A Method for Modeling and Estimating the Probability of Secondary and Derivative Abnormal States in Marine Water Environments

Abstract

The present invention discloses a method for modeling and estimating the secondary and derivative probabilities of abnormal states in the marine water environment, belonging to the field of state modeling and estimation, which includes: obtaining the real-time data stream of marine environmental parameters, performing segmented processing by using a sliding window mechanism, and reducing the data dimension by using principal component analysis. Subsequently, a graph structure model based on time series is constructed to capture the correlation and spatio-temporal dependence relationships among the parameters. A radial basis function kernel is used to perform non-linear mapping on the graph structure model to capture the complex non-linear characteristics among the marine environmental parameters. Finally, a multi-class support vector machine algorithm is used to classify the abnormal states, and the estimation results of the secondary and derivative probabilities of the abnormal states are output in real time. The present invention can not only timely detect abnormal events in the marine environment, but also provide dynamic modeling support for marine environmental monitoring, significantly improving the accuracy and real-time performance of marine environmental monitoring.

Description

Technical Field

[0001] The present invention belongs to the field of state modeling and estimation, and particularly relates to a method for modeling and estimating the probability of secondary and derivative abnormal states in the marine water environment. Background Art

[0002] In the process of modeling and estimating the probability of secondary and derivative abnormal states in the marine water environment, how to effectively perform sliding modeling is a key technical problem. Due to the dynamic change characteristics of marine environmental parameters, traditional static modeling methods are difficult to accurately capture the evolution trend of abnormal states. Although sliding modeling can update model parameters in real time, it still faces many challenges in practical applications. First, the selection of the size of the sliding window directly affects the accuracy and efficiency of modeling. A too small window may make the model sensitive to noise, while a too large window may introduce too much historical data and reduce the real-time performance of the model. Second, sliding modeling needs to consider the correlation and spatio-temporal dependence relationships among marine environmental parameters. Ignoring these relationships may lead to the model misjudging secondary and derivative events of abnormal states. In addition, the high-dimensional and non-linear characteristics of marine environmental data also bring difficulties to sliding modeling. How to effectively perform feature selection and extraction in the sliding window and construct a robust non-linear model is an urgent problem to be solved. To sum up, in the modeling and estimation of the probability of secondary and derivative abnormal states in the marine water environment, how to design an adaptive sliding modeling method that takes into account factors such as window size selection, parameter correlation, data high-dimensionality and non-linearity is a complex and key technical problem that requires further research and exploration. Summary of the Invention

[0003] To solve the above technical problems, the present invention provides a method for modeling and estimating the probability of secondary and derivative abnormal states in the marine water environment, including:

[0004] Obtain a real-time data stream of marine environmental parameters, and use a sliding window mechanism to segment the time series information in the data stream to obtain data within the window;

[0005] Use the principal component analysis method to reduce the dimensionality of the high-dimensional data within the window, extract the main feature components while retaining key information, and obtain the dimensionality-reduced data;

[0006] Construct a graph structure model based on the time series according to the dimensionality-reduced data, capture the correlation and spatio-temporal dependence relationships among marine environmental parameters based on the graph structure model, and judge the interaction intensity among the marine environmental parameters;

[0007] Use a radial basis function kernel to perform non-linear mapping on the graph structure model, construct a non-linear mapping relationship, capture the non-linear characteristics among the marine environmental parameters, and obtain a non-linear mapping result;

[0008] Based on the non - linear mapping results, the support vector machine algorithm is used to classify abnormal states and determine whether there are abnormal events in the current window; and through the sliding window mechanism, the secondary and derivative probability estimation results of the abnormal state are output in real time.

[0009] Preferably, a real - time data stream of ocean environmental parameters is obtained, and the sliding window mechanism is used to segment the time - series information in the data stream. The process of obtaining the data within the window includes:

[0010] The data stream is segmented by using the sliding window mechanism. According to the preset window size and sliding step, the data stream is divided into multiple time windows;

[0011] For the data within each time window, time - series information is extracted. The time - series information includes the change trend characteristics and numerical range characteristics of each ocean environmental parameter within the current window;

[0012] According to the time - series information, a time - series analysis algorithm is used to model and predict the data within the window, and a prediction result of the ocean environmental parameter is obtained;

[0013] The prediction results within each time window are summarized and fused to obtain a comprehensive prediction result for the entire data stream;

[0014] The prediction result is compared with a preset threshold. If the prediction result exceeds the normal range, the early - warning mechanism is triggered to send an alarm message to the relevant department, indicating the possible abnormal situation of the ocean environment;

[0015] The real - time data stream is continuously monitored, the sliding window moves forward continuously, the data within the new time window is analyzed and predicted, and the size and sliding step of the sliding window are dynamically adjusted according to the historical data and prediction results.

[0016] Preferably, the principal component analysis method is used to reduce the dimension of the high - dimensional data within the window, extract the main feature components while retaining key information. The process of obtaining the reduced - dimension data includes:

[0017] The high - dimensional feature information of the data within the window is obtained, a principal component analysis model is constructed, and the high - dimensional data is processed by the principal component analysis model for dimension reduction;

[0018] By using the principal component analysis method, the high - dimensional data within the window is decomposed by features, the main feature components of the data are extracted, and the number of principal components is determined according to the eigenvalue size;

[0019] According to the number of principal components, the original high-dimensional data is subjected to dimensionality reduction mapping, and the data is mapped from the high-dimensional space to the low-dimensional space. During the dimensionality reduction process, the key information and characteristic information of the data are retained by retaining the principal components;

[0020] Reconstruct the data after dimensionality reduction to obtain reconstructed data with the same dimension as the original data, and compare it with the original data to evaluate the degree of information retention of the dimensionality reduction process;

[0021] According to the effect of dimensionality reduction, adjust and optimize the parameters of the principal component analysis model, adjust the number of principal components and optimize the data preprocessing method;

[0022] The optimized principal component analysis model is applied to the adjusted window data to obtain the final dimensionality reduction processing result.

[0023] Preferably, the process of capturing the correlation and spatiotemporal dependency between the ocean environment parameters based on the graph structure model and determining the interaction strength between the ocean environment parameters includes:

[0024] Based on the reduced-dimensional marine environmental parameter data, a time series graph structure model is constructed, in which nodes represent marine environmental parameters and edges represent the correlation and spatiotemporal dependency between parameters.

