Marine AI analysis and marine disaster prediction method and system based on multi-source heterogeneous data

Abstract

The invention relates to a marine AI analysis and marine disaster prediction method and system based on multi-source heterogeneous data. The method comprises the following steps: performing data matrix construction and information amount calculation on ocean monitoring data to obtain a monitoring index weight matrix, and performing inter-sample similarity calculation and neighborhood radius calculation on the monitoring index weight matrix to obtain a dimension reduction feature data set; inputting the dimension reduction feature data set into a three-layer neural network structure, and performing parallel training on each monitoring point to obtain a distributed prediction model and a training output result; establishing an n-level state space according to a training output result, and obtaining a layered early warning threshold set through Markov decision; and inputting the real-time monitoring feature data into the distributed prediction model for abnormal state prediction to obtain an abnormal state prediction result, comparing the abnormal state prediction result with the layered early warning threshold set, and outputting an early warning level signal. According to the method, the accuracy of ocean AI analysis and ocean disaster prediction is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-source heterogeneous data processing, and in particular to a method and system for marine AI analysis and marine disaster prediction based on multi-source heterogeneous data. Background Art

[0002] The multi-source heterogeneous data generated by the marine environment monitoring system has the characteristics of high dimensionality, multiple types, and strong real-time performance. Traditional data analysis methods are difficult to effectively handle these complex data features. Existing marine disaster prediction methods mainly rely on a single data source for analysis and cannot fully utilize the rich information contained in multi-source data, resulting in insufficient prediction accuracy and untimely early warning responses.

[0003] The main challenges facing the current analysis of marine monitoring data are how to extract key features from massive heterogeneous data and how to build accurate prediction models. Although deep learning methods have made significant progress in the field of data analysis, there are still problems such as high computational complexity and weak model generalization ability when processing multi-source heterogeneous data in the ocean, which makes it difficult to meet the needs of real-time warning. In view of the rapid spread of marine disasters, the real-time warning system needs to consider both prediction accuracy and response timeliness. Existing warning methods often use fixed thresholds for judgment, lack the ability to adapt to dynamic changes in the marine environment, and are prone to omissions or false alarms. In addition, different levels of marine disasters require different warning strategies. How to establish a reasonable multi-level warning mechanism is also an urgent problem to be solved. Summary of the invention

[0004] The main purpose of the present invention is to provide a method and system for ocean AI analysis and ocean disaster prediction based on multi-source heterogeneous data. The present invention improves the accuracy of ocean AI analysis and ocean disaster prediction.

[0005] To achieve the above objectives, the present invention provides a method for ocean AI analysis and ocean disaster prediction based on multi-source heterogeneous data, comprising the following steps: Constructing a data matrix and calculating the amount of information on the marine monitoring data to obtain a monitoring indicator weight matrix, and calculating the similarity between samples and the neighborhood radius on the monitoring indicator weight matrix to obtain a reduced-dimensional feature data set; The dimension-reduced feature data set is input into a three-layer neural network structure, and each monitoring point is trained in parallel to obtain a distributed prediction model and training output results; Establishing an n-level state space according to the training output results, and obtaining a hierarchical warning threshold set through Markov decision making; The real-time monitoring feature data is input into the distributed prediction model to predict the abnormal state, and the abnormal state prediction result is obtained. The abnormal state prediction result is compared with the hierarchical warning threshold set to output a warning level signal.

[0006] The present invention also provides an ocean AI analysis and ocean disaster prediction system based on multi-source heterogeneous data, comprising: A calculation module is used to construct a data matrix and calculate the amount of information for the marine monitoring data to obtain a monitoring indicator weight matrix, and to calculate the similarity between samples and the neighborhood radius of the monitoring indicator weight matrix to obtain a reduced dimension feature data set; A parallel training module is used to input the dimension reduction feature data set into a three-layer neural network structure, perform parallel training on each monitoring point, and obtain a distributed prediction model and training output results; A decision module, used to establish an n-level state space according to the training output results, and obtain a hierarchical warning threshold set through Markov decision making; The output module is used to input the real-time monitoring feature data into the distributed prediction model to predict the abnormal state, obtain the abnormal state prediction result, compare the abnormal state prediction result with the hierarchical warning threshold set, and output the warning level signal.

[0007] In summary, the technical solution provided by the present invention effectively reduces the data dimension, retains key feature information, and reduces computing resource consumption through a feature extraction method that combines the Critic weighting method with the neighborhood rough set; adopts a distributed neural network structure for parallel training, reduces the model training time, and improves the system's ability to process large-scale data; a multi-level early warning mechanism based on the Markov decision process enables the system to dynamically adjust the early warning strategy according to different degrees of abnormality, thereby reducing the false alarm rate; by updating the state transition probability and the early warning threshold in real time, the system's adaptability to dynamic changes in the marine environment is enhanced; the hierarchical early warning threshold structure is adopted to make the early warning decision more targeted, facilitate timely adoption of corresponding disaster prevention and mitigation measures, and thereby improve the accuracy of marine AI analysis and marine disaster prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 This is a schematic diagram of the steps of a method for ocean AI analysis and ocean disaster prediction based on multi-source heterogeneous data in one embodiment of the present invention; Figure 2 It is a structural block diagram of an ocean AI analysis and ocean disaster prediction system based on multi-source heterogeneous data in one embodiment of the present invention.

[0009] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0010] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0011] Reference Figure 1 , this embodiment provides an ocean AI analysis and ocean disaster prediction method based on multi-source heterogeneous data, including the following steps: S1, constructing a data matrix and calculating the amount of information for the marine monitoring data to obtain a monitoring indicator weight matrix, and calculating the similarity between samples and the neighborhood radius of the monitoring indicator weight matrix to obtain a reduced dimension feature data set; Among them, by collecting marine monitoring data, including key indicators such as temperature, salinity, dissolved oxygen, pH value and chlorophyll concentration, and sorting and arranging these data in the order of collection time, an initial data matrix is ​​formed, in which each row represents the sampling data at different time points, and each column corresponds to a specific monitoring indicator. In order to avoid the deviation of subsequent analysis due to different dimensions of different monitoring indicators, the initial data is standardized, and all data are mapped into a unified numerical range to obtain a standardized data matrix. The data volatility of each monitoring indicator is calculated. By analyzing the fluctuation amplitude of the monitoring data, the discrete degree of each indicator is evaluated, that is, their changes in time and space. The size of the discrete degree is related to the sensitivity of the indicator to the overall system state. The correlation between the monitoring indicators is analyzed to evaluate the dependency between different indicators. By comparing the data change trends of each two monitoring indicators, the correlation between them is calculated. Combining the volatility and correlation analysis results of the monitoring indicators, the comprehensive information content of each monitoring indicator is calculated. The higher the information content of an indicator, the more it indicates that the indicator has significant volatility and is relatively independent of other indicators. This information content can quantify the importance of each indicator in the overall system. Normalize the comprehensive information of each monitoring indicator to form a weight matrix. Each element of the weight matrix represents the weight ratio of the corresponding monitoring indicator in the overall analysis. The weight distribution can highlight the most important indicators while weakening the secondary factors that have little impact on the system state. Calculate the similarity between samples based on the weight matrix. Similarity analysis compares monitoring data at different time points or spatial locations to evaluate their closeness in comprehensive status. Based on the similarity between samples, a clustering algorithm is applied to determine the neighborhood radius of data distribution, achieve feature dimensionality reduction of the original data, and generate a reduced dimensionality feature data set.

