Error correction system and method for marine environment observation data

Abstract

The invention discloses an error correction system and method for marine environment observation data. The error correction system comprises a data acquisition module, a compensation modeling module, an error analysis module, a data compensation module, a data fusion module and a feedback optimization module. The data acquisition module acquires ADCP flow measurement, GPS, IMU, temperature, pressure and water quality monitoring data. And the error analysis module detects a drift error by using Kalman filtering and a hidden Markov model, and analyzes an environmental error based on a long short-term memory network. The data compensation module optimizes a compensation model and improves data precision. And the data fusion module fuses multi-sensor data by adopting a particle filter and a Bayesian optimization algorithm. The feedback optimization module dynamically adjusts the weight of the compensation model according to the error correction data, and improves the adaptability of the system. According to the invention, the accuracy of ocean current monitoring data is improved, and more reliable data support is provided for underwater environment research.

Description

Technical Field

[0001] The present invention relates to the technical field of marine water quality monitoring, and particularly to an error correction system and method for marine environmental observation data. Background Art

[0002] Marine environmental monitoring plays an important role in the fields of marine resource development, marine ecological protection, and navigation safety. The accurate measurement of ocean current data is crucial for applications such as marine environmental research, maritime navigation, search and rescue operations, and pollution monitoring. Currently, common ocean current monitoring devices include buoy monitoring stations, unmanned underwater vehicles, and underwater sensor arrays, which are usually equipped with an acoustic Doppler current profiler (ADCP), a global positioning system (GPS), an inertial measurement unit (IMU), and water quality sensors, etc.

[0003] However, during the long-term operation of existing ocean current monitoring systems, affected by factors such as the complexity of the marine environment and sensor drift, the measured data has errors. For example, the ADCP is affected by water body turbulence, attitude changes, and instrument errors, and may generate drift errors; the cumulative errors of the GPS and IMU will also reduce the positioning accuracy; in addition, changes in environmental factors such as temperature and pressure may have an unstable impact on water quality sensing data. Currently, the commonly used error correction methods mainly rely on single data sources or simple empirical models, and cannot make full use of multi-source data for error compensation, resulting in limited correction effects.

[0004] Therefore, there is an urgent need for a system that can comprehensively utilize multi-source data and combine advanced algorithms to effectively correct drift errors and environmental errors, so as to improve the accuracy and reliability of ocean current monitoring data. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technology, the purpose of the present invention is to provide an error correction system and method for marine environmental observation data, which is used to improve the accuracy and reliability of ocean current monitoring data.

[0006] To achieve the above purpose, the present invention provides the following technical solutions: An error correction system for marine environmental observation data, applicable to water quality sensing devices on buoy monitoring stations, unmanned underwater vehicles, and underwater sensor arrays, includes: A data acquisition module, configured to collect in real time the ADCP flow measurement data, GPS data, IMU data, temperature data, pressure data on the water quality sensing device, and collect the water quality monitoring data of the water quality sensing device; A compensation modeling module, configured to obtain a plurality of historical training data of the water quality sensing device, and construct an initial compensation model based on each of the historical training data; An error analysis module, connected to the data acquisition module, is used to perform anomaly detection on the ADCP flow measurement data, the GPS data, and the IMU data according to the improved Kalman filter algorithm and the hidden Markov model to obtain drift errors, and perform anomaly detection on the temperature data and the pressure data according to the long short-term memory network to obtain environmental errors; A data compensation module, connected to the error analysis module, the compensation modeling module, and the data acquisition module, is used to correct the errors of the initial compensation model according to the drift error and the environmental error to obtain an optimized compensation model, and the optimized compensation model is used to perform error compensation on the water quality monitoring data to obtain compensated monitoring data; A data fusion module, connected to the data compensation module, is used to fuse and optimize the compensated monitoring data of multiple the sensing devices according to the particle filter algorithm and the Bayesian optimization algorithm to obtain error correction data; A feedback optimization module, connected to the data fusion module and the data compensation module, is used to process the compensated monitoring data and the error correction data to obtain deviation feedback data, and dynamically adjust the weight parameters of the optimized compensation model based on the deviation feedback data to obtain the adjusted optimized compensation model.

[0007] Further, the historical training data includes historical error data, historical true water quality data, and historical sensing monitoring data, and the compensation modeling module includes: A preprocessing unit, used to perform data preprocessing on each of the historical error data, each of the historical true water quality data, and each of the historical sensing monitoring data to obtain multiple preprocessed historical error data, multiple historical true water quality data, and multiple historical sensing monitoring data; An introduction unit, used to introduce a CNN-LSTM hybrid model as the original model; A training unit, connected to the introduction unit and the preprocessing unit, is used to use each of the preprocessed historical error data and each of the historical sensing monitoring data as inputs, and use the corresponding historical true water quality data as outputs to retrain the original model to obtain the initial compensation model.

[0008] Further, the error analysis module includes a drift error detection unit, and the drift error detection unit includes: A preliminary detection subunit, used to perform a preliminary estimate of the drift error according to the improved Kalman filter algorithm to obtain a preliminary error, and then analyze the time series characteristics of the preliminary error according to the hidden Markov model and predict the preliminary error trend; A burst detection subunit, used to detect burst errors according to the double sliding window method and the isolation forest algorithm; An enhanced detection subunit, configured to obtain historical data of the ADCP flow measurement data and the IMU data, and input the historical data into a preset bidirectional long short-term memory network to obtain a prediction error; A fusion detection subunit, respectively connected to the preliminary detection subunit, the burst detection subunit, and the enhanced detection subunit, configured to fuse the preliminary error trend, the burst error, and the prediction error to obtain the drift error.