[0025] The time series graph structure model is trained using a graph neural network to learn and capture the complex interaction relationship between marine environmental parameters;

[0026] During the training process, the attention mechanism is introduced to dynamically adjust the connection weights between nodes at different time steps and spatial positions to highlight the mutual influence between key parameters;

[0027] Inference is performed on the trained graph neural network model to obtain the interaction strength between the ocean environment parameters at each time step and spatial position;

[0028] According to the interaction strength, a causal relationship diagram between the marine environmental parameters is constructed to obtain the causal dependency and action path between the parameters;

[0029] For the key nodes and paths in the causal graph, a graph attention network is used to further mine the high-order interaction patterns and long-range dependencies between parameters;

[0030] The prediction results of graph neural network and graph attention network are integrated to judge the interaction strength between different marine environmental parameters.

[0031] Preferably, the process of inferring the trained graph neural network model to obtain the interaction strength between the ocean environment parameters at each time step and spatial position includes:

[0032] Construct the input data structure of the graph neural network model according to the spatio-temporal distribution of ocean environmental parameters, where nodes represent each spatio-temporal position and edges represent the connection relationships between adjacent spatio-temporal positions;

[0033] Feed the input data into the pre-trained graph neural network model, calculate the hidden state vector of each node through forward propagation, and aggregate the information of adjacent nodes through the attention mechanism;

[0034] At the output layer of the graph neural network, use a fully connected layer to map the hidden state vector of each node to the predicted value of the ocean environmental parameter corresponding to the spatio-temporal position;

[0035] Calculate the difference between the predicted value of each spatio-temporal position and the predicted value of the adjacent spatio-temporal position as the interaction strength between the spatio-temporal position and the adjacent position;

[0036] Judge the degree of mutual influence between different ocean environmental parameters according to the magnitude of the interaction strength. The greater the strength, the stronger the correlation between the two parameters;

[0037] Normalize the interaction strength matrix to obtain the relative interaction strength distribution under different spatio-temporal positions and environmental parameter combinations;

[0038] Visualize the normalized interaction strength matrix to generate a heat map or a network diagram.

[0039] Preferably, use a radial basis function kernel to perform a non-linear mapping on the graph structure model, construct a non-linear mapping relationship, and capture the non-linear characteristics between the ocean environmental parameters. The process of obtaining the non-linear mapping result includes:

[0040] Construct a graph structure model according to the ocean environmental parameter data; where nodes represent parameters and edges represent the relationships between parameters;

[0041] Use a radial basis function kernel to perform a non-linear mapping on the graph structure model to capture the non-linear characteristic relationships between ocean environmental parameters; through non-linear mapping, map the original graph structure model to a high-dimensional feature space to obtain the non-linear mapping result;

[0042] Construct a non-linear relationship model between ocean environmental parameters according to the non-linear mapping result;

[0043] Use the support vector machine algorithm to train the non-linear relationship model to obtain the trained non-linear relationship model;

[0044] Use the trained non-linear relationship model to predict new ocean environmental parameter data to obtain the prediction result;

[0045] After comparing the prediction results with the actual observed data and evaluating the prediction performance of the non - linear relationship model, the model is continuously iteratively optimized.

[0046] Preferably, the process of using the trained non - linear relationship model to predict new marine environmental parameter data and obtaining the prediction results includes:

[0047] Construct a non - linear relationship prediction model using a machine learning algorithm based on historical parameter data related to the marine environment;

[0048] During the model training process, optimize the model hyperparameters through the cross - validation method and save the trained non - linear relationship model;

[0049] Obtain new marine environmental parameter data, perform pre - processing operations such as data cleaning and normalization on the data; input the pre - processed new data into the trained non - linear relationship model, and the model automatically performs prediction and analysis calculations;

[0050] Obtain the prediction result data from the model output, and perform post - processing such as anti - normalization and data format conversion on the prediction results.

[0051] Preferably, according to the non - linear mapping result, the process of using the support vector machine algorithm to classify abnormal states and determine whether there are abnormal events in the current window includes:

[0052] Obtain the data within the current time window, pre - process the data to remove noise and irrelevant information, and obtain a pure data set;

[0053] Extract features from the pure data set according to a preset non - linear mapping function to obtain a feature vector that can characterize the abnormal state;

[0054] Input the feature vector into a pre - trained support vector machine model for classification prediction to obtain a judgment result on whether there is an abnormality within the current time window;

[0055] If the judgment result indicates the existence of an abnormality, trigger an abnormal alarm and send the relevant information of the abnormal event to the administrator for further processing;

[0056] If the judgment result indicates the non - existence of an abnormality, continue to obtain the data of the next time window and repeat the above steps to realize continuous monitoring of the abnormal state;

[0057] According to the feedback information of the administrator, update and optimize the support vector machine model in real - time, and store the result data of the abnormal detection in the database to form a historical record.

[0058] Preferably, the process of real - time outputting the secondary and derivative probability estimation results of the abnormal state through the sliding window mechanism includes:

[0059] Obtain real-time monitoring data of the marine environment as the input data source for the sliding window mechanism;

[0060] For the obtained monitoring data, use the sliding window mechanism to divide the data into several time windows;

[0061] For the monitoring data within each time window, determine whether there is an abnormal state through an anomaly detection algorithm;

[0062] If an abnormal state is detected within the current time window, according to the pre-established probability model, estimate the probabilities of secondary anomalies and derivative anomalies occurring in the abnormal state, and use the probabilities as the estimation results of the abnormal state in the current time window;

[0063] As the sliding window continuously moves, the estimation results of the secondary probability and derivative probability of the abnormal state are updated in real time, and the updated estimation results of the abnormal state probability are output in real time.

[0064] Compared with the prior art, the present invention has the following advantages and technical effects:

[0065] The present invention obtains the real-time data stream of marine environment parameters, uses the sliding window mechanism for segmented processing, and reduces the data dimension by using principal component analysis. Subsequently, a graph structure model based on time series is constructed to capture the correlation and spatio-temporal dependence relationship between parameters. The present invention innovatively uses a radial basis function kernel to perform nonlinear mapping on the graph structure model, effectively capturing the complex nonlinear characteristics between marine environment parameters. Finally, the multi-class support vector machine algorithm is used to classify the abnormal state, and the estimation results of the secondary and derivative probabilities of the abnormal state are output in real time. This method can not only timely detect abnormal events in the marine environment, but also provide dynamic modeling support for marine environment monitoring, significantly improving the accuracy and real-time performance of marine environment monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The drawings forming a part of this application are used to provide a further understanding of this application. The schematic embodiments and descriptions thereof of this application are used to explain this application and do not constitute an improper limitation to this application. In the drawings:

[0067] Figure 1 It is a schematic diagram of the overall method flow of the embodiment of the present invention;

[0068] Figure 2 It is a schematic diagram of the simplified step flow of the embodiment of the present invention;

[0069] Figure 3 It is another schematic diagram of the method flow of the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0070] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The present application will be described in detail below with reference to the drawings and in combination with the embodiments.

