Space-time intelligent ocean satellite internet-of-things perception information screening method and device

Abstract

The invention provides a space-time intelligent ocean satellite internet of things perception information screening method and device, and belongs to the technical field of data processing. Marine environment data space-time prediction is carried out based on a block time sequence converter, marine hot spot area mining is carried out based on a density clustering algorithm of mean shift, and marine environment data comprises sensing data of sea surface temperature, salinity, wave height and the like; marine environment risk assessment is carried out based on a BP neural network, information screening feedback control is carried out based on multi-agent reinforcement learning, the efficient information processing requirement of a complex marine environment detection scene is met, and the low-redundancy and high-precision data acquisition and transmission adaptive feedback control effect is achieved. The method can be directly applied to an ocean observation and detection facility networking system in the ocean Internet of Things, can provide powerful support in the aspects of military application and civil application, and has wide and important application prospects and value.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a spatiotemporal intelligent ocean satellite Internet of Things perception information screening method and device, which improves the accuracy of ocean information screening control and reduces information redundancy by meeting the efficient information processing requirements of complex ocean environment detection scenarios. Background Art

[0002] With the growing demand for marine resource development and ecological protection, the Ocean Satellite Internet of Things has achieved all-weather collection of multi-dimensional data such as ocean temperature, salinity, flow rate, and biological activity through a three-dimensional sensing network consisting of low-orbit satellites, buoys, submersibles, underwater sonar, and satellite remote sensing. However, the marine environment is highly dynamic, wide-area, and complex. Perception data is characterized by multi-source heterogeneity (such as buoy point data, sonar sequences, and remote sensing images), large differences in temporal and spatial scales (minutes to days), and significant noise interference (such as ocean current disturbances and equipment signal attenuation). Buoy data collected by satellites is easily affected by surface currents and prone to drift errors. Underwater sonar is limited by deep-sea pressure and signal scattering, resulting in unstable resolution. Although satellite remote sensing has a wide coverage area, it is easily interfered with by meteorological conditions such as clouds and sea fog. Traditional information screening methods are mostly based on single sensor data or static spatiotemporal models, which make it difficult to effectively integrate the spatiotemporal correlation of multi-source data, resulting in insufficient filtering of redundant information and deviations in the extraction of key features. They are also unable to adapt to the real-time analysis needs of dynamic scenarios such as tidal changes and sudden storms, seriously restricting the accuracy of ocean monitoring and disaster warning.

[0003] Current ocean satellite IoT sensing information screening technology faces multiple challenges. On the one hand, data from different devices differ significantly in spatiotemporal resolution, data format, and noise type. Traditional fusion algorithms are prone to information conflicts or insufficient utilization of complementarity due to mismatches in spatiotemporal scales. On the other hand, in dynamic ocean scenarios, data distribution exhibits uneven characteristics due to sudden environmental changes. Existing methods face bottlenecks in dynamic weight allocation and cross-modal association mining. Furthermore, traditional technologies have a weak ability to combine ocean physical mechanisms with data characteristics, making it difficult to effectively suppress environmental noise and extract implicit spatiotemporal evolution laws through data-driven models, resulting in insufficient real-time screening capabilities for emergencies.

[0004] To address the above issues, the present invention proposes a spatiotemporal intelligent ocean satellite IoT sensing information screening method, which has the following innovations and advantages:

[0005] Based on the improved convolutional bidirectional long short-term memory neural network, the time series data repair method adopts the bidirectional sliding average modulus BiMA to pre-interpolate and repair the missing data. The pre-interpolated and repaired data is used to construct the training data set of the convolutional bidirectional long short-term memory neural network CNN-BiLSTM. The fully trained CNN-BiLSTM is used to predict the missing data, and the predicted value is used to update the interpolation result. The obtained interpolation repair value is then smoothed by the BiMA model to improve the global consistency of the data, which can more accurately reflect the real time series change law and provide a more reliable foundation for subsequent data analysis and prediction.