[0012] Each weight value in the weight matrix of the monitoring indicator is weighted with the data of the corresponding monitoring indicator. The weighted feature vector is generated by multiplying the original data of each monitoring indicator with its weight in the weight matrix item by item. The Euclidean distance is calculated for each pair of sample data in the weighted feature vector to evaluate the similarity between samples and generate a sample distance matrix, in which each element represents the distance between two groups of samples in the feature space. The sample distance matrix reflects the degree of difference between different samples. The initial similarity data is normalized to standardize all distance values ​​within a fixed range to obtain the standardized similarity value. The neighborhood radius parameter is set according to the data distribution characteristics of the standardized similarity value. In order to ensure that the selection of the neighborhood radius can reflect the true distribution characteristics of the data, the interquartile range method is used to determine the optimal neighborhood radius value. The interquartile range method extracts the range of data variation by analyzing the median and upper and lower quartiles of the statistical data distribution, thereby setting a reasonable neighborhood radius value that can adapt to the data density. This method is suitable for processing scenarios where the sample distribution has a certain degree of non-uniformity, which helps to improve the accuracy of the neighborhood relationship. The neighborhood relationship between samples is constructed based on the standardized similarity value and the optimal neighborhood radius value. If the similarity value between two samples is less than or equal to the neighborhood radius value, the two samples are considered to belong to the same neighborhood, thus establishing a neighborhood relationship. Through this step, a neighborhood relationship set is generated, which contains all sample pairs that meet the neighborhood conditions. The neighborhood relationship set describes the distribution pattern of data samples in the feature space and their local correlation. The neighborhood relationship set is converted into a binary matrix representation to construct a neighborhood decision matrix. The neighborhood decision matrix is ​​a symmetric matrix in which each element only takes a binary representation. If there is a neighborhood relationship between two samples, the value of the corresponding element of the matrix is ​​1; otherwise, it is 0. Through the neighborhood decision matrix, the neighborhood relationship between samples is described comprehensively and compactly in matrix form. Based on the neighborhood decision matrix, the dependence of the conditional attribute set on the decision attribute set is calculated to generate a reduced dimensionality feature data set. The calculation process of the dependence reflects the importance of the conditional attribute set to the decision attribute set. By screening features with high dependence, redundant information is effectively reduced, and the data expression ability and processing efficiency are improved.

[0013] The upper approximate set and the lower approximate set between samples are calculated according to the neighborhood decision matrix. The division of these two sets is based on the conditional attributes, which respectively represent the range that the samples may belong to a certain category and the range that they clearly belong to a certain category. Through the upper approximate set and the lower approximate set, the conditional attribute space is further divided to form the partition structure of the conditional attribute, revealing the similarities and differences between samples. The positive domain of the conditional attribute space is calculated, and by identifying which samples can be clearly classified into the decision attribute space, the dependency value of the conditional attribute set C on the decision attribute set D is obtained. The dependency value represents the contribution of the conditional attribute set to the target decision classification. The higher the value, the stronger the explanatory power of the conditional attribute for the decision attribute. The subset S of the conditional attribute set C is traversed according to the forward search strategy. The dependency value of each feature subset on the decision attribute set is calculated separately to obtain the contribution of each feature to the decision classification and generate a feature importance sequence. The feature importance sequence is sorted according to the contribution of the feature to the decision. Based on the feature importance sequence, the dependency gain threshold interval is set to screen out the optimal feature subset. The interval setting of the dependency gain threshold is based on the distribution of the feature importance sequence, and equidistant sampling is performed within the interval to generate a candidate threshold set. Each threshold in the candidate threshold set represents a potential screening criterion, and its specific role is to determine which features can be retained in the dimension reduction feature data set. Cross-validation is performed on each threshold in the candidate threshold set, and the pros and cons of each threshold are evaluated by calculating the classification accuracy of feature subsets under different thresholds. In the cross-validation process, the classification accuracy is used as a metric to judge the impact of feature screening on model performance. By comparing the classification accuracy corresponding to all thresholds, the threshold that can maximize the classification accuracy is selected as the optimal threshold. Features are screened according to the optimal threshold. During the screening process, features with feature importance greater than the optimal threshold are retained, and other features are discarded. The screened features are combined into a dimension reduction feature data set.

[0014] S2, input the dimension-reduced feature data set into the three-layer neural network structure, conduct parallel training on each monitoring point, and obtain the distributed prediction model and training output results; Specifically, the feature data set after dimensionality reduction is divided into a training data set and a validation data set to ensure that the generalization performance of the model can be evaluated through the validation set during the training process. The training data set is used to update the model parameters, while the validation data set is used to detect whether the model is overfitting or under-optimized. After the data is divided, the training data set is randomly sampled in batches, and the data is split into multiple small batches to generate neural network training batches. This method helps to improve training efficiency and avoid bias introduced by data order problems. A three-layer neural network structure consisting of an input layer, a hidden layer, and an output layer is constructed. In this structure, the number of nodes in the input layer is set to the feature dimension of the dimensionality reduction feature data set to ensure that the input data can be fully received. The number of nodes in the hidden layer is set to 2e+1, where e is the feature dimension of the input data. This design enhances the ability to capture complex patterns by increasing the nonlinear expression ability of the network. The number of nodes in the output layer is fixed to 1, which is used to predict the value of the target variable. The input layer uses linear transformation to process the input data and maps it to the hidden layer; the hidden layer uses the ReLU activation function for nonlinear transformation, which introduces nonlinear characteristics and reduces the gradient vanishing problem; and the output layer uses the Sigmoid activation function to compress the predicted value to the [0,1] interval for probabilistic output or classification tasks. After the network structure is built, the connection weights of the three-layer neural network are initialized. The weight initialization uses a random method to avoid the network falling into a symmetric state or gradient vanishing. The Adam optimizer is used to optimize the parameters of the neural network. The Adam optimizer achieves fast convergence by combining momentum and adaptive learning rate adjustment techniques, and performs well in dealing with sparse gradient problems. During the training process, the training batch data is input into the neural network, and the back propagation algorithm is used to calculate the gradient of the loss function relative to each network parameter. The weight and bias parameters of the network are updated based on this gradient information, so that the model gradually approaches the optimal state. In order to ensure that the model can avoid overfitting, after each training cycle, the validation data set is input into the updated neural network structure, and the loss value of the validation set is calculated. This process is used to dynamically evaluate the performance of the model on unseen data, thereby providing a basis for triggering the early stopping mechanism. When the validation set loss value fails to decrease further within G consecutive training cycles, the model is considered to have reached a performance bottleneck. At this time, the early stopping mechanism is automatically triggered to stop training to avoid overfitting problems. Through this process, an optimized distributed prediction model and the corresponding training output results are obtained.