[0071] It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer-executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0072] As Figures 1-3 shown, this embodiment provides a method for modeling and estimating the probability of secondary and derivative abnormal states of the marine water environment, including:

[0073] Obtain the real-time data stream of marine environmental parameters, and use the sliding window mechanism to segment the time series information in the data stream to obtain the data within the window;

[0074] Use the principal component analysis method to reduce the dimension of the high-dimensional data within the window, extract the main feature components while retaining the key information, and obtain the dimension-reduced data;

[0075] Construct a graph structure model based on the time series according to the dimension-reduced data, capture the correlation and spatio-temporal dependence relationship between marine environmental parameters based on the graph structure model, and judge the interaction strength between marine environmental parameters;

[0076] Use the radial basis function kernel to perform non-linear mapping on the graph structure model, construct a non-linear mapping relationship, capture the non-linear features between marine environmental parameters, and obtain the non-linear mapping result;

[0077] According to the non-linear mapping result, use the support vector machine algorithm to classify the abnormal state, judge whether there is an abnormal event in the current window; and through the sliding window mechanism, output the estimation result of the probability of secondary and derivative abnormal states in real time.

[0078] Furthermore, the process of obtaining the real-time data stream of marine environmental parameters and using the sliding window mechanism to segment the time series information in the data stream to obtain the data within the window includes:

[0079] Perform segmentation processing on the data stream using the sliding window mechanism, and divide the data stream into multiple time windows according to the preset window size and sliding step;

[0080] Extract the time series information for the data within each time window, and the time series information includes the change trend characteristics and numerical range characteristics of each marine environmental parameter within the current window;

[0081] According to the time series information, a time series analysis algorithm is used to model and predict the data within the window, and the prediction results of the marine environmental parameters are obtained;

[0082] The prediction results within each time window are summarized and fused to obtain the comprehensive prediction results for the entire data stream;

[0083] The prediction results are compared with the preset threshold. If the prediction results exceed the normal range, the warning mechanism is triggered, and alarm information is sent to the relevant departments to prompt the possible abnormal marine environmental conditions;

[0084] Continuously monitor the real-time data stream, the sliding window moves forward continuously, analyze and predict the data within the new time window, and dynamically adjust the size and sliding step of the sliding window according to the historical data and prediction results.

[0085] Specifically, in this embodiment, a sliding window mechanism is adopted to process the data stream for the real-time monitoring of marine environmental parameters. This method can effectively capture the dynamic changes of the marine environment. For example, for parameters such as seawater temperature, salinity, and dissolved oxygen, the window size can be set to 24 hours and the sliding step to 1 hour. In this way, the data analyzed each time contains the information of the most recent day and is updated every hour, which not only ensures the timeliness of the data but also reflects the change trend over a longer period. Within each time window, the system extracts the time series information, including the change trend and numerical range of the parameters. For example, the surface water temperature in a certain sea area shows obvious diurnal changes within a 24-hour window, with the highest temperature reaching 28 °C during the day and the lowest temperature being 22 °C at night, presenting a periodic fluctuation as a whole. This information is crucial for understanding the thermodynamic characteristics of local sea areas.

[0086] When this embodiment is based on the extracted time series information and uses a time series analysis algorithm for modeling and prediction, algorithms such as autoregressive moving average model (ARMA) or long short-term memory network (LSTM) can be used. The ARMA model is suitable for relatively stable marine environments, such as the temperature changes in deep sea areas. The LSTM network is more suitable for dealing with complex non-linear relationships, such as the changes in water quality parameters affected by multiple factors such as tides and climate in coastal areas. Through these algorithms, the changes in marine environmental parameters in the next few hours or even days can be predicted. The summarization and fusion of the prediction results is a key step. For example, when predicting the sea level height, the results of multiple time windows are comprehensively considered. If the short-term prediction shows a rapid rise in the sea level and the long-term trend is also upward, then a higher warning level is given. This multi-scale analysis method can effectively reduce false alarms and improve the accuracy of early warnings.

[0087] The triggering of the early warning mechanism is based on preset thresholds. For example, when the predicted seawater acidity (pH value) is below 7.5 or above 8.5, the system automatically sends an alarm to the marine environment management department. This helps to detect and respond to ocean acidification or alkalization problems in a timely manner, protecting the health of the marine ecosystem. At the same time, it can also dynamically adjust the sliding window parameters according to historical data and prediction results. Increase the window size in calm weather to capture longer-term changes; while reduce the window size and the sliding step before a typhoon to more closely monitor the rapidly changing sea conditions. This adaptive mechanism greatly improves the flexibility and applicability of the system. Through this comprehensive and dynamic analysis method, the real-time monitoring of marine environmental parameters can provide important decision-making support for marine resource management, offshore operation safety, and marine ecological protection. It can not only detect abnormal situations in a timely manner but also predict potential environmental changes, providing sufficient time and basis for relevant departments to take preventive measures.

[0088] Furthermore, the principal component analysis method is used to reduce the dimensionality of the high-dimensional data within the window, extract the main characteristic components while retaining the key information. The process of obtaining the dimensionality-reduced data includes:

[0089] Obtain the high-dimensional characteristic information of the data within the window, construct a principal component analysis model, and perform dimensionality reduction processing on the high-dimensional data through the principal component analysis model;

[0090] Through the principal component analysis method, perform eigen decomposition on the high-dimensional data within the window, extract the main characteristic components of the data, and determine the number of principal components according to the eigenvalue size;

[0091] According to the number of principal components, perform dimensionality reduction mapping on the original high-dimensional data, map the data from the high-dimensional space to the low-dimensional space. During the dimensionality reduction process, retain the key information and characteristic information of the data by retaining the principal components;

[0092] Reconstruct the dimensionality-reduced data to obtain reconstructed data with the same dimension as the original data, and compare it with the original data to evaluate the information retention degree of the dimensionality reduction process;

[0093] According to the effect of the dimensionality reduction process, adjust and optimize the parameters of the principal component analysis model, adjust the number of principal components and optimize the data preprocessing method;

[0094] Apply the optimized principal component analysis model to the adjusted window data to obtain the final dimensionality reduction result.

