Marine meteorological data quality control method based on cross-parameter correlation network

Abstract

The invention provides a marine meteorological data quality control method based on a cross-parameter correlation network, which belongs to the technical field of marine meteorology, and comprises the following steps: constructing a cross-parameter correlation network model comprising an air temperature and air pressure correlation rule, a humidity and air temperature linear correlation rule and a wind speed and air pressure gradient extraction correlation rule; a sliding window algorithm is adopted to calculate a correlation coefficient in real time and identify correlation abnormity, a multivariate abnormity detection mechanism of four dimensions of parameter threshold overrun, spatio-temporal change rate abnormity, probability density distribution offset and correlation verification failure is established, a fuzzy comprehensive evaluation method is adopted to calculate a comprehensive abnormity index, and data abnormity is judged. And the authenticity of the abnormal data is confirmed in combination with a multi-sensor cross validation mechanism, and finally the abnormal data is restored by adopting a virtual sensor data reconstruction algorithm based on correlation network reverse calculation, so that the technical problem of poor abnormal data detection effect in the prior art is solved.

Description

Technical Field

[0001] The present invention belongs to the field of marine meteorological technology, and in particular relates to a marine meteorological data quality control method based on a cross-parameter correlation network. Background Art

[0002] In the field of marine meteorological observations, traditional data quality control techniques primarily employ methods such as single-parameter threshold testing, temporal consistency testing, and spatial consistency testing to control the quality of marine meteorological observation data. These methods are widely used in marine environmental monitoring systems such as marine meteorological stations, buoy observation networks, and ship-based automatic weather stations. They identify anomalous data by setting fixed parameter threshold ranges and rate-of-change limits. However, traditional quality control methods exhibit significant drawbacks when dealing with the complex and volatile meteorological parameters of the marine environment. These drawbacks primarily include ignoring the inherent correlations between marine meteorological parameters, failing to effectively identify anomalous data that meets single-parameter threshold requirements but violates inter-parameter correlation patterns, and lacking the ability to dynamically adjust parameter correlations across different sea areas and seasonal conditions. In current marine meteorological data quality control practices, traditional methods rely solely on single-dimensional anomaly determination criteria, making it difficult to accurately identify the complex, nonlinear relationships between multiple meteorological parameters in the marine environment, such as temperature, pressure, humidity, and wind speed. This leads to frequent missed and false detections of anomalous data. In other words, existing techniques for marine meteorological data quality control suffer from insufficient accuracy in identifying multi-parameter correlations, resulting in poor anomalous data detection. Summary of the Invention

[0003] In view of this, the present invention provides a marine meteorological data quality control method based on a cross-parameter correlation network, which can solve the technical problem in the existing technology that the accuracy of multi-parameter correlation relationship recognition in the marine meteorological data quality control process is insufficient, resulting in poor abnormal data detection effect.

[0004] The present invention is implemented as follows: the present invention provides a marine meteorological data quality control method based on a cross-parameter correlation network, including constructing a marine meteorological parameter correlation network model, obtaining temperature and pressure correlation rules, humidity and temperature linear correlation rules and wind speed and pressure gradient square root correlation rules, and forming a cross-parameter correlation network structure; implementing sliding window correlation coefficient calculation, and calculating the correlation coefficient of each parameter in the cross-parameter correlation network structure in real time, and triggering a correlation anomaly flag when the real-time correlation coefficient deviates from the standard correlation coefficient by more than a preset threshold; establishing a multi-dimensional abnormal signal detection mechanism, and detecting parameter threshold exceeding signals, spatiotemporal change rate abnormal signals, probability density distribution offset signals and relative The failure signal of the correlation check is identified and quantified to generate the corresponding intensity value; the adaptive weight optimization algorithm is executed to calculate the dynamic weight value through the adaptive weight optimization model of ocean parameters; the fuzzy comprehensive evaluation calculation is applied to perform weighted fusion operation on the intensity value and the dynamic weight value to calculate the comprehensive anomaly index. When the comprehensive anomaly index exceeds the preset threshold, it is determined to be a data anomaly; multi-sensor cross-validation processing is implemented to determine the authenticity of the data anomaly and calculate the sensor reliability score; abnormal data repair and reconstruction processing is performed, and the virtual sensor data reconstruction algorithm is adopted to solve the estimated value of the abnormal parameter through the correlation network inverse calculation equation, and the repaired ocean meteorological parameter data is output.

[0005] Among them, the steps of constructing the marine meteorological parameter correlation network model are to use the iterative regression algorithm to analyze and process the historical marine meteorological data to obtain the temperature and pressure correlation rules, and at the same time establish the humidity and temperature linear correlation rules and the wind speed and pressure gradient square root correlation rules to form a cross-parameter correlation network structure containing 6 groups of core parameters. The parameter nodes in the cross-parameter correlation network structure are connected by the correlation coefficient.

[0006] Among them, the cross-parameter correlation network structure is specifically a network topology model based on the principles of marine meteorology, which is a correlation between multiple meteorological parameters. It describes the coordinated change law of various meteorological elements in the marine environment by quantifying the mathematical relationship between different parameters.

[0007] Among them, the temperature and pressure correlation rule is specifically a piecewise negative correlation mathematical relationship between temperature and pressure obtained by iterative calculation using least squares regression analysis and polynomial fitting algorithm on a large amount of historical marine meteorological observation data. The final correlation coefficient and applicable temperature range are determined through more than 1,000 iterative optimizations.

[0008] The linear correlation rule between humidity and temperature is specifically a linear correlation relationship obtained by statistically analyzing the observation data of relative humidity and temperature in the marine environment through a linear regression algorithm, and the correlation coefficient is determined by the Pearson correlation coefficient calculation method.

[0009] The wind speed and pressure gradient square root correlation rule is specifically a square root function relationship between wind speed and pressure gradient established based on the principles of fluid mechanics, and the function parameters are determined by nonlinear fitting analysis of measured wind speed data and pressure gradient data of the same period.

[0010] Among them, the step of implementing the sliding window correlation coefficient calculation is to use a sliding window algorithm with a time window length of 30 minutes to calculate the correlation coefficient of each parameter in the cross-parameter correlation network structure in real time. When the calculated real-time correlation coefficient deviates from the corresponding standard correlation coefficient in the temperature and pressure correlation rule, the humidity and temperature linear correlation rule, and the wind speed and pressure gradient square root correlation rule by more than 25%, the correlation anomaly flag is triggered.

[0011] Among them, the steps of establishing a multi-dimensional abnormal signal detection mechanism are to identify and quantify the four dimensions of parameter threshold exceeding signal, spatiotemporal change rate abnormal signal, probability density distribution offset signal and correlation check failure signal respectively, and generate corresponding parameter threshold exceeding intensity value, spatiotemporal change rate abnormal intensity value, probability density distribution offset intensity value and correlation check failure intensity value.

[0012] Among them, the steps of executing the adaptive weight optimization algorithm are to calculate the dynamic weight values ​​of the parameter threshold exceeding intensity value, the spatiotemporal change rate abnormal intensity value, the probability density distribution offset intensity value and the correlation verification failure intensity value respectively through the ocean parameter adaptive weight optimization model. The ocean parameter adaptive weight optimization model adjusts the hierarchical fusion weight parameters of the model based on the current sea area water depth data, geographic latitude data and seasonal temperature difference data.

[0013] Among them, the steps of applying fuzzy comprehensive evaluation calculation and processing are to perform weighted fusion operation on the parameter threshold exceeding intensity value, the spatiotemporal change rate abnormal intensity value, the probability density distribution offset intensity value and the correlation verification failure intensity value with the dynamic weight value, and calculate the comprehensive anomaly index through the fuzzy comprehensive evaluation operation rules. When the comprehensive anomaly index exceeds the preset threshold of 0.7, it is judged as data anomaly.

[0014] Among them, the steps of implementing multi-sensor cross-validation processing are to start the multi-sensor cross-validation mechanism for the identified data anomalies, determine the authenticity of the data anomalies through comparative analysis of redundant sensor observation data, and at the same time call the sensor health assessment function to process the sensor output data variance, data drift trend coefficient, cross-validation consistency index, sensor working time and environmental corrosion degree assessment value to calculate the sensor reliability score.