[0015] S3, establish an n-level state space based on the training output results, and obtain a hierarchical warning threshold set through Markov decision making; It should be noted that, according to the training output results of the distributed prediction model, the abnormal degree of the ocean state is divided into levels. The basis for the division is the range of abnormal values ​​output by the model or the actual severity of the abnormal event. For example, the abnormal degree is divided into n levels such as low abnormality, moderate abnormality and high abnormality. Each level corresponds to a clear state identifier, and an n-level state space is established to represent the ocean state under different abnormal levels. After the state space is established, the changes between different states are dynamically analyzed, and the transition between states is recorded by counting the frequency of state changes at adjacent moments. At each time point, the change between the current state and the state at the next moment is recorded as a state transition. By accumulating these transition times, the total number of transitions of each state to other states is obtained. The transition ratio of each state is calculated based on the number of transitions to form the transition probability data between states. The state transition probability data reflects the possibility of the system transitioning from one state to another. The transition probability data is organized into a matrix form to generate a state transition probability matrix. The rows of the matrix represent the current state, the columns represent the possible states at the next moment, and each element in the matrix represents the probability value of the corresponding state transition. The matrix structure shows the dynamic evolution law of the system between different states. Set up a reward mechanism. By analyzing historical data or actual needs, reward scores are assigned to the warning performance of different states. For states that can accurately and timely trigger warnings, positive rewards are given to encourage such behavior; for states that produce delayed warnings or false alarms, negative rewards are given to reflect their costs. These reward scores constitute the state reward function, which guides the Markov decision algorithm to optimize the response behavior of the warning system. The state transition probability matrix and the state reward function are input into the Markov decision algorithm, and the decision value of each state is calculated for multiple rounds of iterations. The Markov decision algorithm gradually optimizes the value of each state by seeking a balance between state transitions and reward scores until these values ​​converge and stabilize. The converged state value reflects the relative importance and optimal behavior choice of each state in the overall warning decision. According to the converged optimal decision value, the corresponding warning threshold is set for each state. The setting of the warning threshold is based on the abnormality of the state, and is arranged in order from low to high to ensure that the warning system can gradually upgrade the response intensity. The hierarchical warning threshold set effectively guides the warning system to trigger the corresponding response level when facing different degrees of abnormality through a clear hierarchical structure. The resulting hierarchical warning threshold set covers all levels from low anomalies to high anomalies, and takes into account the accuracy of the warning and the timeliness of the response.

[0016] S4, input the real-time monitoring feature data into the distributed prediction model to predict the abnormal state, obtain the abnormal state prediction result, and compare the abnormal state prediction result with the hierarchical warning threshold set to output the warning level signal.

[0017] Specifically, feature extraction is performed on the real-time monitoring feature data, and the feature dimension is reduced according to the same feature dimension as the dimension reduction feature data set to obtain a real-time feature vector. The real-time feature vector is input into the distributed prediction model. The distributed prediction model uses the parameters optimized in the training process to perform forward propagation calculation on the input data through its three-layer neural network structure to generate an abnormal state prediction value. The prediction value is a quantitative assessment of the current marine monitoring state, indicating the degree of abnormality of the current environment, and the range is normalized to the [0,1] interval. The abnormal state prediction value is compared with the thresholds of each level defined in the hierarchical warning threshold set, and the abnormal level to which the current state belongs is determined by judging the threshold interval range in which the prediction value falls. The abnormal level is divided from low to high, including low abnormality, moderate abnormality and high abnormality, etc., reflecting the severity of the current state. Through this step, the abnormal state prediction result is generated. The state transition probability matrix is ​​updated according to the abnormal state prediction result. The state transition data is dynamically updated by recording the number of transitions from the current state to the next state and performing statistical analysis on these transition data. The update can reflect the latest real-time state changes, so that the state transition probability matrix can adapt to the new monitoring data in real time, thereby more accurately describing the dynamic behavior of the system. The updated state transition data is input into the Markov decision process. The Markov decision algorithm recalculates the state value function of each state based on the new state transition probability matrix and the previously defined state reward function. Through multiple rounds of iterative calculations, the state value function finally converges to a stable optimal state value. Based on the newly calculated state value, the hierarchical warning threshold set is dynamically adjusted to adapt to changes in the current environment, and an updated hierarchical warning threshold set is obtained. The corresponding warning level signal is generated according to the abnormal level corresponding to the abnormal state prediction result. For example, a low abnormality triggers a green signal to remind the system that it is in a safe state; a moderate abnormality triggers a yellow signal to indicate potential risks; a high abnormality triggers a red signal to warn that there is a major risk and an immediate response is required.

[0018] In one example, the data matrix is ​​constructed and the information volume is calculated for the marine monitoring data to obtain the monitoring indicator weight matrix, and the similarity between samples and the neighborhood radius are calculated for the monitoring indicator weight matrix to obtain a reduced dimension feature data set, including: The monitoring data of ocean temperature, salinity, dissolved oxygen, pH value, and chlorophyll are arranged in the order of collection time, and an initial data matrix with h rows of samples and f columns of features is established for the monitoring data; Perform maximum and minimum value normalization operations on the data in the initial data matrix to obtain a standardized data matrix, and calculate the standard deviation of the standardized data matrix according to the monitoring indicator column, perform a discrete degree analysis on the data samples of each monitoring indicator, and obtain the standard deviation data of f monitoring indicators; The standardized data matrix is ​​paired with the monitoring indicators for correlation analysis, and the Pearson correlation coefficient between the two monitoring indicators is calculated to obtain the f-order correlation coefficient matrix; The information content of each monitoring indicator is calculated based on the standard deviation data and the correlation coefficient matrix, and the standard deviation of each monitoring indicator is multiplied by the linear combination value of the correlation coefficient of the monitoring indicator relative to other monitoring indicators to obtain the monitoring indicator information content data; Normalize the monitoring indicator information data, divide the information of each monitoring indicator by the sum of the information of all monitoring indicators, and obtain the monitoring indicator weight matrix; The similarity between samples and the neighborhood radius are calculated for the monitoring indicator weight matrix to generate a reduced-dimensional feature data set.

