A multi-source data fusion correction method for ocean observation station

Abstract

The application provides a marine observation station multi-source data fusion correction method, and belongs to the technical field of marine observation.The layered interpolation correction model containing a weighted moving average algorithm and a spatial collaborative interpolation algorithm is established, an adaptive threshold adjustment mechanism of the marine environment based on a closed-loop feedback system is constructed, a multi-factor coupling drift prediction model of a Kalman filter combined with a long short-term memory network is used, a gradual degradation fusion strategy switched according to environmental conditions is designed, a biological pollution gradual detection correction model using a liquid neural network structure is constructed, a multi-source data quality evaluation system based on data continuity, numerical stability, consistency and historical reliability is established, high-precision adaptive fusion correction of marine environment multi-source sensor data is realized, and the technical problem that marine environment multi-source sensor data is difficult to realize high-precision adaptive fusion correction is solved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of ocean observation, and specifically relates to a multi-source data fusion correction method for an ocean observation station. BACKGROUND

[0002] As an important infrastructure for ocean environment monitoring, ocean observation stations widely use arrays of multiple types of sensors to monitor real-time parameters such as temperature, humidity, wind speed, salinity, and pressure. Traditional techniques mainly rely on single sensor data or simple data averaging methods to obtain ocean environmental information, and play an important role in the fields of ocean engineering, ocean resource development, and ocean environment protection. However, traditional data processing methods have obvious limitations in the face of complex ocean environments. Single sensors are easily affected by environmental interference and produce abnormal data. Simple data averaging methods cannot effectively identify and correct sensor drift, and lack self-adaptive ability to environmental changes. In the current ocean observation system, due to the complexity and variability of the ocean environment, sensors deployed for a long time may encounter problems such as biological contamination, equipment aging, and environmental interference, leading to a decline in data quality. However, existing technologies lack effective multi-source data fusion mechanisms to comprehensively utilize the complementary information of different sensors, and cannot dynamically adjust data processing strategies according to environmental conditions. That is, there is a technical problem in the prior art that ocean environment multi-source sensor data is difficult to achieve high-precision adaptive fusion correction. SUMMARY

[0003] Therefore, the present application provides a multi-source data fusion correction method for an ocean observation station, which can solve the technical problem of high-precision adaptive fusion correction of ocean environment multi-source sensor data in the prior art.

[0004] The application is implemented in the following manner: the application provides a marine observation station multi-source data fusion correction method, which comprises the following steps: arranging a multi-type sensor array at a marine observation station, including temperature sensors, humidity sensors, wind speed sensors, salinity sensors and pressure sensors, the sensor array is distributed in a three-dimensional grid structure, and each sensor node is configured with a data acquisition module and a wireless transmission module; a layered interpolation correction model is established, a weighted moving average algorithm is used for single-point instantaneous anomalies, the weight function is a time decay exponential function, a spatial collaborative interpolation algorithm is used for regional continuous anomalies, and the weight is inversely proportional to the square of the distance between sensor nodes; an adaptive threshold adjustment mechanism for marine environment is constructed, the corrected data and the field calibration data are compared and analyzed through a closed-loop feedback system, and the abnormal detection threshold range of each type of sensor is automatically adjusted according to the change of environmental conditions; a multi-factor coupled drift prediction model is established, a Kalman filter is combined with a long short-term memory network structure, and input parameters include temperature gradient, salinity concentration gradient and time series data; a progressive degradation fusion strategy is designed, a four-level fusion mode switching mechanism is established according to weather conditions and sensor working states, and a comprehensive environmental evaluation function is used to calculate the fusion mode level; a biological pollution gradual detection correction model is constructed, a liquid neural network structure is used to monitor the slow change of sensor response characteristics, model input includes sensor historical response curve, biological attachment density estimation value and environmental parameter sequence, and output includes pollution degree evaluation value and correction coefficient; and a multi-source data quality evaluation system is established, the data quality of each sensor node is scored in real time, the scoring parameters include data continuity, numerical stability, consistency with adjacent sensors and historical reliability, the fusion weight is allocated according to the quality score, and the correction model parameters are updated.

[0005] In the step of distributing the sensor array in a three-dimensional grid structure, the distance between adjacent sensors is 5m to 15m, the sampling frequency of the data acquisition module configured in each sensor node is 1Hz, and the transmission frequency of the wireless transmission module is 915MHz.

[0006] In the weighted moving average algorithm, the weight function is a time decay exponential function, the decay constant of the time decay exponential function is determined by a historical fluctuation period, the last four effective data points are used for correction, the time decay exponential function is in the form of negative exponential decay, the decay constant is determined according to the historical fluctuation period, the historical fluctuation period is obtained by spectral analysis of sensor data in the past 12 months, and the decay constant is in the range of 0.1 to 0.3.

[0007] In the spatial collaborative interpolation algorithm, the inverse distance weighting method is used, the distance index in the weight calculation formula is the Euclidean distance, the weight coefficient is greater than 0.8 when the distance between sensor nodes is less than 3m, and the weight coefficient is less than 0.1 when the distance is greater than 20m.

[0008] The closed-loop feedback system comprises a data comparison module, a deviation analysis module and a threshold updating module.

[0009] The marine environment adaptive threshold adjustment mechanism specifically expands the humidity sensor threshold range by 10% to 15% during the high-salt-spray period.

[0010] The temperature gradient is calculated from the measurement values of adjacent temperature sensors, the salinity concentration gradient is calculated from the measurement values of adjacent salinity sensors, and the time series data is the historical measurement sequence of each sensor.

[0011] The comprehensive environment evaluation function is specifically established based on wind speed measurement, wave height measurement and visibility measurement, the wave height measurement is obtained by converting pressure sensor data, the visibility measurement is obtained by optical sensor, the comprehensive environment evaluation function is a weighted linear combination of wind speed measurement, wave height measurement and visibility measurement, and the weight coefficients are 0.4, 0.4 and 0.2 respectively.

[0012] The four-level fusion mode switching mechanism specifically sets the value range of the comprehensive environment evaluation function to 0 to 100, when the value of the comprehensive environment evaluation function is less than 25, the first-level fusion mode is selected, when the value of the comprehensive environment evaluation function is in the interval of 25 to 50, the second-level fusion mode is selected, when the value of the comprehensive environment evaluation function is in the interval of 50 to 75, the third-level fusion mode is selected, and when the value of the comprehensive environment evaluation function is greater than 75, the fourth-level fusion mode is selected.

[0013] The multi-factor coupled drift prediction model specifically uses Kalman filter for state prediction and error correction, and uses long short-term memory network for learning drift pattern, the long short-term memory network comprises an input layer, two hidden layers and an output layer, the number of hidden layer neurons is 64 and 32 respectively, the activation function uses hyperbolic tangent function, the network output is drift prediction value, and the drift prediction value is used for correcting sensor measurement data.

[0014] The biological attachment density estimation value is obtained by image analysis of underwater camera equipment.

[0015] Specifically, the training dataset for the multi-factor coupled drift prediction model is established by collecting long-term sensor monitoring data from different sea areas and seasons, recording the drift change curve of each sensor, and simultaneously recording the corresponding temperature gradient change sequence, salinity concentration gradient change sequence, and time label. A database of the correspondence between sensor drift and environmental factor changes is established, containing more than 50,000 historical records, covering a complete monitoring period of no less than 3 years.

