Real-time deep sea subsurface buoy monitoring system

Abstract

The invention provides a real-time deep sea subsurface buoy monitoring system, which belongs to the technical field of deep sea measurement, and comprises a control chip, a multi-parameter sensor array, a data storage device and a power supply, high-precision monitoring is realized through the following steps: recording ocean data acquired by a multi-parameter sensor; correcting the sound wave propagation model and calculating the relative position of the subsurface buoy; constructing a three-dimensional data matrix and realizing data compression; identifying abnormal data by using a matrix disorder index; processing the time sequence data by adopting a sliding window weighted average method and eliminating jump; carrying out distributed preprocessing on the multi-source heterogeneous data and carrying out parallel processing by applying a matrix partitioning technology; accumulated errors are corrected by combining historical data and applying a Kalman filtering algorithm, and the problems of environment interference and sensor drift are solved through a deep sea environment disturbance compensation network model and a self-adaptive weighted error optimization function.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep sea measurement, and in particular relates to a real-time deep sea buoy monitoring system. Background Art

[0002] Deep-sea submersible monitoring systems are essential tools for marine scientific research and environmental monitoring. Traditional technologies primarily deploy multi-parameter sensor arrays to collect multi-dimensional data such as ocean temperature, salinity, and currents. Existing submersible systems typically utilize a timed sampling, local storage, and periodic data transmission model, playing a fundamental role in the global ocean observation network. While traditional deep-sea submersible technology is relatively mature in shallow waters and nearshore areas, its application in the complex environments of the deep sea is significantly limited.

[0003] Traditional deep-sea buoy monitoring systems have many technical defects: first, sensors are prone to measurement deviations in the deep-sea high-pressure, low-temperature and high-salinity environment, and as the operating time increases, sensor drift and aging phenomena intensify, resulting in a continuous decline in measurement data accuracy; second, there is coupling interference between multi-parameter sensors, and the measurement error characteristics vary under different environmental conditions, making it difficult to establish a unified calibration model; third, traditional data processing methods have high computational complexity and low processing efficiency, and cannot meet the needs of real-time processing of large-scale data.

[0004] Existing technologies primarily use simple filtering algorithms and periodic manual calibration to address measurement errors. These methods struggle to cope with the complex interference and long-term sensor drift found in deep-sea environments. In particular, they lack effective means to address nonlinear and dynamically changing measurement errors caused by environmental factors. This leads to a significant degradation in data quality over long periods of time and makes it impossible to support the continuous and reliable acquisition of high-precision ocean data. In other words, existing technologies face technical challenges in deep-sea environments, where sensor measurement data is subject to multiple sources of interference, resulting in reduced accuracy and low data processing efficiency. Summary of the Invention

[0005] In view of this, the present invention provides a real-time deep-sea buoy monitoring system, which can solve the technical problems in the prior art that sensor measurement data in deep-sea environments is interfered with by multiple sources, resulting in reduced accuracy and low data processing efficiency.

[0006] The present invention is implemented as follows: the present invention provides a real-time deep-sea buoy monitoring system including a control chip, a multi-parameter sensor array, a data storage device and a power supply; the control chip is electrically connected to the multi-parameter sensor array, the data storage device and the power supply respectively; the multi-parameter sensor array is used to collect ocean temperature stratification data, salinity distribution data and current velocity data; the data storage device is used to record all collected monitoring data and system operating status parameters; the power supply is used to provide stable power for the entire system; a monitoring data recording module is provided in the control chip to process the raw data collected by the multi-parameter sensor array; the monitoring data recording module introduces an error optimization function based on an adaptive weighted algorithm and a deep-sea environmental disturbance compensation network model to identify error sources and propagation paths, establish an association model between environmental parameters and measurement deviations, dynamically adjust data fusion weights, eliminate coupling interference and data offset between different sensors, suppress error accumulation effects, obtain optimized data and save them to the data storage device.

[0007] Among them, the monitoring data recording module is used to perform the following steps: recording ocean temperature stratification data, salinity distribution data and current velocity data collected by the multi-parameter sensor array; using the temperature stratification data collected by the multi-parameter sensor array to correct the sound wave propagation model, and using trilateration to calculate the precise relative position between each buoy based on the signal round-trip delay time and the received signal strength index; constructing the collected ocean data into a three-dimensional data matrix, extracting the main change patterns through matrix singular value decomposition, eliminating data redundancy, and achieving data compression through low-rank matrix approximation.

[0008] Among them, the monitoring data recording module is also used to perform the following steps: apply the matrix sliding window method to process the time series data, establish the covariance matrix of the multi-dimensional measurement data, calculate the matrix disorder index, and identify abnormal data patterns; based on the correlation model of environmental parameters and measurement deviations, analyze the impact of environmental factors on the measurement accuracy of the sensor; continuously monitor the data flow, perform time series analysis on the collected raw data, identify the continuity error trend, and use the sliding window weighted average method to eliminate data jumps; optimize the data transmission network structure between buoys based on the minimum spanning tree algorithm to minimize the communication energy consumption of the multi-buoy system.

[0009] Among them, the monitoring data recording module is also used to perform the following steps: distributed preprocessing of the collected multi-source heterogeneous data, parallel processing of large-scale data using matrix block computing technology, extraction of key features, compression of repeated redundant information, and reduction of data storage volume; based on the historical data in the data storage device and combined with real-time measurement values, the Kalman filter algorithm is used to model and predict the cumulative errors, thereby realizing online correction and optimization of the measurement data.

[0010] Among them, the sliding window weighted average method specifically sets a data window with a length of 30 minutes on the time series, assigns a time-attenuated weight to each data point in the window, and the data point closer to the current moment has a greater weight, and replaces the original measurement value with the weighted average value. The window slides forward every 10 seconds, continuously updating the data series, effectively eliminating short-term random fluctuations.

[0011] Among them, the Kalman filter algorithm specifically establishes a system state space model, models the measurement noise and system noise as Gaussian white noise respectively, and iteratively calculates the system state estimate and covariance matrix through prediction steps and update steps. The prediction step uses the state at the previous moment to estimate the current state, and the update step corrects the prediction result in combination with the current measurement value, and finally outputs the optimal state estimate. The Kalman filter algorithm is suitable for processing the cumulative error caused by sensor drift.

[0012] Among them, the continuity error trend identification specifically establishes a normal fluctuation range model for each sensor data, calculates the first-order difference and second-order difference of 20 adjacent data points, and when the difference value of 5 consecutive data points exceeds 3 times the standard deviation of the normal range and shows a monotonically increasing or monotonically decreasing trend, it is determined that there is a continuity error and the error repair process is triggered.

[0013] Among them, the matrix singular value decomposition specifically constructs the ocean environment parameters into a three-dimensional tensor of dimension m×n×t, where m represents the horizontal spatial dimension, n represents the vertical depth dimension, and t represents the time dimension. The tensor decomposition method is used to decompose the original high-dimensional data into the product form of three low-dimensional matrices, retaining the eigenvectors corresponding to the most significant k singular values, and reconstructing the compressed low-rank approximate matrix to achieve data dimensionality reduction and retain the main change patterns. The compression ratio reaches 15% of the original data volume.

[0014] Among them, the matrix disorder index specifically calculates the ratio of the condition number of the measurement data covariance matrix to the matrix spectral norm, quantifies the uniformity of the distribution of the matrix eigenvalues, and the larger the condition number, the higher the data disorder. When the matrix disorder index exceeds the preset threshold of 10, the anomaly detection process is triggered.

[0015] The correlation model between environmental parameters and measurement deviations specifically utilizes a multivariate regression analysis method, with temperature, pressure, and salinity as independent variables and various measurement deviations as dependent variables, to fit the influence function of environmental factors on measurement accuracy. When the environmental parameter changes exceed a preset threshold, the correction coefficient is automatically applied to calibrate the original measurement value. In the process of establishing and applying the correlation model, a pre-trained deep-sea environmental disturbance compensation network model is introduced. The specific structure of the deep-sea environmental disturbance compensation network model is composed of a six-layer deep neural network, including two input encoding layers, three hidden processing layers, and one output decoding layer. The input encoding layer is responsible for receiving and standardizing environmental parameter data, the hidden processing layer is embedded with a sparse attention mechanism to capture the long-term correlation and short-term fluctuation characteristics between different environmental parameters, and the output decoding layer generates a correction coefficient matrix for various sensors.

[0016] Compared to existing technologies, the present invention provides a real-time deep-sea buoy monitoring system. Through multi-layered data processing and error optimization, this system achieves high-precision acquisition and efficient processing of sensor data in deep-sea environments. The system integrates a multi-parameter sensor array, an intelligent control chip, a data storage device, and a stable power supply, creating a complete solution for monitoring marine environmental parameters.

[0017] This system overcomes the core defects of traditional technologies: on the one hand, it models environmental parameters and measurement deviations through a deep-sea environmental disturbance compensation network model, achieves precise calibration under different environmental conditions, and effectively eliminates the impact of environmental factors on measurement accuracy; on the other hand, it applies advanced data processing methods such as matrix singular value decomposition, adaptive weighted algorithm and Kalman filtering to effectively identify and eliminate sensor drift, coupling interference and random fluctuations, ensuring data quality in long-term operation; at the same time, through distributed preprocessing and matrix block computing technology, it significantly improves data processing efficiency, reduces storage requirements, and realizes efficient processing of large-scale data.

[0018] This invention solves the technical problem of reduced accuracy of sensor measurement data in deep-sea environments due to multi-source interference, establishes a complete technical path from error identification, environmental impact modeling to data optimization processing, enabling the system to maintain high-precision measurement capabilities in complex deep-sea environments, and guarantees system performance through efficient data processing algorithms, providing technical support for long-term continuous monitoring of deep-sea environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the method of the present invention.