[0006] A spatiotemporal prediction method for marine environmental data based on an improved block time series transformer is proposed. Marine environmental data such as sea surface temperature and salinity are cut into local blocks according to time and space dimensions. Multi-scale features are captured through the block time series transformer PatchTST, and multimodal fusion is performed with multi-feature data of the marine environment. The sparse attention mechanism is used to model long-range spatiotemporal dependencies, and physical constraints such as ocean dynamics equations are introduced in the decoding stage to improve the rationality of prediction. Finally, future spatiotemporal series are generated through recursive block partitioning to achieve efficient and accurate prediction of marine environmental parameters, reducing error accumulation while taking into account computational efficiency.

[0007] An ocean hotspot area mining method based on the improved mean shift density clustering algorithm is proposed. The mean shift algorithm MS is used to adaptively determine the initial cluster center, and the local density gradient is calculated through a sliding window and the centroid position is iteratively updated to complete coarse-grained clustering to capture high-density areas. The coarse clustering results are then refined in combination with the density clustering algorithm DBSCAN, and finally ocean hotspot areas with significant statistical characteristics are extracted.

[0008] A marine environmental risk assessment method based on the improved sine-cosine algorithm and BP neural network is proposed. The sine-cosine algorithm is used to optimize the BP network structure, and the initial weights and thresholds of the BP neural network are globally optimized. The algorithm convergence is enhanced by combining the adaptive parameter adjustment strategy. Finally, real-time monitoring data is input and the risk probability distribution is output through the Sigmoid activation function to achieve high-precision and strong robust dynamic risk assessment of the marine environment.

[0009] Based on the improved information screening feedback control method of multi-agent reinforcement learning, the problem of collaborative processing of multi-source information is modeled as a Markov game process. A distributed decision-making and centralized training framework is adopted, and key information screening between agents is realized through a two-level attention mechanism. It is finally applied to real-time data filtering, anomaly detection and collaborative control in dynamic environments, meeting the efficient information processing needs of complex marine environment detection scenarios, and achieving an adaptive feedback control effect that significantly reduces the transmission load while ensuring data validity.

[0010] Through these innovations, the present invention can solve problems such as weak spatiotemporal correlation of multi-source heterogeneous data, low efficiency of redundant information filtering, and poor adaptability to dynamic scenarios in marine satellite IoT perception, realize spatiotemporal collaborative modeling and dynamic weight optimization of multimodal ocean data, significantly improve the extraction accuracy of key features in marine environmental monitoring, the timeliness of sudden disaster warnings, and the real-time processing capabilities of resource-constrained devices in edge computing scenarios, and provide highly reliable data support for marine resource development, ecological protection, and disaster prevention and control. Summary of the Invention

[0011] In view of this, the present invention provides a spatiotemporal intelligent ocean satellite Internet of Things perception information screening method and device to eliminate or improve the defects existing in the prior art, such as weak spatiotemporal correlation of multi-source heterogeneous data in ocean satellite Internet of Things perception, low efficiency of redundant information filtering, poor adaptability to dynamic scenarios, etc., to realize spatiotemporal collaborative modeling, dynamic weight optimization and lightweight intelligent screening of multimodal ocean data, thereby improving the accuracy of ocean information screening control and reducing information redundancy.

[0012] The present invention provides a time series data repair method based on an improved convolutional bidirectional long short-term memory neural network. The method uses a bidirectional moving average (BiMA) model to pre-interpolate and repair missing data, uses a CNN-BiLSTM model to predict time series data, and smoothes the interpolated repair values obtained using the BiMA model to achieve smooth interpolation and repair of time series data. The method includes the following steps:

[0013] Pre-interpolation stage: First, use the two-way sliding average model to pre-interpolate the time series data of ocean satellite IoT perception, build an adaptive window with an initial length of m at the left end of the time series, and calculate the mean x of the m observations in this window. l If the number of observations in the window is less than m, then expand the window to the right until the number of observations in the window is equal to m, and calculate the mean x l The left end of the expanded time series is k x l Similarly, a window of length m is constructed at the right end of the time series, and the mean x of the m observations in this window is calculated. r If the number of observations in the window is less than m, expand the window to the left until the number of observations in the window is equal to m, and calculate the mean x r The right end of the expanded time series is k x r .

