Construction method and device of multi-source ocean data deep learning analysis and prediction platform

Abstract

The invention relates to the technical field of deep learning, and discloses a multi-source ocean data deep learning analysis and prediction platform construction method and device, and the method comprises the steps: collecting historical multi-source ocean environment parameter data, and carrying out the classification preprocessing, and obtaining a sudden change monitoring data set, a periodic monitoring data set, and a statistical monitoring data set; performing data decomposition and mask encryption processing to obtain a local training data set and a distributed training data set; constructing a local prediction model based on the local training data set, and constructing a distributed prediction model based on the distributed training data set; training and iteratively optimizing the local prediction model and the distributed prediction model by adopting a federated learning framework to obtain a target prediction model; and deploying the target prediction model to a real-time calculation unit, performing multi-source marine data analysis on the real-time multi-source marine environment parameter data, and outputting a marine environment parameter prediction result, thereby improving the adaptability and prediction accuracy of the model to marine environment changes.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a method and device for constructing a multi-source ocean data deep learning analysis and prediction platform. Background Art

[0002] Traditional ocean data analysis methods have difficulty processing large-scale, multi-source heterogeneous data, and have obvious deficiencies in data security, real-time performance, and prediction accuracy.

[0003] In the field of marine environmental parameter prediction, existing analysis platforms generally have problems such as insufficient data classification, single prediction model, and low computing efficiency. In particular, when dealing with sudden changes in the marine environment, traditional methods cannot respond in time and accurately predict, which seriously affects the actual effect of marine environmental monitoring. In addition, marine data has obvious privacy sensitivity. How to achieve efficient distributed computing while protecting data security has become a technical problem that needs to be solved urgently. Summary of the invention

[0004] The present invention provides a method and device for constructing a multi-source ocean data deep learning analysis and prediction platform, which improves the model's adaptability to changes in the ocean environment and its prediction accuracy.

[0005] In a first aspect, the present invention provides a method for constructing a multi-source ocean data deep learning analysis and prediction platform, the method for constructing a multi-source ocean data deep learning analysis and prediction platform comprising: Collect historical multi-source marine environmental parameter data and perform classification preprocessing to obtain mutation monitoring data sets, periodic monitoring data sets and statistical monitoring data sets; Performing data decomposition and mask encryption processing based on the mutation monitoring data set, the period monitoring data set, and the statistical monitoring data set to obtain a local training data set and a distributed training data set; Building a local prediction model based on the local training data set, and building a distributed prediction model based on the distributed training data set; Using a federated learning framework to perform parameter fusion training and iterative optimization on the local prediction model and the distributed prediction model to obtain a target prediction model; The target prediction model is deployed to a real-time computing unit, and multi-source ocean data analysis is performed on real-time multi-source ocean environment parameter data to output ocean environment parameter prediction results.

[0006] In a second aspect, the present invention provides a device for constructing a multi-source ocean data deep learning analysis and prediction platform, the device for constructing a multi-source ocean data deep learning analysis and prediction platform comprising: The acquisition module is used to collect historical multi-source marine environmental parameter data and perform classification preprocessing to obtain mutation monitoring data sets, periodic monitoring data sets and statistical monitoring data sets; A decomposition module, used for performing data decomposition and mask encryption processing based on the mutation monitoring data set, the periodic monitoring data set and the statistical monitoring data set to obtain a local training data set and a distributed training data set; A construction module, configured to construct a local prediction model based on the local training data set, and to construct a distributed prediction model based on the distributed training data set; A training module, used for performing parameter fusion training and iterative optimization on the local prediction model and the distributed prediction model using a federated learning framework to obtain a target prediction model; The output module is used to deploy the target prediction model to the real-time computing unit, perform multi-source ocean data analysis on the real-time multi-source ocean environment parameter data, and output the ocean environment parameter prediction results.

[0007] The third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the above-mentioned method for constructing a multi-source ocean data deep learning analysis and prediction platform.

[0008] A fourth aspect of the present invention provides a computer-readable storage medium, which stores instructions that, when executed on a computer, enable the computer to execute the method for constructing the above-mentioned multi-source ocean data deep learning analysis and prediction platform.

[0009] In the technical solution provided by the present invention, by classifying and preprocessing multi-source marine environmental parameter data, the data is divided into three categories: mutation monitoring, periodic monitoring and statistical monitoring, thereby realizing refined data management and improving the pertinence and efficiency of subsequent analysis; by adopting data decomposition and mask encryption processing technology, while ensuring data security, the coordination of local computing and distributed computing is realized, and the problem of data privacy protection is effectively solved; the local prediction model and distributed prediction model constructed based on deep neural networks fully utilize the advantages of different computing architectures and enhance the feature extraction and prediction capabilities of the model; the introduction of a federated learning framework for parameter fusion training effectively solves the model training problem in a distributed environment while ensuring the privacy security of the data; the prediction model is iteratively optimized through deep reinforcement learning technology, thereby improving the model's adaptability to changes in the marine environment and the prediction accuracy; the deployment method combining online computing and offline computing is adopted, which not only ensures the rapid processing of real-time data, but also makes full use of historical data for trend analysis, thereby realizing comprehensive optimization of prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0011] Figure 1 Schematic diagram of the steps of a method for constructing a multi-source ocean data deep learning analysis and prediction platform in an embodiment of the present invention; Figure 2 It is a schematic diagram of the structure of a device for constructing a multi-source ocean data deep learning analysis and prediction platform in an embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

[0012] Embodiments of the present invention provide a method and apparatus for constructing a multi-source ocean data deep learning analysis and prediction platform. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products, or devices.

[0013] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , an embodiment of a method for constructing a multi-source ocean data deep learning analysis and prediction platform in an embodiment of the present invention includes: Step S1, collecting historical multi-source marine environmental parameter data and performing classification preprocessing to obtain mutation monitoring data sets, periodic monitoring data sets and statistical monitoring data sets; It is understandable that the execution subject of the present invention may be a device for constructing a multi-source ocean data deep learning analysis and prediction platform, or a terminal or a server, which is not specifically limited here. The embodiment of the present invention is described by taking a server as an execution subject as an example.

[0014] Specifically, the integrity analysis of the collected historical multi-source marine environmental parameter data is carried out to evaluate whether the sampling time series of temperature data, salinity data and water level data are missing or discontinuous. A data evaluation matrix is ​​constructed to quantify the integrity status of each data. The calculation process of the matrix includes counting the number of valid data points, data coverage and the distribution law of sampling time for each parameter in a specific time period, forming the basis for integrity evaluation. The timeliness analysis of the data evaluation matrix is ​​carried out to evaluate the time validity of historical data. The deviation value between the sampling time interval and the preset sampling period is calculated. By calculating the difference between the sampling time point of each parameter data and the ideal sampling time point, a timeliness index set is obtained to reflect whether the data collection conforms to the expected time law. In this way, data points with abnormal time distribution are screened out. Based on the timeliness index set, mutation feature extraction is carried out on the historical multi-source marine environmental parameter data, and the change rate of data between adjacent time points is calculated. By analyzing these change rates, parameter values ​​that fluctuate significantly in a short period of time are identified. For example, when the water level data rises or falls abnormally between adjacent time points, this change is marked as mutation feature data. Similarly, sudden and drastic changes in temperature or salinity data will also be identified as mutation features. The extracted mutation feature data is marked as a mutation monitoring data set to characterize significant abnormal events in historical data. Periodic analysis is performed on the data that are not marked as mutation features in the historical multi-source marine environmental parameter data to mine the sampling periodic features of the data, thereby revealing potential periodic laws. The periodic components of these data are extracted through spectrum analysis, time series decomposition or other signal processing methods. For example, tidal data show significant periodic laws. By extracting these periodic features, the relevant data are marked as periodic monitoring data sets to capture the periodic dynamics in the data. For the remaining unmarked data in the historical multi-source marine environmental parameter data, its potential statistical laws are mined based on statistical feature analysis. The statistical distribution characteristics of historical data are calculated, including indicators such as mean, variance, skewness, kurtosis, and the distribution form of the data (such as normal distribution, lognormal distribution, etc.). Through these statistical feature analyses, a comprehensive statistical feature data set is constructed to reflect the stability or variability of the data in the long-term trend. The statistical feature data are marked as statistical monitoring datasets to supplement the mutation monitoring datasets and cycle monitoring datasets.