[0016] Specifically, the liquid neural network structure comprises a reservoir computing unit and a readout layer. The reservoir computing unit consists of 100 liquid neurons, and the connection weights between neurons are dynamically adjusted based on three parameters: sensor deployment density, bio-attachment cycle, and seawater temperature. The leakage rate parameter of the reservoir neurons is set to 0.3, and the time constant is set to 20ms. The readout layer adopts a linear regression structure, and the output dimension is the number of sensors.

[0017] Specifically, the training dataset for the biofouling gradual detection and correction model is established by collecting long-term monitoring data from sensors in different seasons and sea areas, recording the response characteristic change curves of each sensor, and periodically photographing the surface condition of the sensors using underwater cameras. The degree of biofouling is quantified through image recognition technology, and a database of the correspondence between the sensor response characteristic change curves and the estimated biofouling density is established. The database contains more than 10,000 historical records.

[0018] The scoring algorithm of the multi-source data quality assessment system specifically adopts a weighted average method, with data continuity score weighting 0.3, numerical stability score weighting 0.25, consistency score weighting with neighboring sensors weighting 0.25, historical reliability score weighting 0.2, and the final quality score range from 0 to 100. Sensor data with scores below 60 are not included in the fusion calculation.

[0019] The data continuity is obtained by calculating the data missing rate, the numerical stability is obtained by calculating the standard deviation, the consistency with neighboring sensors is obtained by calculating the correlation coefficient, and the historical reliability is obtained by calculating the historical failure frequency.

[0020] This invention solves the technical problem of achieving high-precision adaptive fusion correction of multi-source sensor data in the marine environment by constructing a hierarchical interpolation correction model, an adaptive threshold adjustment mechanism for the marine environment, and a multi-factor coupled drift prediction model. The invention employs a hierarchical correction strategy combining a weighted moving average algorithm and a spatial collaborative interpolation algorithm, effectively overcoming the shortcomings of traditional methods in accurately identifying and correcting different types of abnormal data. A closed-loop feedback system enables dynamic adjustment of the anomaly detection threshold, avoiding the insufficient adaptability of fixed threshold settings under environmental changes. Simultaneously, a prediction model combining a Kalman filter and a long short-term memory network solves the problem of traditional methods' inability to effectively predict sensor drift. The progressive degradation fusion strategy and the gradual detection and correction model for biological pollution established in this invention fundamentally improve the system's data processing capabilities under harsh environmental conditions, and the multi-source data quality assessment system ensures the reliability and accuracy of the fusion results. In summary, this invention solves the technical problem mentioned in the background art of achieving high-precision adaptive fusion correction of multi-source sensor data in the marine environment. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a schematic diagram of the multi-factor coupled drift prediction model involved in the present invention.

[0023] Figure 3 This is a schematic diagram of the biological pollution gradual detection correction model involved in the present invention.

[0024] Figure 4 This is a diagram illustrating the fusion mode switching process during the passage of a typhoon in Example 2.

[0025] Figure 5 This is a diagram of the sensor's gradual detection process of biological contamination in Example 2.

[0026] Figure 6 This is a comparison chart of data quality before and after applying the present invention in Example 2. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0028] like Figure 1 The diagram shown is a flowchart of a multi-source data fusion and correction method for marine observation stations provided by this invention. This method includes the following steps:

[0029] S01. A multi-type sensor array is deployed at the marine observation station, including temperature sensors, humidity sensors, wind speed sensors, salinity sensors and pressure sensors. The sensor array is distributed in a three-dimensional grid structure, with the spacing between adjacent sensors being 5m to 15m. Each sensor node is equipped with a data acquisition module and a wireless transmission module. The sampling frequency of the data acquisition module is 1Hz, and the transmission frequency of the wireless transmission module is 915MHz.

[0030] S02. Establish a hierarchical interpolation correction model. For single-point instantaneous anomalies, a weighted moving average algorithm is adopted, with the weight function being a time decay exponential function. Four valid data points are taken before and after the correction. For regional continuous anomalies, a spatial collaborative interpolation algorithm is adopted. The weight is inversely proportional to the square of the distance between sensor nodes. The decay constant of the time decay exponential function is determined by the historical fluctuation period.

[0031] S03. Construct an adaptive threshold adjustment mechanism for the marine environment. Through a closed-loop feedback system, compare and analyze the corrected data with the field calibration data. Automatically adjust the abnormal detection threshold range of various sensors according to changes in environmental conditions. During periods of high salt spray incidence, the threshold range of the humidity sensor is expanded by 10% to 15%. The field calibration data is obtained through regular manual calibration.

[0032] S04. Establish a multi-factor coupled drift prediction model, using a Kalman filter combined with a long short-term memory network structure. The input parameters include temperature gradient, salinity concentration gradient and time series data. The temperature gradient is calculated from the measurement values ​​of adjacent temperature sensors, the salinity concentration gradient is calculated from the measurement values ​​of adjacent salinity sensors, and the time series data is the historical measurement sequence of each sensor.

[0033] S05. Design a progressive degradation fusion strategy, establish a four-level fusion mode switching mechanism based on weather conditions and sensor operating status, calculate the fusion mode level through a comprehensive environmental assessment function, which is based on wind speed measurement, wave height measurement and visibility measurement, the wave height measurement is obtained through pressure sensor data conversion, and the visibility measurement is obtained through an optical sensor.

[0034] S06. Construct a gradual detection and correction model for biological pollution. Use a liquid neural network structure to monitor the slow changes in the sensor response characteristics. The model input is the sensor's historical response curve, the estimated value of biological attachment density, and the environmental parameter sequence. The output is the pollution level assessment value and the correction coefficient. The estimated value of biological attachment density is obtained through image analysis of underwater camera equipment.

[0035] S07. Establish a multi-source data quality assessment system to score the data quality of each sensor node in real time. The scoring parameters include data continuity, numerical stability, consistency with neighboring sensors, and historical reliability. Based on the quality score, allocate fusion weights and update and correct model parameters. The quality score is used to generate the final fused data output.

[0036] The time decay exponential function of the weighted moving average algorithm is negative exponential decay, and the decay constant is determined based on the historical fluctuation cycle. The historical fluctuation cycle is obtained by spectral analysis of sensor data over the past 12 months, and the decay constant ranges from 0.1 to 0.3 to ensure that the weight of recent data is higher than that of long-term data.

[0037] The spatial collaborative interpolation algorithm adopts an inverse distance weighting method. The distance index in the weight calculation formula is Euclidean distance. When the distance between sensor nodes is less than 3m, the weight coefficient is greater than 0.8, and when the distance is greater than 20m, the weight coefficient is less than 0.1, ensuring a reasonable reflection of spatial correlation.

[0038] The closed-loop feedback system includes a data comparison module, a deviation analysis module, and a threshold update module. The data comparison module performs a comparison analysis every hour. The deviation analysis module calculates the root mean square error between the corrected data and the on-site calibration data. When the root mean square error exceeds a preset threshold, the threshold update module is triggered to adjust the abnormal detection threshold range.