[0009] Further, the error analysis module includes an environmental error detection unit, and the environmental error detection unit includes: An error pre-screening subunit, configured to calculate the mean, variance, skewness, and kurtosis of the temperature and pressure data, and obtain screened data after removing abnormal data based on the 3σ principle; A modeling and prediction subunit, connected to the error pre-screening subunit, configured to obtain historical temperature data, historical pressure data, and historical environmental errors, construct a unidirectional long short-term memory network based on the historical temperature data, the historical pressure data, and the historical environmental errors, and input the screened data into the unidirectional long short-term memory network to predict the environmental error; A model adjustment subunit, connected to the modeling and prediction subunit, configured to calculate the credibility of the environmental error according to Bayesian inference and dynamically adjust the weights of the unidirectional long short-term memory network based on the credibility.

[0010] Further, the data compensation module includes: An error feature extraction unit, configured to align the drift error, the environmental error, and the water quality monitoring data in time, and extract error features of the drift error and the environmental error according to the principal component analysis method and the empirical mode decomposition method; An incremental learning unit, connected to the error feature extraction unit, configured to, based on the incremental learning method, enable the initial compensation model to online learn the extracted error features and perform error correction to obtain a corrected compensation model; A dynamic optimization unit, connected to the incremental learning unit, configured to dynamically adjust the hyperparameters of the corrected compensation model according to the Bayesian optimization algorithm to obtain the optimized compensation model.

[0011] Further, the data fusion module includes: A multi-modal feature extraction unit, configured to align the compensated monitoring data of multiple the sensing devices in time, and extract sensing features of each of the compensated monitoring data according to the principal component analysis method and the deep learning feature extraction algorithm; A weight adjustment unit, connected to the multi-modal feature extraction unit, is configured to calculate the feature importance of each of the sensing features according to information entropy, and dynamically adjust the weight of the corresponding compensation monitoring data of the sensing features according to the feature importance and the Bayesian optimization algorithm; A weighted fusion unit, connected to the weight adjustment unit, is configured to perform weighted fusion on the adjusted weights and the corresponding compensation monitoring data to obtain preliminary corrected data; A timing optimization unit, connected to the weighted fusion unit, is configured to perform timing optimization on the preliminary corrected data according to the particle filter algorithm to obtain error-corrected data.

[0012] Further, the feedback optimization module includes: A deviation calculation unit, configured to calculate deviation feedback data from the compensation monitoring data and the error-corrected data according to the Euclidean distance; A compensation processing unit, connected to the deviation calculation unit, is configured to process the deviation feedback data to obtain an improvement in compensation accuracy; A model optimization unit, connected to the compensation processing unit, is configured to dynamically adjust the weight parameters of the optimized compensation model according to the improvement in compensation accuracy to obtain the adjusted optimized compensation model.

[0013] Further, the calculation formula of the deviation feedback data is configured as: ; wherein, is used to represent the deviation feedback data, is used to represent the compensation monitoring data, is used to represent the error-corrected data, is used to represent the total number of the deviation feedback data.

[0014] An error correction method for ocean environment observation data, applied to the above-mentioned error correction system for ocean environment observation data, includes: Step S1, a data acquisition module collects in real time the ADCP flow measurement data, GPS data, IMU data, temperature data, pressure data on the water quality sensing device, and collects the water quality monitoring data of the water quality sensing device; Step S2, a compensation modeling module obtains a plurality of historical training data of the water quality sensing device, and constructs an initial compensation model based on each of the historical training data; Step S3, an error analysis module performs anomaly detection on the ADCP flow measurement data, the GPS data, and the IMU data according to the improved Kalman filter algorithm and the hidden Markov model to obtain drift errors, and performs anomaly detection on the temperature data and the pressure data according to the long short-term memory network to obtain environmental errors; Step S4, the data compensation module corrects the errors of the initial compensation model according to the drift error and the environmental error to obtain an optimized compensation model, and the optimized compensation model is used to perform error compensation on the water quality monitoring data to obtain compensated monitoring data; Step S5, the feedback optimization module processes the compensated monitoring data and the error correction data to obtain deviation feedback data, and dynamically adjusts the weight parameters of the optimized compensation model based on the deviation feedback data to obtain the adjusted optimized compensation model.

[0015] Advantages of the present invention: By acquiring ADCP flow measurement data, GPS data, IMU data, temperature data, pressure data and water quality monitoring data, the present invention comprehensively utilizes a variety of sensor information, effectively reduces the errors that may be brought by a single data source, and improves the integrity and reliability of the data; The present invention also uses the Kalman filter and the hidden Markov model to detect the drift error, and combines the long short-term memory network to analyze the environmental error, which can accurately identify data anomalies caused by sensor drift, environmental changes, etc., and improve the accuracy of error detection; The present invention dynamically optimizes the compensation model according to the drift error and the environmental error to ensure the real-time and effectiveness of error correction, makes the sea current monitoring data more accurate and stable, and then dynamically adjusts the compensation model based on the error correction data to realize the adaptive optimization of the compensation model, and improves the applicability and long-term stability of the system in different environments; In addition, the present invention adopts the particle filter algorithm and the Bayesian optimization algorithm to fuse and optimize the data of multiple sensing devices, reduce the measurement deviation between sensors, and improve the credibility and accuracy of the final error correction data; In summary, the present invention effectively improves the accuracy and reliability of the sea current monitoring data, and provides higher-quality data support for marine environmental research, navigation safety, pollution monitoring and underwater operations. Description of the drawings

[0016] Figure 1 is a schematic structural diagram of the sea current drift error correction system in the present invention; Figure 2 is a schematic structural diagram of the compensation modeling module in the present invention; Figure 3 is a schematic structural diagram of the error analysis module in the present invention; Figure 4 is a schematic structural diagram of the data compensation module in the present invention; Figure 5 is a schematic structural diagram of the data fusion module in the present invention; Figure 6 is a schematic structural diagram of the feedback optimization module in the present invention; Figure 7 It is a flowchart of the steps of the ocean current drift error correction method in the present invention.