[0014] Prediction phase: A CNN-BiLSTM training dataset is constructed using pre-interpolated time series data and a rolling prediction approach. This involves using the first n data points to predict the n+1th data point, where n is the length of the time series fluctuation period. The time series data is converted into two-dimensional segments using a sliding window and fed into the CNN network. Local temporal features are extracted using a convolutional layer, and dimensionality reduction is achieved through a pooling layer. Subsequently, the BiLSTM network receives the feature sequence output by the CNN and simultaneously captures long-term forward and reverse dependencies using bidirectional long-short-term memory units, combining contextual information to model complex time series patterns. Finally, a fully connected layer maps the BiLSTM hidden state to the prediction result. Model training is completed through loss function optimization and backpropagation, enabling time series prediction of missing values in the time series data segments.

[0015] Iteration phase: To ensure that the model focuses more on fitting actual observations during training, allowing it to learn more realistic time series features, the network loss calculation method is adjusted. If the label is an actual observation, the original loss is used. If the label is an estimate, the loss is multiplied by t, where t∈(0,1) represents the confidence level in the estimate. The mean squared error (MSE) function is used to calculate the model loss. During the iteration process, the model with the lowest loss value is retained and its predictions are used to update the interpolation results.

[0016] During the smoothing phase, a fixed or dynamic window is first slid along the forward and reverse directions of the time series data to calculate the moving average of the data within the window. The forward sliding window calculates the local mean based on historical data to eliminate random noise, while the reverse sliding window uses the current point and subsequent adjacent points to capture potential trends. A weighted average is used to fuse the bidirectional calculation results, balancing the lag bias of the unidirectional moving average while preserving the overall data trend and local details. This effectively reduces padding errors, improves the consistency of the time series curve, and provides a low-noise, highly stable data foundation for subsequent analysis.

[0017] The two-way sliding average is characterized by constructing an adaptive window with an initial length of m at the left end of the time series and calculating the mean x of the m observations in this window. l If the number of observations in the window is less than m, then expand the window to the right until the number of observations in the window is equal to m, and calculate the mean x l The left end of the expanded time series is k x l Similarly, a window of length m is constructed at the right end of the time series, and the mean x of the m observations in this window is calculated. r If the number of observations in the window is less than m, expand the window to the left until the number of observations in the window is equal to m, and calculate the mean x r The right end of the expanded time series is k xr .

[0018] The convolutional bidirectional long short-term memory neural network is characterized by using a multi-scale convolution kernel group to extract local spatial features of the input data in parallel, outputting a feature map after compression through a pooling layer, inputting the feature map into a bidirectional long short-term memory network (BiLSTM), capturing long-term dependencies in forward and reverse time series respectively, generating a temporal latent state vector, adopting a gating mechanism to fuse the bidirectional latent states, and dynamically allocating feature weights through a self-attention layer to enhance the representation of key temporal nodes, mapping the weighted feature input into a fully connected layer to a target dimension, generating a final prediction result through an anti-normalization module, and suppressing overfitting through a Dropout layer. The network parameters are optimized in combination with a backpropagation algorithm to minimize the prediction error.

[0019] The present invention provides a spatiotemporal prediction method for marine environmental data based on an improved block time series transformer. The method utilizes a block mechanism to cut marine environmental data into local blocks according to time and space dimensions, utilizes the channel independence mechanism of PatchTST to capture multi-scale features, and combines the multi-feature data of the marine environment for multimodal fusion. The method utilizes a sparse attention mechanism to model long-range spatiotemporal dependencies, introduces physical constraints such as ocean dynamics equations in the decoding stage to improve the rationality of prediction, and finally generates future spatiotemporal sequences through recursive block partitioning to achieve efficient and accurate prediction of marine environmental parameters. The method comprises the following steps:

[0020] Sequence data block: We represent the i-th one-dimensional sequence as Each sequence is independently input into the Transformer backbone network. (i) First, it is divided into blocks, which can be overlapping or non-overlapping. The block length is recorded as P, and the non-overlapping area between two consecutive blocks is recorded as S. After the block process, a block sequence will be generated. Among them, N represents the number of blocks generated,