[0020] Figure 2 Schematic diagram of the system composition in Example 1. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0022] The present invention provides a real-time deep-sea buoy monitoring system comprising a control chip, a multi-parameter sensor array, a data storage device, and a power supply; the control chip is electrically connected to the multi-parameter sensor array, the data storage device, and the power supply, respectively; the multi-parameter sensor array is used to collect ocean temperature stratification data, salinity distribution data, and current velocity data; the data storage device is used to record all collected monitoring data and system operating status parameters; the power supply is used to provide stable power for the entire system; and the sampling frequency of the multi-parameter sensor array is once every 5 minutes.

[0023] The control chip is provided with a monitoring data recording module, such as Figure 1 As shown, the monitoring data recording module is used to perform the following steps:

[0024] S01, record the ocean temperature stratification data, salinity distribution data, and ocean current velocity data collected by the multi-parameter sensor array;

[0025] S02, using the temperature layered data collected by the multi-parameter sensor array to calibrate the acoustic wave propagation model, and using trilateration to calculate the precise relative positions of the buoys based on the round-trip signal delay time and the received signal strength index;

[0026] S03. Construct the collected ocean data into a three-dimensional data matrix, extract the main change patterns through matrix singular value decomposition, eliminate data redundancy, and achieve data compression through low-rank matrix approximation;

[0027] S04. Apply the matrix sliding window method to process the time series data, establish the covariance matrix of the multidimensional measurement data, calculate the matrix disorder index, and identify abnormal data patterns;

[0028] S05. Based on the correlation model between environmental parameters and measurement deviation, analyze the impact of environmental factors on sensor measurement accuracy;

[0029] S06. Continuously monitor data streams, perform time series analysis on collected raw data, identify continuity error trends, and use a sliding window weighted average method to eliminate data jumps;

[0030] S07. Optimize the data transmission network structure between buoys based on the minimum spanning tree algorithm to minimize the communication energy consumption of the multi-buoy system;

[0031] S08. Perform distributed pre-processing on the collected multi-source heterogeneous data, apply matrix block computing technology to parallel process large-scale data, extract key features, compress repeated redundant information, and reduce data storage capacity;

[0032] S09. Based on the historical data in the data storage device and in combination with the real-time measurement value, the Kalman filter algorithm is applied to model and predict the cumulative error, thereby realizing online correction and optimization of the measurement data, obtaining the optimized data and saving it to the data storage device.

[0033] Among them, the sliding window weighted average method in S06 specifically sets a data window with a length of 30 minutes on the time series, assigns a time decay weight to each data point in the window, and the data point closer to the current moment has a greater weight, and replaces the original measurement value with the weighted average value. The window slides forward every 10 seconds, continuously updates the data series, and effectively eliminates short-term random fluctuations.

[0034] Among them, the Kalman filter algorithm in S09 specifically establishes a system state space model, models the measurement noise and system noise as Gaussian white noise respectively, and iteratively calculates the system state estimate and covariance matrix through the prediction step and the update step. The prediction step uses the state at the previous moment to estimate the current state, and the update step corrects the prediction result in combination with the current measurement value, and finally outputs the optimal state estimate. The Kalman filter algorithm is suitable for processing the cumulative error caused by sensor drift.

[0035] Among them, the continuity error trend identification in S06 specifically establishes a normal fluctuation range model for each sensor data, calculates the first-order difference and second-order difference of 20 adjacent data points, and when the difference value of 5 consecutive data points exceeds 3 times the standard deviation of the normal range and shows a monotonically increasing or monotonically decreasing trend, it is determined that there is a continuity error and the error repair process is triggered.

[0036] Among them, the matrix singular value decomposition in S03 specifically constructs the marine environmental parameters into a three-dimensional tensor of dimension m×n×t, where m represents the horizontal dimension of space, n represents the vertical depth dimension, and t represents the time dimension. The tensor decomposition method is used to decompose the original high-dimensional data into the product form of three low-dimensional matrices, retaining the eigenvectors corresponding to the most significant k singular values, and reconstructing the compressed low-rank approximation matrix to achieve data dimensionality reduction and retain the main change pattern. The compression ratio reaches 15% of the original data volume.

[0037] Among them, the matrix disorder index in S04 is specifically calculated by calculating the ratio of the condition number of the measurement data covariance matrix to the matrix spectral norm, and quantifying the uniformity of the distribution of the matrix eigenvalues. The larger the condition number, the higher the data disorder. When the matrix disorder index exceeds the preset threshold of 10, the anomaly detection process is triggered.

[0038] Among them, the association model between the environmental parameters and measurement deviations in S05 specifically utilizes the multivariate regression analysis method, takes temperature, pressure and salinity as independent variables, and various measurement deviations as dependent variables, to fit the influence function of environmental factors on measurement accuracy. When the environmental parameter changes exceed the preset threshold, the correction coefficient is automatically applied to calibrate the original measurement value; in the process of establishing and applying the association model, a pre-trained deep-sea environmental disturbance compensation network model is introduced. The specific structure of the deep-sea environmental disturbance compensation network model is composed of a six-layer deep neural network, including two input coding layers, three hidden processing layers and one output decoding layer, wherein the input coding layer is responsible for receiving and standardizing environmental parameter data, the hidden processing layer is implanted with a sparse attention mechanism to capture the long-term correlation and short-term fluctuation characteristics between different environmental parameters, and the output decoding layer generates a correction coefficient for various sensors. number matrix; the steps of establishing the training data set in the pre-training process of the deep-sea environmental disturbance compensation network model specifically include collecting millions of deep-sea environmental parameter samples and corresponding sensor measurement error data from the global ocean observation system, cleaning, labeling and enhancing the collected original data, constructing a comprehensive training set covering different sea areas, different depths and different seasons, and dividing it into training set, verification set and test set in a ratio of 8:1:1; the steps of pre-training the deep-sea environmental disturbance compensation network model specifically include first using an unsupervised learning method to pre-train the environmental parameter data to learn the intrinsic correlation rules between the environmental parameters, and then using a supervised learning method to train the mapping relationship between the environmental parameters and the measurement error, and then adapting the model to the sea area and depth range through the transfer learning method, and finally applying reinforcement learning technology to optimize the model parameters to minimize the long-term accumulated error.

[0039] An error optimization function based on an adaptive weighted algorithm is introduced into the monitoring data recording module. The error optimization function is used to perform adaptive weighted optimization processing on the original data collected by the multi-parameter sensor array to eliminate coupling interference and data offset between different sensors. The input includes the original sensor data matrix, historical data deviation coefficient, environmental parameter influencing factor, sensor health status index and system operation time. The output is an optimized high-precision measurement data set. The error optimization function identifies the source and propagation path of the error by establishing a multi-layer error propagation model, and dynamically adjusts the data fusion weight according to the performance characteristics of different sensors under different environmental conditions to suppress the error accumulation effect.

[0040] Among them, the distributed preprocessing in S08 specifically performs data noise reduction, outlier detection and feature extraction operations in the control chip of each potential target, compressing the original data to less than 25% of the original volume while retaining key information, thereby improving the overall data processing efficiency of the system; the matrix block calculation technology divides the large-scale data matrix into multiple sub-matrix blocks, and performs matrix operations in parallel on the multi-core processing unit in the control chip. The sub-matrix calculation results are merged into the final result through the divide-and-conquer method, reducing the calculation complexity and accelerating the large-scale data processing process.

[0041] Among them, the trilateration method in S02 specifically selects three reference buoys with known positions, measures the sound wave propagation time from the target buoy to the reference buoy, calculates the distance from the target buoy to each reference buoy based on the corrected seawater sound speed, establishes a set of three-variable quadratic equations, and solves the three-dimensional coordinates of the target buoy through the Newton iteration method, with a positioning accuracy of 0.5 meters.

[0042] The minimum spanning tree algorithm in S07 specifically models the multi-buoy system as an undirected weighted graph, with each buoy as a node in the graph and the communication links between buoys as edges. The edge weights are determined by the communication energy consumption and link quality. The Kruskal algorithm is used to construct a minimum spanning tree covering all buoys to ensure that the total communication energy consumption is minimized. The minimum spanning tree structure is dynamically updated when the relative positions of the buoys change, ensuring real-time communication efficiency.

[0043] The specific implementation of the above steps is described in detail below.

[0044] The hardware of this real-time deep-sea buoy monitoring system includes a control chip, a multi-parameter sensor array, a data storage device, and a power supply. The control chip uses a low-power ARM Cortex-M7 processor with a main frequency of 240MHz, built-in 1MB flash memory and 512KB SRAM, and has a floating-point unit and digital signal processing functions to perform all system control and data processing tasks. The control chip is connected to the multi-parameter sensor array via the SPI bus and connected to the I 2The C bus connects to the data storage device and to the power supply via the power management interface. The multi-parameter sensor array consists of a temperature sensor, a conductivity sensor, a pressure sensor, and an acoustic Doppler sensor. The temperature sensor is a platinum resistance type with a measurement range of -5°C to 35°C and an accuracy of ±0.01°C; the conductivity sensor is an inductive type with a measurement range of 0 to 70 mS / cm and an accuracy of ±0.005 mS / cm; the pressure sensor is a piezoresistive type with a measurement range of 0 to 600 bar and an accuracy of ±0.1% of full scale; the acoustic Doppler sensor operates at a frequency of 600 kHz, with a velocity range of ±5 m / s and an accuracy of ±0.5 cm / s. The data storage device uses 128GB of industrial-grade solid-state memory, supporting high-speed data writing and cyclical storage. The power supply is a lithium-ion battery pack with a total capacity of 50 Ah. Equipped with an intelligent power management unit, it enables low-power operation mode switching, ensuring continuous system operation for over six months. The entire system is enclosed in a titanium alloy pressure-resistant casing, with a pressure resistance rating of 6,000 meters.

