, and transmission.

[0176] Specifically, after the multi-dimensional parameter collaborative inversion module 3 outputs high-precision multi-dimensional parameter inversion profile data, the hierarchical data encoding unit 41 first divides the data into core layer data, feature layer data and redundant layer data according to the profile layering type and the degree of drastic change in parameter gradient, so that the inversion results have been layered according to information value before entering the communication link.

[0177] Subsequently, the communication status monitoring subunit 421 continuously detects the current signal-to-noise ratio, bandwidth, and disconnection duration, and sends the judgment results to the communication status adaptive unit 43 and the closed-loop feedback unit 52.

[0178] When the communication status is good, the system maintains encrypted sampling of the key layer and outputs complete inversion results, while simultaneously transmitting core layer data and feature layer data. When the communication status deteriorates to a normal communication status, the system maintains encrypted sampling of the key layer, extends the sampling period of the stable layer, and prioritizes scheduling core layer data into the real-time transmission link, while transferring feature layer data to the local cache. When the communication status further deteriorates to a very poor communication status, the system prioritizes retaining the sampling and result output of the key layer and the abnormal mutation layer, suspends active uploading, writes the data to be transmitted into the local encrypted cache, and improves the retention level of core layer data and high-precision multidimensional parameter inversion profile data through the local storage scheduling unit 54.

[0179] In this process, communication status feedback not only changes the transmission object, but also directly causes changes in sampling rhythm, changes in the granularity of real-time output of inversion results, and changes in local storage strategy. Therefore, a clear technical correlation and causal response relationship is formed between communication status feedback and sampling / inversion / transmission strategy.

[0180] During the transmission execution phase, the redundancy fault-tolerant verification unit 44 adds a CRC-32 checksum and an XOR redundancy backup packet to the core layer data, and adds a CRC-16-CCITT checksum to the feature layer data. If the receiving end detects the loss of a single block of core layer data, it directly recovers it through the XOR redundancy backup packet; if it detects the loss of feature layer data, it initiates a retransmission request block by block. If the link is interrupted or the bandwidth drops sharply, the breakpoint recording subunit 422 records the completed transmission blocks and their status according to the breakpoint log file structure; after the link is restored, the incremental resume transmission subunit 423 only retransmits the incomplete blocks according to the breakpoint log file.

[0181] Thus, this embodiment realizes end-to-end collaborative autonomy, from communication state awareness, closed-loop decision-making, sampling and result output adjustment, hierarchical coding and fault-tolerant transmission, to breakpoint resumption and local retention.

[0182] IV. Technical Effects; Compared with Embodiment 1 and Embodiment 2, this embodiment further has the following technical effects: First, by using the communication status feedback signal as the closed-loop trigger source, the weak communication status perception is directly mapped to the adjustment of sampling strategy, the adjustment of the output mode of inversion result, and the switching of transmission strategy, thereby constructing a collaborative autonomous architecture that runs through the entire link of sensing, sampling, inversion and transmission. Secondly, under extremely poor communication conditions, the system does not simply reduce the overall workload, but prioritizes the preservation of information acquisition capabilities of the critical layer and the abnormal mutation layer, and simultaneously increases the priority of saving core layer data and high-precision multi-dimensional parameter inversion profile data, thereby improving the robustness of information acquisition under extreme environments. Third, through a joint mechanism of hierarchical data encoding, CRC check, XOR redundancy recovery, breakpoint log management and incremental resume transmission, the core layer data’s priority transmission capability, priority recovery capability and final complete reachability in weak communication environments are improved. Fourth, through the local storage scheduling unit 54 and the breakpoint log file structure, the long-term traceability, recoverability and retransmission of key observation results are realized, thereby improving the security and integrity of key results data in long-term unattended marine observation scenarios.

[0183] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A system for the collaborative inversion of multi-dimensional parameters of a vertical profile of a marine area, characterized in that, include: The sensor data self-compensation module is used to perform in-situ drift compensation on the raw observation data and output high-fidelity observation data. The profile adaptive sampling module is connected to the sensor data self-compensation module. It is used to calculate vertical gradient data based on the high-fidelity observation data and divide the profile layer type. Based on the closed-loop feedback control logic, it performs differentiated sampling for different layer types and outputs a simplified sampling dataset. The multidimensional parameter collaborative inversion module is connected to the profile adaptive sampling module. It is used to receive the simplified sampling dataset, drive a lightweight hybrid neural network model with physical constraint terms containing the ocean hydrodynamic conservation equation, and perform coupling and decoupling through adaptive weight allocation to obtain high-precision multidimensional parameter inversion profile data. An underwater weak communication optimization module, connected to the multi-dimensional parameter collaborative inversion module, includes: A hierarchical data encoding unit is used to encode the high-precision multidimensional parameter inversion profile data into core layer data and feature layer data according to the importance level of the high-precision multidimensional parameter inversion profile data. The breakpoint resume management unit is used to generate breakpoint logs and perform incremental resume transmission of incomplete data segments after communication is restored. The communication status adaptive unit is connected to the hierarchical data encoding unit and the breakpoint resume management unit, respectively, and is used to dynamically adjust the transmission and caching strategy of the core layer data and the feature layer data according to the underwater acoustic communication quality, and output the communication status feedback signal. The edge collaborative control module is connected to the sensor data self-compensation module, the profile adaptive sampling module, the multi-dimensional parameter collaborative inversion module, and the underwater weak communication optimization module, respectively. It is used to receive the communication status feedback signal and the profile feature feedback signal output by the multi-dimensional parameter collaborative inversion module, and to uniformly schedule the sampling strategy parameters of the profile adaptive sampling module and the transmission parameters of the underwater weak communication optimization module.