[0019] In this example, the monitoring data of ocean temperature, salinity, dissolved oxygen, pH value and chlorophyll are arranged in the order of collection time. These data are constructed into an initial data matrix. Assume that the number of monitoring samples is , the number of monitoring indicators is , then the initial data matrix The dimension is The elements in the matrix Indicates The sample in In order to eliminate the influence of the numerical range of different monitoring indicators on subsequent analysis, the initial data matrix Perform the maximum and minimum value normalization operation. The normalization formula is: ; in, is the original data value, and They are The minimum and maximum values ​​in the column (i.e. monitoring indicator), is the standardized data value. Map all data to the interval [0,1], and the standardized matrix is ​​recorded as . To standardize the data matrix Calculate the standard deviation of each monitoring indicator in. The standard deviation is an important indicator to measure the degree of data dispersion, and its calculation formula is: ; in, It is The standard deviation of the monitoring indicators, It is The mean of the column indicates the average value of the monitoring indicator. is the sample size. Through this step, we get The standard deviation data of the monitoring indicators reflects the volatility of each indicator. The correlation analysis between the monitoring indicators is carried out. Pair the monitoring indicators in two pairs and calculate the Pearson correlation coefficient between them. The calculation formula of the correlation coefficient is: ; in, It is and The correlation coefficient of the monitoring indicators is in the range of [-1,1]. The closer to 1, the stronger the positive correlation, and the closer to -1, the stronger the negative correlation. By calculating all possible pairs, a The correlation coefficient matrix ,in Based on standard deviation data And the correlation coefficient matrix , calculate the information amount of each monitoring indicator. The calculation formula of information amount is: ; in, Indicates The amount of information of each monitoring indicator, is the standard deviation of the indicator, is the correlation between this indicator and other monitoring indicators, It represents the average correlation between this indicator and all other indicators. Through this formula, the information volume comprehensively considers the volatility and independence of the indicator. In order to make the information volume comparable, the information volume data is normalized to obtain the monitoring indicator weight matrix The normalization formula is: ; in, It is The weight of the monitoring indicators, the normalized weight matrix Reflects the relative importance of each indicator to the system status. Using the weight matrix , calculate the similarity between samples. The similarity between samples is expressed by weighted Euclidean distance, and its formula is: ; in, It is a sample and samples Based on these distances, the neighborhood radius is calculated by a density clustering algorithm (such as DBSCAN) and a reduced-dimensional feature dataset is generated.

[0020] In one example, the similarity between samples and the neighborhood radius are calculated for the monitoring indicator weight matrix to generate a reduced dimension feature data set, including: The weight values ​​in the monitoring indicator weight matrix are weighted and calculated with the data of the corresponding monitoring indicators to obtain a weighted feature vector; The Euclidean distances of the sample data in the weighted feature vector are calculated pairwise, and the sample distance matrix is ​​constructed to obtain the initial similarity data, and the initial similarity data is normalized to obtain the standardized similarity value; The neighborhood radius parameter is set based on the data distribution characteristics of the standardized similarity value, and the interquartile range method is used to determine the optimal neighborhood radius value; Construct sample neighborhood relationships based on the standardized similarity value and the optimal neighborhood radius value. If the similarity value between samples is less than or equal to the neighborhood radius value, a neighborhood relationship is established to obtain a neighborhood relationship set. The neighborhood relationship set is converted into a binary matrix representation, and a neighborhood decision matrix representing the neighborhood relationship between samples is constructed; The dependence of the condition attribute set on the decision attribute set is calculated based on the neighborhood decision matrix to obtain a reduced-dimensional feature data set.

[0021] In this example, the weight values ​​in the monitoring indicator weight matrix are weighted with the corresponding monitoring indicator data to obtain the weighted feature vector. Suppose the monitoring indicator weight matrix is ,in Indicates The weight of the monitoring indicators, and Assume that the standardized monitoring data matrix is ,in Indicates The sample in The value under the monitoring indicator, the matrix dimension is ,in is the sample size, is the number of monitoring indicators. The calculation formula of weighted eigenvector is: ; in, It is The weighted feature vector value of a sample reflects the comprehensive characteristics of the sample on all monitoring indicators. After calculating all samples, the weighted feature vector set is obtained. Based on the weighted feature vector set , calculate the Euclidean distance between sample data pairs, and thus construct a sample distance matrix. and samples The Euclidean distance calculation formula between them is: ; in, Representation sample and samples The weighted Euclidean distance of , reflects the similarity or difference between the two in the feature space. By calculating the distance of all sample pairs, a symmetric sample distance matrix is ​​constructed , where the matrix dimensions are , each element Represents the corresponding distance between samples. For the sample distance matrix Normalization is performed to ensure that the distance value is within a uniform range, which is convenient for subsequent similarity calculation. The normalization formula is: ; in, is the normalized similarity value, with a value range of [0,1]. The normalized matrix Represents the standardized similarity value matrix, reflecting the similarity relationship between samples. Based on the distribution characteristics of the standardized similarity value matrix, the neighborhood radius parameter is set to determine the neighborhood relationship between samples. The interquartile range method is used to determine the optimal neighborhood radius value. The upper and lower quartiles of the similarity value are calculated. and , and the interquartile range : ; According to the interquartile range method, the optimal neighborhood radius is set to: ; in, Represents the neighborhood radius value, which is used to define the neighborhood relationship between samples. and the normalized similarity value , construct a neighborhood relationship set between samples. and The similarity value of Less than or equal to , then it is considered that there is a neighborhood relationship between them. The neighborhood relationship set is expressed as: ; The neighborhood relationship set Converted into a binary matrix ,in Representation sample and There is a neighborhood relationship Indicates that there is no neighborhood relationship. The constructed binary matrix That is the neighborhood decision matrix. After obtaining the neighborhood decision matrix, calculate the dependence of the condition attribute set on the decision attribute set. Suppose the condition attribute set is , the decision attribute set is , the definition of dependency is: ; in, represents the dependency of the conditional attribute set on the decision attribute set, is the number of positive domain samples of the decision attribute on the conditional attribute, is the total number of samples. The dependency value reflects the importance of the conditional attribute set to the classification of the decision attribute. Through the above steps, a dimension reduction feature data set is generated.