[0015] Step S2: performing data decomposition and mask encryption processing based on the mutation monitoring data set, the periodic monitoring data set, and the statistical monitoring data set to obtain a local training data set and a distributed training data set; Specifically, the mutation monitoring dataset, periodic monitoring dataset and statistical monitoring dataset are input into the data privacy measurement module for data sensitivity calculation, and a privacy protection level matrix is ​​generated to reflect the demand intensity of different data in terms of privacy protection. The calculation is based on the sensitivity characteristics of the data, application scenarios and potential privacy leakage risk assessment. For example, the abnormal events contained in the mutation monitoring dataset have high sensitivity, so a higher privacy protection level is required, while the statistical monitoring dataset has relatively low sensitivity due to its historical statistical distribution. According to the generated privacy protection level matrix, the mutation monitoring dataset is decomposed and the data correlation coefficient is calculated. The data is divided into key local data and key shared data, where key local data includes content with high sensitivity and strong local characteristics, while key shared data includes content that can be distributed and shared under the premise of protecting privacy. The calculation of the correlation coefficient is the core of this division. Based on the data change rate at adjacent time points, spatial distribution consistency and the causal relationship between mutation characteristics, the local characteristics and shareable characteristics of key data are effectively extracted. For the periodic monitoring dataset, data segmentation and data importance weight calculation are performed to identify the criticality of data in periodic regularity analysis. Through frequency domain analysis and periodic feature extraction, the data is divided into regular local data and regular shared data. Regular local data mainly reflects regional or node-specific periodic features, while regular shared data contains periodic laws that are common across regions and nodes. At the same time, time series feature decomposition and time series correlation calculation are performed based on the statistical monitoring data set to extract background local data and background shared data. Time series feature decomposition uses signal processing methods to separate long-term trends, periodic components and random fluctuations, while time series correlation calculation further extracts background characteristics by measuring the correlation strength between different time points. Key local data, regular local data and background local data are locally fused to form a local training data set. The fusion process is based on a feature weighting algorithm to ensure that the characteristics of various types of data are fully retained, and the training efficiency of local data is improved through data standardization and feature engineering. The local training data set is used for the training of single-node or single-region models, and its characteristics fully reflect the dynamic changes of the local marine environment. At the same time, dynamic mask generation technology is used to protect the privacy of key shared data, regular shared data and background shared data. Dynamic mask generation generates corresponding mask templates for each type of data according to the privacy protection level matrix. These templates ensure that data privacy is not leaked by adding noise, data perturbation or encryption technology. After the mask processing is completed, the data is input into the dynamic affine transformation module, which further protects the mask data by calculating the linear transformation parameter matrix. The specific process of affine transformation includes linear mapping, rotation and translation of the original data, which effectively prevents unauthorized data recovery or reverse inference. The transformed data is marked as transformed encrypted data.The transformed encrypted data is distributed to multiple distributed nodes to form a distributed training data set. In the distributed node allocation process, the load balancing algorithm and data partitioning strategy are used to ensure balanced data distribution on each node, while taking into account the limitations of network bandwidth and computing resources to optimize training efficiency. The distributed training data set can support multi-node joint training, and realize cross-region and cross-node model parameter sharing and optimization through the federated learning framework.

[0016] Step S3: constructing a local prediction model based on the local training data set, and constructing a distributed prediction model based on the distributed training data set; Specifically, the local training data set is input into the local feature extraction network for training to obtain the local feature matrix. The local feature extraction network adopts a three-layer fully connected layer structure design, in which each fully connected layer uses the ReLU activation function to enhance the nonlinear expression ability of the model, and introduces a batch normalization layer to accelerate convergence and improve training stability. Based on the local feature matrix, a local time series processing unit is constructed. The local time series processing unit is based on the long short-term memory network (LSTM) and is used to process the dynamic characteristics of time series data. The design of the long short-term memory network includes three main structures: input gate, forget gate and output gate. The input gate is used to select important input features, the forget gate is used to filter irrelevant or redundant information, and the output gate is used to generate the output of the current time step. Through this structure, the local time series processing unit can capture the dynamic characteristics of short-term fluctuations and long-term trends in the data and obtain local time series features. The local time series features are input into the local prediction network for training. The structure of the local prediction network includes four layers of residual convolution modules and one layer of fully connected output layer. The residual convolution module can effectively alleviate the gradient vanishing problem of the deep network, and extract multi-scale features through convolution operations. The output layer is used to generate the final prediction results. In this process, the local feature extraction network, the local time series processing unit and the local prediction network together constitute a local prediction model. Through collaborative optimization, the model can fully adapt to the characteristics of local training data and accurately predict local marine environmental parameters. At the same time, the data sharding strategy is adopted for the processing of distributed training data sets. By dividing the distributed training data sets into multiple data subsets and assigning each data subset to the corresponding distributed nodes, the load balancing and parallel acceleration of computing tasks are achieved. In each distributed node, the data subset is input into the distributed feature extraction network for feature learning. The network adopts a transformer encoder structure, and its core consists of a multi-head self-attention layer and a feedforward network layer. The multi-head self-attention layer captures the long-distance dependencies in the data by calculating the weight distribution between different features, while the feedforward network layer enhances the feature expression ability through nonlinear transformation. Through this process, the distributed feature extraction network can generate a high-dimensional distributed feature matrix that fully reflects the characteristics of distributed data. The distributed feature matrix aggregates features between nodes through the parameter server architecture, and the parameter server architecture adopts an asynchronous update mechanism to improve the efficiency and robustness of distributed training. In this architecture, each node asynchronously sends and receives feature update information to generate a global feature vector. Based on the global feature vector, a distributed prediction network is constructed. The distributed prediction network adopts a multi-layer perceptron structure, which contains three hidden layers and one output layer. In this network, the global feature vector is mapped to the final prediction result through hierarchical nonlinear transformation and feature combination.The distributed feature extraction network, parameter server architecture and distributed prediction network together constitute a distributed prediction model, which has the advantage of being able to utilize the distributed computing power of multiple nodes and simultaneously realize efficient processing and prediction of massive distributed data. The local prediction model is deployed on the local server to make high-precision predictions of local marine environmental parameters, while the distributed prediction model is deployed on a distributed node cluster to process multi-source data and complex prediction tasks over a wide area.