[0045] The specific implementation of step S01 is that the control chip sends data acquisition instructions to the multi-parameter sensor array every 5 minutes according to a preset sampling frequency, activating each sensor to measure ocean parameters. The temperature sensor collects temperature data at multiple depth measurement points to form a vertical temperature distribution profile; the conductivity sensor calculates seawater salinity based on the temperature data; the pressure sensor measures water depth; and the acoustic Doppler sensor measures the speed and direction of the ocean current. The control chip reads the output values ​​of each sensor via the SPI interface at a rate of 100kbps, adds a timestamp and geographic location information to each measurement value, and constructs a structured data record. The control chip stores the collected raw data and the initially processed standardized data simultaneously in a data storage device, using partitioned storage to ensure that the raw data and processed data are stored separately to facilitate subsequent analysis and verification. This step aims to establish a complete raw data set of ocean environmental parameters to provide basic data support for subsequent analysis.

[0046] In step S02, the control chip first uses temperature stratification data to calculate the sound velocity profile at different depths. This sound velocity calculation uses an empirical formula recommended by UNESCO, taking temperature, salinity, and depth as input parameters and outputting sound velocity values ​​at different depths. The control chip then modifies the acoustic wave propagation model based on the sound velocity profile and uses ray tracing to simulate the propagation path of sound waves in stratified seawater. The system then measures the round-trip propagation time and received signal strength of the acoustic wave signal through an acoustic communication network between the buoys. Based on the modified sound velocity and the measured sound wave propagation time, the control chip calculates the distance between the buoys. When using trilateration to determine the relative position of the buoys, three reference buoys with known positions are selected, and the distances from the target buoy to each of these three reference buoys are calculated, establishing a system of three quadratic equations. The control chip solves this system of equations using the Newton iteration method to calculate the three-dimensional coordinates of the target buoy. The iteration terminates when the position change is less than 0.05 meters or the number of iterations reaches 10. This method combines accurate sound velocity profiles and sound wave propagation models, with a positioning accuracy of up to 0.5 meters, effectively solving the problem of precise positioning of buoy arrays in deep-sea environments.

[0047] The specific implementation of step S03 is that the control chip reorganizes the collected ocean data into a three-dimensional data matrix based on spatial and temporal dimensions. The horizontal spatial dimension represents the position distribution of different buoys, the vertical depth dimension represents the depth distribution of each measuring point, and the temporal dimension represents the different sampling moments. The control chip then applies a matrix singular value decomposition algorithm to the constructed three-dimensional tensor for dimensionality reduction. First, the three-dimensional tensor is expanded into a two-dimensional matrix, and then the matrix's singular values ​​and corresponding left and right singular vectors are calculated. The singular values ​​are sorted from largest to smallest, and the top k largest singular values ​​and corresponding singular vectors are selected. The value of k is adaptively determined based on the data characteristics, generally set so that the retained singular value energy accounts for 95% of the total energy. The control chip then reconstructs a low-rank approximation matrix using the retained k singular values ​​and corresponding singular vectors to achieve data compression. This method can effectively extract the main change patterns of marine environmental data, eliminate redundant information, and achieve a compression rate of up to 85%, while retaining key environmental characteristics and change trends, providing a compact and information-rich data representation for subsequent analysis.

[0048] The specific implementation method of step S04 is that the control chip applies the matrix sliding window method to process the time series data. The time window length is set to 60 minutes, and the window contains multidimensional measurement data of 12 consecutive sampling points. The window slides one sampling point each time, that is, 5 minutes, and the covariance matrix is ​​calculated for the data in each window. The covariance matrix reflects the correlation and fluctuation characteristics between the various measurement parameters. The control chip calculates the ratio of the condition number of the covariance matrix to the matrix spectral norm as an indicator of the matrix disorder. The condition number is defined as the ratio of the maximum singular value to the minimum non-zero singular value, and the spectral norm is defined as the maximum singular value. When the disorder index exceeds the preset threshold of 10, it indicates that there is an abnormal pattern in the data. The control chip further analyzes the components of the abnormal feature vector to locate the parameters and time points where the abnormality occurs. This method utilizes the inherent correlation between multidimensional data, and can effectively identify situations such as sensor failure, measurement anomalies or environmental mutations, thereby improving the reliability of system monitoring data.

[0049] In step S05, the control chip analyzes the impact of environmental factors on sensor measurement accuracy based on a correlation model between environmental parameters and measurement deviations. This model employs a multivariate regression analysis method, with temperature, pressure, and salinity as independent variables and various measurement deviations as dependent variables. The control chip uses historical data to fit the environmental factor influence function and establish a quantitative relationship model between the parameters. When an environmental parameter change exceeds a set threshold, such as a temperature change exceeding 0.5°C, a pressure change exceeding 5 bar, or a salinity change exceeding 0.2 PSU, the control chip automatically applies a correction factor to calibrate the original measurement value. During this model application process, the control chip incorporates a pre-trained deep-sea environmental disturbance compensation network model. This model consists of a six-layer deep neural network, comprising two input encoding layers, three hidden processing layers, and one output decoding layer. The input encoding layer receives and normalizes environmental parameter data, the hidden processing layer incorporates a sparse attention mechanism to capture long- and short-term characteristics between environmental parameters, and the output decoding layer generates a correction coefficient matrix for each sensor type. This method can adapt to complex and changing deep-sea environments and improve the accuracy and reliability of measurement data under various environmental conditions.

[0050] The specific implementation of step S06 involves the control chip continuously monitoring the data stream and analyzing the collected raw data in real time. First, a normal fluctuation range model is established for each sensor data type, calculating the mean and standard deviation of the historical data, defining the normal fluctuation range as the mean ± 3 times the standard deviation. The control chip then calculates the first-order and second-order differences of 20 adjacent data points in real time, analyzing the data change trend and acceleration. If the difference values ​​of five consecutive data points exceed 3 times the normal range and show a monotonically increasing or decreasing trend, a continuity error is determined, triggering the error correction process. The control chip then smooths the abnormal data using a sliding window weighted average method, setting a 30-minute data window containing six consecutive measurement points. Within the window, each data point is assigned a time-decay weight using an exponential decay function, with data points closer to the current time receiving greater weights. The control chip then calculates the weighted average to replace the raw measurement value, sliding the window forward every 10 seconds to continuously update the data series. This method effectively eliminates short-term random fluctuations and continuity errors, improving the smoothness and reliability of the data time series.

[0051] The specific implementation of step S07 involves the control chip optimizing the inter-buoy data transmission network structure based on a minimum spanning tree algorithm. The multi-buoy system is modeled as an undirected weighted graph, with each buoy as a node and inter-buoy communication links as edges. Edge weights are determined by both communication energy consumption and link quality. Communication energy consumption is proportional to the square of the distance between buoys, while link quality is related to the degree of signal interference and the characteristics of the underwater acoustic channel. The control chip constructs a minimum spanning tree using the Kruskal algorithm. This algorithm first sorts all edges by weight from smallest to largest, then adds edges to the spanning tree one by one, skipping edges that would form loops until all nodes are connected. The resulting minimum spanning tree minimizes total communication energy consumption while ensuring data connectivity between all buoys. When the relative position of a buoy changes by more than a preset threshold of 1 meter, the control chip recalculates the communication link weights and updates the minimum spanning tree structure. This method dynamically adapts to changes in the position of the buoy array, ensuring real-time communication efficiency while minimizing system communication energy consumption and extending the operating life of the buoy network.

[0052] The specific implementation of step S08 involves distributed preprocessing of the collected multi-source heterogeneous data by the control chip. Data noise reduction, outlier detection, and feature extraction are performed in each target's control chip. Data noise reduction utilizes the wavelet transform method, selecting the db4 wavelet basis function. The original signal undergoes a four-layer decomposition, and the high-frequency coefficients are soft-thresholded to reconstruct the signal, effectively removing measurement noise. Outlier detection utilizes the 3σ principle, marking data outside the range of ±3 standard deviations from the mean as outliers. Feature extraction calculates time-domain statistical features and frequency-domain energy distribution characteristics, preserving key information. The control chip utilizes matrix block computing technology to process large-scale data, partitioning the data matrix into multiple sub-matrix blocks and performing matrix operations in parallel on the multi-core processing unit. The sub-matrix size is optimized based on the control chip's cache capacity, typically 64×64. The control chip uses a divide-and-conquer approach to merge the sub-matrix calculation results, achieving efficient large-scale data processing. This method compresses the original data to less than 25% of its original volume while preserving key information, significantly improving system data processing efficiency and storage utilization.

[0053] The specific implementation of step S09 involves the control chip applying a Kalman filter algorithm to model and predict cumulative errors based on historical data and real-time measurements in the data storage device, thereby achieving online correction and optimization of the measurement data. First, the control chip establishes a system state-space model, whose state variables include the measured values ​​and drift rates of each sensor, and the observed variables are the actual sensor output values. The control chip models the measurement noise and system noise as Gaussian white noise, respectively, and estimates the noise covariance matrix based on the historical sensor performance. The Kalman filter algorithm iteratively calculates the system state through a prediction step and an update step. The prediction step estimates the current state and its covariance based on the previous state and state transition matrix. The update step calculates the Kalman gain, corrects the prediction result based on the current measurement value, and outputs the optimal state estimate. The control chip dynamically adjusts the noise covariance matrix to adapt to the error characteristics under different operating conditions. After completing the Kalman filter processing, the control chip timestamps the optimized data and stores it in a binary structured format in the optimized data partition of the data storage device. The storage format includes data header information (including timestamp, location information, and correction parameters) and data body (including the optimized measurement values ​​and their confidence intervals). At the same time, the control chip retains the correspondence between the original data and the optimized data, facilitating subsequent tracing and verification. For data generated during long-term operation, the system adopts a rolling update strategy. When the storage capacity approaches the threshold (90% by default), it prioritizes retaining the latest data and key historical data nodes to ensure the system's continued operation. This algorithm is particularly suitable for processing cumulative errors caused by sensor drift. It can track system status changes in real time, provide optimal estimates, significantly improve the accuracy and reliability of long-term monitoring data, and provide stable and reliable data support for the buoy monitoring system.