2. The multi-dimensional parameter collaborative inversion system for vertical marine profiles according to claim 1, characterized in that, The sensor data self-compensation module includes a drift modeling unit, which is used to construct a drift mathematical model that couples temperature parameters, pressure parameters and salinity parameters, and to perform online drift correction on the original observation data based on the drift mathematical model. The input variables of the drift mathematical model include temperature deviation, static pressure deviation, salinity deviation, and continuous working time. The output of the drift mathematical model is the sensor output drift. The coupling coefficient in the drift mathematical model is predetermined through offline calibration.

3. The multi-dimensional parameter collaborative inversion system for vertical marine profiles according to claim 1, characterized in that, The sensor data self-compensation module includes a steady-state layer identification unit and a hysteresis correction unit; The steady-state layer identification unit is used to calculate the vertical gradient value of water parameters at each depth layer in the original observation data in real time, and identify the continuous depth interval where the vertical gradient value of water parameters is less than a preset gradient threshold as a uniform steady-state layer, and use the average value of the observation data in the uniform steady-state layer as the natural reference benchmark for the in-situ drift compensation. The hysteresis correction unit is used to perform dynamic hysteresis correction on the data after in-situ drift compensation using a second-order inertial element, so as to eliminate the profile depth offset distortion introduced by sensor response hysteresis.

4. The multi-dimensional parameter collaborative inversion system for marine vertical profiles according to claim 1, characterized in that, The profile adaptive sampling module includes a gradient calculation unit, a layer determination unit, a dynamic control unit, and a data order preservation unit; The gradient calculation unit is used to calculate the vertical gradient value of the temperature, salinity and density parameters in the high-fidelity observation data layer by layer. The layer determination unit is connected to the gradient calculation unit and is used to compare the vertical gradient value of each depth layer with a preset first gradient threshold and a preset second gradient threshold respectively. When the vertical gradient value is greater than the preset first gradient threshold, it is determined to be a critical layer. When the vertical gradient value is less than the preset second gradient threshold, it is determined to be a stable layer. When the vertical gradient value is between the preset first gradient threshold and the preset second gradient threshold, it is determined to be a transition layer. When the vertical gradient value jumps between adjacent depth layers by more than a preset abrupt change threshold, it is determined to be an abnormal abrupt change layer. The dynamic control unit is connected to the hierarchical determination unit and is used to execute an encrypted sampling strategy for the critical layer, a sparse sampling strategy for the stable layer, and a fixed-point retesting strategy for the abnormal mutation layer based on the closed-loop feedback control logic. The data ordering unit is connected to the dynamic control unit and is used to extract feature points in each depth layer whose information density exceeds a preset density threshold and remove redundant sampling points, so as to output the simplified sampling dataset while preserving the complete morphological features of the profile.

5. The multi-dimensional parameter collaborative inversion system for marine vertical profiles according to claim 1, characterized in that, The multidimensional parameter collaborative inversion module includes a mechanism constraint unit and a lightweight network unit; The mechanism constraint unit is used to construct the ocean fluid continuity equation, temperature-salinity diffusion equation and water body radiation transfer equation as physical constraint terms, and to embed the physical constraint terms into the loss function of the lightweight hybrid neural network model in the form of regularization terms, so that the lightweight hybrid neural network model is guided by the physical constraint terms during training and inference, and outputs initial multidimensional inversion data that meets the requirements of physical conservation and profile self-consistency. The lightweight network unit is used to carry the lightweight hybrid neural network model, which adopts a hybrid structure of convolutional neural network and lightweight Transformer. The storage amount of model parameters is less than a preset storage threshold to support real-time inference on embedded microcontrollers.

6. The multi-dimensional parameter collaborative inversion system for vertical marine profiles according to claim 5, characterized in that, The multi-dimensional parameter collaborative inversion module includes a coupling decoupling unit and a profile reconstruction unit; The coupling and decoupling unit is used to calculate the mutual information entropy value between each parameter in the initial multidimensional inversion data, and to assign corresponding inversion weights to each parameter according to the mutual information entropy value, so as to eliminate the strong coupling interference between temperature parameter, salinity parameter, turbidity parameter and flow velocity parameter, and obtain decoupled multidimensional parameter profile data. The profile reconstruction unit is connected to the coupling and decoupling unit and is used to perform full-depth interpolation and smoothing on the decoupled multidimensional parameter profile data to output high-precision multidimensional parameter inversion profile data without depth jumps or physical violations.