[0017] Reference numerals: 1, data acquisition module; 2, compensation modeling module; 21, preprocessing unit; 22, introduction unit; 23, training unit; 3, error analysis module; 31, drift error detection unit; 311, preliminary detection subunit; 312, burst detection subunit; 313, enhanced detection subunit; 314, fusion detection subunit; 32, environmental error detection unit; 321, error initial screening subunit; 322, modeling prediction subunit; 323, model adjustment subunit; 4, data compensation module; 41, error feature extraction unit; 42, incremental learning unit; 43, dynamic optimization unit; 5, data fusion module; 51, multi-modal feature extraction unit; 52, weight adjustment unit; 53, weighted fusion unit; 54, timing optimization unit; 6, feedback optimization module; 61, deviation calculation unit; 62, compensation processing unit; 63, model optimization unit. Specific implementation manners

[0018] The present invention will be further described in detail below with reference to the drawings and embodiments. The same components are denoted by the same reference numerals. It should be noted that the terms "front", "rear", "left", "right", "upper" and "lower" used in the following description refer to the directions in the drawings, and the terms "bottom surface" and "top surface", "inner" and "outer" refer to the directions towards or away from the geometric center of a specific component, respectively.

[0019] Example 1, referring to Figure 1 , which is the first embodiment of the present invention. This embodiment provides an error correction system for ocean environment observation data, which can effectively improve the accuracy and reliability of ocean current monitoring data and is applicable to water quality sensing devices on buoy monitoring stations, unmanned underwater vehicles, and underwater sensor arrays, including: A data acquisition module 1, configured to collect ADCP flow measurement data, GPS data, IMU data, temperature data, pressure data on the water quality sensing device in real time, and collect water quality monitoring data of the water quality sensing device; A compensation modeling module 2, configured to obtain a plurality of historical training data of the water quality sensing device and construct an initial compensation model based on each historical training data; An error analysis module 3, connected to the data acquisition module 1, configured to perform anomaly detection on the ADCP flow measurement data, GPS data, and IMU data according to the improved Kalman filter algorithm and the hidden Markov model to obtain drift errors, and perform anomaly detection on the temperature data and pressure data according to the long short-term memory network to obtain environmental errors; The data compensation module 4 is connected to the error analysis module 3, the compensation modeling module 2, and the data acquisition module 1. It is used to correct the errors of the initial compensation model according to the drift error and the environmental error to obtain an optimized compensation model, and the optimized compensation model is used to compensate the errors of the water quality monitoring data to obtain compensated monitoring data; The data fusion module 5 is connected to the data compensation module 4. It is used to fuse and optimize the compensated monitoring data of multiple sensing devices according to the particle filter algorithm and the Bayesian optimization algorithm to obtain error-corrected data; The feedback optimization module 6 is connected to the data fusion module 5 and the data compensation module 4. It is used to process the compensated monitoring data and the error-corrected data to obtain deviation feedback data, and dynamically adjust the weight parameters of the optimized compensation model based on the deviation feedback data to obtain an adjusted optimized compensation model.

[0020] Working principle of Embodiment 1: The data acquisition module 1 is deployed on the buoy monitoring station, the autonomous underwater vehicle, and the underwater sensor array, and is responsible for real-time acquisition of ADCP flow measurement data, GPS data, IMU data, temperature data, pressure data, and water quality monitoring data. The ADCP is used to measure the sea current velocity and direction, the GPS provides position reference, the IMU monitors the device attitude change, and the temperature and pressure sensors record the environmental parameters.

[0021] The compensation modeling module 2 trains the initial compensation model based on long-term historical data. In the laboratory environment, a large amount of historical training data is used, and a compensation model based on a deep neural network is adopted to enable it to have preliminary error correction capabilities.

[0022] The error analysis module 3 uses the Kalman filter and the hidden Markov model to detect errors in the ADCP, GPS, and IMU data. The GPS signal may be affected by sea waves or signal occlusion, resulting in short-term drift. In this embodiment, the HMM is used to identify abnormal states and combined with the KF for smoothing processing. In addition, in this embodiment, the long short-term memory network is used to analyze the temperature and pressure data to identify the influence of environmental factors on the sensor readings, such as the influence of deep-sea high pressure on the ADCP flow measurement accuracy.

[0023] The data compensation module 4 dynamically optimizes the initial compensation model based on the error analysis results. For example, for the ADCP drift error, the system uses the Bayesian optimization method to adjust the compensation model parameters to make the compensation result more accurate.

[0024] The data fusion module 5 uses the particle filter algorithm and the Bayesian optimization algorithm to fuse and optimize the data of multiple sensing devices. For example, the underwater sensors at different depths may measure slightly deviated flow velocity data, and the system performs weighted averaging through the particle filter algorithm to make the final error-corrected data more stable.

[0025] The feedback optimization module 6 utilizes the error correction data after data fusion to dynamically adjust the weight parameters of the optimization compensation model. For example, if it is found that the IMU data has a drift trend in the system, the influence of the IMU data on the final calculation result is adaptively reduced, and the weights of the GPS and ADCP data are increased.