[0054] The error optimization function, based on an adaptive weighting algorithm, is used to optimize the raw data collected by a multi-parameter sensor array. The function's inputs include the raw sensor data matrix, historical data deviation coefficients, environmental parameter influencing factors, sensor health index, and system runtime; its output is an optimized, high-precision measurement data set. The optimization function first establishes a multi-layer error propagation model to identify error sources and propagation paths. Error sources include inherent sensor error, environmental interference error, accumulated drift error, and random noise. Propagation path analysis considers the coupling relationship between sensors and the error amplification effect in the data processing flow. The control chip dynamically adjusts data fusion weights based on the performance characteristics of different sensors under different environmental conditions. Weight adjustment is based on the sensor's historical accuracy, stability under current environmental conditions, consistency with related parameter measurements, and drift trends related to runtime. Weight assignment uses an exponentially weighted moving average method, assigning higher weights to sensors with better historical performance. This function effectively suppresses the effects of error accumulation, achieves high-precision optimization of measurement data, and improves the system's long-term operational stability and data reliability.

[0055] Furthermore, the deep-sea environmental disturbance compensation network model is the core component for achieving high-precision deep-sea monitoring. Its structural design and training methods need to fully consider the complexity and variability of the deep-sea environment. The model adopts a six-layer deep neural network structure, including two input coding layers, three hidden processing layers, and one output decoding layer, which realizes the precise mapping from environmental parameters to sensor correction coefficients within the network. The input coding layer first receives multi-dimensional environmental parameter data including temperature, pressure, salinity, sound speed, and ocean current speed. Through standardization, the parameters of different dimensions are mapped to a unified numerical range to reduce the impact of the magnitude difference between parameters on network training. The first input coding layer contains 256 neurons and uses the ReLU activation function to enhance the nonlinear expression ability. The second input coding layer further extracts features and reduces the dimension to a 128-dimensional feature vector. The hidden processing layer is the core computational unit of the network. The first hidden layer contains 512 neurons, responsible for capturing the basic correlation patterns between environmental parameters. The second hidden layer, with 256 neurons, incorporates a sparse attention mechanism. Using self-attention, it calculates the weights of the impact of different environmental parameters on measurement error, highlighting the influence of key factors. The third hidden layer, with 128 LSTM units, utilizes a long-short-term memory (LSTM) structure to process the dynamic characteristics of environmental parameters over time, capturing both long-term stable patterns and short-term fluctuations, addressing the limitations of traditional networks in processing time series data. The output decoding layer uses a linear activation function to directly output a correction coefficient matrix for each sensor type. The matrix dimensions correspond to the sensor type and the number of parameters. Each element represents the correction coefficient for the corresponding sensor parameter under the given environmental conditions, achieving an end-to-end mapping from environmental state to measurement correction.

[0056] Furthermore, the establishment of a training dataset for the deep-sea environmental disturbance compensation network model is fundamental to ensuring model performance. A comprehensive training sample library was constructed using multi-source data fusion and a stratified sampling strategy. First, millions of deep-sea environmental parameter samples were collected from the Global Ocean Observing System (GOOS), the International Deep-Sea Program (IODP), and various national ocean observation networks. These samples included historical data on environmental parameters such as temperature, pressure, and salinity from different sea areas, depths, and seasons. The deviations between the corresponding sensor measurements and the standard values ​​were also recorded. The raw data were systematically cleaned to remove obvious outliers and incomplete records, and the data were aligned in time and space and standardized. They were then professionally labeled, and oceanographic experts were invited to classify and label the types of environmental interference, and establish a correspondence library between environmental interference and measurement errors. Data enhancement was then performed, and samples under extreme environmental conditions that were difficult to collect directly were artificially synthesized by adding Gaussian noise, wavelet transform, and time scale scaling. The number of samples at the boundary of the training set was expanded, and the generalization ability of the model in abnormal environments was improved. Finally, a comprehensive training set covering different sea areas (including major sea areas such as the Pacific, Atlantic, Indian, and Arctic Oceans), different depths (from the surface to the Hadar Abyss of 11,000 meters), and different seasons (covering seasonal changes and special meteorological events) was constructed, containing approximately 12.5 million environment-error paired samples, divided into training set, validation set, and test set in a ratio of 8:1:1 to ensure the comprehensiveness and representativeness of model training.

[0057] Furthermore, the deep-sea environmental disturbance compensation network model is trained using a multi-stage strategy. First, unsupervised learning is used to pre-train environmental parameter data. An autoencoder structure is used to allow the network to autonomously learn the inherent correlations between environmental parameters and capture the main patterns of deep-sea environmental changes. A supervised learning approach is then used to train the mapping between environmental parameters and measurement errors. Mean squared error is used as the loss function, and parameters are updated using the Adam optimizer. The learning rate is dynamically adjusted using a cosine annealing strategy. The pre-trained model is then adapted to the target sea area and depth range through transfer learning. The network's front-layer parameters are frozen, and only the weights of the back-layers are fine-tuned to improve the model's accuracy in the target sea area. Finally, reinforcement learning techniques are used to optimize the model parameters. A reward function is designed to minimize long-term cumulative error. A deep Q-network framework is used to fine-tune the network parameters, enabling the model to predict and compensate for changes in sensor performance over the long term. Through this training process, the model not only accurately compensates for measurement errors under static environmental conditions, but also adapts to complex error patterns caused by dynamic environmental changes and sensor performance degradation over time, providing a stable and reliable data quality assurance mechanism for deep-sea buoy systems.

[0058] The mathematical model or calculation process involved in the present invention is described in detail below.

[0059] The sound velocity calculation in step S02 adopts the following empirical formula, which is specifically expressed as follows:

[0060] c=1449.2+4.6T-0.055T 2 +0.00029T 3 +(1.34-0.01T)(S-35)+0.016D;

[0061] Where c is the speed of sound in m / s; T is the temperature in °C; S is the salinity in PSU; and D is the depth in m.

[0062] The parameters are obtained as follows: T is directly measured by a temperature sensor; S is calculated by combining the conductivity value measured by a conductivity sensor with the temperature; and D is calculated by dividing the pressure value measured by a pressure sensor by the acceleration of gravity and the seawater density. This formula accounts for the combined effects of temperature, salinity, and depth on the speed of sound. The temperature term is fitted with a cubic polynomial to reflect the nonlinear relationship, the salinity term incorporates the temperature-salinity interaction, and the depth term represents the pressure effect. This formula accurately describes the variations in sound speed under different ocean environments. This formula is based on the Chen-Millero formula recommended by UNESCO.

[0063] The trilateration method in step S02 involves a set of three quadratic equations, which is specifically expressed as follows:

[0064]

[0065] Where (x, y, z) is the three-dimensional coordinate of the target buoy; (x i ,y i , z i ) is the known coordinate of the i-th reference buoy, i = 1, 2, 3; d i is the distance from the target buoy to the i-th reference buoy.

[0066] The parameter acquisition method is: refer to the latent marker coordinates (x i ,y i , z i ) is a known value determined by measurement when the system is initially deployed; the distance d i By measuring the round-trip time t i Calculated from the speed of sound c, d i =c·t i / 2. When Newton iteration method is used to solve the equations, the objective function is constructed:

[0067]

[0068] The iteration formula is:

[0069]

[0070] Where J(x, y, z) is the Jacobian matrix:

[0071]

[0072] This set of equations is based on the Euclidean distance definition and uses three known reference points to determine the spatial position of an unknown point. It is a mathematical expression of spherical triangulation positioning and is suitable for precise positioning in three-dimensional space.

[0073] The matrix singular value decomposition (SVD) in step S03 is applied to the three-dimensional data tensor. First, the three-dimensional tensor Expand to a two-dimensional matrix Then perform SVD decomposition:

[0074] A=U∑V T ;

[0075] Where, is the left singular vector matrix; is a diagonal matrix of singular values; is the right singular vector matrix.

[0076] Select the first k largest singular values ​​and the corresponding singular vectors to reconstruct the low-rank approximate matrix:

[0077]

[0078] Where, Contains the first k columns of U; is a diagonal matrix containing the first k singular values; Contains the first k columns of V.

[0079] The parameter k is determined based on the energy ratio:

[0080]

[0081] Where σ i is the i-th singular value; r is the rank of A. This method uses the low-rank approximation property of the matrix to retain the main change pattern of the data, eliminate noise and redundant information, and achieve data compression and dimensionality reduction.

[0082] The matrix disorder index calculation in step S04 involves the ratio of the condition number of the covariance matrix to the matrix spectral norm. (n is the number of sampling points, p is the parameter dimension), the covariance matrix C is calculated as follows:

[0083]

[0084] Where, is the matrix consisting of the means of each column of X.

[0085] The condition number of the covariance matrix is:

[0086]

[0087] Where λ max (C) and λ min (C) are the largest and smallest non-zero eigenvalues ​​of C, respectively.