[0022] In one example, the dependency of the condition attribute set on the decision attribute set is calculated based on the neighborhood decision matrix to obtain a reduced dimension feature data set, including: According to the neighborhood decision matrix, the upper approximate set and the lower approximate set between samples are calculated to obtain the conditional attribute space partition, and the positive domain calculation is performed on the conditional attribute space partition to obtain the dependency value of the conditional attribute set C on the decision attribute set D; The subset S of the condition attribute set C is traversed according to the forward search strategy, and the dependency value of each feature subset on the decision attribute set is calculated to obtain the feature importance sequence; Based on the feature importance sequence, the dependency gain threshold interval is set, and equidistant sampling is performed within the interval to obtain a candidate threshold set; Each threshold in the candidate threshold set is cross-validated, and the classification accuracy of the feature subset under different thresholds is calculated to obtain the optimal threshold. The features are screened according to the optimal threshold, and the features with importance greater than the threshold are combined into a dimensionality reduction feature data set.

[0023] In this example, the upper and lower approximate sets between samples are calculated based on the neighborhood decision matrix, thereby dividing the conditional attribute space and completing the positive domain calculation. Suppose the neighborhood decision matrix is ,in Representation sample and samples Whether there is a neighborhood relationship (1 means yes, 0 means no). Assume that the sample set is , each sample has a decision attribute and the lower approximate set is defined as follows: ; ; in, Representation and Sample A set of samples with the same decision attributes. By dividing the conditional attribute space into upper approximation and lower approximation, After completing the upper approximation and lower approximation calculations, the positive domain set The size of the conditional attribute set determines For decision attribute set The dependency formula is defined as: ; in, is the number of samples in the positive domain set, is the total number of samples. Dependence value The larger the value, the more conditional attribute set The higher the importance of the decision attribute classification, the higher the importance of the condition attribute set. All subsets of Traverse according to the forward search strategy and calculate the dependency value of each subset on the decision attribute set. Assume , by stepwise selection of subsets through forward search , calculate their dependency values ​​in turn: ; in, Is a subset The size of the positive domain set. The dependency values ​​of all subsets form a feature importance sequence, and each value in the sequence reflects the contribution of the corresponding feature subset to the decision attribute classification. After obtaining the feature importance sequence, set the dependency gain threshold interval to select appropriate features from the sequence. Set the maximum value of the feature importance sequence to be , the minimum value is , then the threshold interval is defined as: ; Generate a set of candidate thresholds by equidistant sampling within this interval. Each threshold Represents a feature screening criterion, selecting the importance greater than To evaluate the effect of each candidate threshold, the cross-validation method is used to calculate the classification accuracy of feature subsets under different thresholds. Cross-validation is divided into the following steps: divide the data into training set and validation set, and Filter feature subsets , and then use the training set to train the model, and use the validation set to calculate the classification accuracy. Suppose the classification accuracy is , by comparing all candidate thresholds value, select the threshold that gives the highest classification accuracy As the optimal threshold. , filter out the values ​​with importance greater than A subset of features , and use it as a dimension reduction feature dataset.

[0024] After the features are screened according to the optimal threshold to obtain a reduced dimension feature data set and before the reduced dimension feature data set is divided into a training data set and a validation data set, the method also includes: inputting the reduced dimension feature data set into a principal component analysis model, calculating a feature covariance matrix and a feature vector matrix, performing a linear transformation on the reduced dimension feature data set to obtain an initial projection space coordinate; calculating a set of k nearest neighbor points for each data point in the initial projection space coordinate, constructing a local weight matrix based on the Euclidean distance between the data point and its nearest neighbor point, and using a Gaussian kernel function to calculate the similarity value between pairs of data points to obtain a local structural feature matrix; calculating the intra-class scattering and inter-class scattering for the initial projection space coordinates, minimizing the projection distance of intra-class data points and maximizing the projection distance of inter-class data points, constructing a global scattering matrix, and obtaining a global structural feature matrix; constructing a dual-constraint objective function based on the local structural feature matrix and the global structural feature matrix, setting local structure preservation weight parameters α and global ... The local structure maintains the weight parameter β, and the objective function is solved by the alternating optimization algorithm to obtain the optimal projection direction; the importance score of the projection coordinates is calculated according to the optimal projection direction, and the corresponding re-weighting coefficient is assigned to each dimension of the initial projection space coordinates, the weight coefficient of the important coordinates is set to a value greater than 1, and the weight coefficient of the redundant coordinates is set to a value less than 1, so as to obtain the coordinate weighting coefficient matrix; the coordinate weighting coefficient matrix is ​​matrix multiplied with the initial projection space coordinates to obtain the re-weighted projection subspace; a similarity graph is constructed for the re-weighted projection subspace, the eigenvalues ​​and eigenvectors of the graph Laplacian matrix are calculated, and the K-means algorithm is used to cluster the eigenvectors to obtain the subspace clustering results; the subspace clustering results are used as new feature dimensions, and feature splicing is performed with the dimensionality reduction feature data set, and the spliced ​​feature matrix is ​​normalized to obtain the enhanced feature data set, which is used for subsequent neural network training.

[0025] In one example, the reduced dimension feature data set is input into a three-layer neural network structure, and each monitoring point is trained in parallel to obtain a distributed prediction model and training output results, including: The dimension-reduced feature data set is divided into a training data set and a validation data set, and the training data set is randomly sampled in batches to obtain a neural network training batch; Construct a three-layer neural network structure, which includes an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is equal to the feature dimension of the reduced feature data set, the number of nodes in the hidden layer is 2e+1, and the number of nodes in the output layer is 1. The input layer uses linear transformation to process input data, the hidden layer uses ReLU activation function for nonlinear transformation, and the output layer uses Sigmoid activation function to map the output value to the [0,1] interval to obtain the prediction result. Initialize the connection weights in the three-layer neural network structure, use the Adam optimizer to optimize the parameters, input the neural network training batch data into the three-layer neural network structure, use the back propagation algorithm to calculate the gradient, and update the network parameters; After each training cycle, the validation data set is input into the updated neural network for verification. When the validation set loss value does not decrease for G consecutive cycles, the early stopping mechanism is triggered to obtain the distributed prediction model and training output results.