[0026] In actual tests, this embodiment was deployed on a buoy monitoring station and an AUV in a certain sea area. By comparing the corrected sea current data of this system with high-precision reference measurement data (such as shipborne high-precision ADCP measurement data), it was found that this system can effectively reduce the drift error, improve the data accuracy by about 30%, and reduce the flow measurement error to within ±0.05 m / s, showing a significant improvement compared with the traditional single-sensor error correction method (the error is about ±0.15 m / s).

[0027] This embodiment can be widely applied to fields such as marine environmental monitoring, marine scientific research, maritime traffic navigation safety, underwater archaeology, and marine resource development, providing reliable technical support for accurate sea current data.

[0028] In summary, this embodiment proposes a sea current drift error correction system based on multi-source data fusion, which can significantly improve the accuracy and reliability of sea current monitoring data, and has better adaptability and real-time performance compared with traditional error compensation methods.

[0029] Preferably, the historical training data includes historical error data, historical true water quality data, and historical sensing monitoring data. Refer to Figure 2 , then the compensation modeling module 2 includes: A preprocessing unit 21 for performing data preprocessing on each historical error data, each historical true water quality data, and each historical sensing monitoring data to obtain multiple preprocessed historical error data, multiple preprocessed historical true water quality data, and multiple preprocessed historical sensing monitoring data; An introduction unit 22 for introducing a CNN-LSTM hybrid model as the original model; A training unit 23, connected to the introduction unit 22 and the preprocessing unit 21, for using each preprocessed historical error data and each historical sensing monitoring data as inputs and the corresponding historical true water quality data as outputs to retrain the original model to obtain an initial compensation model.

[0030] Specifically, in this embodiment, the preprocessing unit 21 performs data cleaning and format conversion on the historical error data, historical true water quality data, and historical sensing monitoring data to improve the stability and accuracy of model training. The main processing flow of the preprocessing unit 21 is as follows: Data denoising: Using the Savitzky-Golay filtering method to smooth the original data and remove high-frequency noise; Data Standardization: Standardize all input data (historical error data, historical sensor monitoring data) to meet the input requirements of the CNN-LSTM model; Data Alignment: Synchronize historical error data, historical true water quality data, and historical sensor monitoring data based on timestamps to ensure the temporal consistency of data input; Data Splitting: Divide the data into a training set and a test set according to the ratio of 80% for training and 20% for testing.

[0031] The CNN-LSTM hybrid model consists of a CNN part and an LSTM part. The CNN part extracts features from historical sensor monitoring data and historical error data to identify the spatial correlation of the data. The LSTM part processes temporal data, captures the patterns of historical error data changing over time, and improves the model's prediction ability for time series data.

[0032] Combination method of the CNN-LSTM hybrid model: First, use CNN to extract local features of the data (e.g., the spatial relationship between historical error data and historical sensor monitoring data); then, pass the features output by CNN to LSTM for time series modeling to learn the error change pattern; finally, combine the fully connected layer to output the final error compensation result.

[0033] The training process of the model by the training unit 23 includes: Model Initialization: Load the CNN-LSTM structure and set parameters such as the number of network layers, the size of the convolutional kernel, and the number of LSTM units.

[0034] Forward Propagation: Input historical error data and historical sensor monitoring data, extract features through CNN and then pass them to LSTM, and finally output the predicted water quality data through the fully connected layer.

[0035] Loss Calculation: Use the mean squared error (MSE) as the loss function to calculate the error between the model output value and the historical true water quality data.

[0036] Backward Propagation and Optimization: Use the Adam optimizer to update the weights of CNN and LSTM to minimize the loss function.

[0037] Model Training: Iteratively train until convergence, and finally obtain an initial compensation model for error compensation.

[0038] The initial compensation model is finally tested at the buoy monitoring station in a certain sea area: Collected 10,000 groups of historical error data, historical sensor monitoring data, and historical true water quality data for 6 months for training and testing.

[0039] Model performance: The MSE error decreased by 40%. Compared with the traditional BP neural network model, CNN-LSTM has better time series modeling ability; the error of water quality monitoring data decreased by 25%, effectively improving the accuracy of water quality data; the computing efficiency increased by 30%. By using CNN for dimensionality reduction to extract features, the efficiency of LSTM in processing time series data is improved. Through this embodiment, it can be proved that the initial compensation model trained based on CNN-LSTM can effectively reduce errors and improve the reliability of ocean water quality monitoring.

[0040] Preferably, referring to Figure 3 , the error analysis module 3 includes a drift error detection unit 31, and the drift error detection unit 31 includes: A preliminary detection subunit 311, configured to preliminarily estimate the drift error according to the improved Kalman filter algorithm to obtain a preliminary error, and then analyze the time series characteristics of the preliminary error according to the hidden Markov model, and predict the preliminary error trend; A burst detection subunit 312, configured to detect the burst error according to the double sliding window method and the isolation forest algorithm; An enhanced detection subunit 313, configured to obtain the historical data of the ADCP flow measurement data and the IMU data, and input them into a preset bidirectional long short-term memory network to obtain a predicted error; A fusion detection subunit 314, respectively connected to the preliminary detection subunit 311, the burst detection subunit 312, and the enhanced detection subunit 313, configured to fuse the preliminary error trend, the burst error, and the predicted error to obtain the drift error.

[0041] Specifically, in this embodiment, the extended Kalman filter (EKF), the hidden Markov model (HMM), the double sliding window method, the isolation forest algorithm, and the bidirectional long short-term memory network (Bi-LSTM) are used for ocean current drift error detection.