[0088] The matrix spectral norm is:

[0089] ||C||2=λ max (C);

[0090] The matrix disorder index is defined as:

[0091]

[0092] When δ>10, it is determined to be an abnormal data pattern. This indicator reflects the stability of the covariance structure of multidimensional data. The larger the condition number, the more uneven the data distribution, and the higher the indicator value, the more serious the data disorder.

[0093] The multiple regression analysis method in step S05 establishes a correlation model between environmental parameters and measurement deviations:

[0094] ΔM i =β 0i +β 1i T+β 2i P+β 3i S+β 4i T·P+β 5i T·S+β 6i P·S+β 7i T 2 +β 8i P 2 +β 9i S 2 +ε i ;

[0095] Where, ΔM i is the measurement deviation of the i-th sensor; T is temperature, unit is °C; P is pressure, unit is bar; S is salinity, unit is PSU; β ji is the regression coefficient; ε is the random error term, which obeys the normal distribution

[0096] The parameter acquisition method is: regression coefficient β ji Estimated by the least squares method based on historical calibration data:

[0097]

[0098] Where X is the design matrix containing environmental parameters and their interaction terms and quadratic terms; ΔM i is the corresponding measurement deviation vector. This model takes into account the main effect, interaction effect and nonlinear effect of environmental parameters, and can fully reflect the complex influence of environmental factors on measurement accuracy.

[0099] The sliding window weighted average method in step S06 is applied to data smoothing:

[0100]

[0101] Where, is the smoothed value at time t; x t-i is the original observation value at time ti; L is the window length, which is 30 minutes and contains 6 sampling points; w i is the weight, using exponential decay function: w i =e -αi , where α is the attenuation factor and its value is 0.5.

[0102] Continuity error trend identification is based on first-order and second-order difference analysis:

[0103] Δ 1 x t =x t -x t-1 ;

[0104] Δ 2 x t =Δ 1 x t -Δ 1 x t-1 =x t -2x t-1 +x t-2 ;

[0105] When the difference values ​​of five consecutive points meet the following conditions, it is determined to be a continuity error:

[0106] And sgn(Δ 1 x t )=sgn(Δ 1 x t-1 )=...=sgn(Δ 1 x t-4 );

[0107] Where, is the standard deviation of the first-order difference; sgn() is the sign function. This method combines sliding window smoothing and differential analysis techniques to effectively identify and correct trend anomalies in the data.

[0108] The minimum spanning tree algorithm in step S07 uses the Kruskal algorithm to optimize the communication network. The communication link weight is defined as:

[0109]

[0110] Where w ij is the weight of the communication link between buoys i and j; d ij is the distance between two buoys; SNR ij is the signal-to-noise ratio; SNR max is the maximum signal-to-noise ratio; P loss，ij is the transmission loss; α, β, and γ are weight coefficients, which are 0.5, 0.3, and 0.2 respectively.

[0111] The total weight of the minimum spanning tree constructed by Kruskal's algorithm is:

[0112] W MST =∑ (i，j)∈MST w ij ;

[0113] Where MST is the edge set contained in the minimum spanning tree. This algorithm minimizes the total communication cost and optimizes the network topology to minimize communication energy consumption.

[0114] The wavelet transform in step S08 is used to reduce data noise:

[0115]

[0116] Where W ψ f(a, b) is the wavelet coefficient of the signal f(t); ψ(t) is the wavelet basis function; a is the scale parameter; and b is the translation parameter.

[0117] Soft thresholding is:

[0118]

[0119] In the formula, λ is the threshold value, which is σ is the standard deviation of noise, and n is the signal length. This method uses the time-frequency localization characteristics of wavelet transform to effectively separate signal and noise and achieve data denoising.

[0120] The Kalman filter algorithm in step S09 is based on the state space model:

[0121] x k =Fx k-1 +Bu k +wk ;

[0122] z k =Hx k +v k ;

[0123] Where x k is the state vector at time k, including the measurement value and drift rate of each sensor; z k is the observation vector; F is the state transfer matrix; B is the control input matrix; u k is the control vector; H is the observation matrix; w k is the process noise, which obeys the distribution N(0, Q); v k is the measurement noise, which follows the distribution N(0, R).

[0124] Kalman filter prediction steps:

[0125]

[0126] P k|k-1 =FP k-1|k-1 F T +Q;

[0127] Update steps:

[0128] K k =P k|k-1 H T (HP k|k-1 H T +R) -1 ;

[0129]

[0130] P k|k =(IK k H)P k|k-1 ;

[0131] Where, is the prior state estimate; is the posterior state estimate; P k|k-1 is the prior error covariance; P k|k is the posterior error covariance; K k is the Kalman gain. This algorithm achieves the optimal estimation of the system state through prediction-correction iteration and is particularly suitable for dealing with cumulative errors.

[0132] The error optimization function based on the adaptive weighted algorithm is defined as follows:

[0133]

[0134] β i =α1ehist，i +α2e env，i +α3e cons，i +α4e time，i ;

[0135] Where, is the optimized measurement value; x i is the original measurement value of the i-th sensor; w i is the weight of sensor i; β i Score the error; e hist，i Score historical accuracy; e env，i Score the environmental stability; e cons，i Score consistency; e time，i is the time drift score; α1 to α4 are balance coefficients, which are 0.4, 0.3, 0.2, and 0.1 respectively.

[0136] The scoring calculation method is:

[0137]

[0138] e time，i =γ·t / t max ;

[0139] Where M is the number of historical calibration samples; x i，j is the measurement value of sensor i at sample j; x ref，j is the reference value of sample j; σ j is the standard deviation of sample j; p k is the current environment parameter; p ref，k is the reference environmental condition; c k is the environmental parameter weight; N is the number of sensors; t is the running time; t max is the maximum design runtime; γ is the time weight coefficient. This function dynamically adjusts sensor weights by comprehensively evaluating the impact of various factors to achieve high-precision data fusion.

[0140] Specifically, the principle of this invention is to solve the problem of sensor measurement data accuracy and processing efficiency in deep-sea environments by building a multi-level data quality assurance mechanism and an efficient data processing framework. Its technical principles are mainly reflected in the following aspects:

[0141] First, to address the issue of deep-sea environmental interference, this paper innovatively introduces a deep-sea environmental disturbance compensation network model. This model utilizes a six-layer deep neural network structure. The input encoding layer receives standardized environmental parameters, the hidden processing layer captures the long-term and short-term correlations between environmental parameters, and the output decoding layer generates a sensor correction coefficient matrix. The model is pre-trained on millions of deep-sea environmental parameter samples. Unsupervised learning captures the inherent correlations between environmental parameters. Supervised learning then establishes a mapping from environmental parameters to measurement errors. Finally, transfer learning and reinforcement learning are used to adapt and optimize the model to specific scenarios, fundamentally addressing the impact of environmental factors on measurement accuracy.

[0142] Secondly, to address sensor drift and coupling interference, the present invention designs an error optimization function based on an adaptive weighted algorithm. This function establishes a multi-layer error propagation model, identifies error sources and propagation paths, and dynamically adjusts data fusion weights to effectively suppress the cumulative effect of errors. Simultaneously, the system applies a sliding window weighted averaging method and a Kalman filter algorithm to process time series data. The former sets a 30-minute data window, assigns time-decaying weights, and updates the data every 10 seconds to eliminate short-term random fluctuations. The latter establishes a state-space model, models noise as Gaussian white noise, and iteratively calculates the optimal state estimate through prediction and update. This method is particularly suitable for handling cumulative errors caused by sensor drift.

[0143] Thirdly, to address data processing efficiency, the present invention applies matrix singular value decomposition (SVD) technology to construct ocean environmental parameters into a three-dimensional tensor. By retaining the most significant eigenvectors through tensor decomposition, the system achieves data dimensionality reduction to 15% of the original data volume while preserving the main variation patterns. Simultaneously, the system employs distributed preprocessing and matrix block computing techniques to perform data noise reduction and feature extraction within each buoy control chip, compressing the data to less than 25% of its original volume. Multi-core parallel computing accelerates the processing process.

[0144] Furthermore, the minimum spanning tree algorithm is used to optimize the inter-buoy communication network structure, minimizing total communication energy consumption. Trilateration is used to locate the buoys with an accuracy of 0.5 meters. The coordinated application of these technologies forms a logically complete and technologically advanced deep-sea monitoring solution, enabling the system to achieve high-precision and efficient continuous monitoring in complex deep-sea environments.

[0145] A specific embodiment 1 of the present invention is provided below. The specific implementation of each step in this embodiment 1 is described in detail as follows.

[0146] The specific implementation of step S01 is the same as above and will not be repeated here.

[0147] The specific implementation of step S02 is that the control chip first calculates the sound velocity profile at different depths using the temperature layer data. The sound velocity calculation uses the following empirical formula:

[0148] c=1449.2+4.6T-0.055T 2 +0.00029T 3 +(1.34-0.01T)(S-35)+0.016D;

[0149] Where c is the speed of sound in m / s; T is the temperature in °C; S is the salinity in PSU; and D is the depth in meters. This formula accounts for the combined effects of temperature, salinity, and depth on the speed of sound. The temperature term is fitted with a cubic polynomial to reflect the nonlinear relationship, the salinity term incorporates the temperature-salinity interaction, and the depth term accounts for the pressure effect.

[0150] The control chip modifies the acoustic wave propagation model based on the sound velocity profile and uses ray tracing to simulate the propagation path of the sound wave in stratified seawater. The system then measures the round-trip propagation time of the acoustic wave signal and the received signal strength through the acoustic communication network between the buoys. Based on the modified sound velocity and the measured sound wave propagation time, the control chip calculates the distance between the buoys: d i =c·t i / 2, where d i is the distance, t i The round trip time.