[0015] Among them, the steps of performing abnormal data repair and reconstruction processing are to use the virtual sensor data reconstruction algorithm for the confirmed data anomaly, input the normal parameter values ​​and temperature and pressure correlation rules, humidity and temperature linear correlation rules, wind speed and pressure gradient square root correlation rules in the cross-parameter correlation network structure, solve the estimated values ​​of the abnormal parameters through the reverse calculation equation of the correlation network, and output the repaired marine meteorological parameter data to complete the quality control processing flow.

[0016] Among them, the ocean parameter adaptive weight optimization model is specifically a multi-layer neural network model based on a hierarchical attention network architecture, which includes four main components: input layer, feature extraction layer, attention mechanism layer and output layer. The attention mechanism layer adopts a multi-level attention structure to allocate weights to different sea area characteristics and seasonal characteristics.

[0017] The sensor health evaluation function is used to calculate the working status reliability score of the marine meteorological sensor. The input includes the sensor output data variance, data drift trend coefficient, cross-validation consistency index, sensor working time and environmental corrosion degree assessment value. The output is a sensor reliability score between 0 and 1. When the sensor reliability score is lower than 0.6, it means that the sensor needs maintenance and calibration.

[0018] Among them, the virtual sensor data reconstruction algorithm is specifically a data reconstruction algorithm that infers missing or abnormal parameter values ​​through mathematical operations based on the correlation relationship between parameters. The normal parameter values ​​are specifically credible marine meteorological parameter observation values ​​verified by quality control in the cross-parameter correlation network structure.

[0019] Among them, the correlation network reverse calculation equation is specifically a set of mathematical equations for solving the estimated values ​​of abnormal parameters based on known normal parameter values ​​and correlation rules. The estimated values ​​of the abnormal parameters are specifically the inferred parameter values ​​obtained by solving the correlation network reverse calculation equation to replace the abnormal data.

[0020] The present invention solves the technical problem of poor abnormal data detection effect caused by insufficient accuracy of multi-parameter correlation relationship identification in the process of marine meteorological data quality control by constructing a cross-parameter correlation network model and combining a multi-dimensional anomaly detection mechanism and an adaptive weight optimization algorithm. The present invention adopts the temperature and pressure correlation rules, humidity and temperature linear correlation rules and wind speed and pressure gradient square root correlation rules based on historical data analysis to establish a cross-parameter correlation network, monitors the changes in the correlation coefficients between parameters in real time through a sliding window algorithm, and combines the comprehensive analysis of four dimensions of parameter threshold exceeding the limit, spatiotemporal change rate anomaly, probability density distribution offset and correlation verification failure to effectively improve the accuracy and reliability of abnormal data identification. The present invention dynamically adjusts the detection weights of each dimension according to environmental characteristics such as sea depth, geographical latitude and seasonal temperature difference through an ocean parameter adaptive weight optimization model, and calculates the comprehensive anomaly index using a fuzzy comprehensive evaluation method, thereby overcoming the technical defects of the traditional method in inaccurate parameter correlation relationship identification in a complex marine environment. In summary, the present invention solves the technical problem of poor abnormal data detection effect caused by insufficient accuracy of multi-parameter correlation relationship identification in the process of marine meteorological data quality control mentioned in the background technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a flow chart of the method of the present invention.

[0022] Figure 2 This is a time series change diagram of marine meteorological parameters in Example 2.

[0023] Figure 3 Schematic diagram of the cross-parameter correlation network structure in Example 2.

[0024] Figure 4 This is a statistical chart of the multi-dimensional anomaly detection results in Example 2. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0026] like Figure 1 FIG. 1 is a flow chart of a method for quality control of marine meteorological data based on a cross-parameter correlation network provided by the present invention. The method comprises the following steps: S01. Construct a marine meteorological parameter correlation network model, use an iterative regression algorithm to analyze and process historical marine meteorological data to obtain temperature and pressure correlation rules, and simultaneously establish humidity and temperature linear correlation rules and wind speed and pressure gradient square root correlation rules to form a cross-parameter correlation network structure containing 6 groups of core parameters, in which each parameter node is connected by a correlation coefficient; S02. Calculate the sliding window correlation coefficient. Use a sliding window algorithm with a time window length of 30 minutes to calculate the correlation coefficient of each parameter in the cross-parameter correlation network structure in real time. When the calculated real-time correlation coefficient deviates from the corresponding standard correlation coefficient in the temperature and pressure correlation rule, the humidity and temperature linear correlation rule, and the wind speed and pressure gradient square root correlation rule by more than 25%, a correlation anomaly flag is triggered. S03. Establish a multi-dimensional abnormal signal detection mechanism to identify and quantify the four dimensions of parameter threshold exceeding signal, spatiotemporal change rate abnormal signal, probability density distribution offset signal, and correlation check failure signal, and generate corresponding parameter threshold exceeding intensity value, spatiotemporal change rate abnormal intensity value, probability density distribution offset intensity value, and correlation check failure intensity value; S04. Execute an adaptive weight optimization algorithm to calculate dynamic weight values ​​for the parameter threshold over-limit intensity value, the spatiotemporal change rate anomaly intensity value, the probability density distribution offset intensity value, and the correlation check failure intensity value, respectively, using an ocean parameter adaptive weight optimization model. The ocean parameter adaptive weight optimization model adjusts the hierarchical fusion weight parameters of the model based on current sea area water depth data, geographic latitude data, and seasonal temperature difference data. S05. Apply fuzzy comprehensive evaluation calculation processing to perform weighted fusion operation on the parameter threshold over-limit intensity value, the spatiotemporal change rate abnormal intensity value, the probability density distribution offset intensity value, and the correlation check failure intensity value with the dynamic weight value, and calculate the comprehensive anomaly index using the fuzzy comprehensive evaluation operation rule. When the comprehensive anomaly index exceeds the preset threshold of 0.7, it is determined that the data is abnormal; S06. Implement multi-sensor cross-validation processing. Activate the multi-sensor cross-validation mechanism for the identified data anomaly. Determine the authenticity of the data anomaly through comparative analysis of redundant sensor observation data. Simultaneously, call the sensor health assessment function to process the sensor output data variance, data drift trend coefficient, cross-validation consistency index, sensor operating time, and environmental corrosion degree assessment value to calculate the sensor reliability score. S07. Execute abnormal data repair and reconstruction processing, use virtual sensor data reconstruction algorithm for confirmed data anomalies, input normal parameter values ​​in the cross-parameter correlation network structure and the temperature and pressure correlation rules, humidity and temperature linear correlation rules, and wind speed and pressure gradient square root correlation rules, solve the estimated values ​​of abnormal parameters through the correlation network reverse calculation equation, and output the repaired marine meteorological parameter data to complete the quality control processing flow.

[0027] Among them, the cross-parameter correlation network structure refers to a network topology model based on the principles of marine meteorology that is based on the correlation between multiple meteorological parameters. It describes the coordinated change patterns of various meteorological elements in the marine environment by quantifying the mathematical relationships between different parameters. The temperature and pressure correlation rule refers to the piecewise negative correlation mathematical relationship between temperature and pressure obtained by iteratively calculating a large amount of historical marine meteorological observation data using least squares regression analysis and polynomial fitting algorithms. When the temperature is higher than 25°C, the correlation coefficient is negative and the absolute value fluctuates between 0.5 and 0.7. The final correlation coefficient and applicable temperature range are determined through more than 1,000 iterative optimizations. The humidity and temperature linear correlation rule refers to the linear correlation relationship obtained by statistically analyzing the observation data of relative humidity and temperature in the marine environment through a linear regression algorithm. The correlation coefficient is determined by the Pearson correlation coefficient calculation method.