[0042] Among them, the preliminary detection subunit 311 uses the improved Kalman filter to fuse the ADCP data and the IMU data, estimate the drift error, and output the preliminary error value.

[0043] Then, the hidden Markov model (HMM) is used to calculate the time series characteristics of the preliminary error, such as autocorrelation, trend change rate, etc.; the future trend of the error is also predicted through the HMM to obtain the preliminary error trend.

[0044] The burst detection subunit 312 sets a short-term window and a long-term window by using the double sliding window method, and calculates the mean and standard deviation of the drift error respectively; when the mean of the short-term window exceeds the 3σ threshold of the long-term window, the burst error point is marked.

[0045] The burst detection subunit 312 also trains an unsupervised anomaly detection model using the Isolation Forest algorithm to identify abnormal drift errors; at the same time, it combines the sliding window method to eliminate normal errors and retain burst error points.

[0046] The enhanced detection subunit 313 uses a bidirectional long short-term memory network for time series prediction, and this process includes: Training the Bi-LSTM model with historical ADCP and IMU data as inputs; Predicting the error change trend within a short time window to obtain predicted error data.

[0047] The fusion detection subunit 314 calculates the final drift error based on a weighted fusion method, and the calculation formula is configured as: Where, represents the final drift error, represents the preliminary error trend, represents the burst error, represents the predicted error, are the first fusion coefficient, the second fusion coefficient, and the third fusion coefficient set in advance, which are dynamically adjusted by the Bayesian optimization algorithm to ensure the optimal weights of different errors.

[0048] Preferably, the error analysis module 3 includes an environmental error detection unit 32, and the environmental error detection unit 32 includes: An error preliminary screening subunit 321, which is used to calculate the mean, variance, skewness, and kurtosis of temperature and pressure data, and obtain the screened data after eliminating abnormal data based on the 3σ principle; A modeling and prediction subunit 322, connected to the error preliminary screening subunit 321, which is used to obtain historical temperature data, historical pressure data, and historical environmental errors, construct a unidirectional long short-term memory network based on the historical temperature data, historical pressure data, and historical environmental errors, and input the screened data into the unidirectional long short-term memory network to predict the environmental error; A model adjustment subunit 323, connected to the modeling and prediction subunit 322, which is used to calculate the credibility of the environmental error according to Bayesian inference and dynamically adjust the weights of the unidirectional long short-term memory network based on the credibility.

[0049] Specifically, in this embodiment, the error preliminary screening subunit 321 calculates the mean, variance, skewness, and kurtosis of temperature and pressure data, analyzes the data distribution, and then eliminates abnormal values based on the 3σ principle: if the data exceeds the mean ±3σ, it is considered abnormal, and the abnormal data points are eliminated and recorded. The output of the error preliminary screening subunit 321 is the screened data after eliminating abnormalities.

[0050] The modeling and prediction subunit 322 trains a unidirectional long short-term memory network. The model adjustment subunit 323 first calculates the credibility of the environmental error according to Bayesian inference: calculates the confidence interval of the prediction error and sets the credibility threshold; then dynamically adjusts the LSTM weights: if the credibility is low, increases the weight of historical data to make the model more sensitive to long-term trends; if the credibility is high, maintains the current model parameters to improve the real-time prediction ability. The output of the model adjustment subunit 323 is the adjusted predicted value of the environmental error.

[0051] The unidirectional long short-term memory network is finally experimented at the buoy monitoring station in a certain sea area: Data source: Collect 6 months of ADCP, GPS, IMU, temperature, and pressure data, a total of 20,000 groups of data; Detection effect: The detection accuracy of the drift error is improved by 35% (the isolation forest algorithm effectively reduces the false detection rate); The detection error of the environmental error is reduced by 20% (Bayesian optimization improves the prediction credibility of LSTM); The overall error compensation accuracy is improved by 30%, and the water quality monitoring data is more credible.

[0052] Through this embodiment, it can be proved that the error detection method integrating multiple algorithms can effectively improve the reliability of marine environmental monitoring.

[0053] Preferably, referring to Figure 4 , the data compensation module 4 includes: An error feature extraction unit 41, which is used to align the drift error, environmental error, and water quality monitoring data in time, and extract error features from the drift error and environmental error according to the principal component analysis method and the empirical mode decomposition method; An incremental learning unit 42, connected to the error feature extraction unit 41, is used to make the initial compensation model online learn the extracted error features and perform error correction based on the incremental learning method to obtain a corrected compensation model; A dynamic optimization unit 43, connected to the incremental learning unit 42, is used to dynamically adjust the hyperparameters of the corrected compensation model according to the Bayesian optimization algorithm to obtain an optimized compensation model.

[0054] Specifically, this embodiment is used to align the error data in time and extract error features through the principal component analysis method and the empirical mode decomposition method to ensure that the initial compensation model can make full use of the error patterns.

[0055] The error feature extraction unit 41 uses the linear interpolation method to align different data sources in time to ensure that all data has the same time dimension. Then, the dynamic time warping (DTW) method is used to match the time series data to improve the alignment accuracy of data with different sampling rates.

[0056] The error feature extraction unit 41 extracts the main error information by calculating the eigenvectors of the drift error and the environmental error, reduces the data dimension, and improves the calculation efficiency.

[0057] The error feature extraction unit 41 also decomposes the drift error and the environmental error into intrinsic mode functions (IMFs) with different frequencies, removes high-frequency noise, and improves the interpretability of the error patterns.