[0021] Transformer encoding: via trainable linear projection W P ∈R D×P , and a learnable additive positional encoding W pos ∈R D×N , mapping the blocks to a potential representation space of dimension D, the data input to the Transformer encoder is represented as in Then each attention head h=1,...,H of the multi-head attention transforms it into a query matrix Bond Matrix and the value matrix in The attention output is obtained using the scaled product as follows:

[0022]

[0023] After the attention output is processed by the normalization layer and the feedforward network layer with residual connection, the generated result is represented as z (i) ∈R D×N , and finally a fully connected layer with a linear head is used to obtain the prediction result, as shown in the following formula:

[0024]

[0025] Physical constraint decoding and recursive prediction: In the decoding stage, the ocean dynamics equations are introduced as physical constraints. Through residual connection or loss function design, the residual after decoupling the model prediction results and the physical laws is jointly optimized to ensure that the predicted velocity field, temperature field and other parameters conform to the basic laws of fluid mechanics. For example, when predicting the ocean velocity field data, assuming that the initial predicted velocity field output by the model decoder is u p , pressure field p p The physical residuals are calculated using the Navier-Stokes equations, constraining the predicted values to conform to the laws of fluid mechanics: Then introduce the physical residual to correct the predicted value, u final =u p -αR m , p final =p p -βR c Subsequently, a recursive block generation strategy is adopted to take the current prediction block as the input of the next time step and gradually generate future spatiotemporal sequences.

[0026] The block mechanism is characterized in that the model divides the input sequence into smaller subsequences, and then inputs these subsequences into the encoder in the Transformer backbone structure to calculate the self-attention between these blocks as a whole, thereby reducing the spatial and temporal complexity of the model and enabling the use of longer input sequences to extract meaningful temporal relationships.

[0027] The channel independence mechanism is characterized in that the PatchTST model applies attention weights to each channel separately, and the multivariate time series input data is regarded as a multi-channel signal, where each individual time series represents a different channel that encapsulates a specific signal.

[0028] The sparse attention mechanism is characterized by selective attention, focusing on and calculating only the relationships between elements that have a significant impact on model performance, while ignoring those with less significant impact. Dynamic pruning preserves key attention heads and local hashing maps high-dimensional features to low-dimensional buckets, significantly reducing the complexity of the attention mechanism.

[0029] The present invention provides a method for mining ocean hotspots based on an improved mean-shift density clustering algorithm. Based on the density gradient directional iteration mechanism of the mean-shift algorithm, the neighborhood radius parameter of DBSCAN is dynamically modified to capture high-density areas. The rough clustering results are further refined in combination with DBSCAN to ultimately extract ocean hotspots with significant statistical characteristics. The method comprises the following steps:

[0030] DBSCAN single-shot clustering: A point is randomly selected as the current observation point. The ε-neighborhood of the current point is then calculated based on the given neighborhood radius eps and the minimum number of points MinPts. If the number of points in the ε-neighborhood is at least MinPts, the current point is marked as a core point, and the points in these neighborhoods are placed in the candidate pool of core points. Otherwise, the point is marked as a noise point. For each core point, all points in its eps-neighborhood (including other core points) are connected to form a cluster. For each core point, the algorithm examines all points in its ε-neighborhood. For each newly discovered core point, if it is already in the candidate pool of core points, it is ignored; if not, it is added to the core point pool, and this expansion process is recursively repeated for the points in its ε-neighborhood. Through this process, all density-reachable core points can be discovered and grouped into the same cluster. Boundary points are assigned to the cluster of the core point to which they are connected.

[0031] Mean-Shift secondary clustering: randomly select a point as the center point, find all points within the bandwidth from the center point, record them as set M, and consider these points to belong to cluster C, that is, the density function calculation area in the first iteration is recorded as cluster C, and the access frequency of these points belonging to cluster C is +1; calculate the vector from the center point to each element in set M, add these vectors to get the offset vector Shift; move the center point along the direction of the offset vector, the moving distance is ||Shift||; repeat the steps until the size of the offset vector meets the set threshold requirement, record the center point at this time, and continue to repeat until all points are marked and classified.