[0039] The comprehensive environmental assessment function is a weighted linear combination of wind speed measurement, wave height measurement, and visibility measurement, with weighting coefficients of 0.4, 0.4, and 0.2, respectively. The value range of the comprehensive environmental assessment function is 0 to 100. When the comprehensive environmental assessment function value is less than 25, it corresponds to the first-level fusion mode; when the comprehensive environmental assessment function value is in the range of 25 to 50, it corresponds to the second-level fusion mode; when the comprehensive environmental assessment function value is in the range of 50 to 75, it corresponds to the third-level fusion mode; and when the comprehensive environmental assessment function value is greater than 75, it corresponds to the fourth-level fusion mode.

[0040] like Figure 2As shown, the multi-factor coupled drift prediction model employs a Kalman filter combined with a long short-term memory (LSTM) network structure. The Kalman filter is responsible for state prediction and error correction, while the LSM network is responsible for learning drift patterns. The LSM network includes an input layer, two hidden layers, and an output layer. The number of neurons in the hidden layers is 64 and 32, respectively. The activation function is the hyperbolic tangent function, and the network output is the drift prediction value, which is used to correct the sensor measurement data. The steps for establishing the training dataset for the multi-factor coupled drift prediction model specifically include collecting long-term sensor monitoring data from different sea areas and seasons, recording the drift change curve of each sensor, and simultaneously recording the corresponding temperature gradient change sequence, salinity concentration gradient change sequence, and time label. A database of the correspondence between sensor drift and environmental factor changes is established, containing more than 50,000 historical records covering a complete monitoring period of no less than 3 years. Each record includes sensor type identification, deployment coordinates, monitoring start time, and drift time series. The training process for the multi-factor coupled drift prediction model includes: temperature gradient time series, salinity concentration gradient time series, ambient temperature time series, ambient humidity time series, and manually calibrated records; specifically, the training dataset is divided into a training set, a validation set, and a test set in a 7:2:1 ratio; a sliding window time series segmentation method is used to ensure time continuity; the training process is divided into a Kalman filter parameter initialization stage and a long short-term memory network training stage; in the Kalman filter parameter initialization stage, the initial values ​​of the state transition matrix, observation matrix, process noise covariance matrix, and observation noise covariance matrix are determined by the maximum likelihood estimation method; in the long short-term memory network training stage, the backpropagation algorithm is used to optimize the network parameters, the loss function is the mean squared error, the learning rate is set to 0.001, the batch size is set to 32, the training cycle is 1000 iterations, the model performance is evaluated on the validation set every 100 iterations, and training is stopped early when the loss no longer decreases after 50 consecutive iterations; finally, the model's generalization ability is evaluated on the test set.

[0041] like Figure 3As shown, the biofouling gradual detection and correction model adopts a liquid neural network structure. The network includes a reservoir computing unit and a readout layer. The reservoir computing unit consists of 100 liquid neurons. The connection weights between neurons are dynamically adjusted based on three parameters: sensor deployment density, biofouling cycle, and seawater temperature. The sensor deployment density is calculated by the number of sensors per unit area. The biofouling cycle is obtained through periodic analysis of historical biofouling density estimates. The seawater temperature is measured by a temperature sensor. The learning rate parameter in the connection weight update formula is directly proportional to the sensor deployment density and inversely proportional to the biofouling cycle. The temperature influence coefficient... The leakage rate parameter of the reservoir neuron was set to 0.3 and the time constant to 20ms using sigmoid function mapping. The readout layer adopted a linear regression structure, with the output dimension being the number of sensors. Each output corresponds to a contamination correction coefficient for a sensor, which is used to correct the sensor measurement data. The steps for establishing the training dataset of the biocontamination gradual detection and correction model specifically include collecting long-term monitoring data of sensors in different seasons and sea areas, recording the response characteristic change curve of each sensor, and periodically photographing the sensor surface condition using underwater camera equipment. The degree of biofouling is quantified through image recognition technology to establish the sensor... A database mapping response characteristic change curves to estimated biofouling density is established. This database contains over 10,000 historical records, covering the complete cycle of the four seasons. Each record includes sensor type identification, deployment location coordinates, monitoring duration, response characteristic parameter sequence, estimated biofouling density sequence, ambient temperature data sequence, ambient humidity data sequence, ambient salinity data sequence, and comparative data before and after manual cleaning. The training steps for the biofouling gradual detection correction model specifically include dividing the training dataset into training and validation sets at an 8:2 ratio, and employing time-series cross-validation to ensure the model's generalization ability. The training process is divided into a pre-training phase and a fine-tuning phase. In the pre-training phase, a large amount of unlabeled sensor time-series data is used to train the internal representation capability of the reservoir computing unit. The training objective is to minimize the reconstruction error. The optimization algorithm adopts the adaptive moment estimation algorithm. The pre-training period is 500 epochs. In the fine-tuning phase, labeled contamination level data is used to train the readout layer parameters. The loss function adopts a linear combination of mean squared error and regularization term. The regularization coefficient is set to 0.01. The fine-tuning period is 200 epochs. After each epoch, the model performance is evaluated on the validation set. Training is stopped early when the validation loss no longer decreases after 20 consecutive epochs.

[0042] The scoring algorithm of the multi-source data quality assessment system adopts a weighted average method. The weight of data continuity is 0.3, the weight of numerical stability is 0.25, the weight of consistency with neighboring sensors is 0.25, and the weight of historical reliability is 0.2. The final quality score ranges from 0 to 100. Sensor data with a score below 60 are not included in the fusion calculation. The data continuity is obtained by calculating the data missing rate, the numerical stability is obtained by calculating the standard deviation, the consistency with neighboring sensors is obtained by calculating the correlation coefficient, and the historical reliability is obtained by calculating the historical failure frequency.

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

[0044] The specific implementation of step S01 involves systematically arranging multi-type sensor arrays within the observation station area according to a three-dimensional grid structure, based on the geographical environment and monitoring needs of the marine observation station. First, the spatial layout of the sensor array is determined, with temperature sensors, humidity sensors, wind speed sensors, salinity sensors, and pressure sensors distributed in a regular grid. The spacing between adjacent sensors is controlled within the range of 5m to 15m. This spacing setting is based on the spatial correlation principle of marine environmental parameters, ensuring sufficient spatial representativeness of sensor data while avoiding redundant arrangement. Each sensor node is equipped with a data acquisition module, using a sampling frequency of 1Hz for real-time data acquisition. This frequency setting considers the time-scale characteristics of marine environmental parameter changes and can effectively capture the dynamic changes in the marine environment. A wireless transmission module is also configured, with an operating frequency set to 915MHz. This frequency has good penetration and anti-interference capabilities, suitable for the wireless data transmission requirements of the marine environment. The sensor nodes adopt a combined power supply method of solar power system and lithium battery backup power supply to ensure the continuous and stable operation of the system.

[0045] The specific implementation of step S02 involves constructing a hierarchical interpolation correction model and employing differentiated correction strategies for different types of data anomalies. For single-point instantaneous anomalies, a weighted moving average algorithm is used for correction. This algorithm is based on time series analysis principles and estimates normal values ​​by analyzing valid data points before and after the anomaly point. The weighting function adopts a time decay exponential function to ensure that recent data has a higher weight. Four valid data points before and after the anomaly are used for correction calculations. The number of data points is set based on the time correlation analysis results of marine environmental parameters. The decay constant of the time decay exponential function is determined by analyzing the spectral analysis of sensor data from the past 12 months to determine historical fluctuation cycles. The decay constant is set to a range of 0.1 to 0.3 to ensure the timeliness and accuracy of the correction results. For regional continuous anomalies, a spatial collaborative interpolation algorithm is used. This algorithm is based on geostatistical principles and utilizes spatial correlation for data correction. The weight calculation adopts an inverse distance weighting method, where the weight is inversely proportional to the square of the Euclidean distance between sensor nodes. When the distance between sensor nodes is less than 3m, the weight coefficient is greater than 0.8, and when the distance is greater than 20m, the weight coefficient is less than 0.1, ensuring that spatial correlation is reasonably reflected in the correction process.