[0017] Step S4: using a federated learning framework to perform parameter fusion training and iterative optimization on the local prediction model and the distributed prediction model to obtain a target prediction model; Specifically, the network parameters in the local prediction model and the network parameters in the distributed prediction model are input into the parameter parsing unit for structural separation. The function of the parameter parsing unit is to deconstruct the parameter structure of the two types of models. By analyzing the hierarchical characteristics and functional modules of their network architecture, the parameters of the local prediction model are extracted as a local parameter matrix, and the parameters of the distributed prediction model are extracted as a distributed parameter matrix. The local parameter matrix is ​​subjected to hierarchical weight calculation to generate a local hierarchical parameter set. The calculation process is weighted according to the hierarchical structure of the model and the importance of the weight. For example, the parameters closer to the output layer have higher weights because they directly affect the prediction results, while the parameters closer to the input layer are adjusted according to the importance of feature extraction. At the same time, the node parameter consistency calculation is performed on the distributed parameter matrix. Based on the parameter differences and similarities between distributed nodes, the distributed consistency parameter set is generated by calculating the consistency indicators of the parameters between nodes, such as mean, variance or covariance. The distributed consistency parameter set reflects the common characteristics of the distributed model on different nodes and ensures that the distributed training can capture the global data characteristics. A parameter mapping relationship is established based on the local hierarchical parameter set and the distributed consistency parameter set to obtain a federated learning training framework. The establishment of parameter mapping relationships aims to ensure that different model parameters can work in coordination under the federated learning framework, such as by matching the gradient information of similar parameters or aligning different parameters through regularization constraints. On this basis, the parameters in the federated learning training framework are globally optimized to generate a global training strategy set. The global optimization process relies on distributed optimization algorithms, such as FedAvg or a variant algorithm based on gradient descent. The goal is to obtain a parameter strategy set with global generalization capabilities through multiple rounds of iterative optimization. The global training strategy set is input into the differential privacy processing unit for noise injection to protect the data privacy of nodes participating in federated learning. The differential privacy processing unit ensures that sensitive data will not be leaked during training by introducing random noise in gradient updates or model parameters. The secure parameter update strategy generated by noise injection is used for secure exchange and update of parameters in the federated learning framework. After obtaining the secure parameter update strategy, it is back-propagated to generate the training parameter set of each node. This process combines the global training strategy and the local data characteristics of each node to ensure that the updated parameters can reflect the balance between global characteristics and local characteristics. The generated training parameter sets are deployed to the corresponding data node training units for training iterations. On each node, the performance of the local model is gradually improved through multiple rounds of local training and parameter updates. At the same time, the training results between nodes are aggregated into the global model through the parameter sharing mechanism of the federated learning framework. After multiple training iterations, a fused prediction parameter set is obtained. Based on the fused prediction parameter set, the prediction state space, parameter adjustment strategy, and objective function are iteratively optimized.The prediction performance of the model is improved by dynamically adjusting the model structure and optimization objectives, such as redefining the loss function of the prediction task or adjusting the learning rate of parameter updates. Through multiple rounds of optimization iterations, a target prediction model is generated, which can achieve accurate prediction of multi-source ocean data on a global scale, while taking into account the dual needs of data privacy protection and prediction performance improvement under the synergy of local and distributed nodes.

[0018] The fused prediction parameter set is input into the state encoding module, which encodes the parameter state according to the prediction accuracy of the parameter and the computing resource consumption, generates a state vector matrix, and describes the performance and resource requirements of the parameters in the current state of the model. The action space is constructed for the state vector matrix. The core of the action space is to design the action decision set based on the parameter update amplitude and the learning rate change range. The action space matrix is ​​generated by systematically analyzing the possibility of different parameter adjustments and the dynamic change range of the learning rate. The action space matrix defines all possible operations in the model optimization process, including adjusting weights, changing learning rates, or modifying parameter update methods, providing a basis for strategy selection for the reinforcement learning algorithm. The reward function model is constructed based on the state vector matrix and the action space matrix. The design of the reward function is the key in the reinforcement learning process. Its goal is to guide action selection by measuring the effect of model optimization. The reward function sets calculation rules based on prediction error and resource utilization. The prediction error reflects the performance improvement of the model under different actions, while the resource utilization measures the consumption level of computing resources during the optimization process. By combining the two, the reward function can ensure the rationality of resource consumption while improving the model accuracy, balancing performance and efficiency. The reward function model is input into the deep Q network for optimization structure design. In the design of the deep Q network (DQN), a Q value calculation module is constructed based on a convolutional neural network (CNN) and a fully connected network (FCN). The state features are extracted through the convolution layer, and the Q value of each state-action pair is calculated through the fully connected layer. The core goal of this network is to evaluate the long-term rewards of different actions in the current state, thereby providing guidance for action selection. Through this design, the basic Q network structure is obtained. In order to improve the stability and convergence speed of the network, a dual network expansion design is performed on the Q network structure. By copying the basic Q network, a target Q network and an evaluation Q network are constructed respectively to form a dual Q network system. In this system, the target Q network is used to generate a stable target value, while the evaluation Q network is used to dynamically update the current Q value estimate. The dual network system can effectively alleviate the non-stationary problem in reinforcement learning and ensure the continuous improvement of model performance during training. The dual Q network system is input into the experience replay module for sample training. The experience replay module provides randomly sampled training samples by storing the state transition sequence and reward value sequence to avoid the influence of data correlation on model training. In each training, the parameters are updated based on the data of experience playback, and the accuracy of action selection and model performance are improved by optimizing the Q value calculation module to generate a training parameter set. The exploration strategy is designed for the training parameter set. In order to achieve optimized action selection, the epsilon-greedy strategy is adopted. In the action selection process, the balance mechanism of exploration and utilization is combined. By calculating the selection probability of different actions, the ratio of exploring new actions and utilizing existing optimal actions is dynamically adjusted. This strategy can prevent the model from falling into the local optimum and ensure sufficient exploration near the optimal solution to generate the optimal action strategy.The optimal action strategy is deployed into the prediction model architecture to optimize the model parameters. By introducing the optimized strategy into the prediction model, the parameter update method, learning rate and model structure are dynamically adjusted to finally obtain the target prediction model.

[0019] Step S5: deploy the target prediction model to the real-time computing unit, perform multi-source ocean data analysis on the real-time multi-source ocean environment parameter data, and output the ocean environment parameter prediction results.

[0020] Specifically, the target prediction model is input into the model decomposition unit for online and offline decomposition. The model decomposition unit divides the target prediction model into an online computing model and an offline computing model by analyzing the computational complexity and task real-time performance of the model structure. The online computing model focuses on real-time data processing and can quickly respond to the dynamic changes of the marine environment, while the offline computing model performs in-depth analysis and trend mining on historical data to provide insights into environmental characteristics over a long period of time. The online computing model is deployed in real time. Based on the computing resource allocation strategy, the online computing model is deployed to the corresponding real-time computing nodes, which have high computing power and low latency characteristics, such as edge computing devices or high-performance servers, so as to build an efficient real-time computing framework. The framework optimizes the distribution and dynamic scheduling of computing resources to achieve rapid processing and feature extraction of real-time multi-source marine environmental parameter data. At the same time, the offline computing model is deployed in a distributed manner, and the model is allocated to multiple historical data nodes based on the data storage strategy. The offline computing framework relies on distributed storage systems, such as HDFS (Hadoop Distributed File System) or cloud storage platforms, to process large-scale historical data sets. By distributing the model on each node, the offline computing framework can efficiently utilize the computing and storage resources of multiple nodes to conduct in-depth analysis of long-term marine environmental data and identify periodic laws, trend changes and potential anomalies in the data. After completing the online and offline deployment of the model, the newly collected real-time multi-source marine environmental parameter data is input into the data diversion module for flow direction determination. The data diversion module classifies the data according to the timestamp attribute of the data, divides the latest real-time data into online processing data streams, and divides non-real-time data or delayed data into offline processing data streams. In this way, the data flow direction can be precisely controlled so that the online computing model and the offline computing model process their own exclusive data streams respectively. For the online processing data stream, the real-time computing framework starts the feature extraction task and performs real-time feature calculations on core environmental parameters such as temperature parameters, salinity parameters and water level parameters. The feature extraction process includes steps such as time series analysis, mutation feature identification and short-term trend prediction. Its goal is to generate real-time prediction results and quickly reflect the latest dynamics of the marine environment. Since the online computing model is deployed in low-latency real-time computing nodes, this processing can be completed with extremely high efficiency, thus meeting the needs of real-time applications. At the same time, the offline processed data stream is transmitted to the offline computing framework for historical feature analysis. The offline computing framework uses the offline computing model to deeply mine the long-term accumulated marine environmental data, and generates historical prediction results through long-term trend analysis methods (such as moving average, empirical mode decomposition, periodic feature extraction, etc.). The real-time prediction results and historical prediction results are input into the result fusion module for weighted fusion operation. The result fusion module generates the final marine environmental parameter prediction results by weighted average or dynamic adjustment according to the weight ratio of the two types of results.The weight distribution is optimized based on the relative importance of real-time predictions and historical predictions and the dynamic changes of the current environment, ensuring that the fusion results can both respond quickly to short-term changes and capture the stability of long-term trends.