[0095] Specifically, in this embodiment, dimensionality reduction through principal component analysis can effectively extract the main features of high-dimensional data. In marine environmental monitoring, high-dimensional data composed of multiple parameters such as temperature, salinity, dissolved oxygen, etc. may be faced. Through principal component analysis, these parameters can be reduced to a few principal components, which not only retains the key information but also simplifies the data structure. Taking marine water quality monitoring as an example, assume there are 10 parameters such as temperature, salinity, dissolved oxygen, pH value, turbidity, etc. Through principal component analysis, it may be found that the first 3 principal components can explain 85% of the total variance of the data. This means that these 3 principal components can be used to replace the original 10 parameters, greatly reducing the data dimension. In actual operation, first, the original data needs to be standardized to eliminate the dimensional differences between different parameters. Then calculate the covariance matrix and solve its eigenvalues and eigenvectors. The size of the eigenvalues reflects the importance of the corresponding principal components, and the number of principal components to be retained can be determined according to the cumulative contribution rate. For example, if the cumulative contribution rate of the first 4 principal components reaches 90%, these 4 principal components can be selected to be retained. After dimensionality reduction, it is necessary to evaluate the degree of information retention. In this embodiment, one method is to reconstruct the dimensionality-reduced data back to the original dimension and calculate the mean square error with the original data. If the error is small, it indicates that the dimensionality reduction effect is good. Another method is through visualization, project the data into the space composed of the first two or three principal components, and observe the distribution of the data. The optimization of the principal component analysis model involves multiple aspects. For example, different data preprocessing methods such as min-max normalization or Z-score normalization can be tried to see which method can obtain a better dimensionality reduction effect. In addition, the selection criteria of the principal components can be adjusted, such as trying different cumulative contribution rate thresholds, or using methods such as the scree plot to determine the optimal number of principal components. In marine environmental monitoring, principal component analysis can not only be used for data dimensionality reduction but also help to discover the potential relationships between parameters. For example, it may be found that the first principal component is highly correlated with water temperature and dissolved oxygen, which implies that there may be some physical or biological connections between these two parameters. This discovery can provide clues for further marine ecological research. The data after dimensionality reduction through principal component analysis is more conducive to visualization and interpretation. For example, the data of different sea areas or different periods can be projected onto a two-dimensional plane to visually observe their distribution and clustering. This visualization method can help to quickly identify outliers or special patterns and provide strong support for marine environmental monitoring and early warning. Generally speaking, principal component analysis provides a powerful and flexible tool for the processing of marine environmental data. It can not only effectively reduce the data dimension, simplify the subsequent analysis process, but also help to extract key information from complex multi-dimensional data, discover potential data structures and rules. This is of great significance for improving the efficiency and accuracy of marine environmental monitoring.

[0096] Furthermore, based on the graph structure model, the correlation and spatiotemporal dependency between the marine environmental parameters are captured, and the process of judging the interaction intensity between the marine environmental parameters includes:

[0097] Based on the reduced-dimensional marine environmental parameter data, a time series graph structure model is constructed, in which nodes represent marine environmental parameters and edges represent the correlation and spatiotemporal dependency between parameters.

[0098] A graph neural network is used to train the time series graph structure model to learn and capture the complex interactions between marine environmental parameters;

[0099] During the training process, the attention mechanism is introduced to dynamically adjust the connection weights between nodes at different time steps and spatial positions to highlight the mutual influence between key parameters;

[0100] Inference is performed on the trained graph neural network model to obtain the interaction strength between the ocean environment parameters at each time step and spatial position;

[0101] According to the interaction intensity, a causal relationship diagram between marine environmental parameters is constructed to obtain the causal dependency and action path between the parameters;

[0102] For the key nodes and paths in the causal graph, a graph attention network is used to further mine the high-order interaction patterns and long-range dependencies between parameters;

[0103] The prediction results of graph neural network and graph attention network are integrated to judge the interaction strength between different marine environmental parameters.

[0104] Specifically, a time series graph structure model is constructed, in which nodes represent different marine environmental parameters, such as water temperature, salinity, dissolved oxygen, etc., and edges represent the correlation and spatiotemporal dependency between these parameters. For example, there may be a negative correlation between water temperature and salinity, because high temperature usually leads to water evaporation, thereby increasing salinity.

[0105] The application of graph neural networks (GNNs) enables the capture of these complex interactions. In this embodiment, through the learning of historical data, GNNs can identify how water temperature changes affect dissolved oxygen content or the relationship between ocean current speed and plankton distribution. The introduction of the attention mechanism further enhances the performance of the model, enabling it to dynamically adjust the weights between different parameters. For example, near the equator, the impact of water temperature on other parameters may be more significant, so the model will assign a higher weight to the water temperature node. In the inference stage, the specific interaction intensity can be obtained. Suppose at a certain time step, the model shows that the impact intensity of water temperature on dissolved oxygen is 0.8, while the impact intensity on salinity is 0.5. This provides a quantitative basis for understanding the relative importance between parameters. Based on these results, a causal relationship graph can be constructed to clearly show the causal chain between parameters. For example, it may be found that an increase in water temperature leads to a decrease in dissolved oxygen, which in turn affects fish distribution, forming a complete causal chain.

[0106] The introduction of the graph attention network (GAT) enables the exploration of deeper interaction patterns. It may reveal the non-linear relationships that exist between certain parameters. For example, the complex relationship between water temperature and the number of plankton may exhibit different patterns within a specific temperature range. GAT can also help discover long-range dependencies, such as finding that water temperature changes several months ago have a significant impact on the current fish distribution.

[0107] By integrating the results of GNN and GAT, the interactions between ocean environmental parameters can be comprehensively judged. This method can not only reveal direct causal relationships but also discover indirect and complex mutual influences. For example, it may be found that ocean acidification not only directly affects the growth of shellfish but also indirectly affects the entire food chain by changing the structure of the plankton community. This comprehensive analysis helps to understand the internal mechanism of the marine ecosystem and provides a scientific basis for marine resource management and environmental protection.

[0108] Furthermore, the process of performing inference on the trained graph neural network model to obtain the interaction intensity between ocean environmental parameters at each time step and spatial location includes:

[0109] According to the spatio-temporal distribution of ocean environmental parameters, construct the input data structure of the graph neural network model, where nodes represent each spatio-temporal location and edges represent the connection relationships between adjacent spatio-temporal locations;

[0110] Feed the input data into the pre-trained graph neural network model, calculate the hidden state vector of each node through forward propagation, and aggregate the information of adjacent nodes through the attention mechanism;

[0111] At the output layer of the graph neural network, use a fully connected layer to map the hidden state vector of each node to the predicted value of the ocean environmental parameter corresponding to the spatio-temporal location;

[0112] Calculate the difference between the predicted values at each spatio-temporal position and the predicted values at adjacent spatio-temporal positions as the interaction intensity between the spatio-temporal position and the adjacent positions.