[0046] The specific implementation of step S03 involves constructing a marine environment adaptive threshold adjustment mechanism, achieving dynamic optimization of the anomaly detection threshold through closed-loop feedback control. The system establishes a closed-loop feedback system comprising three core components: a data comparison module, a deviation analysis module, and a threshold update module. The data comparison module performs a comparison analysis hourly, comparing the corrected sensor data with field calibration data obtained through periodic manual calibration. The deviation analysis module uses statistical analysis methods to calculate the root mean square error (RMSE) between the corrected data and the field calibration data; this error index comprehensively reflects the accuracy level of the data correction. When the RMS error exceeds a preset threshold, the threshold update module is triggered to automatically adjust the anomaly detection threshold range of various sensors. Based on the seasonal variation characteristics of the marine environment, the threshold range of the humidity sensor is expanded by 10% to 15% during periods of high salt spray incidence. This adjustment range is determined based on the statistical analysis results of historical environmental data, effectively adapting to the dynamic changes in the marine environment.

[0047] The specific implementation of step S04 involves establishing a multi-factor coupled drift prediction model, employing a hybrid architecture combining a Kalman filter and a long short-term memory network to achieve accurate sensor drift prediction. Input parameters include three key types of information: temperature gradient, salinity concentration gradient, and time-series data. The temperature gradient is obtained by calculating the difference between measurements from adjacent temperature sensors, the salinity concentration gradient by calculating the difference between measurements from adjacent salinity sensors, and the time-series data consists of historical measurement sequences from each sensor. The Kalman filter is responsible for state prediction and error correction, and based on linear dynamic systems theory, it can achieve optimal state estimation even in noisy environments. The long short-term memory network is responsible for learning complex drift patterns. This network structure includes an input layer, two hidden layers, and an output layer. The number of neurons in the hidden layers is set to 64 and 32, respectively, and the activation function is the hyperbolic tangent function. The network output is the predicted drift value. This hybrid architecture fully utilizes the real-time performance of the Kalman filter and the learning capability of the long short-term memory network to achieve accurate prediction and timely correction of sensor drift.

[0048] The specific implementation of step S05 involves designing a progressive degradation fusion strategy and establishing a four-level fusion mode switching mechanism based on marine environmental conditions and sensor operating status. First, a comprehensive environmental assessment function is established. This function is constructed based on a weighted linear combination of wind speed, wave height, and visibility measurements, with weight coefficients set to 0.4, 0.4, and 0.2, respectively. The weight allocation considers the importance of each environmental factor to the data fusion. Wave height measurements are obtained from pressure sensor data using the hydrostatic pressure principle, and visibility measurements are obtained from optical sensors based on the light scattering principle. The comprehensive environmental assessment function's value range is set from 0 to 100. When the function value is less than 25, it corresponds to the first-level fusion mode, employing high-precision fusion using all sensors; when the function value is between 25 and 50, it corresponds to the second-level fusion mode, employing key sensor fusion; when the function value is between 50 and 75, it corresponds to the third-level fusion mode, employing core sensor fusion; and when the function value is greater than 75, it corresponds to the fourth-level fusion mode, employing basic sensor fusion. This hierarchical strategy is based on reliability theory, ensuring system robustness through degradation fusion under harsh environmental conditions.

[0049] The specific implementation of step S06 involves constructing a gradual biofouling detection and correction model, employing a liquid neural network structure to monitor the slow changes in sensor response characteristics. The liquid neural network comprises two main components: a reservoir computation unit and a readout layer. The reservoir computation unit consists of 100 liquid neurons, a number chosen based on a balance between sensor array size and computational complexity. The connection weights between neurons are dynamically adjusted based on three key parameters: sensor density, biofouling cycle, and seawater temperature. Sensor density is obtained by calculating the number of sensors per unit area; the biofouling cycle is obtained through periodic analysis of historical biofouling density estimates; and seawater temperature is obtained directly from temperature sensors. The leakage rate parameter of the reservoir neurons is set to 0.3, and the time constant is set to 20 ms. These parameter settings are optimized based on the dynamic characteristics of the liquid neural network. The readout layer employs a linear regression structure, with the output dimension equal to the number of sensors, and each output corresponding to a biofouling correction coefficient for one sensor. The estimated biofouling density is obtained by acquiring images using underwater cameras and performing image analysis using computer vision technology. This method enables a quantitative assessment of the degree of biofouling.

[0050] The specific implementation of step S07 involves establishing a multi-source data quality assessment system to quantitatively evaluate the data quality of each sensor node in real time. The scoring algorithm employs a weighted average method, comprehensively considering four key indicators: data continuity, numerical stability, consistency with neighboring sensors, and historical reliability. The data continuity score weight is set at 0.3, obtained by calculating the data missing rate, reflecting the completeness of sensor data acquisition; the numerical stability score weight is set at 0.25, obtained by calculating the standard deviation of the measurement data, reflecting the stability of sensor measurements; the consistency score weight with neighboring sensors is set at 0.25, obtained by calculating the correlation coefficient, reflecting the spatial consistency of sensor data; and the historical reliability score weight is set at 0.2, obtained by statistically analyzing historical failure frequencies, reflecting the long-term reliability level of the sensor. The final quality score range is set from 0 to 100. Sensor data with scores below 60 are not included in the fusion calculation; this threshold is determined based on the actual needs of data quality control. Fusion weights are dynamically allocated based on the quality scores, and model parameters are updated and corrected in real time to achieve adaptive optimization of the data fusion process.

[0051] The key technical ideas of this invention are analyzed as follows. The first key technical idea is a multi-factor coupled drift prediction model. This model, through the organic combination of a Kalman filter and a long short-term memory network, achieves accurate prediction and active correction of sensor drift. Compared to traditional passive calibration methods, this technology can actively predict sensor drift trends based on environmental factors such as temperature gradients and salinity concentration gradients, significantly improving the foresight and accuracy of data correction. The second key technical idea is a biological contamination gradual detection and correction model, which uses a liquid neural network structure to monitor the gradual change process of sensor response characteristics. Compared to traditional periodic cleaning and maintenance methods, this technology can monitor the gradual impact of biological contamination in real time, compensate for measurement deviations caused by contamination through dynamic correction coefficients, significantly reducing maintenance costs and improving data quality. The third key technical idea is a progressive degradation fusion strategy, which intelligently switches fusion modes according to environmental conditions and sensor status. Compared to traditional fixed fusion modes, this technology can ensure continuous system operation under harsh environmental conditions through degradation fusion, significantly improving the system's environmental adaptability and robustness. The synergistic effect of these three key technological approaches has created an intelligent and adaptive ocean observation data processing system. Compared with existing technologies, it has achieved a technological leap from passive correction to active prediction, from regular maintenance to real-time monitoring, and from fixed modes to dynamic adjustments, comprehensively improving the quality and reliability of ocean observation data.