[0021] In the embodiment of the present invention, by classifying and preprocessing multi-source marine environmental parameter data, the data is divided into three categories: mutation monitoring, periodic monitoring and statistical monitoring, thereby achieving refined data management and improving the pertinence and efficiency of subsequent analysis; by adopting data decomposition and mask encryption processing technology, while ensuring data security, the coordination of local computing and distributed computing is achieved, and the problem of data privacy protection is effectively solved; the local prediction model and distributed prediction model constructed based on deep neural networks fully utilize the advantages of different computing architectures and enhance the feature extraction and prediction capabilities of the model; the introduction of a federated learning framework for parameter fusion training effectively solves the model training problem in a distributed environment while ensuring the privacy security of the data; the prediction model is iteratively optimized through deep reinforcement learning technology, thereby improving the model's adaptability to changes in the marine environment and prediction accuracy; the deployment method combining online computing and offline computing is adopted, which not only ensures the rapid processing of real-time data, but also makes full use of historical data for trend analysis, thereby achieving comprehensive optimization of prediction results.

[0022] In a specific embodiment, the process of executing step S1 may specifically include the following steps: Conduct integrity analysis on the collected historical multi-source marine environmental parameter data, and calculate the sampling time series integrity of temperature data, salinity data, and water level data to obtain a data evaluation matrix; Perform timeliness analysis on the data evaluation matrix and calculate the deviation between the sampling time interval and the preset sampling period to obtain a timeliness index set; Based on the timeliness index set, the mutation characteristics of historical multi-source marine environmental parameter data are extracted, and the data change rate at adjacent time points is calculated to obtain the mutation characteristic data, which is marked as the mutation monitoring data set; Perform periodic analysis on the data that are not marked as mutation characteristics in the historical multi-source marine environmental parameter data, calculate the sampling periodic characteristics of the data, obtain periodic characteristic data, and mark the periodic characteristic data as a periodic monitoring data set; Statistical feature analysis is performed on the remaining unlabeled data in the historical multi-source marine environmental parameter data, and the historical statistical distribution characteristics are calculated to obtain statistical feature data, which are then marked as statistical monitoring data sets.

[0023] Specifically, the integrity of historical multi-source marine environmental parameter data is analyzed to evaluate the integrity of the data sampling time series. Assuming that the collected parameters include temperature ,salinity and water level , the continuity of the data is measured by calculating the proportion of valid data points for each parameter. Defining integrity indicators for: ; in, quantity, By calculating , and , forming a data evaluation matrix: ; This matrix is ​​used to quantify the integrity of different parameters as the basis for subsequent timeliness analysis. Timeliness analysis is performed on the data evaluation matrix to evaluate the stability of the data sampling time interval. The actual sampling time interval for each parameter is defined as ,in For the The timestamp of the sampling points, the preset sampling period is . Calculate the time interval deviation value The formula is: ; Timeliness Index It is defined as the average value of the deviation over all valid data points: ; By calculation , and , forming a timeliness index set to reflect the sampling time consistency of different parameters. Based on the timeliness index set, the data is subjected to mutation feature extraction. The key to mutation feature extraction is to calculate the rate of change of data between adjacent time points. : ; in, . Set a mutation threshold when When the time point is considered to have a mutation, it is marked as mutation feature data. These data are classified into the mutation monitoring data set to capture parameter values ​​that have changed significantly in a short period of time. For data not marked as mutation features, periodic analysis is performed. The time domain data is converted into frequency domain data through discrete Fourier transform to extract the periodic characteristics of the data. Define Fourier transform The amplitude is: in, and The frequency The corresponding real and imaginary parts. is the frequency with the maximum amplitude, period pass The periodic data is marked as a periodic monitoring data set to describe regular changes. The remaining unlabeled data is subjected to statistical feature analysis. Statistical feature analysis includes calculating the mean ,variance , skewness and Kurtosis Other characteristic indicators: ; ; These indicators are used to quantify the distribution characteristics of data and mark data with stable mean and small variance as statistical monitoring datasets.

[0024] In a specific embodiment, the process of executing step S2 may specifically include the following steps: The mutation monitoring data set, periodic monitoring data set and statistical monitoring data set are input into the data privacy measurement module to calculate the data sensitivity and obtain the privacy protection level matrix; According to the privacy protection level matrix, the mutation monitoring data set is decomposed and the data correlation coefficient is calculated to obtain key local data and key shared data; Perform data segmentation and data importance weight calculation on the periodic monitoring data set to obtain conventional local data and conventional shared data, and perform time series feature decomposition and time series correlation calculation based on the statistical monitoring data set to obtain background local data and background shared data; Perform local fusion processing on key local data, conventional local data and background local data to obtain a local training data set; Dynamically generate masks for key shared data, conventional shared data and background shared data to obtain mask-processed data; The masked data is input into the dynamic affine transformation module to calculate the linear transformation parameter matrix to obtain the transformed encrypted data, and the transformed encrypted data is distributed to nodes to obtain a distributed training data set.