7. The multi-dimensional parameter collaborative inversion system for vertical marine profiles according to claim 1, characterized in that, The edge collaborative control module includes a data scheduling unit and a closed-loop feedback unit; The data scheduling unit is used to uniformly manage the data flow and control flow between the sensor data self-compensation module, the profile adaptive sampling module, the multi-dimensional parameter collaborative inversion module, and the underwater weak communication optimization module, to ensure that the high-fidelity observation data, the simplified sampling dataset, the high-precision multi-dimensional parameter inversion profile data, and the core layer data and the feature layer data are connected between the modules according to a preset time sequence link. The closed-loop feedback unit is connected to the data scheduling unit and is used to receive the communication status feedback signal output by the underwater weak communication optimization module and the profile feature feedback signal output by the multi-dimensional parameter collaborative inversion module. Based on the underwater acoustic communication quality represented by the communication status feedback signal and the profile structure complexity represented by the profile feature feedback signal, it generates sampling strategy adjustment instructions and transmission strategy adjustment instructions.

8. The multi-dimensional parameter collaborative inversion system for vertical marine profiles according to claim 1, characterized in that, The edge collaborative control module includes a low-power communication instruction unit and a local storage scheduling unit; The low-power communication instruction unit is used to send a communication mode switching instruction or a breakpoint resume trigger instruction to the underwater weak communication optimization module according to the underwater acoustic communication quality, so as to control the underwater weak communication optimization module to enter a data transmission state that matches the current underwater acoustic communication quality. The local storage scheduling unit is used to perform hierarchical storage management of the original observation data, the high-fidelity observation data, the simplified sampling dataset, and the high-precision multidimensional parameter inversion profile data. When the local storage space is detected to be lower than the preset storage threshold, the core layer data and the high-precision multidimensional parameter inversion profile data are retained first, and the non-core data are rolled over and stored in order of data timestamp.

9. The multi-dimensional parameter collaborative inversion system for vertical marine profiles according to claim 4, characterized in that, The underwater weak communication optimization module also includes a redundant fault-tolerant verification unit; The hierarchical data encoding unit is specifically used to divide the high-precision multi-dimensional parameter inversion profile data into the core layer data, the feature layer data, and the redundant layer data according to the profile layer type and the degree of drastic change of parameter gradient corresponding to each data block in the high-precision multi-dimensional parameter inversion profile data. The core layer data is encoded using a lossless compression algorithm, the feature layer data is encoded using a lossy compression algorithm, and the redundant layer data is discarded or periodically packaged. The redundancy fault-tolerant verification unit is connected to the hierarchical data encoding unit and is used to add cyclic redundancy check codes to the encoded core layer data and feature layer data, and to generate an additional XOR redundancy backup packet for the core layer data. When the receiving end detects that the core layer data has been lost, it recovers the lost data through the XOR redundancy backup packet. When the receiving end detects that the feature layer data has been lost, it triggers a retransmission request for the corresponding data block.

10. The multi-dimensional parameter collaborative inversion system for marine vertical profiles according to claim 1, characterized in that, The breakpoint resume management unit includes a communication status monitoring subunit, a breakpoint recording subunit, and an incremental resume subunit. The communication status monitoring subunit is used to monitor the signal-to-noise ratio, bandwidth, and disconnection duration of the underwater acoustic communication link in real time. When the signal-to-noise ratio is lower than a preset signal-to-noise ratio threshold or the bandwidth is lower than a preset bandwidth threshold, a breakpoint recording operation is triggered. The breakpoint recording subunit is connected to the communication status monitoring subunit and is used to record the successfully transmitted data segment number, transmission timestamp and data segment check code in the local storage medium when a breakpoint recording trigger signal is received, forming a breakpoint log file. The incremental resume transmission subunit is connected to the breakpoint recording subunit. After detecting the restoration of the underwater acoustic communication link, it reads the breakpoint log file and, based on the transmitted data segment number recorded in the breakpoint log file, uploads only the remaining data segments that have not been transmitted to the receiving end to achieve incremental resume transmission. The communication state adaptive unit is connected to the incremental transmission subunit and is used to divide the current underwater acoustic communication link into a good communication state, a normal communication state, or a very poor communication state according to the signal-to-noise ratio value and the bandwidth value, and to execute corresponding transmission strategies for different communication states. Specifically, in the good communication state, the core layer data and the feature layer data are transmitted; in the normal communication state, only the core layer data is transmitted and the feature layer data is temporarily stored locally; and in the very poor communication state, active uploading is paused and the data to be transmitted is encrypted and stored locally.

Images