[0113] According to the magnitude of the interaction intensity, judge the degree of mutual influence between different ocean environmental parameters. The greater the intensity, the stronger the correlation between the two parameters.

[0114] Normalize the interaction intensity matrix to obtain the relative interaction intensity distribution under different combinations of spatio-temporal positions and environmental parameters.

[0115] Visualize the normalized interaction intensity matrix to generate a heat map or a network graph.

[0116] Specifically, construct the input data structure of the graph neural network model. Taking seawater temperature, salinity, and dissolved oxygen as examples, each observation station can be regarded as a node in the graph, and the connection relationship between adjacent stations is represented as an edge. This structure can effectively capture the spatial distribution characteristics of the parameters. In model training, the hidden state vector of each node is calculated through forward propagation. For example, for a certain observation station, its hidden state may contain the comprehensive information of the temperature, salinity, and dissolved oxygen at that position. Through the attention mechanism, the model can automatically learn the importance weights between different parameters, so as to more accurately aggregate the information of adjacent nodes. The fully connected layer maps the hidden state vector to the predicted values of specific ocean environmental parameters. Suppose the hidden state vector of a certain station is [0.8, 0.5, 0.3], and after being mapped by the fully connected layer, the predicted temperature value at that position may be 25°C, the salinity is 35‰, and the dissolved oxygen is 6 mg / L.

[0117] Calculating the difference between the predicted values at adjacent spatio-temporal positions can reveal the interaction intensity between parameters. If the predicted temperature values of two adjacent stations are 25°C and 26°C respectively, and the predicted salinity values are 35‰ and 35.5‰ respectively, it may indicate that the spatial variation of temperature is relatively small, while the spatial variation of salinity is relatively large, reflecting the spatial correlation characteristics of different parameters.

[0118] Normalization helps to compare the relative interaction intensities of different parameter combinations. For example, normalizing the interaction intensities of temperature-salinity, temperature-dissolved oxygen, and salinity-dissolved oxygen to the interval [0, 1] respectively, the results may be 0.7, 0.5, and 0.8, indicating that the interaction between salinity and dissolved oxygen is the strongest, followed by temperature-salinity, and the weakest is temperature-dissolved oxygen.

[0119] Visualization is an effective means to intuitively display the dynamic interactions of ocean environmental parameters in this embodiment. Heat maps can represent the interaction intensity with different shades of color. For example, strong interactions are represented by dark red, and weak interactions are represented by light yellow. Network graphs can represent the importance of parameters with the size of nodes and the interaction intensity with the thickness of edges, thus clearly showing the complex relationship network among parameters. This method based on graph neural networks can effectively capture the spatio-temporal dependence relationships of ocean environmental parameters and can better reflect the non-linear interactions among parameters compared with traditional methods. By analyzing the interaction intensity output by the model, researchers can deeply understand the internal mechanisms of the marine ecosystem and provide a scientific basis for marine environmental protection and resource management. For example, discovering a strong correlation between temperature and dissolved oxygen may reveal the potential impact of global warming on the habitat environment of marine organisms and provide an important reference for formulating relevant policies.

[0120] Furthermore, a radial basis function kernel is used to perform non-linear mapping on the graph structure model, construct a non-linear mapping relationship, and capture the non-linear characteristics among ocean environmental parameters. The process of obtaining the non-linear mapping result includes:

[0121] Construct a graph structure model based on ocean environmental parameter data; where nodes represent parameters and edges represent the relationships between parameters;

[0122] Use a radial basis function kernel to perform non-linear mapping on the graph structure model and capture the non-linear characteristic relationships among ocean environmental parameters; through non-linear mapping, map the original graph structure model to a high-dimensional feature space to obtain the non-linear mapping result;

[0123] Construct a non-linear relationship model among ocean environmental parameters according to the non-linear mapping result;

[0124] Use the support vector machine algorithm to train the non-linear relationship model to obtain the trained non-linear relationship model;

[0125] Use the trained non-linear relationship model to predict new ocean environmental parameter data to obtain the prediction result;

[0126] Compare the prediction result with the actual observed data, evaluate the prediction performance of the non-linear relationship model, and then continuously iterate and optimize the model.

[0127] Specifically, collecting ocean environmental parameter data to construct a graph structure model includes collecting the observed values of parameters such as seawater temperature, salinity, and dissolved oxygen at different spatio-temporal points. Taking these parameters as the nodes of the graph and the correlations between parameters as the edges forms the initial graph structure model.

[0128] The application of the radial basis function kernel can effectively capture the non-linear relationships between parameters. For example, there is a complex non-linear correlation between seawater temperature and dissolved oxygen. An increase in temperature will reduce the dissolved oxygen content, but this relationship is not a simple linear one. Through the radial basis function kernel, this complex relationship can be mapped into a high-dimensional space to more accurately describe the interactions between parameters. After non-linear mapping, the originally simple graph structure is transformed into a more complex high-dimensional feature space. In this space, the relationships between ocean environmental parameters become clearer. For example, the surface water temperature and the deep nutrient concentration, which originally seemed unrelated, may show an obvious correlation in the high-dimensional space, reflecting the influence of the water body vertical mixing process. The complex non-linear relationship model constructed based on the non-linear mapping results can more comprehensively describe the ocean environmental system. This model may reveal some unexpected correlations, such as the non-linear relationship between phytoplankton biomass and sea current speed, reflecting the response mechanism of the ecosystem to the physical environment.

[0129] The application of the support vector machine algorithm enables the non-linear relationship model to have prediction capabilities. Through training, the model learns the internal laws between ocean environmental parameters. For example, it may master how to predict the change of chlorophyll concentration based on water temperature, salinity, and nutrient concentration, which is of great significance for studying ocean primary productivity. When using the trained model for prediction, new ocean environmental parameter data can be input, such as the meteorological forecast data for a certain sea area in the next week, and the model can give the prediction results of the water quality parameters in this sea area. This has important application value for ocean environmental monitoring and early warning.