[0025] Specifically, the mutation monitoring dataset, cycle monitoring dataset, and statistical monitoring dataset are input into the data privacy measurement module. The task of the data privacy measurement module is to evaluate the sensitivity of each dataset and calculate the privacy protection level matrix. Suppose each data feature in the dataset is ,in represents the data sample index, Indicates feature dimension. Defines data sensitivity for: ; Among them, Variance Indicates the variance of the feature, reflecting the volatility of the data; is the entropy of the feature, which measures the uncertainty of the data; is a positive value to prevent the denominator from being zero. After calculating the sensitivity of all data features, the privacy protection level matrix is ​​generated : ; in, is the number of data samples, is the feature number. Based on the privacy protection level matrix, the mutation monitoring dataset is decomposed and the data correlation coefficient is calculated. Assume that the mutation monitoring dataset is , where each is a set of eigenvalues. By calculating the correlation coefficient between the features , determine the dependencies between data features: ; in, Representation characteristics and The covariance of and are their standard deviations respectively. According to the size of the correlation coefficient, the data is divided into key local data and key shared data. Data with high correlation and high privacy protection level are marked as key local data, while data with low correlation or low privacy protection level are marked as key shared data. When segmenting the periodic monitoring data set, the data categories are distinguished by calculating the data importance weight. Data importance weight Defined as: ; in, It is a feature The mean of is the standard deviation. Data with higher weights are marked as regular local data, while data with lower weights are marked as regular shared data. Time series feature decomposition is performed based on statistical monitoring data sets, and decomposition methods (such as empirical mode decomposition) are used to divide data into trend terms, periodic terms, and random terms. Time series data is defined as , decomposed into: ; in, is a trend item, is a periodic term, is a random term. By calculating the time series correlation of these components, background local data and background shared data are generated. The key local data, regular local data and background local data are fused to generate a local training dataset. The local training dataset is used for local model training to ensure data privacy. Dynamic mask generation is performed on key shared data, regular shared data and background shared data. The mask value Generated by adding noise: ; in, The mean is 0 and the variance is The masked data is input into the dynamic affine transformation module. The affine transformation formula is: ; in, is the linear transformation matrix, is the offset vector. After affine transformation, the data becomes transformed encrypted data. The transformed encrypted data is distributed to distributed nodes, and a distributed training data set is generated through a distributed training mechanism. The distributed training data set is used for joint model training across nodes, which not only ensures data security but also improves computing efficiency, thereby achieving the dual goals of data privacy protection and efficient analysis.

[0026] In a specific embodiment, the process of executing step S3 may specifically include the following steps: The local training data set is input into the local feature extraction network for training to obtain the local feature matrix. The local feature extraction network contains three fully connected layers, each of which uses the ReLU activation function and a batch normalization layer. A local time series processing unit is constructed based on the local feature matrix, and local time series features are obtained through long short-term memory network processing in the local time series processing unit. The long short-term memory network includes input gate, forget gate and output gate structure. The local time series features are input into the local prediction network for training to obtain a local prediction model. The local prediction network includes a 4-layer residual convolution module and a 1-layer fully connected output layer. The local prediction model includes a local feature extraction network, a local time series processing unit, and a local prediction network. Distribute the distributed training data set to nodes, obtain multiple data subsets through data sharding strategy, and assign each data subset to the corresponding distributed node; Input the data subset into the distributed feature extraction network, and obtain the distributed feature matrix through the transformer encoder structure in the distributed feature extraction network. The transformer encoder contains a multi-head self-attention layer and a feedforward network layer. Perform inter-node feature aggregation on the distributed feature matrix and obtain the global feature vector through the parameter server architecture, which adopts an asynchronous update mechanism. A distributed prediction network is constructed based on the global feature vector. A distributed prediction model is obtained through a multi-layer perceptron structure in the distributed prediction network. The multi-layer perceptron contains three hidden layers and one output layer. The distributed prediction model includes a distributed feature extraction network, a parameter server architecture, and a distributed prediction network. The local prediction model is deployed on the local server, and the distributed prediction model is deployed on the distributed node cluster to perform the prediction tasks separately.

[0027] Specifically, the local training data set is input into the local feature extraction network for training. The structure of the local feature extraction network consists of three fully connected layers, each of which uses a ReLU activation function and a batch normalization layer. , the output of the fully connected layer is: ; in, Indicates The output of the layer, is the weight matrix, is the bias vector, ReLU is the activation function. The batch normalization layer normalizes the output of each layer, and the formula is: ; in, and are the batch mean and standard deviation, is a small constant used to avoid the denominator being zero. Through the training of the three-layer fully connected network, the output feature matrix is the local feature matrix, which represents the feature extraction result of local data. A local time series processing unit is constructed based on the local feature matrix. The core of the local time series processing unit is the long short-term memory network (LSTM), which is used to process the dynamic features of time series data. LSTM contains input gate, forget gate and output gate structure, which are used to control the input, forget and output process of information. The input vector , the formula for the input gate is: ; The formula of the forget gate is: ; The formula for the output gate is: ; in, is the Sigmoid activation function, is the weight matrix input to the gate, is the weight matrix from hidden state to gate, is the bias vector. Through the control of these gates, LSTM calculates the time step The hidden state of and cell status , and obtain local time series features. The local time series features are input into the local prediction network for training. The local prediction network consists of 4 layers of residual convolution modules and 1 layer of fully connected output layer. The residual convolution module alleviates the gradient vanishing problem of the deep network through skip connections, and its output formula is: ; in, and is the convolution kernel. After the residual module and the fully connected output layer, the local prediction network generates the prediction result. The local prediction model consists of a local feature extraction network, a local time series processing unit, and a local prediction network. For distributed training data sets, the data is divided into multiple data subsets through a data sharding strategy, and each data subset is assigned to a corresponding distributed node. Each data subset is input into a distributed feature extraction network. The network adopts a transformer encoder structure, and its core is a multi-head self-attention layer and a feedforward network layer. The multi-head self-attention layer calculates the attention score of each input vector, and the formula is: ; in, , and are query, key, and value matrices respectively, is the dimension of the key. Through the multi-head mechanism, the attention calculation is expanded into multiple parallel branches to capture more features. The distributed feature matrix is ​​aggregated between nodes through the parameter server architecture. The parameter server adopts an asynchronous update mechanism to improve the efficiency of distributed training. The aggregated global feature vector is input into the distributed prediction network. The distributed prediction network is composed of a multi-layer perceptron, and its structure is: ; in, and It is The weights and biases of the layers. The distributed prediction network generates prediction results. The distributed prediction model includes a distributed feature extraction network, a parameter server architecture, and a distributed prediction network. The local prediction model and the distributed prediction model are deployed in the local server and the distributed node cluster respectively. The local prediction model focuses on real-time prediction, while the distributed prediction model handles complex distributed data tasks.

[0028] In a specific embodiment, the process of executing step S4 may specifically include the following steps: The network parameters in the local prediction model and the network parameters in the distributed prediction model are input into the parameter parsing unit for structural separation to obtain a local parameter matrix and a distributed parameter matrix; Perform hierarchical weight calculation on the local parameter matrix to obtain a local hierarchical parameter set, and perform node parameter consistency calculation based on the distributed parameter matrix to obtain a distributed consistency parameter set; A parameter mapping relationship is established according to the local level parameter set and the distributed consistency parameter set to obtain a federated learning training framework, and a global optimization operation is performed on the parameters in the federated learning training framework to obtain a global training strategy set; The global training strategy set is input into the differential privacy processing unit for noise injection to obtain the security parameter update strategy, and the security parameter update strategy is back-propagated to obtain the training parameter set of each node; Deploy the training parameter sets to the corresponding data node training units for training iterations to obtain the fusion prediction parameter sets; The prediction state space, parameter adjustment strategy and objective function are iteratively optimized according to the fused prediction parameter set to obtain the target prediction model.