[0028] The wind speed and pressure gradient square root correlation rule refers to a square root function relationship between wind speed and pressure gradient established based on fluid mechanics principles. The function parameters are determined through nonlinear fitting analysis of measured wind speed data and contemporaneous pressure gradient data. The sliding window algorithm refers to a data processing method that continuously moves a fixed-length data window across a time series and performs statistical analysis and calculations on the data within the window. The real-time correlation coefficient refers to the correlation value between parameters calculated within the current time window using the sliding window algorithm. The standard correlation coefficient refers to the correlation benchmark value predefined in the temperature and pressure correlation rule, the humidity and temperature linear correlation rule, and the wind speed and pressure gradient square root correlation rule. The correlation anomaly flag is an anomaly alarm signal generated when the real-time correlation coefficient deviates from the standard correlation coefficient by more than a set threshold. The parameter threshold violation signal is an anomaly warning signal generated when the value of a marine meteorological parameter exceeds the normal range of variation. The spatiotemporal change rate anomaly signal is an anomaly warning signal generated when the rate of change of a marine meteorological parameter in the temporal or spatial dimensions exceeds a reasonable range. The probability density distribution deviation signal is an anomaly warning signal generated when the statistical distribution characteristics of a marine meteorological parameter deviate from the historical normal distribution pattern.

[0029] The correlation check failure signal refers to an anomaly warning signal generated when the correlation between parameters does not conform to the preset rules. The parameter threshold violation intensity value refers to an anomaly intensity assessment value obtained by quantitatively analyzing the degree of parameter threshold violation. The spatiotemporal change rate anomaly intensity value refers to an anomaly intensity assessment value obtained by quantitatively analyzing the degree of spatiotemporal change rate deviation. The probability density distribution offset intensity value refers to an anomaly intensity assessment value obtained by quantitatively analyzing the degree of probability density distribution offset. The correlation check failure intensity value refers to an anomaly intensity assessment value obtained by quantitatively analyzing the degree of correlation check failure. The ocean parameter adaptive weight optimization model is a multi-layer neural network model based on a hierarchical attention network architecture, which is used to dynamically calculate the weight distribution of each dimension of anomaly detection based on ocean area characteristics and seasonal characteristics. The current ocean water depth data refers to the seabed depth measurement value of the quality control processing location. The geographic latitude data refers to the geographic latitude coordinate value of the quality control processing location. The seasonal temperature difference data refers to the historical average temperature change amplitude corresponding to the season of the quality control processing time. The hierarchical fusion weight parameter refers to the weight distribution coefficient when fusing features of different abstraction levels in the ocean parameter adaptive weight optimization model. The dynamic weight value refers to the weight coefficient output by the ocean parameter adaptive weight optimization model used for weighted fusion of multi-dimensional anomaly signals. The fuzzy comprehensive evaluation operation rule uses fuzzy mathematics theory to establish a fuzzy relationship matrix and calculates the comprehensive evaluation results through matrix operations and weight vectors. The comprehensive anomaly index is a numerical indicator calculated using the fuzzy comprehensive evaluation operation rule that represents the overall anomaly level of marine meteorological data. Data anomaly refers to the state determination result of marine meteorological observation data deviating from the normal variation pattern.

[0030] The multi-sensor cross-validation mechanism refers to a verification method that uses observation data from multiple independent sensors to cross-compare and verify the authenticity of abnormal data. Redundant sensor observation data refers to marine meteorological observation data collected by multiple sensors deployed at the same or adjacent locations for cross-validation. The sensor health assessment function refers to a mathematical function model used to assess sensor operating status and data reliability. The sensor output data variance refers to the statistical variance of sensor output data under stable environmental conditions. The data drift trend coefficient refers to the degree of systematic drift in sensor output data over time. The cross-validation consistency index is a quantitative assessment of the consistency between multiple sensor observations. The sensor operating time refers to the cumulative operating time of the sensor from deployment to the current moment. The environmental corrosion assessment value is a quantitative assessment of the degree of corrosion caused by the marine environment on sensor equipment. The sensor reliability score is the score output by the sensor health assessment function that represents the reliability of the sensor operating status. The virtual sensor data reconstruction algorithm is a data reconstruction algorithm that uses mathematical operations to infer missing or abnormal parameter values ​​based on the correlation between parameters. The normal parameter value refers to the credible marine meteorological parameter observation value verified by quality control in the cross-parameter correlation network structure. The correlation network inverse calculation equations refer to a set of mathematical equations established based on known normal parameter values ​​and correlation rules, used to solve the estimated values ​​of abnormal parameters. The estimated values ​​of abnormal parameters are the inferred parameter values ​​obtained by solving the correlation network inverse calculation equations and used to replace the abnormal data. The repaired marine meteorological parameter data refers to the complete marine meteorological data set output after processing by the virtual sensor data reconstruction algorithm. The quality control process refers to the complete marine meteorological data quality control process, from anomaly detection to data repair.

[0031] The sensor health assessment function is used to calculate the working status reliability score of the marine meteorological sensor. The input includes the sensor output data variance, data drift trend coefficient, cross-validation consistency index, sensor working time and environmental corrosion degree assessment value. The output is the sensor reliability score between 0 and 1. When the sensor reliability score is lower than 0.6, it indicates that the sensor needs maintenance and calibration.

[0032] The specific structure of the ocean parameter adaptive weight optimization model is a multi-layer neural network model based on a hierarchical attention network architecture, which includes four main components: an input layer, a feature extraction layer, an attention mechanism layer, and an output layer. The attention mechanism layer adopts a multi-level attention structure to assign weights to different sea area characteristics and seasonal characteristics. The hierarchical fusion weight parameters of the model are dynamically adjusted according to the three key parameters of the current sea area water depth data, geographic latitude data, and seasonal temperature difference data. The current sea area water depth data affects the weight distribution ratio of the stability of the ocean environment, the geographic latitude data affects the weight distribution ratio of climate characteristics, and the seasonal temperature difference data affects the weight distribution ratio of time changes. The training data set establishment step of the ocean parameter adaptive weight optimization model specifically includes collecting marine meteorological observation data from different sea areas and seasons around the world as the basic data source, performing quality screening and outlier removal on the collected raw data to form a clean data set, classifying and labeling the data according to the geographical location of the sea area, water depth information and seasonal time to establish a multi-dimensional labeling system, using the abnormal data identification results labeled by marine meteorological experts as the target label of supervised learning, dividing the data set into training set, validation set and test set according to the ratio of 8 to 1 to 1, and standardizing all input features to ensure the stability of model training. The training step of the ocean parameter adaptive weight optimization model specifically includes iteratively updating the model parameters using the stochastic gradient descent optimization algorithm, setting the learning rate to 0.001 and dynamically adjusting the learning rate value using the cosine annealing scheduling strategy, using the cross entropy loss function to measure the difference between the model prediction weight and the actual weight, setting the batch size to 64 and performing a complete training process of 200 training cycles, using the validation set to evaluate the model performance and save the optimal model parameters after each training cycle, using the early stopping mechanism to prevent the occurrence of model overfitting during training, and finally verifying the generalization ability and practical application effect of the model through the test set.

[0033] The hierarchical weight adjustment function is used to adjust the hierarchical fusion weight parameters of the ocean parameter adaptive weight optimization model. The hierarchical weight adjustment function is calculated based on the four data of the current sea area water depth data, geographic latitude data, seasonal temperature difference data and data quality assessment score to obtain a comprehensive adjustment factor value. When the comprehensive adjustment factor value is in the range of 0 to 0.25, the deep feature-dominated weight distribution strategy is adopted to adjust the hierarchical fusion weight parameters of the model. When the comprehensive adjustment factor value is in the range of 0.25 to 0.5, the middle feature balanced weight distribution strategy is adopted to adjust the hierarchical fusion weight parameters of the model. When the comprehensive adjustment factor value is in the range of 0.5 to 0.75, the shallow feature-priority weight distribution strategy is adopted to adjust the hierarchical fusion weight parameters of the model. When the comprehensive adjustment factor value is in the range of 0.75 to 1.0, the surface feature enhanced weight distribution strategy is adopted to adjust the hierarchical fusion weight parameters of the model.

[0034] The specific implementation of the above steps is described in detail below.