[0032] The mean shift algorithm is characterized in that for a given d-dimensional space R d n sample points x in i (i=1,2...n), for any point x, its Mean-Shift vector is: Among them Sh It represents the data point whose distance from the point in the data set to x is less than the radius of the sphere h, that is: S h (x) = {y: (yx i ) T (yx i )<h 2 The introduction of the drift mean of the Gaussian kernel function can make the center point of the calculated distance have a greater weight, reflecting the characteristic that the shorter the distance, the greater the weight. The improved drift mean is: Where x is the center point; x i is the point within the range; n is the number of points within the range, and g(x) is the negative derivative of the kernel function. The drift process is to iteratively update the position of the sphere center by calculating the drift vector. The update formula is: x′=x+M h .

[0033] The density-based clustering algorithm is characterized by selecting any sample point M in the research object. If it is a core object, a circle is drawn with the sample point M as the center and ε as the radius. This circle forms a cluster X. Then, all points in the cluster are traversed to find the core object. If the sample point P is the object being sought, the sample points in P's ε-neighborhood are also included in the cluster X. The algorithm is recursively executed in sequence until the cluster X can no longer be expanded.

[0034] The present invention provides a marine environmental risk assessment method based on an improved sine-cosine algorithm to optimize a BP neural network. The method constructs a BP network structure optimized by the sine-cosine algorithm, globally optimizes the initial weights and thresholds of the BP neural network, and enhances the convergence of the algorithm by combining an adaptive parameter adjustment strategy. Finally, real-time monitoring data is input, and a risk probability distribution is output through a Sigmoid activation function to achieve a high-precision and robust dynamic risk assessment of the marine environment. The method comprises the following steps:

[0035] Determine the BP neural network topology: Select a three-layer BP neural network structure consisting of input, hidden, and output layers. Set the number of input and output nodes based on the number of sample eigenvalues. Determine the optimal number of hidden layer nodes using empirical formulas and the addition / subtraction method. With the number of input, output, and hidden layer nodes determined, calculate the initial weights and thresholds for the BP neural network, which serve as the spatial search dimension for each individual in the sine-cosine algorithm.

[0036] SCA algorithm optimization: The individuals in the SCA algorithm are mapped to the initial weight threshold of the BP neural network. The individuals calculate their fitness values through the fitness function. The fitness function is the mean square error function in the standard BP neural network training algorithm. By continuously comparing the fitness values of individuals at different positions, the positions of the individuals are gradually updated, and the optimal value is found and assigned to the BP network as the initial weight threshold.

[0037] Risk prediction: The SCA algorithm assigns the optimal value found as the weight threshold of the BP network. After network training, the SCA-BP model is constructed. The model is applied to risk level prediction, and the effectiveness of the model prediction is measured through evaluation indicators such as accuracy.

[0038] The sine-cosine algorithm is characterized in that, assuming the population size is N, the search space is n-dimensional, and the position of the i-th individual is expressed as: x i =(x i1 , x i2 ,…,x in ), i = 1, 2, ..., N. First, randomly generate N individual positions x1, x2, ..., x in the n-dimensional solution space. N , the fitness value of the individual is calculated by the objective function, and the fitness value of the i-th individual is expressed as Record the optimal individual position x in the population * :x * =arcsin f(x i ), i = 1, 2, ..., N, the j-th dimension of the i-th individual in the population is updated as follows:

[0039]

[0040] in is the current optimal individual position; r1 is the control parameter that determines the area of the next position (moving direction).

[0041] The BP network is characterized by being composed of an input layer, a hidden layer, and an output layer, and having a forward propagation phase of the signal, from the input layer through the hidden layer, and finally reaching the output layer. It also has a backward propagation phase of the error, from the output layer to the hidden layer, and finally to the input layer, and sequentially adjusting the weights and biases from the hidden layer to the output layer, and the weights and biases from the input layer to the hidden layer; in the forward propagation, the weights from the input layer to the hidden layer are From the hidden layer to the output layer is Where v and w are the weights from the input layer to the hidden layer and from the hidden layer to the output layer, d is the number of inputs, and x is the number of i is the input data, θ h is the activation function. In back propagation, the network parameters are adjusted by calculating the error between the output layer and the expected value, so that the error becomes smaller. The formula is Where E is the error, y k Output layer output result, T k is the true value.