[0052] The detailed structure of the multi-factor coupled drift prediction model includes an integrated architecture of a Kalman filter component and a long short-term memory (LSM) network component. The Kalman filter handles the linear dynamic process, describing the system's dynamic characteristics through the state transition matrix, observation matrix, process noise covariance matrix, and observation noise covariance matrix. The LSM network comprises an input layer, two hidden layers, and an output layer. The hidden layer contains 64 and 32 neurons, respectively, and the activation function is the hyperbolic tangent function. The training dataset for this model involves collecting long-term sensor monitoring data from different sea areas and seasons, recording the drift curves of each sensor, as well as the corresponding temperature gradient change sequences, salinity concentration gradient change sequences, and time labels. A database containing over 50,000 historical records, covering a complete monitoring period of at least three years, is established. The training process consists of three stages: dataset partitioning, Kalman filter parameter initialization, and LSM network training. The training, validation, and test sets are partitioned in a 7:2:1 ratio. The initial parameters of the Kalman filter are determined using maximum likelihood estimation, and the network parameters are optimized using backpropagation.

[0053] The detailed structure of the biofouling gradual detection and correction model is based on a liquid neural network architecture, comprising two core components: a reservoir computation unit and a readout layer. The reservoir computation unit consists of 100 liquid neurons, with connection weights dynamically adjusted based on sensor density, biofouling cycle, and seawater temperature. The leakage rate parameter for the reservoir neurons is set to 0.3, and the time constant is set to 20 ms. The readout layer employs a linear regression structure, with the output dimension being the number of sensors, and each output corresponding to a biofouling correction coefficient for one sensor. The training dataset for this model includes collecting long-term sensor monitoring data from different seasons and sea areas, periodically photographing the sensor surface using underwater cameras, quantifying the degree of biofouling through image recognition technology, and establishing a database containing over 10,000 historical records, covering the complete cycle of spring, summer, autumn, and winter. The training process is divided into a pre-training phase and a fine-tuning phase. The pre-training phase uses a large amount of unlabeled time-series data to train the reservoir computation unit, while the fine-tuning phase uses labeled data with biofouling level tags to train the readout layer parameters.

[0054] It should be noted that the multi-factor coupled drift prediction model, by integrating the influence of multiple environmental factors, can accurately predict the drift behavior of sensors in complex marine environments, solving the problem of insufficient accuracy in traditional single-factor prediction models. Compared with existing simple linear regression prediction methods and correction methods based on empirical formulas, this model has stronger nonlinear modeling capabilities and environmental adaptability. The biofouling gradual detection and correction model, based on the dynamic characteristics of liquid neural networks, can capture the gradual process of biofouling, solving the problem that traditional periodic calibration methods cannot respond to changes in pollution in real time. Compared with existing maintenance methods based on fixed-period cleaning and simple judgment methods based on threshold detection, this model has higher sensitivity and continuity. The synergistic application of the two models enables intelligent control of sensor data quality from multiple dimensions, forming a complete data correction technology system, significantly improving the accuracy and reliability of marine observation data.

[0055] It should be noted that this invention also solves the following technical problems: The difficulty in accurately detecting and compensating for the impact of biofouling on sensors under long-term deployment environments. In the marine environment, sensors inevitably become affected by marine organism attachment after long-term deployment. This biofouling is a gradual process, causing slow changes in sensor response characteristics. Traditional technologies struggle to identify and quantify this gradual impact. This invention constructs a biofouling gradual detection and correction model, employing a liquid neural network structure to monitor the slow changes in sensor response characteristics. Combined with estimates of biofouling density obtained from underwater cameras, it establishes a correlation between changes in sensor response characteristics and the degree of biofouling. Through the collaborative work of the reservoir computing unit and the readout layer, it can accurately assess the degree of contamination and calculate the corresponding correction coefficients, thereby effectively compensating for the impact of biofouling on sensor measurement accuracy. The invention also addresses the lack of flexibility in data fusion strategies for marine observation systems under harsh environmental conditions. The marine environment is highly dynamic and complex. Under different weather conditions and sea states, the operating status and data quality of sensors change significantly. Traditional fixed fusion strategies cannot adapt to such environmental changes. This invention designs a progressive degradation fusion strategy and establishes a comprehensive environmental assessment function based on wind speed, wave height, and visibility. It automatically switches between four fusion modes according to changes in environmental conditions. Under harsh environmental conditions, a more conservative and reliable fusion strategy is adopted, while under favorable environmental conditions, a more refined and efficient fusion method is used. This adaptive fusion strategy ensures that the system can provide reliable data fusion results under various environmental conditions, significantly improving the environmental adaptability and data processing flexibility of the marine observation system.

[0056] Specifically, the principle of this invention is as follows: The core principle of this invention's technical solution, which addresses the difficulty of achieving high-precision adaptive fusion correction of multi-source sensor data in marine environments, lies in constructing a multi-level, multi-dimensional data fusion correction system. First, the hierarchical interpolation correction model distinguishes between single-point instantaneous anomalies and regional continuous anomalies, employing different algorithms in the time and spatial domains for targeted processing. The time decay exponential function ensures high weighting of recent data, while the spatial collaborative interpolation algorithm utilizes the spatial correlation between sensor nodes. This hierarchical processing strategy aligns with the different characteristics and distribution patterns of marine environmental data anomalies. Second, the marine environment adaptive threshold adjustment mechanism achieves dynamic threshold optimization through a closed-loop feedback system. It continuously compares the corrected data with on-site calibration data, automatically adjusting the detection threshold according to changes in environmental conditions. This adaptive mechanism conforms to the time-varying characteristics of the marine environment. The multi-factor coupled drift prediction model combines the state estimation capability of a Kalman filter with the sequence learning capability of a long short-term memory network. Through the comprehensive input of temperature gradient, salinity concentration gradient, and time series data, it can accurately predict the sensor drift trend. This multi-factor coupling method conforms to the multi-factor influence mechanism of sensor drift. A progressive degradation fusion strategy establishes a four-level fusion mode based on environmental conditions. The processing strategy is dynamically switched through a comprehensive environmental assessment function, ensuring the system's stability and reliability under different environmental conditions. The biofouling gradual detection correction model employs a liquid neural network structure, capable of adapting to slow changes in sensor response characteristics. It adjusts the correction coefficient by monitoring changes in biofouling density and environmental parameters; this gradual detection mechanism aligns with the gradual nature of biofouling. The multi-source data quality assessment system provides a scientific basis for fusion weight allocation through comprehensive scoring of data continuity, numerical stability, consistency, and historical reliability, ensuring high accuracy and reliability of the final fusion result.

[0057] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0058] In this embodiment, the specific implementation of step S01 is the same as described above, and will not be repeated in detail here.