[0035] The specific implementation of step S01 is to first use an iterative regression algorithm to perform statistical analysis on historical marine meteorological data of no less than 5 years, and establish a piecewise correlation mathematical model between temperature and air pressure through least squares regression analysis. When the temperature is higher than 25°C, the correlation coefficient is set to a negative value and the absolute value is controlled within the range of 0.5 to 0.7. The final correlation coefficient and applicable temperature range are determined through more than 1000 iterative optimizations. The purpose of this step is to establish a reliable benchmark rule for the correlation between temperature and air pressure. Then, a linear regression analysis is performed on the observed data of relative humidity and temperature using the Pearson correlation coefficient calculation method to establish a linear correlation rule for humidity and temperature, and the correlation coefficient threshold is set between 0.6 and 0.8. Then, based on the principles of fluid mechanics, a nonlinear fitting algorithm is used to establish a square root function relationship between wind speed and pressure gradient. The function parameters are determined through statistical analysis of the measured wind speed data and the pressure gradient data of the same period, and the correlation coefficient threshold is set between 0.4 and 0.6. Finally, six groups of core marine meteorological parameters, including air temperature, air pressure, humidity, wind speed, sea temperature, and salinity, are constructed into a network topology structure. Each parameter node is connected through a quantified correlation coefficient to form a complete cross-parameter correlation network model.

[0036] The specific implementation method of step S02 is to use a sliding window algorithm with a time window length of 30 minutes to perform real-time correlation analysis on each parameter in the cross-parameter correlation network structure. The sliding window moves at a time interval of 5 minutes each time to ensure the continuity and real-time nature of data processing. In each time window, the Pearson correlation coefficient calculation method and the Spearman rank correlation coefficient calculation method are used to perform correlation analysis on the parameter pairs to obtain the real-time correlation coefficient value. The calculated real-time correlation coefficient is compared with the corresponding standard correlation coefficient in the temperature and pressure correlation rule, humidity and temperature linear correlation rule, and wind speed and pressure gradient square root correlation rule established in step S01. When the degree of deviation exceeds 25%, a correlation anomaly identification signal is triggered. The purpose of this step is to identify potential data anomalies by real-time monitoring of correlation changes between parameters. The deviation threshold of 25% is a reasonable limit determined based on the natural variation range of the marine environment and measurement errors.

[0037] The specific implementation of step S03 is to establish a four-dimensional abnormal signal detection mechanism to identify and quantify different types of data anomalies. First, a parameter threshold overrun detection mechanism is established. Based on historical statistical data, the reasonable variation range of each marine meteorological parameter is determined. The temperature range is set to -5°C to 45°C, the air pressure range is set to 980 hPa to 1040 hPa, and the wind speed range is set to 0 to 50 m / s. When the parameter value exceeds the set range, the degree of overrun is calculated and the parameter threshold overrun intensity value is generated. Secondly, a spatiotemporal change rate anomaly detection mechanism is established. Anomalies are identified by calculating the parameter change rate in the time dimension and the gradient change in the spatial dimension. The temporal change rate threshold is set to no more than 10% of the parameter mean value per hour, and the spatial gradient change threshold is set to no more than twice the parameter standard deviation per kilometer. Then, a probability density distribution shift detection mechanism is established. The current data distribution is compared with the historical normal distribution using the chi-square test and the Kolmogorov-Smirnov test. When the test statistic exceeds the significance level of 0.05, a distribution shift anomaly is determined. Finally, a correlation check failure detection mechanism is established. When the correlation between parameters does not meet the correlation rule established in step S01, a correlation check failure signal is generated. The purpose of this step is to comprehensively detect data anomalies from multiple angles and improve the accuracy and reliability of anomaly identification.

[0038] The specific implementation of step S04 is to calculate the dynamic weight values ​​of abnormal signals in each dimension using an ocean parameter adaptive weight optimization model. This model is constructed based on a hierarchical attention network architecture and includes four main components: an input layer, a feature extraction layer, an attention mechanism layer, and an output layer. The input layer receives three key environmental parameters: current sea depth data, geographic latitude data, and seasonal temperature difference data. The feature extraction layer uses a multi-layer perceptron network to perform nonlinear transformation and abstraction on the input features. The attention mechanism layer uses a multi-level attention structure to calculate weights for different sea and seasonal characteristics. The model adjusts the weight distribution ratio of ocean environmental stability based on the current sea depth data, with a weight of 0.7 for relatively stable deep-sea environments and a weight of 0.3 for highly variable shallow-sea environments. The model adjusts the weight distribution ratio of climate characteristics based on geographic latitude data, with a weight of 0.8 for drastic climate changes in low-latitude regions and a weight of 0.4 for relatively stable climate changes in high-latitude regions. The model adjusts the weight distribution ratio of temporal changes based on seasonal temperature difference data, with a weight of 0.9 for seasons with large temperature differences and a weight of 0.5 for seasons with small temperature differences. The purpose of this step is to dynamically adjust the weight distribution of anomaly detection according to the specific sea environment characteristics and improve the adaptability and accuracy of the quality control method.

[0039] The specific implementation method of step S05 is to use a fuzzy comprehensive evaluation calculation method to perform weighted fusion processing on multi-dimensional abnormal signals. First, a fuzzy relationship matrix is ​​established, and the parameter threshold over-limit intensity value, the spatiotemporal change rate abnormal intensity value, the probability density distribution offset intensity value and the correlation check failure intensity value are used as evaluation factors. The triangle membership function is used to convert each intensity value into a fuzzy value between 0 and 1. Then, the dynamic weight value calculated in step S04 is used to form a weight vector, and the comprehensive evaluation result is calculated through matrix operation and weight vector. The specific calculation process is to multiply the abnormal intensity value of each dimension with the corresponding dynamic weight value, and then sum all the product results to obtain a comprehensive abnormal index. When the comprehensive abnormal index exceeds the preset threshold of 0.7, it is determined to be a data abnormality. This threshold is the optimal judgment limit determined based on statistical analysis of a large number of actual application cases. The purpose of this step is to comprehensively consider the abnormal information of multiple dimensions, avoid the limitations of single-dimensional judgment, and improve the accuracy and robustness of abnormal identification.

[0040] The specific implementation of step S06 involves initiating a multi-sensor cross-validation mechanism for identified data anomalies. This mechanism compares and analyzes observation data from multiple redundant sensors deployed in the same or adjacent locations. First, a variance analysis method is used to compare the consistency of observation data from different sensors. A sensor fault is determined when the standard deviation of inter-sensor data differences exceeds twice the measurement accuracy. A sensor health assessment function is then invoked to evaluate the sensor's operating status. This function inputs five parameters: the variance of the sensor output data, the data drift trend coefficient, the cross-validation consistency index, the sensor operating time, and an assessment of the environmental corrosion level. The data variance threshold is set at 1.5 times the normal operating state, the data drift trend coefficient threshold is set at no more than 0.1% per month, and the cross-validation consistency index threshold is set at or above 0.8. Reliability begins to decline when the sensor's operating time exceeds 80% of its design lifespan. The environmental corrosion level is comprehensively assessed based on seawater salinity and temperature. A sensor reliability score is calculated using a weighted average method. A score below 0.6 indicates that the sensor requires maintenance and calibration. The purpose of this step is to verify the authenticity of abnormal data through multi-sensor data cross-validation, avoid misjudgments due to sensor failure, and improve the credibility of quality control results.

[0041] The specific implementation of step S07 is to repair the confirmed abnormal data using a virtual sensor data reconstruction algorithm. This algorithm establishes a mathematical model based on the parameter association relationship in the cross-parameter correlation network structure. First, the normal parameter values ​​related to the abnormal parameters are extracted from the cross-parameter correlation network structure. Then, according to the temperature and pressure correlation rules, humidity and temperature linear correlation rules, and wind speed and pressure gradient square root correlation rules established in step S01, a set of correlation network inverse calculation equations is constructed. The least squares method and Newton-Raphson iteration method are used to solve the equations to obtain the estimated values ​​of the abnormal parameters. To ensure the rationality of the reconstructed data, the estimated results are constrained and checked, requiring that the reconstructed values ​​are within the historical statistical range and that the association relationship with the relevant parameters conforms to physical laws. Finally, the reconstructed parameter values ​​are replaced with the original abnormal data, and the restored complete marine meteorological parameter data set is output. The purpose of this step is to infer the reasonable values ​​of the abnormal parameters through the inherent association relationship between the parameters, ensure the integrity and continuity of the data, and provide a reliable data foundation for subsequent marine meteorological analysis.