[0058] The incremental learning unit 42 enables the error compensation model to have the online learning ability based on the incremental learning method, and can continuously optimize the error correction effect.

[0059] The incremental learning unit 42 allows it to dynamically update the weights; each time the initial compensation model receives new error feature data, it uses the parameter transfer method for fine-tuning instead of retraining the entire model; The model after incremental learning is compared with the initial compensation model. If the new model performs better, it is updated as the correction compensation model; the sliding window method is used to limit the time range of the model memory to avoid overfitting of historical data.

[0060] The dynamic optimization unit 43 uses the Bayesian optimization algorithm to dynamically adjust the hyperparameters of the correction compensation model to further improve the error compensation accuracy.

[0061] The processing steps of Bayesian optimization for hyperparameter adjustment: Set the hyperparameter search space (such as learning rate, number of neural network layers, regularization parameter, etc.).

[0062] Use Gaussian process regression (GPR) to estimate the effects of different hyperparameter configurations.

[0063] Use the Bayesian optimization strategy: during the hyperparameter search process, introduce the expected improvement (EI) criterion, and preferentially try high-potential hyperparameter combinations; calculate the contribution of hyperparameters to the error compensation effect, and gradually optimize the hyperparameter configuration. Iteratively optimize until the optimal parameter combination is found, and the output is the final optimized compensation model.

[0064] The optimized compensation model is experimented at the buoy monitoring station in a certain sea area: Data source: Collect 15,000 groups of data of ADCP, GPS, IMU, temperature, pressure, and water quality monitoring data for 6 months.

[0065] Error compensation effect: After the error features are extracted by the principal component analysis method and the empirical mode decomposition method, the error reduction is 25%; the incremental learning method improves the adaptability of the error compensation model by 30%; the Bayesian optimization further improves the error compensation accuracy by 20%.

[0066] This embodiment proves that the data compensation method combining the principal component analysis method, the empirical mode decomposition method, incremental learning, and Bayesian optimization can effectively improve the error correction accuracy and enhance the reliability of water quality monitoring data.

[0067] Preferably, referring to Figure 5 , the data fusion module 5 includes: A multimodal feature extraction unit 51, configured to perform time alignment on the compensated monitoring data of multiple sensing devices, and extract sensing features from each compensated monitoring data according to the principal component analysis method and the deep learning feature extraction algorithm; A weight adjustment unit 52, connected to the multimodal feature extraction unit 51, configured to calculate the feature importance of each sensing feature according to information entropy, and dynamically adjust the weights of the compensated monitoring data corresponding to the sensing features according to the feature importance and the Bayesian optimization algorithm; A weighted fusion unit 53, connected to the weight adjustment unit 52, configured to perform weighted fusion on the adjusted weights and the corresponding compensated monitoring data to obtain preliminary corrected data; A timing optimization unit 54, connected to the weighted fusion unit 53, configured to perform timing optimization on the preliminary corrected data according to the particle filter algorithm to obtain error corrected data.

[0068] Specifically, in this embodiment, the multimodal feature extraction unit 51 uses the linear interpolation method to align the compensated monitoring data of multiple sensing devices at the same time point, and at the same time calculates the time deviation between different compensated monitoring data; finally, non-linear alignment is performed to improve data consistency.

[0069] The multimodal feature extraction unit 51 calculates the principal components of the data through principal component analysis, and extracts the most representative sensing features; furthermore, through dimensionality reduction processing, redundancy is reduced and the calculation efficiency is improved.

[0070] The multimodal feature extraction unit 51 also uses a convolutional neural network (CNN) to extract spatial features; at the same time, a bidirectional long short-term memory network (Bi-LSTM) is used to identify time series features to improve the timing information modeling ability.

[0071] The weight adjustment unit 52 dynamically adjusts the importance weights of the sensing features based on information entropy calculation and Bayesian optimization: the processing steps for calculating feature importance: calculate the information entropy of each sensing feature; a feature with a high entropy value indicates a high information richness and is assigned a higher weight; Bayesian optimization dynamically adjusts the weights: set the weight search space, use Gaussian process regression (GPR) to estimate the effects of different weight combinations; adopt the expected improvement (EI) criterion to optimize the weights to minimize the error.

[0072] The weighted fusion unit 53 uses a weighted fusion method to perform data fusion according to the adjusted weights, improving the error correction accuracy. The timing optimization unit 54 performs timing optimization using a particle filter algorithm to improve the stability of the fused data.

[0073] The processing steps of particle filter timing optimization include: Initializing the particle set: setting multiple particles to represent different error correction possibilities; randomly initializing the weights of each particle through a Gaussian distribution.

[0074] Updating the particle state: calculating the residual between the observed value and the predicted value; updating the particle weights so that particles with smaller errors obtain higher weights.

[0075] Resampling: particles with low weights are eliminated, and particles with high weights are replicated to reduce sampling errors.

[0076] The output is the final error correction data.

[0077] Preferably, referring to Figure 6 , the feedback optimization module 6 includes: A deviation calculation unit 61 for calculating deviation feedback data obtained by processing compensation monitoring data and error correction data according to the Euclidean distance; A compensation processing unit 62, connected to the deviation calculation unit 61, for obtaining an improved amount of compensation accuracy according to the deviation feedback data; A model optimization unit 63, connected to the compensation processing unit 62, for dynamically adjusting the weight parameters of the learning optimization compensation model according to the improved amount of compensation accuracy to obtain an adjusted optimized compensation model.

[0078] Preferably, the calculation formula of the deviation feedback data is configured as: ; Among them, is used to represent the deviation feedback data, is used to represent the compensation monitoring data, is used to represent the error correction data, is used to represent the total number of deviation feedback data.