[0151] When using trilateration to determine the relative position of the buoy, a quadratic equation system with three variables is established:

[0152]

[0153] Where (x, y, z) is the three-dimensional coordinate of the target buoy; (x i ,y i , z i ) is the known coordinate of the i-th reference buoy, i = 1, 2, 3; d i is the distance from the target buoy to the i-th reference buoy.

[0154] The control chip solves the equations through Newton iteration method and constructs the objective function:

[0155]

[0156] The iteration formula is:

[0157]

[0158] Where J(x, y, z) is the Jacobian matrix:

[0159]

[0160] The iteration termination condition is that the position change is less than 0.05 meters or the number of iterations reaches 10. This method combines accurate sound velocity profiles and sound wave propagation models to achieve a positioning accuracy of 0.5 meters, effectively solving the problem of precise positioning of buoy arrays in deep-sea environments.

[0161] The specific implementation of step S03 is that the control chip reorganizes the collected ocean data into a three-dimensional data matrix according to the spatial and temporal dimensions. The horizontal dimension of the space represents the position distribution of different buoys, the vertical depth dimension represents the depth distribution of each measuring point, and the time dimension represents different sampling moments. The control chip reorganizes the constructed three-dimensional tensor into a three-dimensional matrix according to the spatial and temporal dimensions. Apply the matrix singular value decomposition algorithm to perform dimensionality reduction. First, expand the three-dimensional tensor into a two-dimensional matrix Then perform SVD decomposition: A=U∑V T ;

[0162] Where, is the left singular vector matrix; is a diagonal matrix of singular values; is the right singular vector matrix.

[0163] Singular values ​​are sorted from largest to smallest, and the first k largest singular values ​​and corresponding singular vectors are selected. The value of k is adaptively determined based on the data characteristics, satisfying: Where σ i is the i-th singular value; r is the rank of A.

[0164] The control chip reconstructs the low-rank approximate matrix using the retained k singular values ​​and corresponding singular vectors:

[0165] Where, Contains the first k columns of U; is a diagonal matrix containing the first k singular values; Contains the first k columns of V. This method can effectively extract the main change patterns of marine environmental data and eliminate redundant information with a compression rate of up to 85%, while retaining key environmental features and change trends, providing a compact and information-rich data representation for subsequent analysis.

[0166] The specific implementation of step S04 is to control the chip to apply the matrix sliding window method to the time series data. The time window length is set to 60 minutes, and the window contains multidimensional measurement data of 12 consecutive sampling points. The window slides one sampling point each time, that is, 5 minutes, and the covariance matrix of the data in each window is calculated:

[0167]

[0168] Where, is the multidimensional measurement data matrix within the time window (n is the number of sampling points, p is the parameter dimension); is the matrix consisting of the means of each column of X.

[0169] The control chip calculates the ratio of the condition number of the covariance matrix to the matrix spectral norm as an indicator of matrix disorder:

[0170]

[0171] Where, is the condition number of the covariance matrix; λ max (C) and λ min (C) are the largest and smallest non-zero eigenvalues ​​of C respectively; ||C||2=λ max (C) is the matrix spectral norm.

[0172] When the disorder index exceeds a preset threshold of 10, it indicates an abnormal data pattern. The control chip further analyzes the components of the abnormal feature vector to locate the parameters and time of the abnormality. This method, leveraging the inherent correlations between multidimensional data, can effectively identify sensor failures, measurement anomalies, or sudden environmental changes, thereby improving the reliability of the system's monitoring data.

[0173] The specific implementation of step S05 is to control the chip to analyze the impact of environmental factors on the sensor's measurement accuracy based on the correlation model between environmental parameters and measurement deviations. The model uses a multiple regression analysis method to establish a quantitative relationship between environmental parameters and measurement deviations:

[0174] ΔM i =β 0i +β 1i T+β 2i P+β 3i S+β 4i T·P+β 5i T·S+β 6i P·S+β 7i T 2 +β 8i P 2 +β 9i S 2 +ε i ;

[0175] Where, ΔM i is the measurement deviation of the i-th sensor; T is temperature, unit is °C; P is pressure, unit is bar; S is salinity, unit is PSU; β ji is the regression coefficient; ε i is a random error term, which obeys the normal distribution

[0176] Regression coefficient β jiEstimated by the least squares method based on historical calibration data:

[0177]

[0178] Where X is the design matrix containing environmental parameters and their interaction terms and quadratic terms; ΔM i is the corresponding measurement deviation vector.

[0179] When environmental parameter changes exceed set thresholds—for example, a temperature change exceeding 0.5°C, a pressure change exceeding 5 bar, or a salinity change exceeding 0.2 PSU—the control chip automatically applies correction coefficients to calibrate the original measurement values. This model incorporates a pre-trained deep-sea environmental disturbance compensation network model. This model consists of a six-layer deep neural network, consisting of two input encoding layers, three hidden processing layers, and one output decoding layer. This method can adapt to the complex and changing deep-sea environment, improving the accuracy and reliability of measurement data under various environmental conditions.

[0180] The specific implementation of step S06 is to control the chip to continuously monitor the data stream and analyze the collected raw data in real time. First, a normal fluctuation range model is established for each sensor data, and the mean and standard deviation of the historical data are calculated. The normal fluctuation range is defined as the mean ± 3 times the standard deviation. The control chip calculates the first-order difference and second-order difference of 20 adjacent data points in real time:

[0181] Δ 1 x t =x t -x t-1 ;

[0182] Δ 2 x t =Δ 1 x t -Δ 1 x t-1 =x t -2x t-1 +x t-2 ;

[0183] When the difference values ​​of five consecutive points meet the following conditions, it is determined to be a continuity error:

[0184] And sgn(Δ 1 x t )=sgn(Δ 1 x t-1 )=...=sgn(Δ 1 x t-4 );

[0185] Where, is the standard deviation of the first-order difference; sgn() is the sign function.

[0186] The control chip uses a sliding window weighted average method to smooth abnormal data:

[0187]

[0188] Where, is the smoothed value at time t; x t-i is the original observation value at time ti; L is the window length, which is 30 minutes and contains 6 sampling points; w i is the weight, using exponential decay function: w i =e -αi , where α is the attenuation factor and its value is 0.5.

[0189] The window slides forward every 10 seconds, continuously updating the data series. This method can effectively eliminate short-term random fluctuations and continuity errors, improving the smoothness and reliability of the data time series.

[0190] The specific implementation of step S07 is to control the chip to optimize the data transmission network structure between the submersibles based on the minimum spanning tree algorithm. The multi-submersible system is modeled as an undirected weighted graph, with each submersible as a node in the graph and the communication links between the submersibles as edges. The edge weights are determined by the communication energy consumption and link quality:

[0191]

[0192] Where w ij is the weight of the communication link between buoys i and j; d ij is the distance between two buoys; SNR ij is the signal-to-noise ratio; SNR max is the maximum signal-to-noise ratio; P loss，ij is the transmission loss; α, β, and γ are weight coefficients, which are 0.5, 0.3, and 0.2 respectively.

[0193] The control chip uses Kruskal's algorithm to construct a minimum spanning tree. The algorithm first sorts all edges by weight from smallest to largest, then adds edges to the spanning tree one by one, skipping edges that would form loops until all nodes are connected. The total weight of the generated minimum spanning tree is:

[0194] W MST =∑ (i，j)∈MST w ij ;

[0195] Where MST is the edge set contained in the minimum spanning tree.

[0196] When the relative position of the buoys changes by more than a preset threshold of 1 meter, the control chip recalculates the communication link weights and updates the minimum spanning tree structure. This method dynamically adapts to the position changes of the buoy array, ensuring real-time communication efficiency while minimizing system communication energy consumption and extending the operating life of the buoy network.

[0197] The specific implementation of step S08 is that the control chip performs distributed preprocessing on the collected multi-source heterogeneous data. Data denoising, outlier detection and feature extraction operations are performed in the control chip of each potential target. Data denoising adopts the wavelet transform method:

[0198]

[0199] Where W ψ f(a, b) is the wavelet coefficient of the signal f(t); ψ(t) is the wavelet basis function; a is the scale parameter; and b is the translation parameter.

[0200] Select the db4 wavelet basis function, perform a 4-layer decomposition on the original signal, and apply soft thresholding to process the high-frequency coefficients:

[0201]

[0202] In the formula, λ is the threshold value, which is σ is the standard deviation of noise, and n is the signal length.

[0203] Outlier detection utilizes the 3σ principle, marking data outside the range of ±3 standard deviations from the mean as outliers. Feature extraction calculates time-domain statistical features and frequency-domain energy distribution characteristics, preserving key information. The control chip applies matrix block computing technology to process large-scale data, partitioning the data matrix into multiple sub-matrix blocks and performing matrix operations in parallel on the multi-core processing unit. This method compresses raw data to less than 25% of its original volume while preserving key information, significantly improving system data processing efficiency and storage utilization.

[0204] The specific implementation of step S09 is that the control chip applies the Kalman filter algorithm to perform online correction and optimization of the cumulative error based on historical data and real-time measurement values. First, the system state space model is established:

[0205] x k =Fx k-1 +Bu k +w k ;z k =Hx k +v k ;

[0206] Where x k is the state vector at time k, including the measurement value and drift rate of each sensor; z kis the observation vector; F is the state transfer matrix; B is the control input matrix; u k is the control vector; H is the observation matrix; w k is the process noise, which obeys the distribution N(0, Q); v k is the measurement noise, which follows the distribution N(0, R).

[0207] Kalman filter prediction steps: P k|k-1 =FP k-1|k-1 F T +Q;

[0208] Update step: K k =P k|k-1 H T (HP k|k-1 H T +R) -1 ; P k|k =(IK k H)P k|k-1 ;

[0209] Where, is the prior state estimate; is the posterior state estimate; P k|k-1 is the prior error covariance; P k|k is the posterior error covariance; K k is the Kalman gain.