[0059] The specific implementation of step S02 involves constructing a hierarchical interpolation correction model and using a weighted moving average algorithm to correct single-point instantaneous anomalies. The mathematical expression of this algorithm is as follows:

[0060] ;

[0061] In the formula, For the first Correction values ​​for each outlier data point; For the first Measurement values ​​of one valid data point; For the first The weighting coefficients for each data point; For the time index of abnormal data points; This is the time index for the valid data points. The weighting function uses a time-decay exponential function, specifically expressed as:

[0062] ;

[0063] In the formula, This is the attenuation constant, with a value ranging from 0.1 to 0.3; This represents the time interval between outliers and valid data points. For regional continuous outliers, a spatial collaborative interpolation algorithm is used, employing an inverse distance weighting method. The weight calculation formula is as follows:

[0064] ;

[0065] In the formula, For the first The first abnormal sensor is relative to the first Weighting coefficients for each normal sensor; For the first The first abnormal sensor and the first Euclidean distance between normal sensors; The number of normal sensors involved in the interpolation calculation; Spatial index for abnormal sensors; This is the spatial index for a normal sensor. The formula for calculating Euclidean distance is:

[0066] ;

[0067] In the formula, For the first The coordinates of the abnormal sensor; For the first The coordinate positions of normal sensors. The formula for calculating the correction value of abnormal sensors is:

[0068] ;

[0069] In the formula, For the first Correction values ​​for each abnormal sensor; For the first The measured value of a normal sensor.

[0070] The specific implementation of step S03 involves constructing a marine environment adaptive threshold adjustment mechanism. The deviation analysis module calculates the root mean square error between the corrected data and the on-site calibration data. The calculation formula is as follows:

[0071] ;

[0072] In the formula, This is the root mean square error; To compare the total number of data points; For the first One field calibration data value; For the first One corrected data value; This is the index for the data points being compared. When... Exceeding the preset threshold When this occurs, the threshold update mechanism is triggered, and the new anomaly detection threshold calculation formula is:

[0073] ;

[0074] In the formula, The updated anomaly detection threshold; The original anomaly detection threshold; The adjustment factor ranges from 0.1 to 0.15. This is the preset root mean square error threshold.

[0075] The specific implementation of step S04 involves establishing a multi-factor coupled drift prediction model. The temperature gradient is calculated using measurements from adjacent temperature sensors, and the calculation formula is as follows:

[0076] ;

[0077] In the formula, For position Temperature gradient at that location; For position Temperature measurement value at the location; and They are respectively and Sensor spacing in the direction; and The sensors are respectively in and The grid coordinate index of the direction. Similarly, the formula for calculating the salt concentration gradient is:

[0078] ;

[0079] In the formula, For position The salt concentration gradient at the location; For position The measured salt concentration at the location.

[0080] The specific implementation of step S05 involves designing a progressive degradation fusion strategy. The mathematical expression of the comprehensive environmental evaluation function is as follows:

[0081] ;

[0082] In the formula, This is the value of the comprehensive environmental assessment function, ranging from 0 to 100. Standardized values ​​for wind speed measurements; The standardized value of the wave height measurement; This is a standardized value for visibility measurements. The standardization formula is:

[0083] ;

[0084] In the formula, This is the actual measured wind speed value; and These represent the minimum and maximum wind speeds, respectively. Wave height measurements are obtained through data conversion from pressure sensors; the conversion formula is:

[0085] ;

[0086] In the formula, This is a dynamic pressure measurement value; The density of seawater is approximately 1025 kg / m³. ; The acceleration due to gravity is taken as 9.8 m / s². .

[0087] The specific implementation of step S06 involves constructing a biological pollution gradual detection correction model. The sensor deployment density is calculated by the number of sensors per unit area, using the following formula:

[0088] ;

[0089] In the formula, Determine the density of sensor placement; The total number of sensors within the observation area; This represents the total area of ​​the observation region. The learning rate parameter in the connection weight update formula is calculated as follows:

[0090] ;

[0091] In the formula, The learning rate parameter is used for the connection weights; The base learning rate is set to 0.01. For reference sensor arrangement density; This is the biological attachment cycle; For reference period; Seawater temperature; This is the temperature influence coefficient, with a value of 0.1. for Type activation function. The mathematical expression for the type activation function is:

[0092] ;

[0093] In the formula, For input variables.

[0094] The specific implementation of step S07 is to establish a multi-source data quality assessment system, and the calculation formula for the quality score is as follows:

[0095] ;

[0096] In the formula, For the final quality score; Score the continuity of the data; Assess numerical stability. A score is given for consistency with neighboring sensors; Historical reliability is scored. Data continuity score is calculated based on the data missing rate:

[0097] ;

[0098] In the formula, This represents the number of missing data points. This represents the total number of data points. The numerical stability score is calculated using the standard deviation.

[0099] ;

[0100] In the formula, The standard deviation of the measurement data; The reference standard deviation threshold is used. The consistency score with neighboring sensors is calculated using the correlation coefficient.

[0101] ;

[0102] In the formula, This is the Pearson correlation coefficient with data from neighboring sensors. The formula for calculating the Pearson correlation coefficient is:

[0103] ;

[0104] In the formula, and The first Measurements from the target sensor and nearby sensors at any given time; and These are the average values ​​measured by the two sensors, respectively. The total number of data points to calculate the correlation coefficient; Indexed by time. Historical reliability scores are calculated based on historical failure frequencies:

[0105] ;

[0106] In the formula, The attenuation coefficient is set to 5. This represents the historical failure frequency.

[0107] Among them, the time decay exponential function of the weighted moving average algorithm Based on the principle of exponential decay, weight allocation is achieved through a negative exponential mapping of time distance. This function ensures that recent data has a higher weight while the weight of older data gradually decreases. Compared to the traditional equal-weighted averaging method, this formula can better preserve the temporal characteristics of the data, effectively suppress the impact of outliers on the correction results, and improve the accuracy of single-point outlier correction. The inverse distance weighting formula of the spatial collaborative interpolation algorithm. Based on the first law of geostatistics, namely that spatial correlation weakens with increasing distance, weight allocation is achieved using the reciprocal of the square of the distance, where Euclidean distance... This formula ensures accurate calculation of spatial distance. Compared to traditional simple averaging interpolation methods, it fully utilizes the spatial distribution characteristics of sensors, ensuring that spatial correlation is reasonably reflected in the correction process, and significantly improving the accuracy of regional anomaly correction. Root Mean Square Error Calculation Formula Based on statistical principles, the root mean square error (RMSE) is used to quantify the correction accuracy. This formula comprehensively reflects the overall deviation level between the corrected data and the reference data. Compared to the simple mean absolute error, the RMS error is more sensitive to large deviations, providing a more reliable basis for threshold adjustment. Adaptive Threshold Adjustment Formula Based on feedback control theory, this formula achieves dynamic adjustment of the anomaly detection threshold by using the ratio of error to threshold. Compared to a fixed threshold method, it can dynamically adjust the detection sensitivity based on the actual correction effect, significantly improving the system's adaptability to environmental changes. Temperature gradient calculation formula. Based on numerical differential theory, and through approximate partial derivative calculation using the central difference scheme, this formula can accurately capture the spatial variation characteristics of the temperature field, providing important environmental factor inputs for drift prediction models. Compared to single-point temperature measurements, gradient information contains richer spatial variation information, significantly improving the accuracy of drift prediction. (Comprehensive environmental assessment function) Based on the principle of multi-factor weighted fusion, a comprehensive assessment of multiple environmental parameters is achieved through linear combination, including standardized processing. This ensures consistent processing of parameters with different dimensions. The weight allocation of the function considers the importance of each environmental factor to data fusion. Compared to judging by a single environmental indicator, the comprehensive evaluation function can more comprehensively reflect the complexity of the marine environment, providing a scientific basis for switching fusion modes. Quality score calculation formula. Based on multidimensional quality assessment theory, a comprehensive quantification of data quality is achieved through weighted averaging, where the Pearson correlation coefficient is used. The linear correlation between sensor data was accurately quantified. Type activation function It provides a smooth nonlinear mapping mechanism. The formula comprehensively evaluates the quality of sensor data from four dimensions: continuity, stability, consistency and reliability. Compared with the traditional single quality index, the multi-dimensional evaluation can more accurately identify data quality problems, provide a scientific quantitative basis for the allocation of fusion weights, and significantly improve the reliability of the final fused data.