[0079] Specifically, in this embodiment, the compensation processing unit 62 is used to calculate the improvement amount of error compensation based on the deviation feedback data to ensure that the compensation model can be continuously optimized.

[0080] The compensation processing unit 62 calculates the average error of the most recent time steps using a moving average filter:

[0081] Among them, represents the average error, Indicates the th deviation feedback data.

[0082] Meanwhile, the compensation processing unit 62 sets the compensation accuracy adjustment rule: If is higher than the preset compensation threshold, the compensation amount is increased; if is lower than the preset compensation threshold, the compensation amount is decreased.

[0083] The compensation processing unit 62 calculates the compensation accuracy improvement amount using the exponential weighted method:

[0084] Where: is the smoothing coefficient, which controls the influence of historical data on the current compensation; is the current compensation accuracy improvement amount.

[0085] Based on the compensation accuracy improvement amount, the model optimization unit 63 dynamically adjusts the weight parameters of the error compensation model to continuously adapt to new error information.

[0086] The specific steps for the model optimization unit 63 to adjust and optimize the weight parameters of the compensation model: Dynamically adjust the model weights based on Bayesian optimization: Set the model weight parameters The goal of weight adjustment is to find the optimal weight combination to minimize the error:

[0087] In this process, Gaussian process regression (GPR) is used to estimate the error compensation effect of different weight combinations; meanwhile, the expected improvement strategy is used to adjust the weights to improve the compensation accuracy.

[0088] During the process of the model optimization unit 63 dynamically updating the compensation model: If the error of the optimized compensation model is reduced by more than the threshold, update the model parameters; if the error shows no obvious improvement, keep the original model weights.

[0089] Finally, the output of the model optimization unit 63 is the adjusted optimized compensation model.

[0090] An error correction method for ocean environment observation data, applied to the above-mentioned error correction system for ocean environment observation data, referring to Figure 7 , includes: Step S1, the data acquisition module 1 continuously acquires ADCP flow measurement data, GPS data, IMU data, temperature data, pressure data on the water quality sensing device, and acquires the water quality monitoring data of the water quality sensing device; Step S2: The compensation modeling module 2 obtains multiple historical training data of the water quality sensing device and constructs an initial compensation model based on each historical training data; Step S3: The error analysis module 3 performs anomaly detection on the ADCP flow measurement data, GPS data, and IMU data according to the improved Kalman filter algorithm and the hidden Markov model to obtain drift errors, and performs anomaly detection on the temperature data and pressure data according to the long short-term memory network to obtain environmental errors; Step S4: The data compensation module 4 corrects the errors of the initial compensation model according to the drift errors and environmental errors to obtain an optimized compensation model, and the optimized compensation model is used to perform error compensation on the water quality monitoring data to obtain compensated monitoring data; Step S5: The feedback optimization module 6 processes the compensated monitoring data and the error correction data to obtain deviation feedback data, and dynamically adjusts the weight parameters of the optimized compensation model based on the deviation feedback data to obtain an adjusted optimized compensation model.

[0091] The above is only the preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, multiple improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. An error correction system for marine environmental observation data, applicable to water quality sensing devices on buoy monitoring stations, unmanned underwater vehicles, and underwater sensor arrays, characterized in that, Including: A data acquisition module (1) for real-time acquisition of ADCP flow measurement data, GPS data, IMU data, temperature data, pressure data on the water quality sensing device, and acquisition of water quality monitoring data of the water quality sensing device; A compensation modeling module (2) for obtaining a plurality of historical training data of the water quality sensing device and constructing an initial compensation model based on each of the historical training data; An error analysis module (3) connected to the data acquisition module (1) for performing anomaly detection on the ADCP flow measurement data, the GPS data, and the IMU data according to the improved Kalman filter algorithm and the hidden Markov model to obtain drift errors, and performing anomaly detection on the temperature data and the pressure data according to the long short-term memory network to obtain environmental errors; A data compensation module (4) connected to the data acquisition module (1), the compensation modeling module (2), and the error analysis module (3) for correcting the errors of the initial compensation model according to the drift errors and the environmental errors to obtain an optimized compensation model, and the optimized compensation model is used for error compensation of water quality monitoring data to obtain compensated monitoring data; A data fusion module (5) connected to the data compensation module (4) for fusing and optimizing the compensated monitoring data of multiple sensing devices according to the particle filter algorithm and the Bayesian optimization algorithm to obtain error-corrected data; A feedback optimization module (6) connected to the data fusion module (5) and the data compensation module (4) for processing the compensated monitoring data and the error-corrected data to obtain deviation feedback data, and dynamically adjusting the weight parameters of the optimized compensation model based on the deviation feedback data to obtain the adjusted optimized compensation model.

2. The error correction system for marine environment observation data according to claim 1, characterized in that: The historical training data includes historical error data, historical true water quality data, and historical sensing monitoring data, and the compensation modeling module (2) includes: A preprocessing unit (21) for preprocessing each of the historical error data, each of the historical true water quality data, and each of the historical sensing monitoring data to obtain a plurality of preprocessed historical error data, a plurality of historical true water quality data, and a plurality of historical sensing monitoring data; An introduction unit (22) for introducing a CNN-LSTM hybrid model as the original model; A training unit (23) connected to the introduction unit (22) and the preprocessing unit (21) for using each of the preprocessed historical error data and each of the historical sensing monitoring data as inputs and the corresponding historical true water quality data as outputs to retrain the original model to obtain the initial compensation model.