[0042] The detailed structure of the ocean parameter adaptive weight optimization model consists of four main components: the input layer, feature extraction layer, attention mechanism layer, and output layer. The input layer receives three key environmental parameters: current ocean depth data, geographic latitude data, and seasonal temperature difference data. Each parameter is first normalized and converted to a value between 0 and 1. The feature extraction layer utilizes a three-layer fully connected neural network structure. The first layer contains 64 neurons, the second layer contains 32 neurons, and the third layer contains 16 neurons. The ReLU activation function is used for nonlinear transformations between each layer. The attention mechanism layer adopts a multi-head attention mechanism structure with four attention heads, each with a dimension of 16. The self-attention mechanism calculates the association weights between different input features. The output layer uses a softmax activation function to output the weight distribution coefficients for anomaly detection in four dimensions. The weights sum to 1 and each weight value is between 0 and 1.

[0043] The detailed steps of establishing the model training data set are to first collect marine meteorological observation data from different sea areas and seasons around the world as the basic data source. The data sources include buoy observation stations, ocean survey ships, satellite remote sensing data and coastal observation stations. The data coverage time span is no less than 10 years. Then the collected raw data is quality screened and outliers are removed. In principle, obviously abnormal observations were removed, and time series continuity tests were used to remove records with excessive data missing, forming a clean basic data set. The data was then divided into three categories based on the geographical location of the sea area: tropical, temperate, and cold seas; three categories based on water depth: deep sea, medium-deep sea, and shallow sea; and four categories based on season: spring, summer, autumn, and winter, establishing a multi-dimensional labeling system. Marine meteorology experts were then invited to manually annotate the data. The annotations included abnormal data identification results and corresponding weight allocation recommendations, which served as target labels for supervised learning. Finally, the dataset was randomly divided into training, validation, and test sets in a ratio of 8:1:1, and all input features were standardized to ensure the stability of model training.

[0044] The reason why the adaptive weight optimization model of ocean parameters is suitable for solving the technical problems of the present invention is that the complexity and variability of the marine environment require the quality control method to have dynamic adaptability. The traditional fixed weight quality control method cannot adapt to the particularity of different sea environments and seasonal changes, and is prone to misjudgment or missed judgment. The adaptive weight optimization model of the present invention automatically learns the optimal weight distribution strategy under different environmental conditions through deep learning technology, and can adjust the weight ratio of anomaly detection in each dimension in real time according to the water depth, latitude and seasonal characteristics of the current sea area. Compared with the existing fixed weight method based on expert experience, the model has stronger generalization ability and adaptability, and can handle complex and changeable marine environmental conditions. Compared with the dynamic weight adjustment technology based on simple statistical methods, the model can more accurately capture the complex correlation between different environmental factors through the attention mechanism, and achieve more accurate weight distribution. The advantage of this model is that it can automatically learn the optimal weight distribution pattern from a large amount of historical data, avoiding the subjectivity and limitations of manually setting weights, improving the objectivity and accuracy of the quality control method, and providing a more intelligent and precise solution for marine meteorological data quality control.

[0045] The key technical ideas of the present invention mainly include four aspects: cross-parameter correlation network construction, multi-dimensional abnormal signal detection mechanism, ocean parameter adaptive weight optimization model and virtual sensor data reconstruction algorithm. The cross-parameter correlation network construction technology overcomes the limitations of traditional single-parameter quality control methods by establishing a mathematical correlation model between marine meteorological parameters. It can improve the accuracy of anomaly identification by utilizing the inherent physical relationship between parameters, and has stronger logical consistency and reliability than existing independent parameter verification methods. The multi-dimensional abnormal signal detection mechanism can more comprehensively identify various types of data anomalies than traditional single-dimensional or two-dimensional detection methods through comprehensive analysis of four dimensions: parameter threshold exceeding, spatiotemporal change rate anomaly, probability density distribution offset and correlation verification failure, effectively reducing the missed detection rate and false detection rate. The ocean parameter adaptive weight optimization model uses deep learning technology to achieve dynamic weight adjustment. Compared with the fixed weight method, it can optimize the detection strategy according to specific environmental conditions, improving the adaptability and accuracy of the quality control method under different sea areas and seasonal conditions. The virtual sensor data reconstruction algorithm repairs abnormal data based on the parameter correlation network, and can maintain the physical consistency and logical rationality of the data compared with simple interpolation or replacement methods. The synergistic effect of these technical ideas has formed a complete intelligent marine meteorological data quality control system. Through the organic combination of parameter correlation analysis, multi-dimensional detection, dynamic weight optimization and intelligent data reconstruction, compared with existing technologies, it has achieved a technological leap from passive detection to active prevention, from single judgment to comprehensive evaluation, and from fixed strategy to adaptive optimization, significantly improving the intelligence level and practical effect of marine meteorological data quality control.

[0046] It should be noted that the present invention also solves the following technical problem: the existing technology has a technical problem that the inaccurate health status assessment of marine meteorological sensors makes it difficult to determine the reliability of data. Traditional sensor status monitoring methods mainly rely on single data variance analysis or simple threshold comparison to judge the working status of the sensor. This method cannot fully reflect the actual health of the sensor in a complex marine environment. In particular, when facing harsh conditions such as high salinity, high humidity, and strong corrosiveness in the marine environment, the performance of the sensor will gradually degrade as the working time increases. Traditional methods are difficult to accurately evaluate this gradual performance degradation process. By establishing a sensor health assessment function, the present invention comprehensively considers five key parameters: sensor output data variance, data drift trend coefficient, cross-validation consistency index, sensor working time, and environmental corrosion degree assessment value. It uses a multi-dimensional fusion method to calculate the sensor reliability score, which can accurately reflect the actual working status of the sensor in the marine environment, provide reliable sensor status information for data quality control, and effectively solve the technical problem of inaccurate sensor health status assessment.

[0047] In addition, the present invention also solves the technical problem in the prior art that the accuracy of repairing abnormal marine meteorological data is not high, resulting in the inability to guarantee data integrity and continuity. Traditional data repair methods usually use simple interpolation algorithms or mean filling methods to process missing or abnormal marine meteorological data. These methods do not fully consider the physical correlation between marine meteorological parameters. The repaired data often does not conform to the actual change law of the marine environment, resulting in a decrease in data quality. The present invention uses a virtual sensor data reconstruction algorithm, based on the normal parameter values ​​in the cross-parameter correlation network and the established correlation rules, and uses the correlation network inverse calculation equation to solve the estimated values ​​of the abnormal parameters. This repair method based on the physical correlation between parameters can ensure that the repaired data conforms to the principles of marine meteorology, maintains the physical consistency and logical rationality of the data, significantly improves the accuracy and reliability of abnormal data repair, and effectively guarantees the integrity and continuity of marine meteorological data.

[0048] Specifically, the present invention addresses the technical issue of inaccurate identification of multi-parameter correlations in marine meteorological data quality control. Its fundamental principle lies in establishing a multi-dimensional anomaly detection system based on a cross-parameter correlation network, which fully considers the inherent physical correlations and dynamic variation characteristics between marine meteorological parameters. First, an iterative regression algorithm analyzes historical marine meteorological data to establish temperature-pressure correlation rules, humidity-temperature linear correlation rules, and wind speed-pressure gradient square root correlation rules. These rules reflect the objective physical relationships between meteorological elements in the marine environment and provide a scientific basis for anomaly detection. Second, a sliding window algorithm is used to calculate the correlation coefficients between parameters in real time. Correlation anomalies are determined by the degree of deviation from the standard correlation coefficient. This dynamic monitoring mechanism can promptly detect anomalous data that violates the normal correlation between parameters. Third, a multi-dimensional anomaly detection mechanism is established, encompassing four dimensions: parameter threshold violations, spatiotemporal rate anomalies, probability density distribution shifts, and correlation check failures. By quantifying the anomaly intensity values ​​of each dimension, the degree of data anomaly is comprehensively assessed, avoiding the limitations of single-dimensional determination. Finally, an adaptive ocean parameter weight optimization model based on a hierarchical attention network architecture was employed. The weight distribution of each dimension was dynamically adjusted based on environmental characteristics such as sea depth, geographic latitude, and seasonal temperature differences. A comprehensive anomaly index was calculated using fuzzy comprehensive evaluation algorithms, enabling adaptive identification of parameter correlations under varying ocean environmental conditions. This technical solution, consistent with the principles of marine meteorology, accurately reflects the coordinated changes in various meteorological elements in the ocean environment, effectively improving the accuracy and reliability of anomaly data detection.

[0049] A specific embodiment 1 of the present invention is provided below. The specific implementation of each step in this embodiment 1 is described in detail as follows.