[0042] The Sigmoid activation function is characterized by introducing nonlinearity into the model to prevent the neural network from being unable to solve the linear inseparable problem due to linear mapping. The original cascade features are reconstructed to obtain an enhanced feature map of cascade fusion, the feature value input is set to x, e is a natural constant, and the Sigmoid activation layer maps the feature value to between [0, 1], Sigmoid (x) = 1 / (1 + e^(-x)); the feature value obtained in this way can be regarded as the importance parameter corresponding to each feature channel, and finally the obtained parameter is applied to the original feature map by the inner product method to obtain an enhanced representation of the feature.

[0043] The present invention provides an information screening feedback control method based on improved multi-agent reinforcement learning. The method adopts a multi-agent reinforcement learning algorithm, deploys a node agent for each sensing node, describes the marine meteorological and hydrological information screening feedback control problem as a Markov intelligent game process, and uses a distributed decision-making centralized training method to obtain a decision network. This method realizes the transmission of full information of key data and the sending of statistical summaries of daily data, thereby completing the adaptive feedback control of the entire information screening process. The method comprises the following steps:

[0044] Multi-agent system architecture deployment: Node agents are deployed at each sensing node to collect and perform preliminary local data screening. Coordinating agents are set up in areas with the same risk level to achieve cross-node collaboration. Global optimization agents are set up to dynamically adjust network-wide screening strategies. Node agents directly receive preset regional risk levels and generate data priorities based on device status.

[0045] Dynamic screening rules driven by risk levels: Node agents perform differentiated collection based on risk levels: high-risk areas only transmit raw data of critical events exceeding the threshold, while medium- and low-risk areas use compressed sensing technology to reduce redundancy; at the same time, a priority score is calculated for each piece of data, and full information is uploaded for high-scoring data, while statistical summaries are sent for low-scoring data.

[0046] Markov intelligent game modeling: The marine meteorological and hydrological information screening feedback control problem is modeled as a partially observable Markov game POSG, and the state vector s is defined. t =[E t , H t , C t ], where E t is the environmental observation data, H t To encode historical information, C t To control the target constraints. Each agent corresponds to a region or a type of sensor node, and the local action Indicates whether to filter the node data. The reward function is designed as a multi-objective reward r t=αr accuracy +βr efficiency -γr cost , where α, β, and γ are dynamically adjusted according to task requirements.

[0047] Multi-agent reinforcement learning: proposes a knowledge distillation transfer mechanism and designs a teacher-student network architecture. After the teacher network is trained centrally, the global strategy is distilled to the local strategy network of each agent through KL divergence loss. The Priority Experience Replay (PER) mechanism is used, based on the TD error δ t =|Q target -Q pred |The experience samples are graded and high-value transitions are replayed first. The centralized training-distributed execution (CTDE) framework is adopted. During the training phase, the central coordinator integrates the global state S t With joint action A t , the QMIX algorithm is used to decompose the global Q value to each agent.

[0048] Feedback control and dynamic optimization: The intelligent agent dynamically adjusts observation weights based on regional importance. The inner loop adjusts data screening actions in real time based on the MARL decision network. The outer loop dynamically updates the reward function weights α, β, and γ based on the system state, automatically adjusting the collection frequency and transmission priority of data according to real-time changes in risk levels. This ensures that high-value information can be aggregated to the data center in a timely and effective manner, facilitating data analysis and risk prediction and prevention.