[0130] The comparison between the prediction results and the actual observed data is a key step in evaluating the model performance. For example, the probability of red tide occurrence predicted by the model can be compared with the actual observed results to calculate the prediction accuracy. If it is found that the prediction deviation of the model is relatively large in some cases, such as during a typhoon passing through, it is necessary to re-examine the model structure and may need to introduce more parameters related to extreme weather to optimize the model. Through this series of steps, not only a model that can describe the complexity of the ocean environment is constructed, but also the ability to predict future changes is given to it. The advantage of this method is that it can capture complex relationships that are difficult to describe by traditional linear models, providing a powerful tool for ocean science research and environmental management.

[0131] Furthermore, the process of using the trained non-linear relationship model to predict new ocean environmental parameter data and obtain prediction results includes:

[0132] Construct a non-linear relationship prediction model using machine learning algorithms based on historical parameter data related to the ocean environment;

[0133] During the model training process, optimize the model hyperparameters through the cross-validation method and save the trained non-linear relationship model;

[0134] Obtain new marine environmental parameter data, and perform preprocessing operations such as data cleaning and normalization on the data; input the preprocessed new data into the trained non - linear relationship model, and the model automatically performs predictive analysis and calculation;

[0135] Obtain the prediction result data from the model output, and perform post - processing such as anti - normalization and data format conversion on the prediction results.

[0136] Specifically, the construction of the non - linear relationship prediction model for marine environmental parameter data in this embodiment is a complex task. First, historical data needs to be collected, such as seawater temperature, salinity, dissolved oxygen, etc. This data may come from different marine observation stations or satellite remote sensing. For example, in a certain area, there may be daily seawater temperature, salinity, and dissolved oxygen data for the past 10 years.

[0137] In this embodiment, the support vector machine (SVM) is used to handle non - linear relationships. Specifically, in marine environment prediction, SVM is used to capture the complex relationships between parameters. For example, the radial basis function (RBF) kernel can be used to establish a non - linear relationship model between seawater temperature and dissolved oxygen. Neural networks, especially deep learning models, are also powerful tools for processing marine environmental data in this embodiment. For example, in this embodiment, the long short - term memory network (LSTM) is used to effectively capture the time - series characteristics of marine parameters, and LSTM is used to predict the change of seawater temperature in the next week.

[0138] During the model training process, cross - validation is an important method. For example, in this embodiment, 5 - fold cross - validation is used to optimize the kernel function parameters of SVM or the number of hidden layers of LSTM. This helps to improve the generalization ability of the model and avoid overfitting.

[0139] After training is completed, preprocessing of new data is required. This includes data cleaning, such as removing outliers. For example, if the seawater temperature suddenly rises by 10 degrees on a certain day, this may be a measurement error and needs to be excluded. Data normalization is also important, which can unify parameters of different scales into the same range, such as normalizing both seawater temperature and salinity to between 0 and 1.

[0140] After model prediction, post-processing is required. If normalization was performed previously, denormalization is needed now. For example, if the normalized seawater temperature predicted by the model is 0.6, it needs to be converted back to the actual temperature value, such as 24°C. Finally, the prediction results can be applied to multiple fields. In terms of marine disaster warning, if the model predicts a sharp rise in seawater temperature within the next three days, it may indicate the formation of a tropical cyclone, and relevant departments can take precautions in advance. In terms of marine resource assessment, if the model predicts a continuous decline in the dissolved oxygen content in a certain sea area, it may affect fish resources, and the fishery department can adjust the fishing plan accordingly. Through this method, not only can the changes of individual marine parameters be predicted, but also the mutual influence between multiple parameters can be analyzed, so as to more comprehensively understand and predict the changes of the marine environment. This is of great significance for fields such as marine resource management and climate change research.

[0141] Furthermore, according to the non-linear mapping results, the support vector machine algorithm is used to classify abnormal states. The process of judging whether there is an abnormal event in the current window includes:

[0142] Obtain the data within the current time window, preprocess the data, remove noise and irrelevant information, and obtain a pure data set;

[0143] According to the preset non-linear mapping function, extract features from the pure data set to obtain a feature vector that can represent the abnormal state;

[0144] Input the feature vector into the pre-trained support vector machine model for classification prediction to obtain the judgment result on whether there is an abnormality in the current time window;

[0145] If the judgment result indicates the existence of an abnormality, trigger an abnormal alarm and send the relevant information of the abnormal event to the administrator for further processing;

[0146] If the judgment result indicates the non-existence of an abnormality, continue to obtain the data of the next time window and repeat the above steps to achieve continuous monitoring of the abnormal state;

[0147] According to the feedback information of the administrator, the support vector machine model is updated and optimized in real time, and the result data of the abnormal detection is stored in the database to form a historical record.

[0148] Specifically, obtaining the data within the current time window is the first step of abnormal detection. For example, in marine environmental monitoring, parameters such as water temperature, salinity, and dissolved oxygen may be collected. These raw data often contain noise and irrelevant information and need to be preprocessed. The preprocessing may include removing outliers, filling in missing values, etc. Taking the water temperature data as an example, there may be abnormally high or low readings due to sensor failures, which need to be identified and removed. The selection of the non-linear mapping function is crucial for feature extraction.

[0149] In this embodiment, in marine environmental monitoring, wavelet transform is used as the non - linear mapping function. Wavelet transform can effectively capture the time - frequency characteristics of signals and is suitable for processing non - stationary marine environmental data. Through wavelet transform, feature vectors that can characterize abnormal states can be extracted, such as abnormal energy distribution in certain frequency bands. The multi - class support vector machine (SVM) model is a powerful classifier suitable for dealing with high - dimensional feature spaces. In this embodiment of marine environmental anomaly detection, different types of abnormal states (such as red tides, oil spills, water pollution, etc.) can be regarded as different classes. The SVM model distinguishes these classes by finding the optimal hyperplane to achieve accurate classification of abnormal states. When an anomaly is detected, the system will trigger an alarm and notify the administrator. For example, if there is a possibility of detecting a red tide, the system will immediately send an alarm to the marine environmental management department, including information such as the location, time, and possible impact range of the anomaly. This enables the management department to take timely measures, such as dispatching monitoring vessels for on - site investigation or issuing early warning information.

[0150] The real - time update and optimization of the model are the keys to maintaining the efficient operation of the system. The feedback information from the administrator, such as the results of on - site investigation, can be used to adjust the model parameters. For example, if the system misreports a red tide event, this error sample can be added to the training set and the model can be retrained to improve the recognition accuracy for similar situations.