[0050] The specific implementation of step S01 is to use an iterative regression algorithm to construct a cross-parameter correlation network model. The temperature and pressure correlation rules are expressed by a piecewise correlation function as follows: ; Where, is the correlation coefficient between temperature and pressure, dimensionless; is the temperature value, in °C; are linear coefficients, and their units are ℃ and ℃ ; is the constant term coefficient, dimensionless; is the coefficient of the quadratic term, in °C ; is the error correction term, dimensionless, ranging from -0.02 to 0.02 and -0.03 to 0.03 respectively. The linear correlation rule between humidity and temperature is expressed as: Where, is the humidity-temperature correlation coefficient, dimensionless; is the linear slope coefficient, in °C ; is the intercept constant, dimensionless; is a random error term, dimensionless, ranging from -0.05 to 0.05. The square root correlation rule of wind speed and pressure gradient is expressed as: Where, is the correlation coefficient of wind speed and pressure gradient, dimensionless; is the pressure gradient, in Pa / km; is the square root coefficient, unit is (Pa / km) ; is the offset constant, dimensionless; is the nonlinear error term, dimensionless, ranging from -0.08 to 0.08.

[0051] The parameter acquisition method is: It is obtained through iterative regression using the least squares method, including step 1: collecting historical temperature and pressure observation data pairs for no less than 5 years; step 2: using the gradient descent algorithm to perform iterative optimization for more than 1000 times; step 3: determining the optimal parameter combination through cross-validation. The Pearson correlation coefficient calculation method is used to obtain it, including step 1: statistical analysis of the linear relationship between humidity and temperature observation data; step 2: fitting the straight line parameters using a linear regression algorithm. The nonlinear fitting algorithm is used to obtain the data, including step 1: calculating the scattered distribution of wind speed data and the square root of the pressure gradient; step 2: using the Levenberg-Marquardt algorithm to perform nonlinear parameter estimation. The environmental data statistics are adopted, including step 1: collecting historical environmental change data of the sea area; step 2: calculating the environmental stability index as a correction factor.

[0052] The specific implementation of step S02 is to use a sliding window algorithm to calculate the real-time correlation coefficient. The real-time correlation coefficient calculation formula is: ; Where, for The real-time correlation coefficient at the moment, dimensionless; For the The parameter observation value at each moment, the unit depends on the specific parameter; is the average value of the parameter in the sliding window, and its unit is the same as that of the corresponding parameter; is the sliding window length, which is set to 6 data points; The current index. The triggering conditions for correlation anomaly identification are: Where, is the corresponding standard correlation coefficient, dimensionless.

[0053] The specific implementation of step S03 is to establish a multi-dimensional abnormal signal detection mechanism, and the numerical calculation formula of the parameter threshold over-limit intensity is: ; Where, is the parameter threshold exceeding strength value, dimensionless; is the observed parameter value, and the unit depends on the parameter type; is the parameter threshold boundary, and its unit is the same as that of the observed parameter; is the normal variation range of the parameter, and the unit is the same as the observed parameter; is the environmental correction factor, dimensionless, ranging from 0.8 to 1.2. The numerical calculation formula for the abnormal intensity of spatiotemporal change rate is: Where, is the value of the abnormal intensity of the spatiotemporal change rate, dimensionless; is the time rate of change of the parameter, in parameter units per hour; is the spatial rate of change of the parameter, in parameter units per kilometer; is the spatial coordinate; is the spatial weight coefficient, dimensionless, ranging from 0.5 to 1.5. The numerical calculation formula for the probability density distribution offset intensity is: Where, is the probability density distribution offset intensity value, dimensionless; is the probability density function of the current data; is the probability density function of historical normal data; is the parameter value variable. The numerical calculation formula for the correlation check failure strength is: Where, The failure strength value for correlation verification is dimensionless; is the number of parameter pairs; For the Real-time correlation coefficients of parameter pairs; For the The expected correlation coefficient of parameter pairs; For the The importance weight of each parameter pair, dimensionless.

[0054] The parameter acquisition method is: It is obtained by geographic information analysis, including step 1: analyzing the topographic and geomorphic characteristics of the sea area; step 2: determining the spatial weight coefficient according to the complexity of the sea area. The expert evaluation method is used to obtain the parameters, including step 1: inviting marine meteorological experts to score the importance of the parameters; step 2: statistically processing the expert scoring results to obtain the weight coefficient.

[0055] The specific implementation of step S04 is to execute the adaptive weight optimization algorithm. The dynamic weight value calculation adopts the output of the ocean parameter adaptive weight optimization model. The hierarchical weight adjustment function is expressed as: ; Where, is the dynamic weight vector, dimensionless; The current sea depth data, in meters; It is the geographic latitude data, the unit is degree; is the seasonal temperature difference data, in °C; is the data quality assessment score, dimensionless, ranging from 0 to 1; is the neural network mapping function. The calculation formula of the comprehensive adjustment factor is: Where, is the value of the comprehensive adjustment factor, dimensionless; is the maximum reference water depth, set to 6000m; is the maximum seasonal temperature difference, set to 30°C. The formula for adjusting the hierarchical fusion weight parameter is: Where, is the adjusted hierarchical fusion weight parameter, dimensionless; is the basic weight parameter, dimensionless; is the adjustment coefficient, dimensionless, ranging from -0.5 to 0.5.

[0056] The specific implementation of step S05 is to apply fuzzy comprehensive evaluation calculation processing, and the comprehensive abnormality index calculation formula is: ; Where, is the comprehensive anomaly index, dimensionless; is the transpose of the dynamic weight vector; is the anomaly intensity vector; For the The dynamic weight value of each dimension, dimensionless; For the The dimensionless abnormal intensity value of the dimension, where , , , The fuzzy relationship matrix is ​​expressed as: ; Where, For the The evaluation factor The membership degree of each evaluation level is dimensionless. The triangular membership function is expressed as: Where, is the triangle membership function value, dimensionless; is the input variable; are the left boundary, peak point and right boundary parameters of the triangular function.

[0057] The specific implementation of step S06 is to implement multi-sensor cross-validation processing, and the sensor health evaluation function is expressed as: ; Where, Score the sensor reliability, dimensionless; The variance of the sensor output data, the unit depends on the sensor type; is the data drift trend coefficient, dimensionless; is the cross-validation consistency index, dimensionless; The sensor working time, in hours; The design life of the sensor, in hours; is the environmental corrosion degree assessment value, dimensionless; is the weighting coefficient, dimensionless, satisfying , the specific values ​​are 0.2, 0.25, 0.3, 0.15, and 0.1 respectively. The formula for calculating the variance analysis statistic is: Where, is the variance analysis statistic, dimensionless; is the between-group mean square; is the within-group mean square; is the number of sensor groups; For the Number of sensors in a group; For the group sensor observation mean; is the overall observation mean; For the Group sensor observations; is the total number of sensors.

[0058] The parameter acquisition method is: The statistical method is used to obtain the data, including step 1: collecting the output data of the sensor in a stable environment for 24 consecutive hours; step 2: calculating the sample variance of the data sequence. The trend analysis method is used to obtain the drift trend coefficient, including step 1: performing linear regression analysis on the historical output data of the sensor; step 2: calculating the slope of the regression line as the drift trend coefficient. The correlation analysis method is used to obtain the data, including step 1: calculating the Pearson correlation coefficient of the target sensor and the redundant sensor observation data; step 2: averaging the correlation coefficients of multiple sensors. The experimental measurement is used to obtain the value, including step 1: regularly detecting the corrosion degree of the sensor housing; step 2: calculating the corrosion degree assessment value based on the corrosion area ratio, the calculation formula is ,in is the corrosion area, is the total surface area of ​​the sensor, is the material corrosion sensitivity coefficient.

[0059] The specific implementation of step S07 is to perform abnormal data repair and reconstruction processing, and the correlation network reverse calculation equation group is expressed as: ; Where, are the temperature, humidity, and wind speed parameters to be solved, in units of ℃, %, and m / s respectively; is the target air pressure value, in Pa; are the known normal humidity and wind speed observation values, in % and m / s respectively; is the dimensionless correlation coefficient between each parameter pair. The estimated value of the abnormal parameter is solved by Cramer's rule: Where, is the estimated value of the temperature anomaly parameter, in °C; is the coefficient matrix; is the matrix after replacing the first column of the coefficient matrix with a constant vector; is the matrix determinant operation. The matrix determinant calculation formula is: Where, is a matrix Middle Rank Elements of a column.