[0049] The Markov game process is characterized by describing the relationship between cooperation and competition among multiple agents. Assume that in a Markov game process involving n agents, each agent i observes the environment state S as O i , choose to take action A i The strategy is described as π i :O i ×A i , the next transformation of the environment is T:S×A1×…A n Depends on the actions of all agents. Each agent obtains a benefit r based on the state of the environment, that is, its own actions. i :S×A i , the total benefit of each agent γ is the discount rate, t is any time, and T is the termination time. The final goal is to find the expected reward of each agent under the connection strategy so that the strategy selected by any agent is the best when the strategies of other agents are determined.

[0050] The multi-agent reinforcement learning is characterized in that a centralized critic network is established for each node agent of the perception node, which can obtain global information, including the global state and the actions of all agents, and give the corresponding value function Q i (x,a1,…,a n Each agent's actor network only needs to make decisions based on local observation information. This enables distributed control of multiple agents, allowing sensing nodes in high-risk areas to transmit key information in a timely manner, while sensing nodes in low-risk areas collect and synthesize information, reducing data redundancy and significantly reducing transmission load while ensuring data validity.

[0051] The beneficial effects of the present invention are at least:

[0052] The present invention provides a spatiotemporal intelligent ocean satellite Internet of Things perception information screening method and device, belonging to the field of data processing technology. The method is based on an improved convolutional bidirectional long short-term memory neural network time series data repair method, providing a more reliable basis for subsequent data analysis and prediction. A spatiotemporal prediction method for ocean environmental data based on an improved block time series transformer captures multi-scale features through the block time series transformer to achieve efficient and accurate prediction of ocean environmental parameters. A method for mining ocean hotspot areas based on an improved mean shift density clustering algorithm uses a density clustering algorithm and a mean shift algorithm to extract ocean hotspot areas with significant statistical characteristics. A method for assessing ocean environmental risk based on an improved sine-cosine algorithm to optimize BP neural network constructs a BP network structure optimized using the sine-cosine algorithm to achieve high-precision and robust dynamic risk assessment of the ocean environment. An information screening feedback control method based on an improved multi-agent reinforcement learning adopts a distributed decision-making and centralized training framework to achieve low-redundancy and high-precision adaptive feedback control effects.

[0053] Additional advantages, objects, and features of the present invention will be set forth in part in the following description and will become apparent to those skilled in the art upon examination of the following or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained by the structures particularly pointed out in the description and drawings.

[0054] Those skilled in the art will understand that the purposes and advantages that can be achieved by the present invention are not limited to the above specific descriptions, and the above and other purposes that can be achieved by the present invention will be more clearly understood based on the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The drawings described herein are used to provide a further understanding of the present invention, constitute a part of this application, and do not constitute a limitation of the present invention. In the drawings:

[0056] Figure 1 The present invention is a schematic diagram of the steps of a spatiotemporal intelligent ocean satellite Internet of Things perception information screening method in one embodiment of the present invention.

[0057] Figure 2 The figure is a flow chart of a spatiotemporal intelligent ocean satellite IoT sensing information screening method in one embodiment of the present invention.

[0058] Figure 3 The figure is a flowchart of a time series data repair method based on a neural network in one embodiment of the present invention.

[0059] Figure 4 The figure is a flow chart of a method for spatiotemporal prediction of marine environmental data according to an embodiment of the present invention.

[0060] Figure 5 The figure is a flow chart of a method for mining ocean hotspot areas according to one embodiment of the present invention.

[0061] Figure 6 The figure is a flow chart of a method for marine environmental risk assessment in one embodiment of the present invention.

[0062] Figure 7 The figure is a flow chart of an information screening feedback control method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments and the accompanying drawings. Here, the exemplary embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0064] It should also be noted that, in order to avoid obscuring the present invention due to unnecessary details, the accompanying drawings only show structures and / or processing steps closely related to the solutions according to the present invention, while other details that are not closely related to the present invention are omitted.

[0065] It should be emphasized that the term "include / comprises" when used herein refers to the existence of features, elements, steps or components, but does not exclude the existence or addition of one or more other features, elements, steps or components.

[0066] It should also be noted that, unless otherwise specified, the term "connection" herein may refer not only to a direct connection but also to an indirect connection involving an intermediate.

[0067] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In the accompanying drawings, the same reference numerals represent the same or similar components, or the same or similar steps.