[0029] Specifically, the network parameters in the local prediction model and the network parameters in the distributed prediction model are input into the parameter parsing unit for structural separation. Assume that the parameter set of the local prediction model is ,in It is The weight matrix of the layer, is the bias vector; the parameter set of the distributed prediction model is ,in and Respectively represent The weights and biases of the layer. Through the parameter parsing unit, these two types of parameters are separated into local parameter matrices and the distributed parameter matrix . For the local parameter matrix Perform layer weight calculation to obtain the local layer parameter set. The layer weight calculation formula is: ; in, It is The Frobenius norm of the layer weight matrix represents the importance weight of the layer parameters. is the total number of layers in the local model. The local level parameter set reflects the contribution of each layer parameter in the model by weighted distribution of each layer weight. Based on the distributed parameter matrix , perform node parameter consistency calculation to obtain a distributed consistency parameter set. Node parameter consistency measures the consistency by calculating the parameter differences on multiple distributed nodes. The formula is: ; in, Indicates Layer in Node Parameters, All nodes The mean of the layer parameters, is the number of distributed nodes. The distributed consistency parameter set is generated by aggregating the consistency calculation results, reflecting the degree of global coordination in distributed training. According to the local level parameter set and the distributed consistency parameter set, a parameter mapping relationship is established to obtain the federated learning training framework. In the federated learning framework, the parameter mapping rule is used to align the local model parameters with the distributed model parameters. The parameter mapping relationship is defined by the weight ratio allocation formula as: ; in, It is Layer local parameters and The mapping relationship weights of layer distributed parameters, and is the adjustment coefficient used to balance the impact of local and distributed parameters. Based on the parameter mapping relationship, the parameters in the federated learning framework are globally optimized to generate a global training strategy set. Global optimization uses an update formula based on gradient descent: ; in, is the learning rate, is the gradient of the global loss function to the parameter. The global training strategy set is input into the differential privacy processing unit to enhance privacy protection through noise injection. Noise injection uses the Laplace noise mechanism, and the formula is: ; in, is the noise distribution, is the sensitivity, is the privacy budget. Through differential privacy processing, a security parameter update strategy is generated. The security parameter update strategy is back-propagated to generate the training parameter set of each node. The parameter update formula for each node is: ; in, It is The local loss function of the node. Deploy the training parameter set to the corresponding data node training unit, perform multiple rounds of training iterations, and obtain the fused prediction parameter set. According to the fused prediction parameter set, optimize the prediction state space, parameter adjustment strategy, and objective function. The objective function form is: ; in, is the node weight, which is optimized by , and finally generate the target prediction model.

[0030] In a specific embodiment, the execution step iteratively optimizes the prediction state space, parameter adjustment strategy and objective function according to the fusion prediction parameter set to obtain the target prediction model, which may specifically include the following steps: The fusion prediction parameter set is input into the state encoding module, and the state encoding is performed according to the prediction accuracy of the parameters and the computing resource consumption to obtain the state vector matrix; The action space is constructed for the state vector matrix, and the action decision set is constructed based on the parameter update amplitude and the learning rate variation range to obtain the action space matrix; The reward function is constructed based on the state vector matrix and the action space matrix, and the reward calculation rules are set according to the prediction error and resource utilization to obtain the reward function model; The reward function model is input into the deep Q network for optimized structure design, and the Q value calculation module is constructed based on the convolutional neural network and the fully connected network to obtain the Q network structure; Perform dual network expansion design on the Q network structure, copy the Q network structure to build the target Q network and the evaluation Q network, and obtain a dual Q network system; The dual Q network system is input into the experience replay module for sample training, and the parameters are updated based on the state transition sequence and reward value sequence to obtain the training parameter set; An exploration strategy is designed for the training parameter set, and the action selection probability is calculated based on the epsilon-greedy strategy to obtain the optimal action strategy. The optimal action strategy is then deployed into the prediction model architecture for parameter optimization configuration to obtain the target prediction model.

[0031] Specifically, the fusion prediction parameter set is input into the state encoding module, and state encoding is performed according to the prediction accuracy of the parameters and the computing resource consumption. Assume that the fusion prediction parameter set is , where each Represents a prediction parameter. Define the state encoding as a vector ,in It is The prediction accuracy of the parameters, is the consumption of corresponding computing resources. The calculation formula is: ; in, yes The prediction error, MaxError is the maximum error of all parameters, Resource It is calculated The amount of resources required, MaxResource is the maximum value of resource consumption. The state encoding module converts each Convert to the corresponding state vector, and combine all state vectors to get the state vector matrix . For the state vector matrix Construct the action space and design the action decision set based on the parameter update amplitude and learning rate change range. Assume that the update amplitude of each parameter is , the learning rate is The action set is defined as ,in Expressing the The first parameter to be executed actions, which are calculated as: ; in, It is The update range, It is By enumerating the possible update amplitudes and learning rate ranges, the action space matrix is ​​generated , each row corresponds to all possible actions for a parameter. Based on the state vector matrix and the action space matrix , and construct the reward function. The goal of reward function design is to balance prediction error and resource utilization and define the reward value for: ; in, Is to perform an action The prediction error after is the resource utilization rate, and is a weight parameter used to adjust the importance of both. The reward function model is input into the deep Q network (DQN) for optimized structure design. DQN uses a convolutional neural network (CNN) to extract state features and calculate the Q value through a fully connected network. Assuming the input state is s, the output of the convolutional layer is: ; in, is the convolution kernel, * indicates the convolution operation, is the bias term, is an activation function (such as ReLU). The fully connected network calculates the Q value based on the convolutional features: ; in, is the weight of the fully connected layer, is the bias. The Q network structure is extended with dual networks to construct the target Q network and the evaluation Q network respectively. The target Q network is used to generate a stable target value, and the formula is: ; in, It’s an instant reward. is the discount factor, is the next state, is the output of the target Q network. The evaluation Q network is used to update the Q value online, and its update formula is: ; in, is the learning rate. The dual Q network system is input into the experience replay module for sample training. The experience replay module stores the state transition sequence , avoiding data correlation through random sampling. During the training process, the Q value is updated based on the stored sample sequence to generate a training parameter set. The exploration strategy is designed for the training parameter set, and the epsilon-greedy strategy is used for action selection. The probability of action selection is: ; in, is the exploration probability, is the size of the action space. Through this strategy, a balance is achieved between exploring new actions and using the best actions. The optimal action strategy is deployed to the prediction model architecture, and the parameter updates are optimized based on the selected actions, and the prediction state space and objective function configuration are adjusted to generate the target prediction model with the best performance.

[0032] In a specific embodiment, the process of executing step S5 may specifically include the following steps: The target prediction model is input into the model decomposition unit for online and offline decomposition to obtain an online calculation model and an offline calculation model. The online calculation model is responsible for real-time data processing, and the offline calculation model is responsible for historical data analysis. Perform real-time deployment of the online computing model, and deploy the online computing model to the corresponding real-time computing node based on the computing resource allocation strategy to obtain a real-time computing framework; Distributed deployment of offline computing models is performed. Based on the data storage strategy, the offline computing models are deployed to the corresponding historical data nodes to obtain an offline computing framework. The newly collected real-time multi-source ocean environment parameter data is input into the data diversion module for flow direction determination, and the real-time multi-source ocean environment parameter data is divided into an online processing data flow and an offline processing data flow based on the data timestamp; Perform real-time feature calculation on the online processing data stream, extract features of temperature parameters, salinity parameters and water level parameters based on the real-time calculation framework, and obtain real-time prediction results; Perform historical feature analysis on the offline processed data stream, perform long-term trend analysis on the historical data based on the offline computing framework to obtain historical prediction results, and input the real-time prediction results and historical prediction results into the result fusion module for weighted fusion operation to obtain the prediction results of marine environmental parameters.