3. The error correction system for marine environment observation data according to claim 1, characterized in that: The error analysis module (3) includes a drift error detection unit (31), and the drift error detection unit (31) includes: A preliminary detection subunit (311) for preliminarily estimating the drift error according to the improved Kalman filter algorithm to obtain a preliminary error, and then analyzing the time series characteristics of the preliminary error according to the hidden Markov model and predicting to obtain a preliminary error trend; A burst detection subunit (312) for detecting burst errors according to a double sliding window method and an isolation forest algorithm; An enhanced detection subunit (313) for obtaining historical data of the ADCP flow measurement data and the IMU data and inputting them into a preset bidirectional long short-term memory network to obtain prediction errors; A fusion detection subunit (314) respectively connected to the preliminary detection subunit (311), the burst detection subunit (312) and the enhanced detection subunit (313) for fusing the preliminary error trend, the burst error and the prediction error to obtain the drift error.

4. The error correction system for marine environmental observation data according to claim 1, characterized in that: The error analysis module (3) includes an environmental error detection unit (32), and the environmental error detection unit (32) includes: An error preliminary screening subunit (321) for calculating the mean, variance, skewness and kurtosis of the temperature and pressure data, and obtaining the screened data after removing abnormal data based on the 3σ principle; A modeling prediction subunit (322) connected to the error preliminary screening subunit (321) for obtaining historical temperature data, historical pressure data and historical environmental errors, constructing a unidirectional long short-term memory network based on the historical temperature data, the historical pressure data and the historical environmental errors, and inputting the screened data into the unidirectional long short-term memory network to predict the environmental error; A model adjustment subunit (323) connected to the modeling prediction subunit (322) for calculating the credibility of the environmental error according to Bayesian inference and dynamically adjusting the weights of the unidirectional long short-term memory network based on the credibility.

5. The error correction system for marine environment observation data according to claim 1, characterized in that: The data compensation module (4) includes: An error feature extraction unit (41) for time-aligning the drift error, the environmental error and the water quality monitoring data, and extracting error features of the drift error and the environmental error according to a principal component analysis method and an empirical mode decomposition method; An incremental learning unit (42) connected to the error feature extraction unit (41) for online learning the error features extracted by the initial compensation model based on an incremental learning method and performing error correction to obtain a corrected compensation model; A dynamic optimization unit (43) connected to the incremental learning unit (42) for dynamically adjusting the hyperparameters of the corrected compensation model according to a Bayesian optimization algorithm to obtain the optimized compensation model.

6. The error correction system for marine environment observation data according to claim 1, characterized in that: The data fusion module (5) includes: A multi-modal feature extraction unit (51) for time-aligning the compensated monitoring data of multiple the sensing devices, and extracting sensing features of each the compensated monitoring data according to a principal component analysis method and a deep learning feature extraction algorithm; A weight adjustment unit (52) connected to the multi-modal feature extraction unit (51) for calculating the feature importance of each the sensing feature according to information entropy, and dynamically adjusting the weights of the compensated monitoring data corresponding to the sensing feature according to the feature importance and a Bayesian optimization algorithm; A weighted fusion unit (53), connected to the weight adjustment unit (52), is configured to perform weighted fusion based on the adjusted weights and the corresponding compensation monitoring data to obtain preliminary correction data; A timing optimization unit (54), connected to the weighted fusion unit (53), is configured to perform timing optimization on the preliminary correction data according to the particle filter algorithm to obtain error correction data.

7. The error correction system for marine environment observation data according to claim 1, wherein: The feedback optimization module (6) includes: A deviation calculation unit (61), configured to calculate deviation feedback data according to the Euclidean distance between the compensation monitoring data and the error correction data; A compensation processing unit (62), connected to the deviation calculation unit (61), is configured to process the deviation feedback data to obtain an improvement in compensation accuracy; A model optimization unit (63), connected to the compensation processing unit (62), is configured to dynamically adjust and learn the weight parameters of the optimized compensation model according to the improvement in compensation accuracy to obtain the adjusted optimized compensation model.

8. The error correction system for ocean environmental observation data according to claim 7, characterized in that: The calculation formula of the deviation feedback data is configured as: ； Among them, is used to represent the deviation feedback data, is used to represent the compensation monitoring data, is used to represent the error correction data, is used to represent the total number of the deviation feedback data.

9. An error correction method for ocean environment observation data, applied to the error correction system for ocean environment observation data according to any one of claims 1-8, characterized in that, including: Step S1, a data acquisition module (1) real-time acquires the ADCP flow measurement data, GPS data, IMU data, temperature data, pressure data on the water quality sensing device, and acquires the water quality monitoring data of the water quality sensing device; Step S2, a compensation modeling module (2) obtains multiple historical training data of the water quality sensing device and constructs an initial compensation model based on each of the historical training data; Step S3, an error analysis module (3) performs anomaly detection on the ADCP flow measurement data, the GPS data, and the IMU data according to the improved Kalman filter algorithm and the hidden Markov model to obtain drift errors, and performs anomaly detection on the temperature data and the pressure data according to the long short-term memory network to obtain environmental errors; Step S4, a data compensation module (4) corrects the errors of the initial compensation model according to the drift errors and the environmental errors to obtain an optimized compensation model, and the optimized compensation model is used to perform error compensation on the water quality monitoring data to obtain compensation monitoring data; Step S5, a feedback optimization module (6) processes the deviation feedback data according to the compensation monitoring data and the error correction data, and dynamically adjusts the weight parameters of the optimized compensation model based on the deviation feedback data to obtain the adjusted optimized compensation model.

Images