[0060] The principle and effect of each formula are explained below. Based on the physical principle of the nonlinear relationship between air temperature and air pressure in marine meteorology, the variation law of correlation in different temperature ranges is accurately described by piecewise function form, in which the low temperature range adopts a linear relationship. Reflects the stable thermodynamic state, and the high temperature section uses a quadratic function Describing complex nonlinear thermal convection processes, it can more accurately reflect the complex thermodynamic processes in the marine environment than the traditional single linear correlation model, significantly improving the accuracy and applicability of correlation modeling. Real-time correlation coefficient calculation formula Using the sliding window Pearson correlation coefficient algorithm, the molecular part Calculate the covariance between parameters, denominator part Standardization is performed to capture the real-time changes in the relationship between parameters through a dynamic time window. Compared with the static correlation analysis method of a fixed time period, it can timely detect abnormal fluctuations in parameter relationships, providing a sensitive and reliable abnormality detection basis for real-time quality control. Parameter threshold over-limit intensity calculation formula Combined with the degree of parameter deviation and environmental correction factors By quantifying the degree of over-limit and taking into account the environmental impact, an accurate assessment of the abnormal intensity can be achieved. Compared with the simple over-limit judgment method, it can provide a more refined quantification of the abnormal degree and provide a reliable numerical basis for subsequent comprehensive evaluation. Based on the partial differential theory, the changing characteristics of parameters in time and space dimensions are comprehensively considered, and the gradient calculation is performed. Capturing the dynamic change pattern of parameters, compared with the traditional single-dimensional change rate detection method, it can fully reflect the spatiotemporal evolution of the marine environment and effectively identify various types of dynamic anomalies. Using integral distance It measures the difference between the current data distribution and the historical normal distribution, and realizes the quantitative comparison of distribution characteristics based on statistical principles. Compared with simple mean or variance comparison methods, it can fully reflect the overall changes in data distribution and provide a rigorous mathematical basis for statistical anomaly detection. Through weighted linear combination The fusion of multi-dimensional abnormal information is achieved, and the importance of each dimension is adjusted according to environmental characteristics in combination with a dynamic weight optimization strategy. Compared with the comprehensive evaluation method with fixed weights, it can adapt to the detection needs under different sea areas and seasonal conditions, and significantly improves the accuracy and adaptability of comprehensive judgment. Sensor health evaluation function Based on the multi-factor comprehensive evaluation theory, through the exponential function and Describes the exponential decay law of sensor performance, the reciprocal function Reflecting the negative impact of variance on reliability, the linear term It directly reflects the consistency contribution and can fully reflect the working status of the sensor compared to the health evaluation method of a single indicator, providing a reliable basis for weight distribution for cross-validation. The correlation network inverse calculation equation group is based on linear algebra theory and parameter correlation network structure, and realizes the calculation and reconstruction of abnormal parameters through matrix operations. The coefficient matrix reflects the correlation relationship between parameters, the constant vector contains known parameter information, and the Cramer's law solution process The uniqueness of the solution is ensured by the determinant ratio. Compared with simple interpolation or mean replacement methods, it can maintain the physical consistency between parameters, ensure that the reconstructed data conforms to the laws of marine meteorology, and provide a scientific and reliable repair method for data integrity.

[0061] To better understand and implement the present invention, Example 2 of a specific application scenario is provided below: A technical team deployed multiple sets of marine meteorological observation equipment in a certain sea area and needed to perform real-time quality control on the collected marine meteorological data. The technical team used a marine meteorological data quality control method based on a cross-parameter correlation network to perform quality control on 30 consecutive days of observation data in July 2024.

[0062] The technical team first constructed a network model for the correlation between marine meteorological parameters and used an iterative regression algorithm to analyze and process the historical marine meteorological data of the sea area over the past three years. After 1200 iterative optimization calculations, the temperature and pressure correlation rules were determined: when the temperature is higher than 25°C, the correlation coefficient is -0.62, and the applicable temperature range is 25°C to 35°C. At the same time, a linear correlation rule for humidity and temperature was established, and the correlation coefficient was determined to be 0.78 through the Pearson correlation coefficient calculation method. The wind speed and pressure gradient square root correlation rule is established based on the principles of fluid mechanics, and the function parameters are determined through nonlinear fitting analysis. A cross-parameter correlation network structure is formed, which includes six core parameters: air temperature, air pressure, humidity, wind speed, wind direction, and sea surface temperature. Each parameter node is connected by a correlation coefficient.

[0063] During the real-time data processing phase, the technical team implemented a sliding window correlation coefficient calculation, using a sliding window algorithm with a 30-minute time window to calculate the correlation coefficient for each parameter in the cross-parameter correlation network structure in real time. At 10:30 AM on July 15th, the calculated real-time correlation coefficient for air temperature and pressure was -0.48, a 22.6% deviation from the standard correlation coefficient of -0.62, which did not exceed the 25% threshold. However, at 11:00 AM, the real-time correlation coefficient dropped to -0.42, a 32.3% deviation, exceeding the 25% threshold and triggering a correlation anomaly flag.

[0064] The technical team established a multi-dimensional abnormal signal detection mechanism to identify and quantify four dimensions. The detection results at 11:00 showed that the observed temperature was 33.8°C, exceeding the upper limit of the normal range of 32.5°C, generating a parameter threshold overlimit intensity value of 0.85. The wind speed suddenly changed from 8.2m / s at the previous moment to 15.7m / s, with a change rate of 91.5%, exceeding the reasonable range by 40%, generating a spatiotemporal change rate anomaly intensity value of 0.92. The probability density distribution of humidity data deviated from the historical normal distribution pattern by 0.31, generating a probability density distribution deviation intensity value of 0.76. The real-time correlation coefficient between temperature and humidity in the correlation check was 0.52, deviating 33.3% from the standard value of 0.78, generating a correlation check failure intensity value of 0.88.

[0065] The technical team implemented an adaptive weight optimization algorithm to calculate dynamic weight values ​​through the ocean parameter adaptive weight optimization model. The input was the current sea depth data of 2850m, the geographical latitude data of 18.5° and the seasonal temperature difference data of 12.3℃. The model was processed based on the hierarchical attention network architecture. Figure 3 As shown, the attention mechanism layer assigns weights to different features. The weight of depth data affecting ocean environmental stability is 0.35, the weight of latitude data affecting climate characteristics is 0.28, and the weight of seasonal temperature differences affecting temporal variation is 0.37. The calculated dynamic weights are: parameter threshold exceeding 0.24, spatiotemporal change rate anomaly 0.31, probability density distribution shift 0.22, and correlation check failure 0.23.

[0066] During the fuzzy comprehensive evaluation calculation and processing phase, the technical team performed a weighted fusion operation on the anomaly intensity value and the dynamic weight value. Using the fuzzy comprehensive evaluation operation rules to establish a fuzzy relationship matrix, the calculated comprehensive anomaly index was 0.85, exceeding the preset threshold of 0.7 and determining a data anomaly. The hierarchical weight adjustment function calculated the comprehensive adjustment factor value of 0.68 based on the current sea depth data, geographic latitude data, seasonal temperature difference data, and data quality assessment scores, falling within the range of 0.5 to 0.75. A shallow-feature-prioritized weight allocation strategy was used to adjust the model's hierarchical fusion weight parameters.

[0067] The technical team implemented multi-sensor cross-validation and activated the multi-sensor cross-validation mechanism for identified data anomalies. Three sets of redundant sensors were deployed in the sea area, numbered A01, A02, and A03. The cross-comparison analysis results are shown in Table 1: Table 1 Sensor cross-validation comparison data table

[0068] By comparing and analyzing redundant sensor observation data, it was determined that the temperature and wind speed data of sensor A01 were abnormal. The technical team used the sensor health assessment function to evaluate each sensor, as shown in Table 2: Table 2 Sensor health evaluation results

[0069] The reliability score of sensor A01 was 0.58, which was lower than the 0.6 threshold, indicating that maintenance and calibration were required. The reliability scores of sensors A02 and A03 were 0.89 and 0.87, respectively, indicating that they were in good working condition.