[0210] The control chip dynamically adjusts the noise covariance matrix to adapt to the error characteristics under different operating conditions. The Kalman filter algorithm is particularly suitable for processing the cumulative errors caused by sensor drift. It can track system state changes in real time, provide optimal estimates, and significantly improve the accuracy and reliability of long-term monitoring data.

[0211] The error optimization function based on the adaptive weighted algorithm introduced in the control chip is used to optimize the raw data collected by the multi-parameter sensor array:

[0212] Where, is the optimized measurement value; x i is the original measurement value of the i-th sensor; w i is the weight of sensor i, which is calculated by the following formula: β i =α1e hist，i +α2e env，i +α3e cons，i +α4e time，i ;

[0213] Where, β i Score the error; ehist，i Score historical accuracy; e env，i Score the environmental stability; e cons，i Score consistency; e time，i is the time drift score; α1 to α4 are balance coefficients, which are 0.4, 0.3, 0.2, and 0.1 respectively.

[0214] The scoring calculation method is:

[0215]

[0216] e time，i =γ·t / t max ;

[0217] Where M is the number of historical calibration samples; x i，j is the measurement value of sensor i at sample j; x ref，j is the reference value of sample j; σ j is the standard deviation of sample j; p k is the current environment parameter; p ref，k is the reference environmental condition; c k is the environmental parameter weight; N is the number of sensors; t is the running time; t max is the maximum design running time; γ is the time weight coefficient.

[0218] This function comprehensively evaluates the influence of various factors and dynamically adjusts the sensor weights to achieve high-precision data fusion, effectively suppress the error accumulation effect, and improve the long-term working stability of the system and data reliability.

[0219] After completing the Kalman filter processing, the control chip timestamps the optimized data and stores it in a binary structured format in the optimized data partition of the data storage device. The storage format includes data header information (including timestamp, location information, correction parameters) and data body (including various optimized measurement values ​​and their confidence intervals). At the same time, the control chip retains the correspondence between the original data and the optimized data to facilitate subsequent tracing and verification. For data generated during long-term operation, the system adopts a rolling update strategy. When the storage capacity approaches the threshold (the default is 90%), the latest data and key historical data nodes are retained first to ensure the system's continuous operation capability. This step significantly improves the accuracy and reliability of long-term monitoring data, providing stable and reliable data support for the buoy monitoring system.

[0220] Optional, such as Figure 2 As shown, the hardware of the real-time deep-sea buoy monitoring system in this embodiment 1 is composed of multiple key components, forming a complete deep-sea environment data acquisition and processing system. The hardware of the system is described in detail below.

[0221] Optionally, the control chip is the core processing unit of the entire system. It adopts a low-power, high-performance ARM Cortex-M7 series processor with a main frequency of up to 400MHz, built-in 2MB flash memory and 512KB SRAM, supports single-precision floating-point unit (FPU) and digital signal processing (DSP) instruction set, is connected to the data storage device through a high-speed SDIO interface, and is connected to the power system through a power management unit to form a complete hardware control network.

[0222] The optional multi-parameter sensor array serves as the system's data acquisition unit, comprised of a variety of high-precision marine environmental parameter sensors. The temperature sensor utilizes a platinum resistance (PT100) structure, with a measurement range of -5°C to 45°C, a resolution of 0.001°C, and long-term stability better than 0.005°C / year. The salinity sensor employs conductivity measurement, with a measurement range of 0-70 PSU and an accuracy of ±0.003 PSU. The pressure sensor utilizes silicon resonant technology, with a range of 0-1200 decibar and an accuracy of 0.01% of full scale. The current sensor utilizes the acoustic Doppler principle, with a measurement range of ±5 m / s and an accuracy of ±1 cm / s. Each sensor is housed in a titanium alloy and sealed with a special resin for corrosion resistance and high pressure resistance. The sensor array utilizes a vertical chain layout, with multiple sensor units installed at equal intervals along the depth dimension. Each unit contains a complete set of temperature, salinity, pressure, and current sensors, enabling simultaneous measurement of water stratification parameters. After preamplification and analog-to-digital conversion, the sensor signals are connected to the control chip via a waterproof connector and shielded cable. To reduce interference between sensors, the system uses time-division multiplexing to control the working timing of each sensor, and uses optoelectronic isolation technology to isolate digital and analog circuits to improve signal acquisition quality.

[0223] The data storage system optionally features a dual-backup architecture. The primary storage unit utilizes a high-reliability solid-state drive (SSD) with a capacity of 1TB and SLC flash memory chips, offering high read and write speeds and a long lifespan. The backup storage unit utilizes a high-capacity 512GB SD card, which regularly synchronizes critical data with the primary storage unit. The power system consists of a primary power supply and a backup power supply. The primary power supply utilizes a high-energy-density lithium-ion battery pack with a total capacity of 500Wh, capable of supporting 30 days of continuous system operation. The backup power supply utilizes a supercapacitor pack with a capacity of 50Wh, providing temporary power in the event of a primary power failure or replacement. The system housing is constructed of high-strength titanium alloy, offering excellent compressive strength and corrosion resistance. Within the housing are multiple independent sealed compartments housing the control unit, sensor unit, power supply unit, and buoyancy control unit. These compartments are interconnected via waterproof connectors and pressure-equalizing valves. An optional acoustic release is located on the top of the housing. Upon receiving an acoustic signal at a preset frequency, it automatically releases the counterweight, allowing the buoy to surface for easy recovery and data retrieval. The optional communication system includes an underwater acoustic communication module and a satellite communication module.

[0224] To better understand and implement the present invention, Example 2 of a specific application scenario is provided below: The water depth in an area ranges from 3,200 to 4,500 meters, with complex sea conditions and significant variations in environmental parameters, presenting a severe test for the stability and reliability of deep-sea monitoring systems. The experiment deployed a monitoring network consisting of eight submersible buoys, spaced approximately 500 to 1,000 meters apart, forming an observation area covering approximately 3 square kilometers. Each buoy was equipped with a complete hardware system, including an ARM Cortex-M7 controller chip, a multi-parameter sensor array, a 128GB solid-state storage device, and a 50Ah lithium-ion battery pack.

[0225] During the experiment, the control chip collected ocean environmental data at a sampling frequency of once every five minutes. The key parameters measured included temperature, salinity, and current velocity at different depths. To verify system performance, each sensor was precisely calibrated before deployment to establish an initial parameter baseline. The system deployment depths and corresponding initial environmental parameters are shown in Table 1:

[0226] Table 1 Buoy deployment depth and initial environmental parameters

[0227]

[0228]

[0229] After one month of operation, the system analyzed the deep-sea sound velocity profile using the sound velocity calculation process. The control chip calculated the sound velocity at each depth using the measured temperature, salinity, and depth data, following an empirical formula recommended by UNESCO. The results showed a significant variation in sound velocity with depth, forming a clear speed gradient with a maximum speed difference of 6.8 m / s. Based on this speed data, the system used ray tracing to refine the sound wave propagation model, significantly improving the positioning accuracy of the buoy.

[0230] When using trilateration to determine the relative positions of buoys, the system selected A01, A03, and A06 as reference buoys, measuring the round-trip times of the acoustic waves to be 1.258 seconds, 0.895 seconds, and 1.147 seconds, respectively. Based on the corrected speed of sound, the distances from buoy A02 to the three reference buoys were calculated to be 926.4 meters, 658.1 meters, and 843.5 meters, respectively. Using the Newton iteration method to solve the quadratic equations, after seven iterations, the positioning error was reduced to 0.042 meters, far exceeding the design target of 0.5 meters. The positioning accuracy of each buoy is shown in Table 2:

[0231] Table 2 Buoy positioning accuracy

[0232] Buoy number Theoretical coordinates (x, y, z) (m) Measured coordinates (x, y, z) (m) Positioning error (m) Number of iterations A01 (0,0,3250) Reference Point - - A02 (500,500,3420) (500.03,499.95,3420.12) 0.14 7 A03 (1000,0,3580) Reference Point - - A04 (1000,500,3750) (1000.08,500.11,3749.92) 0.16 6 A05 (500,1000,3920) (499.87,1000.22,3920.15) 0.29 8 A06 (0,1000,4080) Reference Point - - A07 (0,500,4230) (0.05,499.91,4230.18) 0.21 7 A08 (500,0,4380) (500.18,0.07,4380.04) 0.19 6

[0233] In terms of data compression, the system performed matrix singular value decomposition on three months of accumulated marine environmental data. The original data was constructed as a three-dimensional tensor of 12600×126×8 (number of sampling points × number of depth points × number of parameters), which was expanded into a two-dimensional matrix and then subjected to SVD decomposition. Analysis found that the energy contained in the first 38 singular values ​​accounted for 95.2% of the total energy. The system retained these 38 singular values ​​and the corresponding singular vectors, reconstructing a low-rank approximate matrix, achieving a compression rate of up to 84.7%, while keeping the data reconstruction error less than 0.5%.