[0026] In this example, the feature dataset after dimensionality reduction is divided into a training dataset and a validation dataset to ensure that the model can be optimized through the training data and the generalization performance of the model can be monitored using the validation data. ,in Indicates Sample Features, is the feature dimension. The corresponding target label is ,in Indicates The target value of samples. Divide the data set into a training set and a validation set, and set the ratio of the training set to , then the training set size is , the validation set size is After the division, the training set is , the validation set is Then, we randomly sample the training data set in batches and divide it into multiple small batches to improve the training efficiency and stabilize the gradient calculation. Assume that the batch size is , the training data is divided into batches, each batch is ,in , . Construct a three-layer neural network structure, including input layer, hidden layer and output layer. The number of nodes in the input layer is equal to the feature dimension , the number of nodes in the hidden layer is set to , to increase the network's expressive power, the number of nodes in the output layer is 1, which is used to predict the target value. The forward propagation process of the neural network consists of three stages. In the input layer, the input data Perform linear transformation, the calculation formula is: ; in, is the weight matrix from the input layer to the hidden layer, is the bias vector, is the result of linear transformation. In the hidden layer, The ReLU activation function is used to realize nonlinear transformation, and its calculation formula is: ; in, is the output of the hidden layer. Passed to the output layer for linear transformation and Sigmoid activation processing, the calculation formula is: ; ; in, is the weight matrix from the hidden layer to the output layer, is the bias of the output layer, is the prediction result. After completing the definition of the network structure, all weights , and bias , Initialize with uniform distribution or Xavier initialization to ensure that the parameter values ​​are within a reasonable range. During training, use the Adam optimizer for parameter optimization, which combines the advantages of momentum and adaptive learning rate adjustment to improve convergence speed. The loss function uses cross entropy loss, which is defined as: ; in, is the actual value, is the predicted value, is the batch size. The gradient is calculated by the back-propagation algorithm, and the weights and biases are updated using the Adam optimizer. The gradient update formula is: ; in, is the learning rate. After each training cycle, the validation data set is input into the neural network for validation and the validation loss is calculated. If the validation loss is continuous If it does not decrease within a period, the early stopping mechanism is triggered and training is stopped to prevent overfitting. Finally, the distributed prediction model and training output results are obtained.

[0027] In one example, an n-level state space is established based on the training output results, and a hierarchical warning threshold set is obtained through Markov decision making, including: The training output results of the distributed prediction model are divided into n levels according to the degree of ocean anomaly, namely low anomaly, medium anomaly, and high anomaly. Each level is assigned a corresponding state identifier to obtain an n-level state space; Perform statistical analysis on the state changes at adjacent moments in the n-level state space, record the number of transitions between states, and calculate the proportion according to the total number of transitions to obtain the probability data of transitions between states; The state transition probability data is constructed into a matrix form, where the rows of the matrix represent the current state, the columns represent the state at the next moment, and the values ​​represent the corresponding transition probabilities, thus obtaining the state transition probability matrix; The reward score is set based on the accuracy of marine warning decision-making and the timeliness of warning response. Positive rewards are given to accurate and timely warnings, and negative rewards are given to delayed warnings or false alarms, thus obtaining a state reward function. The state transition probability matrix and state reward function are input into the Markov decision algorithm, and multiple rounds of iterative calculations are performed on each state until the state value converges and stabilizes, and the optimal decision value of each state is obtained; The corresponding warning threshold is set according to the optimal decision value of each state, and the warning threshold is sorted in order from low to high in terms of abnormality to obtain a hierarchical warning threshold set.

[0028] In this example, the training output results of the distributed prediction model are processed and the prediction results are divided into multiple levels according to the degree of ocean anomaly. Level state space. Assume that the output of the prediction model is ,in Indicates The degree of abnormality at a given moment. Based on pre-defined thresholds ,Will Divide into There are 10 levels, each of which corresponds to a status identifier. The classification rules for status identifiers are: ; in, It is The status mark at each moment indicates whether the current abnormality level is low abnormality (status 1), moderate abnormality (status 2), or high abnormality (status 3). ) and other levels. The data at each moment is mapped to the state, and the After constructing the state space, statistical analysis is performed on the state changes at adjacent moments to record the number of transitions between states. Suppose the state sequence is , count each pair of adjacent states Frequency of occurrence ,in and Represent the identifiers of the current state and the next state respectively. By accumulating all the transition times, we can get the frequency matrix of transitions between states. ,in Indicates from the state Transfer to state The total number of transitions of each state is calculated according to its total number of transitions. Indicates from the state Transfer to state The probability of is: ; in, Yes Status The total number of transitions. Through normalized calculation, the state transition probability matrix is ​​constructed , the matrix dimension is , each row represents the current state, each column represents the possible state at the next moment, and the matrix elements Then represents the corresponding transition probability. Combined with the actual needs of marine warning, a reward function is set to quantify the accuracy of warning and response timeliness. Reward function is a two-dimensional matrix, where Indicates from the state Transfer to state The reward value obtained. If the warning is accurate and timely (such as maintaining from low abnormality to low abnormality), a positive reward will be obtained. , while false alarms (such as a shift from a low anomaly to a high anomaly) or delayed warnings (such as a shift from a high anomaly to a low anomaly) receive negative rewards. . The state transition probability matrix And the reward function matrix Enter the Markov decision algorithm to calculate the optimal decision value for each state. The Markov decision algorithm iteratively calculates the state value function , indicating that in the state The optimal value when . The iterative formula of the state value function is: ; in, is a discount factor used to balance short-term and long-term rewards. Through multiple rounds of iterative calculations, the state value function Finally, convergence is achieved to obtain the optimal decision value for each state. , set the corresponding warning threshold The principle of threshold setting is to adjust the state value from low to high, ensuring that the threshold for low abnormal state is the lowest and the threshold for high abnormal state is the highest. These thresholds constitute a hierarchical warning threshold set. .

[0029] In one example, real-time monitoring feature data is input into a distributed prediction model to predict abnormal conditions, obtain abnormal condition prediction results, and compare the abnormal condition prediction results with a hierarchical warning threshold set to output a warning level signal, including: Extract features from the real-time monitoring feature data, perform data dimension reduction processing according to the same feature dimension as the dimension reduction feature data set, and obtain a real-time feature vector; The real-time feature vector is input into the distributed prediction model, and the parameters trained in the three-layer neural network structure are used for forward propagation calculation to obtain the abnormal state prediction value; Compare the abnormal state prediction value with each level threshold in the hierarchical warning threshold set, determine the abnormal level to which the current state belongs based on the threshold interval range, and obtain the abnormal state prediction result; Update the state transition probability matrix according to the abnormal state prediction results, record the transition from the current state to the next predicted state, and obtain the updated state transition data; Input the updated state transition data into the Markov decision process, recalculate the state value function, and dynamically adjust the warning threshold to obtain an updated hierarchical warning threshold set; According to the abnormal level corresponding to the abnormal state prediction result, a warning level signal of the corresponding level is generated.