[0033] Specifically, the target prediction model is input into the model decomposition unit for online and offline decomposition. Assume that the parameter set of the target prediction model is , where each parameter Corresponding to specific prediction tasks or data processing functions. Through the decomposition strategy, the sub-model part with higher real-time processing requirements is extracted as an online computing model, and its parameter set is , which is used to process real-time data streams; and some parameters that are biased towards historical data analysis are extracted as offline calculation models, whose parameter set is , focusing on trend analysis over a long time span. The online computing model is deployed in real-time computing nodes with limited computing resources but fast response speed, and node selection and task allocation are performed through computing resource allocation strategies. Assume that the node set is , each node The computing resource capacity is . Define the resource allocation function Alloc Indicates that the load of the online computing model is distributed to the nodes The resource usage of is, and the allocation conditions are: ; By optimizing the allocation function, we ensure that the online computing model is deployed to the optimal node set under the premise of meeting resource constraints, forming a real-time computing framework. At the same time, the offline computing model adopts distributed deployment, relying on the distributed storage system to distribute the model to multiple historical data nodes. Assuming that the historical data is stored in a distributed file system, the data stored in each node is The total data set is The allocation strategy of the offline computing model is based on the location of data storage and the computing power of the nodes. The optimization formula is: ; Among them, Comm is the communication overhead between the model and the data. By minimizing the communication overhead, the offline computing model is efficiently deployed to the historical data node to build an offline computing framework. After the model is deployed, the newly collected real-time multi-source ocean environment parameter data is input into the data diversion module for flow direction determination. Assume that the input data flow is , where each data point Including feature vectors and timestamp Based on timestamps, the data stream is divided into online processing data streams and offline processing of data streams , the splitting condition is: ; in, is the time window for real-time processing. For the online processing data stream, the real-time computing framework starts the feature calculation task to extract the features of temperature, salinity and water level parameters. The formula for feature extraction is: ; Among them, Extract It is a feature extraction function, including steps such as standardization, time series decomposition and mutation detection. These features are processed through online computing models to generate real-time prediction results: ; At the same time, the offline processed data stream is input into the offline computing framework to perform long-term trend analysis based on historical data. Historical feature analysis includes trend items , Periodic Item and random items The decomposition formula is: ; The offline computation model uses these decomposition features to generate historical forecast results: ; The real-time prediction results and historical forecast results Input the result fusion module to perform weighted fusion operation. Let the fusion weight be and ,satisfy The calculation formula of the fusion result is: ; By adjusting and The value of is used to achieve a balance between real-time and long-term trends, and finally generate prediction results of marine environmental parameters.

[0034] The above describes the method for constructing a multi-source ocean data deep learning analysis and prediction platform in an embodiment of the present invention. The following describes the device for constructing a multi-source ocean data deep learning analysis and prediction platform in an embodiment of the present invention. Figure 2 In one embodiment of the present invention, a device for constructing a multi-source ocean data deep learning analysis and prediction platform includes: The acquisition module is used to collect historical multi-source marine environmental parameter data and perform classification preprocessing to obtain mutation monitoring data sets, periodic monitoring data sets and statistical monitoring data sets; A decomposition module is used to perform data decomposition and mask encryption processing based on the mutation monitoring data set, the periodic monitoring data set and the statistical monitoring data set to obtain a local training data set and a distributed training data set; A construction module, used to build a local prediction model based on a local training data set, and to build a distributed prediction model based on a distributed training data set; The training module is used to perform parameter fusion training and iterative optimization on the local prediction model and the distributed prediction model using a federated learning framework to obtain the target prediction model; The output module is used to deploy the target prediction model to the real-time computing unit, perform multi-source ocean data analysis on the real-time multi-source ocean environment parameter data, and output the ocean environment parameter prediction results.

[0035] Through the collaborative cooperation of the above-mentioned components, by classifying and preprocessing multi-source marine environmental parameter data, the data is divided into three categories: mutation monitoring, periodic monitoring and statistical monitoring, which realizes the refined management of data and improves the pertinence and efficiency of subsequent analysis; the data decomposition and mask encryption processing technology are used to ensure data security while realizing the coordination of local computing and distributed computing, effectively solving the problem of data privacy protection; the local prediction model and distributed prediction model built based on deep neural networks fully utilize the advantages of different computing architectures and enhance the feature extraction and prediction capabilities of the model; the federated learning framework is introduced for parameter fusion training, which effectively solves the model training problem in a distributed environment and ensures the privacy security of the data; the prediction model is iteratively optimized through deep reinforcement learning technology, which improves the model's adaptability to changes in the marine environment and the prediction accuracy; the deployment method combining online computing and offline computing is adopted, which not only ensures the rapid processing of real-time data, but also makes full use of historical data for trend analysis, realizing the comprehensive optimization of prediction results.

[0036] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0037] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0038] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0039] 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.

[0040] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0041] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program codes.

[0042] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a multi-source ocean data deep learning analysis and prediction platform, characterized in that: The method comprises: Collect historical multi-source marine environmental parameter data and perform classification preprocessing to obtain mutation monitoring data sets, periodic monitoring data sets and statistical monitoring data sets; Performing data decomposition and mask encryption processing based on the mutation monitoring data set, the period monitoring data set, and the statistical monitoring data set to obtain a local training data set and a distributed training data set; Building a local prediction model based on the local training data set, and building a distributed prediction model based on the distributed training data set; Using a federated learning framework to perform parameter fusion training and iterative optimization on the local prediction model and the distributed prediction model to obtain a target prediction model; The target prediction model is deployed to a real-time computing unit, and multi-source ocean data analysis is performed on real-time multi-source ocean environment parameter data to output ocean environment parameter prediction results.

2. The method for constructing a multi-source ocean data deep learning analysis and prediction platform according to claim 1, characterized in that: The historical multi-source marine environmental parameter data is collected and classified and pre-processed to obtain a mutation monitoring data set, a periodic monitoring data set and a statistical monitoring data set, including: Conduct integrity analysis on the collected historical multi-source marine environmental parameter data, and calculate the sampling time series integrity of temperature data, salinity data, and water level data to obtain a data evaluation matrix; Performing timeliness analysis on the data evaluation matrix, and calculating the deviation value between the sampling time interval and the preset sampling period to obtain a timeliness index set; Extract mutation characteristics of the historical multi-source marine environmental parameter data based on the timeliness index set, calculate the data change rate at adjacent time points, obtain mutation characteristic data, and mark the mutation characteristic data as a mutation monitoring data set; Performing periodic analysis on the data that are not marked as mutation characteristics in the historical multi-source marine environmental parameter data, calculating the sampling periodic characteristics of the data, obtaining periodic characteristic data, and marking the periodic characteristic data as a periodic monitoring data set; Statistical feature analysis is performed based on the remaining unlabeled data in the historical multi-source ocean environment parameter data, and historical statistical distribution features are calculated to obtain statistical feature data, which are then labeled as statistical monitoring data sets.

3. The method for constructing a multi-source ocean data deep learning analysis and prediction platform according to claim 2, characterized in that: The step of performing data decomposition and mask encryption processing based on the mutation monitoring data set, the periodic monitoring data set, and the statistical monitoring data set to obtain a local training data set and a distributed training data set includes: Inputting the mutation monitoring data set, the period monitoring data set and the statistical monitoring data set into a data privacy measurement module to perform data sensitivity calculation to obtain a privacy protection level matrix; Performing data decomposition and data correlation coefficient calculation on the mutation monitoring data set according to the privacy protection level matrix to obtain key local data and key shared data; Performing data segmentation and data importance weight calculation on the periodic monitoring data set to obtain conventional local data and conventional shared data, and performing time series feature decomposition and time series correlation calculation based on the statistical monitoring data set to obtain background local data and background shared data; Performing local fusion processing on the key local data, the conventional local data and the background local data to obtain a local training data set; Dynamically generate masks for the key shared data, the regular shared data, and the background shared data to obtain mask-processed data; The masked data is input into a dynamic affine transformation module to perform linear transformation parameter matrix calculation to obtain transformed encrypted data, and the transformed encrypted data is distributed to nodes to obtain a distributed training data set.