[0151] Data storage and the establishment of historical records are crucial for long - term analysis and model improvement. By analyzing historical data, seasonal patterns or long - term trends of marine environmental anomalies can be discovered. For example, it may be found that certain sea areas are more prone to red tides in specific seasons, and this information can be used to optimize the monitoring strategy and increase the monitoring frequency during high - risk periods. The operation of the entire system forms a closed - loop: from data collection, pre - processing, feature extraction, to anomaly detection, alarm triggering, administrator feedback, and then to model update and historical data analysis. This closed - loop design ensures that the system can continuously learn and improve to adapt to the complex and changing marine environment. Through continuous optimization, the system can identify potential marine environmental problems earlier and more accurately, providing strong support for marine ecological protection and resource management.

[0152] Furthermore, the process of real - time outputting the probability estimation results of secondary and derivative abnormal states through the sliding window mechanism includes:

[0153] Obtain real - time monitoring data of the marine environment as the input data source for the sliding window mechanism;

[0154] For the obtained monitoring data, use the sliding window mechanism to divide the data into several time windows;

[0155] For the monitoring data within each time window, determine whether there is an abnormal state through an anomaly detection algorithm;

[0156] If an abnormal state is detected within the current time window, then according to the pre-established probability model, estimate the probabilities of secondary anomalies and derivative anomalies occurring in the abnormal state, and use the probabilities as the estimation results of the abnormal state for the current time window;

[0157] As the sliding window continuously moves, the estimation results of the secondary probability and derivative probability of the abnormal state are updated in real time, and the updated estimation results of the abnormal state probability are output in real time.

[0158] Specifically, the real-time monitoring of the marine environment in this embodiment obtains various types of data, such as water temperature, salinity, dissolved oxygen, etc., through the deployed sensor network. These data are collected at a fixed frequency to form a continuous time series. The sliding window mechanism, as an effective method for processing time series data, divides the data stream into overlapping time periods. For example, a window can be set to 30 minutes and slide every 5 minutes, which not only ensures the continuity of the data but also can capture environmental changes in a timely manner. Within each time window, the anomaly detection algorithm plays a key role. Common methods include statistical analysis, machine learning, etc. Taking the detection of water temperature anomalies as an example, a normal temperature range model can be established using historical data, and when the observed value significantly deviates from this range, it is determined as an anomaly. For example, the average water temperature in a certain sea area in summer is 25°C, and the standard deviation is 1.5°C. If the water temperature measured at a certain moment is 30°C, which is significantly beyond the normal range, the system marks it as an abnormal state. Once an anomaly is detected, the system will evaluate the possibility of its causing secondary anomalies and derivative anomalies.

[0159] Secondary anomalies refer to subsequent anomalies directly caused by initial anomalies, while derivative anomalies are indirect chain reactions. Taking water temperature anomalies as an example, it may directly lead to the death of organisms in local sea areas (secondary anomalies), and then trigger the imbalance of the ecosystem (derivative anomalies). The probability model is constructed based on historical data and expert knowledge, considering the interactions between various factors. For example, the model may give the probability of water temperature anomalies causing organism death as 0.7, and the probability of ecosystem imbalance as 0.4. As the sliding window moves continuously, the system continuously updates these probability estimates. This dynamic update mechanism can reflect the real-time changes in the marine environment and provide timely and accurate information support for decision-makers. For example, if water temperature anomalies are detected in multiple consecutive time windows and the degree of anomalies gradually increases, the system will correspondingly increase the probability estimates of secondary and derivative anomalies, possibly rising from the initial 0.7 and 0.4 to 0.9 and 0.6. This method based on the sliding window and probability model has multiple advantages. First, it can timely capture the dynamic changes in the marine environment and not miss important anomaly events. Second, by estimating the probabilities of secondary and derivative anomalies, the system provides a more comprehensive risk assessment, which helps to prevent potential chain reactions. Finally, the real-time updated probability estimates provide valuable data support for the dynamic modeling of the marine environment, helping researchers to deeply understand the complexity and vulnerability of the marine ecosystem.

[0160] Generally speaking, this embodiment organically combines real-time data collection, anomaly detection, and risk assessment, providing strong technical support for marine environment monitoring and protection. It can not only timely discover problems but also predict potential risks, providing a scientific basis for the sustainable utilization of marine resources and the long-term protection of the ecological environment.

[0161] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed in the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for modeling and estimating secondary and derivative probabilities of abnormal marine water environment conditions, characterized in that: include: Acquire the real-time data stream of the ocean environment parameters, and use the sliding window mechanism to segment the time series information in the data stream to obtain the data within the window; Using principal component analysis to reduce the dimensionality of the high-dimensional data in the window, extracting the main characteristic components while retaining key information, and obtaining reduced-dimensional data; Constructing a time series-based graph structure model according to the dimension reduction data, capturing the correlation and spatiotemporal dependency between the marine environmental parameters based on the graph structure model, and determining the interaction intensity between the marine environmental parameters; Using radial basis function kernel to perform nonlinear mapping on the graph structure model, constructing a nonlinear mapping relationship, capturing the nonlinear characteristics between the marine environment parameters, and obtaining a nonlinear mapping result; According to the nonlinear mapping result, a support vector machine algorithm is used to classify the abnormal state and determine whether there is an abnormal event in the current window; And through the sliding window mechanism, the secondary and derivative probability estimation results of the abnormal state are output in real time; The process of using radial basis function kernel to perform nonlinear mapping on the graph structure model, constructing a nonlinear mapping relationship, capturing the nonlinear characteristics between the marine environment parameters, and obtaining the nonlinear mapping result includes: A graph structure model is constructed based on the marine environment parameter data, where nodes represent parameters and edges represent the relationship between parameters. The radial basis function kernel is used to perform nonlinear mapping on the graph structure model to capture the nonlinear characteristic relationship between the marine environmental parameters; through nonlinear mapping, the original graph structure model is mapped to a high-dimensional feature space to obtain a nonlinear mapping result; According to the nonlinear mapping results, a nonlinear relationship model between marine environmental parameters is constructed; Using a support vector machine algorithm to train the nonlinear relationship model to obtain a trained nonlinear relationship model; The trained nonlinear relationship model is used to predict new marine environmental parameter data and obtain prediction results; The prediction results are compared with the actual observed data, and after evaluating the prediction performance of the nonlinear relationship model, the model is continuously iterated and optimized.