[0030] In this example, feature extraction is performed on the monitoring feature data collected in real time. Suppose the original monitoring data matrix is , whose dimensions are ,in is the number of samples collected in real time, is the original feature dimension. In order to compare with the reduced feature dataset Maintain consistent feature dimensions ,Will Projected into the feature space after dimensionality reduction. Let the dimensionality reduction transformation matrix be , then the real-time feature vector The calculation formula is: ; in, is the transformation matrix obtained by a dimensionality reduction algorithm (such as principal component analysis or feature selection), is the feature vector of the real-time monitoring data after dimensionality reduction. Input into the trained distributed prediction model. The model consists of a three-layer neural network structure, including input layer, hidden layer and output layer, and uses the parameters optimized in the training phase for forward propagation calculation. It is The calculation process of the three-layer neural network is as follows: Linear transformation is performed at the input layer: ; in, is the weight matrix from the input layer to the hidden layer, is the bias vector, is the input of the hidden layer, is the number of nodes in the hidden layer (usually ). In the hidden layer, the ReLU activation function is applied for nonlinear transformation: ; The output of the hidden layer is passed to the output layer, and the abnormal state prediction value is obtained through linear transformation and Sigmoid activation function: ; ; in, is the weight matrix from the hidden layer to the output layer, is the bias of the output layer, It is The predicted value of abnormal state of samples. Centrally defined thresholds with hierarchical warning thresholds Compare the size and judge The threshold range in which the current state is located determines the abnormal level : ; in, It is the level indicator of abnormal status, indicating low abnormality (1), moderate abnormality (2) or high abnormality According to the abnormal state prediction results obtained by real-time monitoring, the state transition probability matrix is ​​updated . Assume that the initial state transfer matrix is ,in Indicates from the state Transfer to state By recording the changes in real-time status, the number of transfers is accumulated to , and update the probability matrix by row normalization: ; Updated state transition matrix Reflect the change mode of real-time status. Update the state transfer matrix And the reward function set previously Input to the Markov decision process. The state value function is calculated through multiple rounds of iterations , adjust the optimal decision value of each state. The iterative formula of the state value function is: ; in, is a discount factor that balances short-term and long-term rewards. Iterate until Convergence. According to the converged state value function , dynamically adjust the warning threshold , so that it can adapt to the changing trend of real-time data. According to the level corresponding to the abnormal state prediction result , generating early warning signals of corresponding levels.

[0031] Reference Figure 2 This embodiment provides an ocean AI analysis and ocean disaster prediction system based on multi-source heterogeneous data, including: Calculation module 1 is used to construct a data matrix and calculate the amount of information for the marine monitoring data to obtain a monitoring indicator weight matrix, and to calculate the similarity between samples and the neighborhood radius of the monitoring indicator weight matrix to obtain a reduced dimension feature data set; Parallel training module 2 is used to input the dimension-reduced feature data set into the three-layer neural network structure, perform parallel training on each monitoring point, and obtain a distributed prediction model and training output results; Decision module 3, used to establish n-level state space according to the training output results, and obtain a hierarchical warning threshold set through Markov decision; The output module 4 is used to input the real-time monitoring feature data into the distributed prediction model to predict the abnormal state, obtain the abnormal state prediction result, compare the abnormal state prediction result with the hierarchical warning threshold set, and output the warning level signal.

[0032] In this embodiment, for the specific implementation of each unit in the above system embodiment, please refer to the above method embodiment, which will not be repeated here.

[0033] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0034] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, system, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, system, article or method. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, system, article or method including the element.

[0035] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A marine AI analysis and marine disaster prediction method based on multi-source heterogeneous data, characterized in that: The following steps are involved: Constructing a data matrix and calculating the amount of information on the marine monitoring data to obtain a monitoring indicator weight matrix, and calculating the similarity between samples and the neighborhood radius on the monitoring indicator weight matrix to obtain a reduced-dimensional feature data set; The dimension-reduced feature data set is input into a three-layer neural network structure, and each monitoring point is trained in parallel to obtain a distributed prediction model and training output results; Establishing an n-level state space according to the training output results, and obtaining a hierarchical warning threshold set through Markov decision making; The real-time monitoring feature data is input into the distributed prediction model to predict the abnormal state, and the abnormal state prediction result is obtained. The abnormal state prediction result is compared with the hierarchical warning threshold set to output a warning level signal.

2. The method for ocean AI analysis and ocean disaster prediction based on multi-source heterogeneous data according to claim 1 is characterized in that: The data matrix is ​​constructed and the information amount is calculated for the marine monitoring data to obtain a monitoring indicator weight matrix, and the similarity between samples and the neighborhood radius are calculated for the monitoring indicator weight matrix to obtain a reduced dimension feature data set, including: The monitoring data of ocean temperature, salinity, dissolved oxygen, pH value, and chlorophyll are arranged in the order of collection time, and an initial data matrix with h rows of samples and f columns of features is established for the monitoring data; Performing maximum and minimum value normalization operations on the data in the initial data matrix to obtain a standardized data matrix, and performing standard deviation calculations on the standardized data matrix according to the monitoring indicator columns, performing a discrete degree analysis on the data samples of each monitoring indicator, and obtaining standard deviation data of f monitoring indicators; The standardized data matrix is ​​paired with the monitoring indicators for correlation analysis, and the Pearson correlation coefficient between the monitoring indicators is calculated to obtain an f-order correlation coefficient matrix; Calculate the information content of each monitoring indicator based on the standard deviation data and the correlation coefficient matrix, and multiply the standard deviation of each monitoring indicator by the linear combination value of the correlation coefficient of the monitoring indicator relative to other monitoring indicators to obtain the monitoring indicator information content data; Normalizing the monitoring indicator information data, dividing the information of each monitoring indicator by the sum of the information of all monitoring indicators, to obtain a monitoring indicator weight matrix; The monitoring indicator weight matrix is ​​subjected to sample similarity calculation and neighborhood radius calculation to generate a dimension-reduced feature data set.