[0234] The matrix disorder index calculation is used for anomaly detection. The system sets a 60-minute sliding window to calculate the covariance matrix of multidimensional data and its disorder index. During the monitoring period, the system successfully captured a deep-sea cold water mass crossing event. This event occurred on the 147th day, causing the matrix disorder index to surge from the normal value of 3.2 to 15.7, far exceeding the preset threshold of 10. The system analyzed the feature vectors and determined that the anomaly mainly occurred in the temperature and salinity parameters, thereby identifying the characteristics of the cold water mass. The event capture results are shown in Table 3:

[0235] Table 3 Cold water mass event matrix disorder index table

[0236]

[0237]

[0238] The correlation model between environmental parameters and measurement deviation significantly improved the accuracy of long-term monitoring. After analyzing six months of operating data, the researchers used multiple regression analysis to establish the influence function of environmental factors on measurement accuracy. Temperature changes have the most significant impact on conductivity sensors, and the regression coefficient β 12 =0.182, indicating that for every 1°C change in temperature, the salinity measurement deviation will change by 0.182 PSU. The influence of pressure on the acoustic Doppler sensor is also obvious, and the regression coefficient β 24 =0.008, indicating that every 100 bar pressure change will result in a 0.8 cm / s deviation in flow rate measurement. Based on this model, the system automatically applies the correction factor when the temperature changes by more than 0.5°C or the pressure changes by more than 5 bar, improving measurement accuracy by 38%.

[0239] Continuity error trend identification and sliding window weighted averaging successfully addressed the issue of sensor data jumps. When the temperature sensor on buoy A05 exhibited slight drift, the system detected that the first-order difference values ​​of eight consecutive data points exceeded three standard deviations of the normal range and showed a monotonically increasing trend. The system immediately applied a sliding window weighted averaging method with a 30-minute window length, six sampling points, and an exponential decay function with an α=0.5 weighting. This processing significantly improved data smoothness, effectively suppressed short-term jumps, and achieved a smooth transition in the temperature curve.

[0240] The minimum spanning tree algorithm achieved significant results in optimizing the communication network structure. The system modeled the eight buoys as an undirected weighted graph, calculating the communication link weights based on inter-buoy distances, signal-to-noise ratios, and transmission losses. Using the Kruskal algorithm to construct the minimum spanning tree, the total weight dropped from 1842.5 in the initial network to 962.3, reducing communication energy consumption by 47.8%. When buoy A03 displaced 1.8 meters on the 243rd day due to ocean currents, the system dynamically updated the minimum spanning tree structure, maintaining the optimal state of the communication network.

[0241] For large-scale data processing, the system applies matrix block computing techniques for parallel data processing. The raw data is first subjected to wavelet transform for noise reduction. The db4 wavelet basis function is used for a four-layer decomposition. Soft thresholding is then applied to the high-frequency coefficients to reconstruct the signal. The noise standard deviation is estimated to be 0.035, based on which a threshold λ = 0.087 is calculated, effectively removing measurement noise while preserving signal characteristics. Outlier detection identifies and addresses approximately 1.2% of the total data. After distributed preprocessing, the data volume is compressed to 23.8% of its original volume while retaining all key information.

[0242] The Kalman filter algorithm successfully addressed the cumulative error caused by sensor drift. The state-space model includes the measurement values ​​and drift rates of each sensor. The process noise covariance matrix Q and the measurement noise covariance matrix R are estimated based on the previous two months' operating data. Through iterative prediction and update steps, the system tracks state changes in real time and provides optimal estimates. After 12 months of operation, the temperature sensor drift of the buoy A02 reached 0.08°C. The Kalman filter successfully corrected this drift, maintaining the measurement error within ±0.02°C. The processed, optimized data is stored in a binary structured format on the data storage device, including timestamps, location information, correction parameters, and measurement values ​​with confidence intervals.

[0243] An adaptive weighted error optimization function significantly improves the accuracy of multi-sensor data fusion. The historical accuracy score is based on the previous six months of calibration samples. The environmental stability score considers temperature, pressure, and salinity, with weighting coefficients of 0.5, 0.3, and 0.2, respectively. The consistency score compares the measurement consistency of different sensors, and the time drift score is proportional to the system's operating time, with a coefficient of γ = 0.8. Based on these scores, the system dynamically adjusts the data fusion weights, significantly suppressing error accumulation. The optimized measurement data confidence level increased by 42%.

[0244] Compared to traditional deep-sea monitoring systems, this invention addresses several core technical challenges. Traditional systems primarily rely on fixed parameter models and simple data recording methods, making them incapable of adapting to the complex and changing deep-sea environment and possessing limited data processing capabilities. This invention, by incorporating adaptive algorithms and multidimensional data analysis techniques, achieves environmental adaptability, data compression, and anomaly detection. Using Kalman filtering and an adaptive weighting algorithm, this system successfully controls the cumulative error over long periods of operation within acceptable limits, maintaining stable and reliable performance even in harsh deep-sea environments.

[0245] It should be noted that the variables involved in the present invention are explained in detail as shown in Tables 4 and 5 below.

[0246] Table 4 Variable Explanation Table (Part 1)

[0247]

[0248]

[0249] Table 5 Variable Explanation Table (Part 2)

[0250]

Claims

1. A real-time deep-sea buoy monitoring system, characterized in that: It includes a control chip, a multi-parameter sensor array, a data storage device and a power supply; the control chip is electrically connected to the multi-parameter sensor array, the data storage device and the power supply respectively; the multi-parameter sensor array is used to collect ocean temperature stratification data, salinity distribution data and current velocity data; the data storage device is used to record all collected monitoring data and system operation status parameters; the power supply is used to provide stable power for the entire system; a monitoring data recording module is provided in the control chip to process the raw data collected by the multi-parameter sensor array; the monitoring data recording module introduces an error optimization function based on an adaptive weighted algorithm and a deep-sea environmental disturbance compensation network model to identify the source of error and the propagation path, establish a correlation model between environmental parameters and measurement deviations, dynamically adjust the data fusion weights, obtain optimized data and save it to the data storage device.

2. The real-time deep-sea buoy monitoring system according to claim 1, characterized in that: The monitoring data recording module is used to perform the following steps: record the ocean temperature stratification data, salinity distribution data and current velocity data collected by the multi-parameter sensor array; use the temperature stratification data to calibrate the sound wave propagation model, and calculate the precise relative position of each buoy based on the signal round-trip delay time and the received signal strength index; construct the collected ocean data into a three-dimensional data matrix, extract the main change patterns through matrix singular value decomposition, and achieve data compression through low-rank matrix approximation.

3. The real-time deep-sea buoy monitoring system according to claim 2, characterized in that: The monitoring data recording module is also used to perform the following steps: applying a matrix sliding window method to process the time series data, establishing a covariance matrix of the multidimensional measurement data, calculating a matrix disorder index, and identifying abnormal data patterns; Based on the correlation model between environmental parameters and measurement deviation, the impact of environmental factors on sensor measurement accuracy is analyzed; Continuously monitor data streams, perform time series analysis on collected raw data, identify continuity error trends, and use a sliding window weighted average method to eliminate data jumps; The data transmission network structure between latent buoys is optimized based on the minimum spanning tree algorithm.

4. The real-time deep-sea buoy monitoring system according to claim 3, characterized in that: The monitoring data recording module is also used to perform the following steps: distributed pre-processing of collected multi-source heterogeneous data, parallel processing of large-scale data using matrix block computing technology, extraction of key features, compression of repeated redundant information, and reduction of data storage capacity; Based on the historical data in the data storage device and combined with the real-time measurement values, the Kalman filter algorithm is applied to model and predict the cumulative error, thereby realizing online correction and optimization of the measurement data.

5. The real-time deep-sea buoy monitoring system according to claim 4, characterized in that: The sliding window weighted average method specifically sets a data window of 30 minutes in length on the time series, assigns a time-decay weight to each data point in the window, and the data point closer to the current moment has a greater weight. The weighted average replaces the original measurement value, and the window slides forward every 10 seconds to continuously update the data series.

6. The real-time deep-sea buoy monitoring system according to claim 5, characterized in that: The Kalman filter algorithm specifically establishes a system state space model, models the measurement noise and system noise as Gaussian white noise respectively, and iteratively calculates the system state estimate and covariance matrix through prediction and update steps. The prediction step uses the state at the previous moment to estimate the current state, and the update step combines the current measurement value to correct the prediction result, and finally outputs the optimal state estimate.

7. The real-time deep-sea buoy monitoring system according to claim 6, characterized in that: Continuity error trend identification specifically establishes a normal fluctuation range model for each sensor data, calculates the first-order difference and second-order difference of 20 adjacent data points, and when the difference value of 5 consecutive data points exceeds 3 times the standard deviation of the normal range and shows a monotonically increasing or monotonically decreasing trend, it is determined that a continuity error exists and the error repair process is triggered.

8. The real-time deep-sea buoy monitoring system according to claim 7, characterized in that: Matrix singular value decomposition specifically constructs the ocean environment parameters into a three-dimensional tensor of dimension m×n×t, where m represents the horizontal spatial dimension, n represents the vertical depth dimension, and t represents the time dimension. The tensor decomposition method is used to decompose the original high-dimensional data into the product of three low-dimensional matrices, retaining the eigenvectors corresponding to the most significant k singular values, and reconstructing the compressed low-rank approximate matrix.

9. The real-time deep-sea buoy monitoring system according to claim 8, characterized in that: The matrix disorder index specifically calculates the ratio of the condition number of the measurement data covariance matrix to the matrix spectral norm, quantifying the uniformity of the matrix eigenvalue distribution. The larger the condition number, the higher the data disorder. When the matrix disorder index exceeds the preset threshold of 10, the anomaly detection process is triggered.

10. The real-time deep-sea buoy monitoring system according to claim 9, characterized in that: In the process of establishing and applying the association model, a pre-trained deep-sea environmental disturbance compensation network model was introduced. The specific structure of the deep-sea environmental disturbance compensation network model is composed of a six-layer deep neural network, including two input encoding layers, three hidden processing layers and one output decoding layer. The input encoding layer is responsible for receiving and standardizing environmental parameter data, and the sparse attention mechanism is implanted in the hidden processing layer to capture the long-term correlation and short-term fluctuation characteristics between different environmental parameters. The output decoding layer generates a correction coefficient matrix for various sensors.

Images