2. The method according to claim 1, characterized in that: The real-time data stream of the ocean environment parameters is obtained, and the time series information in the data stream is processed in segments using a sliding window mechanism. The process of obtaining the data in the window includes: The data stream is segmented using a sliding window mechanism, and the data stream is divided into multiple time windows according to a preset window size and sliding step size; Extracting time series information from the data in each time window, wherein the time series information includes the change trend characteristics and value range characteristics of each marine environmental parameter in the current window; Based on the time series information, a time series analysis algorithm is used to model and predict the data in the window to obtain prediction results of the marine environmental parameters; Aggregate and merge the prediction results in each time window to obtain a comprehensive prediction result for the entire data stream; The prediction results are compared with the preset thresholds. If the prediction results exceed the normal range, the early warning mechanism is triggered and an alarm message is sent to relevant departments to indicate possible abnormal marine environment conditions. Continuously monitor the real-time data stream, move the sliding window forward continuously, analyze and predict the data in the new time window, and dynamically adjust the size and sliding step of the sliding window based on historical data and prediction results.

3. The method according to claim 1, characterized in that: The principal component analysis method is used to reduce the dimension of the high-dimensional data in the window, extract the main characteristic components while retaining the key information, and the process of obtaining the reduced-dimensional data includes: Acquire high-dimensional feature information of the data in the window, construct a principal component analysis model, and perform dimensionality reduction processing on the high-dimensional data through the principal component analysis model; Through the principal component analysis method, the high-dimensional data in the window is decomposed to extract the main characteristic components of the data, and the number of principal components is determined according to the size of the eigenvalues; According to the number of principal components, the original high-dimensional data is subjected to dimensionality reduction mapping, and the data is mapped from the high-dimensional space to the low-dimensional space. During the dimensionality reduction process, the key information and characteristic information of the data are retained by retaining the principal components; Reconstruct the data after dimensionality reduction to obtain reconstructed data with the same dimension as the original data, and compare it with the original data to evaluate the degree of information retention of the dimensionality reduction process; According to the effect of dimensionality reduction, adjust and optimize the parameters of the principal component analysis model, adjust the number of principal components and optimize the data preprocessing method; The optimized principal component analysis model is applied to the adjusted window data to obtain the final dimensionality reduction processing result.

4. The method according to claim 1, characterized in that: Based on the graph structure model capturing the correlation and spatiotemporal dependency between the marine environmental parameters, the process of determining the interaction strength between the marine environmental parameters includes: Based on the reduced-dimensional marine environmental parameter data, a time series graph structure model is constructed, in which nodes represent marine environmental parameters and edges represent the correlation and spatiotemporal dependency between parameters. The time series graph structure model is trained using a graph neural network to learn and capture the complex interaction relationship between marine environmental parameters; During the training process, the attention mechanism is introduced to dynamically adjust the connection weights between nodes at different time steps and spatial positions to highlight the mutual influence between key parameters; Inference is performed on the trained graph neural network model to obtain the interaction strength between the ocean environment parameters at each time step and spatial position; According to the interaction strength, a causal relationship diagram between the marine environmental parameters is constructed to obtain the causal dependency and action path between the parameters; For the key nodes and paths in the causal graph, a graph attention network is used to further mine the high-order interaction patterns and long-range dependencies between parameters; The prediction results of graph neural network and graph attention network are integrated to judge the interaction strength between different marine environmental parameters.

5. The method according to claim 4, characterized in that The process of inferring the trained graph neural network model to obtain the interaction strength between the ocean environment parameters at each time step and spatial position includes: According to the spatiotemporal distribution of marine environmental parameters, the input data structure of the graph neural network model is constructed, where the node represents each spatiotemporal position and the edge represents the connection relationship between adjacent spatiotemporal positions; The input data is fed into a pre-trained graph neural network model, the hidden state vector of each node is calculated through forward propagation, and the information of adjacent nodes is aggregated through an attention mechanism; In the output layer of the graph neural network, a fully connected layer is used to map the hidden state vector of each node to the predicted value of the ocean environment parameter at the corresponding spatiotemporal location; Calculate the difference between the predicted value of each spatiotemporal position and the predicted value of the adjacent spatiotemporal position as the interaction strength between the spatiotemporal position and the adjacent position; According to the magnitude of the interaction intensity, the degree of mutual influence between different marine environmental parameters is judged, and the greater the intensity, the stronger the correlation between the two parameters; The interaction intensity matrix is ​​normalized to obtain the relative interaction intensity distribution under different combinations of spatiotemporal positions and environmental parameters; Visualize the normalized interaction intensity matrix to generate a heat map or network diagram.

6. The method according to claim 1, characterized in that The trained nonlinear relationship model is used to predict new marine environmental parameter data. The process of obtaining the prediction results includes: Based on historical parameter data related to the marine environment, a nonlinear relationship prediction model is constructed using a machine learning algorithm; During the model training process, the model hyperparameters are optimized through the cross-validation method, and the trained nonlinear relationship model is saved; Obtain new marine environmental parameter data, perform data cleaning and normalization preprocessing operations on the data; input the preprocessed new data into the trained nonlinear relationship model, and the model automatically performs predictive analysis and calculation; Obtain the prediction result data from the model output and perform post-processing of denormalization and data format conversion on the prediction results.

7. The method according to claim 1, characterized in that According to the nonlinear mapping result, the support vector machine algorithm is used to classify the abnormal state, and the process of judging whether there is an abnormal event in the current window includes: Get the data in the current time window, pre-process the data, remove noise and irrelevant information, and obtain a pure data set; According to a preset nonlinear mapping function, feature extraction is performed on the clean data set to obtain a feature vector capable of characterizing an abnormal state; The feature vector is input into a pre-trained support vector machine model for classification prediction to obtain a judgment result of whether there is an abnormality in the current time window; If the judgment result shows that there is an abnormality, an abnormal alarm is triggered and the relevant information of the abnormal event is sent to the administrator for further processing; If the judgment result shows that there is no abnormality, continue to obtain data of the next time window and repeat the above steps to achieve continuous monitoring of the abnormal state; Based on the administrator's feedback, the support vector machine model is updated and optimized in real time, and the anomaly detection result data is stored in the database to form a historical record.

8. The method according to claim 1, characterized in that The process of outputting secondary and derivative probability estimation results of abnormal states in real time through the sliding window mechanism includes: Obtain real-time monitoring data of the marine environment as the input data source of the sliding window mechanism; For the acquired monitoring data, a sliding window mechanism is used to divide the data into several time windows; For the monitoring data in each time window, anomaly detection algorithms are used to determine whether there is an abnormal state; If an abnormal state is detected in the current time window, the probability of secondary abnormalities and derivative abnormalities occurring in the abnormal state is estimated according to a pre-established probability model, and the probability is used as the abnormal state estimation result of the current time window; As the sliding window continues to move, the estimated results of the secondary probability and the derived probability of the abnormal state are updated in real time, and the updated abnormal state probability estimation results are output in real time.

Images