3. The method for ocean AI analysis and ocean disaster prediction based on multi-source heterogeneous data according to claim 2 is characterized in that: The performing of similarity calculation between samples and neighborhood radius calculation on the monitoring indicator weight matrix to generate a dimension reduction feature data set includes: Performing weighted calculation on the weight values ​​in the monitoring indicator weight matrix and the data of the corresponding monitoring indicator to obtain a weighted feature vector; Calculating the Euclidean distance between the sample data in the weighted feature vector, constructing a sample distance matrix, obtaining initial similarity data, and normalizing the initial similarity data to obtain a standardized similarity value; The neighborhood radius parameter is set based on the data distribution characteristics of the standardized similarity value, and the optimal neighborhood radius value is determined by using the interquartile range method; Constructing a sample neighborhood relationship according to the standardized similarity value and the optimal neighborhood radius value, and establishing a neighborhood relationship if the similarity value between samples is less than or equal to the neighborhood radius value, thereby obtaining a neighborhood relationship set; The neighborhood relationship set is converted into a binary matrix representation, and a neighborhood decision matrix representing the neighborhood relationship between samples is constructed; The dependency of the condition attribute set on the decision attribute set is calculated based on the neighborhood decision matrix to obtain a reduced-dimensional feature data set.

4. The method for ocean AI analysis and ocean disaster prediction based on multi-source heterogeneous data according to claim 3 is characterized in that: The step of calculating the dependency of the condition attribute set on the decision attribute set based on the neighborhood decision matrix to obtain a reduced-dimensional feature data set includes: Calculate the upper approximate set and the lower approximate set between samples according to the neighborhood decision matrix to obtain the conditional attribute space division, and perform positive domain calculation on the conditional attribute space division to obtain the dependency value of the conditional attribute set C on the decision attribute set D; Traversing the subset S of the condition attribute set C according to the forward search strategy, calculating the dependency value of each feature subset on the decision attribute set, and obtaining a feature importance sequence; Setting a dependency gain threshold interval based on the feature importance sequence, performing equidistant sampling within the interval to obtain a candidate threshold set; Each threshold in the candidate threshold set is cross-validated, the classification accuracy of the feature subset under different thresholds is calculated to obtain the optimal threshold, and the features are screened according to the optimal threshold, and the features with importance greater than the threshold are composed of a reduced dimension feature data set.

5. The method for ocean AI analysis and ocean disaster prediction based on multi-source heterogeneous data according to claim 1 is characterized in that: The reduced dimension feature data set is input into a three-layer neural network structure, and each monitoring point is trained in parallel to obtain a distributed prediction model and training output results, including: Dividing the dimension-reduced feature data set into a training data set and a validation data set, and performing batch random sampling on the training data set to obtain a neural network training batch; Construct a three-layer neural network structure, the three-layer neural network structure includes an input layer, a hidden layer and an output layer, wherein the number of nodes in the input layer is equal to the feature dimension of the dimension reduction feature data set, the number of nodes in the hidden layer is 2e+1, and the number of nodes in the output layer is 1; the input layer uses linear transformation to process input data, the hidden layer uses ReLU activation function for nonlinear transformation, and the output layer uses Sigmoid activation function to map the output value to the [0,1] interval to obtain a prediction result; Initializing the connection weights in the three-layer neural network structure, using the Adam optimizer to optimize the parameters, inputting the neural network training batch data into the three-layer neural network structure, using the back propagation algorithm to calculate the gradient, and updating the network parameters; After each training cycle, the verification data set is input into the updated neural network for verification. When the loss value of the verification set does not decrease for G consecutive cycles, the early stopping mechanism is triggered to obtain a distributed prediction model and training output results.

6. The method for ocean AI analysis and ocean disaster prediction based on multi-source heterogeneous data according to claim 1 is characterized in that: The n-level state space is established according to the training output result, and the hierarchical warning threshold set is obtained through Markov decision making, including: The training output results of the distributed prediction model are divided into n levels according to the degree of ocean anomaly, namely, low anomaly, medium anomaly, and high anomaly, and a corresponding state identifier is assigned to each level to obtain an n-level state space; Performing statistical analysis on the state changes at adjacent moments in the n-level state space, recording the number of transitions between states, and performing proportional calculation according to the total number of transitions to obtain the probability data of transitions between states; The state transition probability data is constructed into a matrix form, where the matrix rows represent the current state, the columns represent the state at the next moment, and the values ​​represent the corresponding transition probabilities, to obtain a state transition probability matrix; The reward score is set based on the accuracy of marine warning decision-making and the timeliness of warning response. Positive rewards are given to accurate and timely warnings, and negative rewards are given to delayed warnings or false alarms, thus obtaining a state reward function. Input the state transition probability matrix and the state reward function into the Markov decision algorithm, perform multiple rounds of iterative calculations on each state until the state value converges and stabilizes, and obtain the optimal decision value of each state; The corresponding warning thresholds are set according to the optimal decision values ​​of each state, and the warning thresholds are sorted in order from low to high in terms of abnormality to obtain a hierarchical warning threshold set.

7. The method for ocean AI analysis and ocean disaster prediction based on multi-source heterogeneous data according to claim 1 is characterized in that: The step of inputting the real-time monitoring feature data into the distributed prediction model to perform abnormal state prediction, obtaining an abnormal state prediction result, and comparing the abnormal state prediction result with the hierarchical warning threshold set to output a warning level signal includes: Extracting features from the real-time monitoring feature data, performing data dimension reduction processing according to the same feature dimension as the dimension reduction feature data set, and obtaining a real-time feature vector; The real-time feature vector is input into the distributed prediction model, and the parameters trained in the three-layer neural network structure are used to perform forward propagation calculation to obtain an abnormal state prediction value; Compare the abnormal state prediction value with each level threshold in the hierarchical warning threshold set, determine the abnormal level to which the current state belongs based on the threshold interval range, and obtain the abnormal state prediction result; According to the abnormal state prediction result, the state transition probability matrix is ​​updated, the transition from the current state to the next predicted state is recorded, and the updated state transition data is obtained; Inputting the updated state transition data into a Markov decision process, recalculating the state value function, and dynamically adjusting the warning threshold to obtain an updated hierarchical warning threshold set; According to the abnormal level corresponding to the abnormal state prediction result, a warning level signal of a corresponding level is generated.

8. An ocean AI analysis and ocean disaster prediction system based on multi-source heterogeneous data, characterized in that: For implementing the steps of the method according to any one of claims 1 to 7, the system comprises: A calculation module is used to construct a data matrix and calculate the amount of information for the marine monitoring data to obtain a monitoring indicator weight matrix, and to calculate the similarity between samples and the neighborhood radius of the monitoring indicator weight matrix to obtain a reduced dimension feature data set; A parallel training module is used to input the dimension reduction feature data set into a three-layer neural network structure, perform parallel training on each monitoring point, and obtain a distributed prediction model and training output results; A decision module, used to establish an n-level state space according to the training output results, and obtain a hierarchical warning threshold set through Markov decision making; The output module is used to input the real-time monitoring feature data into the distributed prediction model to predict the abnormal state, obtain the abnormal state prediction result, compare the abnormal state prediction result with the hierarchical warning threshold set, and output the warning level signal.

Images