[0108] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: A certain marine observation station covers an area of ​​approximately 2000... In the relevant sea area, the technical team deployed 125 sensor nodes according to a three-dimensional grid structure, including 25 temperature sensors, 25 humidity sensors, 25 wind speed sensors, 25 salinity sensors, and 25 pressure sensors. The sensor nodes are distributed in a 5×5×5 three-dimensional grid, with a spacing of 10 meters between adjacent sensors. Each sensor node is equipped with a data acquisition module and a wireless transmission module. The data acquisition module uses a sampling frequency of 1Hz for real-time data acquisition, and the wireless transmission module operates at a frequency of 915MHz, ensuring the stability and real-time performance of data transmission.

[0109] In the early stages of system operation, the technical team discovered several types of anomalies in the sensor data. Single-point transient anomalies manifested as individual sensors showing significant deviations from their normal range within a short period, typically lasting less than 10 minutes. Regional continuous anomalies, on the other hand, involved multiple sensors within a specific area simultaneously exhibiting persistent data deviations, usually lasting more than 30 minutes. To address these issues, the team constructed a hierarchical interpolation correction model. For single-point transient anomalies, the system employed a weighted moving average algorithm for correction, setting the decay constant of the time decay exponential function to 0.2, and using four valid data points before and after the time decay to perform the correction calculation. For regional continuous anomalies, the system used a spatial collaborative interpolation algorithm, setting the weighting coefficient to 0.85 when the distance between sensor nodes was less than 3 meters and to 0.08 when the distance was greater than 20 meters.

[0110] The technical team established a closed-loop feedback system to adaptively adjust the anomaly detection threshold. The system performs a data comparison analysis hourly, comparing the corrected sensor data with field calibration data obtained through periodic manual calibration. When the calculated root mean square error exceeds a preset threshold of 0.15, the system automatically triggers a threshold update mechanism. During actual operation, the team found that humidity sensors are more susceptible to the effects of the marine environment during the peak salt spray seasons of spring and summer. Therefore, they expanded the anomaly detection threshold range of the humidity sensor by 12%, effectively improving the system's adaptability to environmental changes.

[0111] As shown in Table 1, the technical team recorded the anomaly detection status of different types of sensors during the 3-month operation period:

[0112] Table 1. Sensor Anomaly Detection Statistics

[0113]

[0114] As shown in Table 1, the humidity sensor had the most abnormal detections, mainly due to the significant impact of salt spray and water vapor in the marine environment. The wind speed sensor had the highest correction success rate, reaching 0.971, indicating that the wind speed sensor has good stability in the marine environment.

[0115] The technical team established a multi-factor coupled drift prediction model to predict sensor drift behavior. The system calculates the temperature and salinity gradients by comparing measurements from adjacent sensors, serving as crucial input parameters for the drift prediction model. A Kalman filter handles state prediction and error correction, while a long short-term memory network learns complex drift patterns. In practical applications, the team found that temperature sensors are prone to drift when seawater temperature changes rapidly; the prediction model can anticipate this drift trend up to two hours in advance and automatically correct for it. Salinity sensors also exhibit similar drift during periods of significant tidal change, and the prediction model effectively predicts and corrects for this drift as well.

[0116] To adapt to the complex and ever-changing marine environment, the technical team designed a progressive degradation fusion strategy. The system establishes a four-level fusion mode switching mechanism based on real-time environmental conditions. The comprehensive environmental assessment function considers three key environmental parameters: wind speed, wave height, and visibility, with weighting coefficients set to 0.4, 0.4, and 0.2, respectively. Under normal sea conditions, the comprehensive environmental assessment function value is typically below 25, and the system employs a level-one fusion mode, with all 125 sensors participating in data fusion calculations. When sea conditions deteriorate and the comprehensive environmental assessment function value reaches 45, the system automatically switches to a level-two fusion mode, selecting only the 80 sensors with the highest quality scores for fusion. Under typhoon conditions, the comprehensive environmental assessment function value may exceed 80, and the system switches to a level-four fusion mode, using only 20 core sensors for basic data fusion.

[0117] like Figure 4 As shown, the technical team recorded the automatic switching process of the system fusion mode during a typhoon. The monitoring data demonstrates that the system can adjust its fusion strategy promptly according to changes in environmental conditions, ensuring reliable observational data even under harsh conditions. Figure 4 It clearly shows the trend of environmental assessment function values ​​over time and the corresponding switching of fusion modes.

[0118] Marine biofouling is a significant factor affecting the long-term stable operation of sensors. The technical team constructed a biofouling gradual detection and correction model, employing a liquid neural network structure to monitor the slow changes in the sensor's response characteristics. The reservoir computing unit consists of 100 liquid neurons, effectively capturing the dynamic characteristics of the biofouling process. The system periodically captures images of the sensor surface using underwater cameras and employs image recognition technology to quantify the degree of biofouling. The team discovered that during the high temperatures of summer, marine biological activity is frequent, and the biofouling density on the sensor surface increases rapidly, averaging 15% per week. The biofouling gradual detection and correction model can monitor this gradual process in real time and automatically calculate corresponding correction coefficients to compensate for the sensor data.

[0119] like Figure 5 As shown in the figure, the technical team demonstrated a typical sensor-based gradual biofouling detection process. The figure reveals that as the biofouling density gradually increases, the sensor's response characteristics change significantly, and the correction model accurately identifies this change and provides corresponding correction coefficients. This gradual detection capability significantly improves the sensor's long-term stability and data quality.

[0120] The technical team established a comprehensive multi-source data quality assessment system to score the data quality of each sensor node in real time. The quality score comprehensively considers four key indicators: data continuity, numerical stability, consistency with neighboring sensors, and historical reliability. During three months of operation, the system recorded the distribution of quality scores for each sensor node. Sensor nodes with quality scores above 90 accounted for 68% of the total, those with scores between 80 and 90 accounted for 23%, those with scores between 60 and 80 accounted for 7%, and those with scores below 60 accounted for 2%. Sensor data with quality scores below 60 were not included in the fusion calculation, ensuring the reliability of the fusion results.

[0121] As shown in Table 2, the technical team compiled statistics on the distribution of sensor fault types across different quality rating ranges:

[0122] Table 2 Distribution of Sensor Fault Types

[0123]

[0124] As shown in Table 2, sensors with lower quality scores mainly exhibit data loss faults, while sensors with higher quality scores primarily show a slight decrease in measurement accuracy. This quality assessment mechanism provides an important basis for system optimization and maintenance.

[0125] After three months of actual operation, the technical team conducted a comprehensive evaluation of the system's overall performance. The system effectively handles various types of data anomalies, including single-point instantaneous anomalies and regional continuous anomalies. The multi-factor coupled drift prediction model significantly improves the accuracy of sensor drift prediction, enabling the system to proactively correct data. The progressive degradation fusion strategy ensures stable operation of the system under various environmental conditions. The biological contamination gradual detection and correction model effectively solves the problem of biological contamination during long-term operation. The multi-source data quality assessment system provides a scientific basis for system optimization.