4. The method for constructing a multi-source ocean data deep learning analysis and prediction platform according to claim 3, characterized in that: The constructing of a local prediction model based on the local training data set and the constructing of a distributed prediction model based on the distributed training data set include: Inputting the local training data set into a local feature extraction network for training to obtain a local feature matrix, wherein the local feature extraction network comprises three fully connected layers, each of which uses a ReLU activation function and a batch normalization layer; Building a local time series processing unit based on the local feature matrix, and obtaining local time series features through long short-term memory network processing in the local time series processing unit, wherein the long short-term memory network includes an input gate, a forget gate, and an output gate structure; Inputting the local time series features into a local prediction network for training to obtain a local prediction model, wherein the local prediction network comprises a 4-layer residual convolution module and a 1-layer fully connected output layer, and the local prediction model comprises a local feature extraction network, a local time series processing unit, and a local prediction network; Distributing the distributed training data set to nodes, obtaining multiple data subsets through a data sharding strategy, and allocating each data subset to a corresponding distributed node; Inputting the data subset into a distributed feature extraction network, and obtaining a distributed feature matrix through a transformer encoder structure in the distributed feature extraction network, wherein the transformer encoder includes a multi-head self-attention layer and a feedforward network layer; Performing inter-node feature aggregation on the distributed feature matrix, and obtaining a global feature vector through a parameter server architecture, wherein the parameter server architecture adopts an asynchronous update mechanism; Building a distributed prediction network based on the global feature vector, obtaining a distributed prediction model through a multi-layer perceptron structure in the distributed prediction network, wherein the multi-layer perceptron includes three hidden layers and one output layer, and the distributed prediction model includes a distributed feature extraction network, a parameter server architecture, and a distributed prediction network; The local prediction model is deployed on a local server, and the distributed prediction model is deployed on a distributed node cluster to perform prediction tasks respectively.

5. The method for constructing a multi-source ocean data deep learning analysis and prediction platform according to claim 4, characterized in that: The method of using a federated learning framework to perform parameter fusion training and iterative optimization on the local prediction model and the distributed prediction model to obtain a target prediction model includes: Inputting the network parameters in the local prediction model and the network parameters in the distributed prediction model into a parameter parsing unit for structural separation to obtain a local parameter matrix and a distributed parameter matrix; Performing hierarchical weight calculation on the local parameter matrix to obtain a local hierarchical parameter set, and performing node parameter consistency calculation based on the distributed parameter matrix to obtain a distributed consistency parameter set; Establishing a parameter mapping relationship according to the local level parameter set and the distributed consistency parameter set to obtain a federated learning training framework, and performing a global optimization operation on the parameters in the federated learning training framework to obtain a global training strategy set; Inputting the global training strategy set into the differential privacy processing unit for noise injection to obtain a security parameter update strategy, and performing back propagation calculation on the security parameter update strategy to obtain a training parameter set for each node; Deploy the training parameter sets to the corresponding data node training units for training iterations to obtain a fusion prediction parameter set; The prediction state space, parameter adjustment strategy and objective function are iteratively optimized according to the fused prediction parameter set to obtain a target prediction model.

6. The method for constructing a multi-source ocean data deep learning analysis and prediction platform according to claim 5, characterized in that: The iterative optimization of the prediction state space, parameter adjustment strategy and objective function according to the fusion prediction parameter set to obtain the target prediction model includes: Inputting the fusion prediction parameter set into a state encoding module, and performing state encoding according to the prediction accuracy of the parameters and the consumption of computing resources to obtain a state vector matrix; Constructing an action space for the state vector matrix, constructing an action decision set based on the parameter update amplitude and the learning rate variation range, and obtaining an action space matrix; Constructing a reward function based on the state vector matrix and the action space matrix, setting a reward calculation rule according to the prediction error and resource utilization, and obtaining a reward function model; The reward function model is input into a deep Q network for optimized structure design, a Q value calculation module is constructed based on a convolutional neural network and a fully connected network, and a Q network structure is obtained; Performing a dual network expansion design on the Q network structure, copying the Q network structure to construct a target Q network and an evaluation Q network, and obtaining a dual Q network system; Input the dual Q network system into the experience replay module for sample training, and update the parameters based on the state transition sequence and the reward value sequence to obtain a training parameter set; An exploration strategy is designed for the training parameter set, and the action selection probability is calculated based on the epsilon-greedy strategy to obtain the optimal action strategy, which is then deployed into the prediction model architecture for parameter optimization configuration to obtain the target prediction model.

7. The method for constructing a multi-source ocean data deep learning analysis and prediction platform according to claim 6, characterized in that: The target prediction model is deployed to a real-time computing unit, and multi-source ocean data analysis is performed on real-time multi-source ocean environment parameter data to output ocean environment parameter prediction results, including: The target prediction model is input into the model decomposition unit for online and offline decomposition to obtain an online calculation model and an offline calculation model, wherein the online calculation model is responsible for real-time data processing and the offline calculation model is responsible for historical data analysis; Performing real-time deployment on the online computing model, deploying the online computing model to the corresponding real-time computing node based on the computing resource allocation strategy, and obtaining a real-time computing framework; Distributed deployment is performed on the offline computing model, and the offline computing model is deployed to the corresponding historical data node based on the data storage strategy to obtain an offline computing framework; Input the newly collected real-time multi-source ocean environment parameter data into the data diversion module for flow direction determination, and divide the real-time multi-source ocean environment parameter data into an online processing data flow and an offline processing data flow based on the data timestamp; Performing real-time feature calculation on the online processing data stream, extracting features of temperature parameters, salinity parameters and water level parameters based on the real-time calculation framework, and obtaining real-time prediction results; Perform historical feature analysis on the offline processed data stream, perform long-term trend analysis on the historical data based on the offline computing framework to obtain historical prediction results, and input the real-time prediction results and the historical prediction results into a result fusion module for weighted fusion operation to obtain marine environment parameter prediction results.

8. A device for constructing a multi-source ocean data deep learning analysis and prediction platform, characterized in that: A method for constructing a multi-source ocean data deep learning analysis and prediction platform as described in any one of claims 1 to 7, wherein the construction device of the multi-source ocean data deep learning analysis and prediction platform comprises: The acquisition module is used to collect historical multi-source marine environmental parameter data and perform classification preprocessing to obtain mutation monitoring data sets, periodic monitoring data sets and statistical monitoring data sets; A decomposition module, used for performing data decomposition and mask encryption processing based on the mutation monitoring data set, the periodic monitoring data set and the statistical monitoring data set to obtain a local training data set and a distributed training data set; A construction module, configured to construct a local prediction model based on the local training data set, and to construct a distributed prediction model based on the distributed training data set; A training module, used for performing parameter fusion training and iterative optimization on the local prediction model and the distributed prediction model using a federated learning framework to obtain a target prediction model; The output module is used to deploy the target prediction model to the real-time computing unit, perform multi-source ocean data analysis on the real-time multi-source ocean environment parameter data, and output the ocean environment parameter prediction results.

9. A computer device, characterized in that: It includes a memory and a processor, the memory stores a computer program that can be run on the processor, and is characterized in that when the processor executes the computer program, it implements the method for constructing a multi-source ocean data deep learning analysis and prediction platform as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the processor executes the method for constructing a multi-source ocean data deep learning analysis and prediction platform as described in any one of claims 1 to 7.

Images