[0070] The technical team performed abnormal data repair and reconstruction, applying a virtual sensor data reconstruction algorithm to the confirmed abnormal data from sensor A01. They input the normal parameter values ​​of sensors A02 and A03 into the cross-parameter correlation network structure, along with the temperature and pressure correlation rules, the humidity-temperature linear correlation rule, and the wind speed and pressure gradient square root correlation rule. By reversing the correlation network equations, they estimated the abnormal parameters: the repaired temperature was 29.4°C and the wind speed was 9.0 m / s.

[0071] During the 30-day quality control process, the system detected 246 data anomalies, including 89 parameter threshold exceeding anomalies, 73 spatiotemporal change rate anomalies, 52 probability density distribution offset anomalies, and 32 correlation check failure anomalies. 198 true anomalies were confirmed through multi-sensor cross-validation, with a false alarm rate of 19.5%. The virtual sensor data reconstruction algorithm successfully repaired 192 abnormal data, with a repair success rate of 97.0%. The relevant charts of the implementation process are shown in the figure below. Figures 2-4 shown.

[0072] The present invention has brought significant progress compared to traditional quality control methods. Traditional methods mainly rely on threshold judgments and simple statistical tests of single parameters, lack correlation analysis between parameters, and are prone to false alarms and missed reports. The present invention, by constructing a cross-parameter correlation network model, makes full use of the physical correlation between marine meteorological parameters, and can identify complex anomalies that are difficult to detect with traditional methods. The multi-dimensional abnormal signal detection mechanism comprehensively evaluates data quality from four perspectives: parameter threshold, spatiotemporal changes, probability distribution, and correlation verification, providing a more comprehensive anomaly recognition capability. The adaptive weight optimization algorithm dynamically adjusts the detection strategy according to the characteristics of the sea area and seasonal characteristics, avoiding the problem of insufficient adaptability caused by fixed weights. The multi-sensor cross-validation mechanism effectively reduces the false alarm rate and improves the accuracy of anomaly recognition. The virtual sensor data reconstruction algorithm infers abnormal data based on the correlation relationship between parameters, ensures the continuity and integrity of the data, and provides a reliable data foundation for subsequent marine environmental analysis.

[0073] It should be noted that the detailed explanations of the variables involved in the present invention are shown in Tables 3 and 4.

[0074] Table 3 Variable Explanation Table (Part I)

[0075] Table 4 Variable Explanation Table (Part II)

[0076] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A method for quality control of marine meteorological data based on a cross-parameter correlation network, characterized in that: It includes constructing a marine meteorological parameter correlation network model, obtaining the temperature and pressure correlation rules, the humidity and temperature linear correlation rules, and the wind speed and pressure gradient square root correlation rules, and forming a cross-parameter correlation network structure; implementing sliding window correlation coefficient calculation, and calculating the correlation coefficient of each parameter in the cross-parameter correlation network structure in real time. When the real-time correlation coefficient deviates from the standard correlation coefficient by more than a preset threshold, the correlation anomaly flag is triggered; establishing a multi-dimensional abnormal signal detection mechanism, identifying and quantifying parameter threshold exceeding signals, spatiotemporal change rate abnormal signals, probability density distribution offset signals, and correlation verification failure signals, and generating corresponding intensity values; executing an adaptive weight optimization algorithm, and calculating dynamic weight values ​​through the marine parameter adaptive weight optimization model; Apply fuzzy comprehensive evaluation calculation processing, perform weighted fusion operation on the intensity value and the dynamic weight value, calculate the comprehensive anomaly index, and determine the data as abnormal when the comprehensive anomaly index exceeds the preset threshold; implement multi-sensor cross-validation processing to determine the authenticity of the data anomaly and calculate the sensor reliability score; perform abnormal data repair and reconstruction processing, adopt virtual sensor data reconstruction algorithm, solve the estimated value of the anomaly parameter through the correlation network inverse calculation equation, and output the repaired marine meteorological parameter data.

2. The method for quality control of marine meteorological data based on a cross-parameter correlation network according to claim 1, characterized in that: The steps of constructing the marine meteorological parameter correlation network model are to use the iterative regression algorithm to analyze and process the historical marine meteorological data to obtain the temperature and pressure correlation rules, and at the same time establish the humidity and temperature linear correlation rules and the wind speed and pressure gradient square root correlation rules to form a cross-parameter correlation network structure containing 6 groups of core parameters. The parameter nodes in the cross-parameter correlation network structure are connected by the correlation coefficient.

3. The marine meteorological data quality control method based on a cross-parameter correlation network according to claim 2 is characterized in that: The cross-parameter correlation network structure is specifically a network topology model based on the principles of marine meteorology, which is a correlation between multiple meteorological parameters. It describes the coordinated change law of various meteorological elements in the marine environment by quantifying the mathematical relationship between different parameters.

4. The method for quality control of marine meteorological data based on a cross-parameter correlation network according to claim 3, characterized in that: The temperature and pressure correlation rule is specifically a piecewise negative correlation mathematical relationship between temperature and pressure obtained by iteratively calculating a large amount of historical marine meteorological observation data using least squares regression analysis and a polynomial fitting algorithm. The final correlation coefficient and applicable temperature range are determined through more than 1,000 iterative optimizations.

5. The method for quality control of marine meteorological data based on a cross-parameter correlation network according to claim 4, characterized in that: The humidity and temperature linear correlation rule is specifically a linear correlation relationship obtained by statistically analyzing the observation data of relative humidity and temperature in the marine environment through a linear regression algorithm, and the correlation coefficient is determined by the Pearson correlation coefficient calculation method.

6. The method for quality control of marine meteorological data based on a cross-parameter correlation network according to claim 5, characterized in that: The wind speed and pressure gradient square root correlation rule is specifically based on the square root function relationship between wind speed and pressure gradient established based on the principles of fluid mechanics, and the function parameters are determined by nonlinear fitting analysis of measured wind speed data and contemporaneous pressure gradient data.

7. The method for quality control of marine meteorological data based on a cross-parameter correlation network according to claim 6, characterized in that: The steps for calculating the sliding window correlation coefficient are as follows: a sliding window algorithm with a time window length of 30 minutes is used to calculate the correlation coefficient of each parameter in the cross-parameter correlation network structure in real time; a correlation anomaly flag is triggered when the calculated real-time correlation coefficient deviates from the corresponding standard correlation coefficient in the temperature and pressure correlation rule, the humidity and temperature linear correlation rule, and the wind speed and pressure gradient square root correlation rule by more than 25%.

8. The method for quality control of marine meteorological data based on a cross-parameter correlation network according to claim 7, characterized in that: The steps for establishing a multi-dimensional abnormal signal detection mechanism are to identify and quantify the four dimensions of parameter threshold exceeding signal, spatiotemporal change rate abnormal signal, probability density distribution offset signal and correlation check failure signal respectively, and generate corresponding parameter threshold exceeding intensity value, spatiotemporal change rate abnormal intensity value, probability density distribution offset intensity value and correlation check failure intensity value.

9. The method for quality control of marine meteorological data based on a cross-parameter correlation network according to claim 8, characterized in that: The steps of executing the adaptive weight optimization algorithm are as follows: specifically, the dynamic weight values ​​are calculated for the parameter threshold exceeding intensity value, the spatiotemporal change rate abnormal intensity value, the probability density distribution offset intensity value and the correlation check failure intensity value respectively through the ocean parameter adaptive weight optimization model; the ocean parameter adaptive weight optimization model adjusts the hierarchical fusion weight parameters of the model based on the current sea area water depth data, geographic latitude data and seasonal temperature difference data.

10. The method for quality control of marine meteorological data based on a cross-parameter correlation network according to claim 9, characterized in that: The steps of applying fuzzy comprehensive evaluation calculation and processing are to perform weighted fusion operation on the parameter threshold exceeding intensity value, the spatiotemporal change rate abnormal intensity value, the probability density distribution offset intensity value and the correlation check failure intensity value with the dynamic weight value, and calculate the comprehensive anomaly index through the fuzzy comprehensive evaluation operation rules. When the comprehensive anomaly index exceeds the preset threshold of 0.7, it is judged as data anomaly.

Images