[0126] like Figure 6 As shown in the figure, the technical team compared the changes in data quality before and after adopting the technical solution of this invention. It is clearly evident from the figure that the system's data accuracy, stability, and reliability were significantly improved after adopting the multi-source data fusion and correction method. Especially under harsh environmental conditions, traditional methods often fail to provide reliable observation data, while the technical solution of this invention ensures the continuous availability of data through intelligent fusion strategies and correction mechanisms.

[0127] This invention represents a significant technological advancement over traditional ocean observation methods. Traditional methods rely primarily on single sensors for measurement; when a sensor malfunctions or malfunctions, the entire monitoring system fails. This invention, however, utilizes multi-source data fusion technology, allowing the system to compensate and correct for sensor malfunctions using data from other sensors, greatly improving its fault tolerance. Traditional methods primarily handle sensor drift through periodic manual calibration, which is not only costly but also unable to detect and address drift issues promptly. This invention, through a multi-factor coupled drift prediction model, achieves proactive prediction and real-time correction of sensor drift, significantly improving long-term data stability. Traditional methods lack adaptability to complex and changing marine environments, often employing fixed data processing strategies. This invention, through a progressive degradation fusion strategy, automatically adjusts the fusion mode based on real-time changes in environmental conditions, ensuring reliable observation data under various conditions. Traditional methods primarily address biocontamination through periodic cleaning, failing to monitor and correct the impact of biocontamination on sensor performance in real time. This invention achieves real-time monitoring and dynamic correction of biological pollution processes through a biological pollution gradual change detection and correction model, which greatly reduces maintenance costs and improves data quality.

[0128] It should be noted that the variables involved in this invention are explained in detail in Table 3.

[0129] Table 3. Variable Explanation Table

[0130]

[0131] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for multi-source data fusion and correction from a marine observation station, characterized in that, include: A multi-type sensor array, including temperature, humidity, wind speed, salinity, and pressure sensors, is deployed at the marine observation station. The sensor array is distributed in a three-dimensional grid structure, with each sensor node equipped with a data acquisition module and a wireless transmission module. A hierarchical interpolation correction model is established. For single-point instantaneous anomalies, a weighted moving average algorithm is used, with the weight function being a time-decay exponential function. For regional continuous anomalies, a spatial collaborative interpolation algorithm is used, with the weights inversely proportional to the square of the distance between sensor nodes. An adaptive threshold adjustment mechanism for the marine environment is constructed. A closed-loop feedback system compares and analyzes the corrected data with the on-site calibration data, adjusting the threshold according to changes in environmental conditions. The abnormal detection threshold ranges of various sensors are dynamically adjusted; a multi-factor coupled drift prediction model is established, employing a Kalman filter combined with a long short-term memory network structure. Input parameters include temperature gradient, salinity concentration gradient, and time series data, with the output being the drift prediction value, which is used to correct sensor measurement data; a progressive degradation fusion strategy is designed, establishing a four-level fusion mode switching mechanism based on weather conditions and sensor operating status, calculating the fusion mode level through a comprehensive environmental assessment function; a gradual biological pollution detection correction model is constructed, using a liquid neural network structure to monitor the slow changes in sensor response characteristics. Model inputs include historical sensor response curves and biofouling density. The estimated values ​​and environmental parameter sequences are output as pollution level assessment values ​​and correction coefficients, which are used to correct sensor measurement data. A multi-source data quality assessment system is established to score the data quality of each sensor node in real time. The scoring parameters include data continuity, numerical stability, consistency with neighboring sensors, and historical reliability. Fusion weights are assigned and correction model parameters are updated based on the quality scores. Specifically, the comprehensive environmental assessment function is established based on wind speed measurements, wave height measurements, and visibility measurements. Wave height measurements are obtained through pressure sensor data conversion, and visibility measurements are obtained through optical sensors. The comprehensive environmental assessment function is based on wind speed measurements... The weighted linear combination of the measured values ​​of wave height and visibility is used, with weighting coefficients of 0.4, 0.4, and 0.2, respectively. Specifically, the four-level fusion mode switching mechanism operates as follows: the comprehensive environmental assessment function ranges from 0 to 100. When the comprehensive environmental assessment function value is less than 25, it corresponds to the first-level fusion mode, employing full-sensor fusion; when the comprehensive environmental assessment function value is between 25 and 50, it corresponds to the second-level fusion mode, employing key sensor fusion; when the comprehensive environmental assessment function value is between 50 and 75, it corresponds to the third-level fusion mode, employing core sensor fusion; and when the comprehensive environmental assessment function value is greater than 75, it corresponds to the fourth-level fusion mode, employing basic sensor fusion.

2. The method for multi-source data fusion and correction of marine observation stations according to claim 1, characterized in that, The sensor array is distributed according to a three-dimensional grid structure, specifically with an adjacent sensor spacing of 5m to 15m, a data acquisition module configured for each sensor node with a sampling frequency of 1Hz, and a wireless transmission module with a transmission frequency of 915MHz.

3. The multi-source data fusion and correction method for marine observation stations according to claim 2, characterized in that, The weighted moving average algorithm specifically uses a time decay exponential function as the weighting function. The decay constant of the time decay exponential function is determined by the historical fluctuation period. Four valid data points are taken before and after the time decay exponential function for correction. The time decay exponential function is in the form of negative exponential decay, and the decay constant is determined according to the historical fluctuation period. The historical fluctuation period is obtained by spectral analysis of sensor data over the past 12 months.

4. The multi-source data fusion and correction method for marine observation stations according to claim 3, characterized in that, The spatial collaborative interpolation algorithm specifically adopts an inverse distance weighting method. In the weight calculation formula, the distance index is Euclidean distance. When the distance between sensor nodes is less than 3m, the weight coefficient is greater than 0.8, and when the distance is greater than 20m, the weight coefficient is less than 0.

1.

5. The multi-source data fusion and correction method for marine observation stations according to claim 4, characterized in that, The closed-loop feedback system specifically includes a data comparison module, a deviation analysis module, and a threshold update module. The data comparison module performs a comparison analysis every hour. The deviation analysis module calculates the root mean square error between the corrected data and the on-site calibration data. When the root mean square error exceeds a preset threshold, the threshold update module is triggered to adjust the abnormal detection threshold range. The on-site calibration data is obtained through manual periodic calibration.

6. The multi-source data fusion and correction method for marine observation stations according to claim 5, characterized in that, Specifically, the adaptive threshold adjustment mechanism for the marine environment expands the threshold range of the humidity sensor by 10% to 15% during periods of high salt spray incidence.

7. The method for multi-source data fusion and correction of marine observation stations according to claim 6, characterized in that, The temperature gradient is calculated from the measurements of adjacent temperature sensors, the salinity gradient is calculated from the measurements of adjacent salinity sensors, and the time series data is the historical measurement sequence of each sensor.

8. The method for multi-source data fusion and correction of marine observation stations according to claim 7, characterized in that, The multi-factor coupled drift prediction model specifically uses a Kalman filter for state prediction and error correction, and a long short-term memory network for learning drift patterns. The long short-term memory network includes an input layer, two hidden layers, and an output layer. The number of neurons in the hidden layers are 64 and 32, respectively, and the activation function is the hyperbolic tangent function.

Images