[0068] In order to solve the contradiction between the high-redundancy transmission of multi-source heterogeneous data in ocean observation and perception and the allocation of sensor network resources, as well as the problem of insufficient accuracy in real-time screening of key information in a dynamic spatiotemporal environment, the present invention provides a spatiotemporal intelligent ocean satellite Internet of Things perception information screening method and device. The method realizes ocean environment time series data repair based on a convolutional bidirectional long short-term memory neural network, performs ocean environment data time series prediction based on a block time series transformer, and mines ocean hotspot areas based on a density clustering algorithm with a mean shift. The ocean environment data includes perception data such as sea surface temperature, salinity, and wave height. The BP neural network is used to perform ocean environment risk assessment and multi-agent reinforcement learning is used to perform information screening and feedback control, which meets the efficient information processing requirements of complex ocean environment detection scenarios and achieves low redundancy and high-precision adaptive feedback control effects.

[0069] like Figure 1 As shown, the method includes the following steps S101 to S105:

[0070] Step S101: Receive user requests and interpolate missing values in the time series of ocean observation data collected by multiple sensors. Pre-interpolate missing data to generate a preliminary complete sequence. Then, use a neural network model to mine spatiotemporal features, iteratively predict and correct the pre-interpolation results, and perform secondary smoothing to eliminate local noise interference and improve the global consistency of the data.

[0071] Step S102: Time series prediction of marine environmental data is performed in the temporal dimension to achieve an accurate assessment of the multi-factor impacts on the marine environment. A block-based time series transformer is used to segment multidimensional data such as sea surface temperature and salinity into local blocks based on the temporal and spatial dimensions. A sparse attention mechanism is used to model long-range spatiotemporal dependencies. This is combined with a constrained decoding process based on ocean dynamics equations to generate future spatiotemporal sequences for each data set.

[0072] Step S103: Spatial hotspot clustering is performed on the high-precision navigation and positioning data in the spatial dimension to achieve regional division and pattern recognition of the maritime space. A clustering algorithm is used to divide the navigation and positioning data into independent clusters based on density, forming hotspot area boundaries. Low-density areas are marked as noise points, achieving spatial hotspot ranking.

[0073] Step S104: Based on the varying requirements for real-time marine environmental data transmission and the acquisition of specific environmental data elements, a marine environmental risk assessment is conducted using the results of time series data prediction and spatial hotspot clustering. Through the neural network structure, fluctuations in time series data and hotspots in spatial regions are input into the risk assessment model as influencing factors, identifying risk outcomes such as abnormal marine weather, ship collisions, and accidents.

[0074] Step S105: Combine ocean observation and perception data with ocean risk assessment results to select key impact information categories. Based on a multi-agent reinforcement learning framework, key node data is selected to achieve low-redundancy, high-precision adaptive feedback control effects.

[0075] In the present invention, the ocean satellite perception data consists of ocean environment time series data and marine high-precision navigation and positioning data.

[0076] like Figure 2 The figure shows a flow chart of a method and device for screening ocean satellite IoT perception information with spatiotemporal intelligence.

[0077] In step S101, the ocean satellite sensing data is preprocessed to fill the data gaps caused by various abnormal situations, and the ocean satellite sensing data is normalized in time series.

[0078] Specifically, observational data from ocean observation and detection platforms often contain anomalies, erroneous information, and useless information due to factors such as inclement weather, sensor signal misalignment, and equipment failure. First, we use physical thresholds and statistical methods to remove outliers that significantly deviate from the normal range. We also combine contextual information to identify systematic errors caused by inclement weather or equipment failure. Second, we use interpolation algorithms to fill data gaps caused by misaligned or missing signals, ensuring the continuity of the time series. We also use sliding average techniques to smooth noisy data. Finally, we standardize or normalize the data to form a standardized dataset that is continuous in time and space and has controllable quality. This provides reliable input for marine environmental data prediction and spatial clustering.

[0079] In step S102, time series prediction is performed on the marine environmental data in the time dimension to achieve an accurate assessment of the impact of multiple factors on the marine environment.

[0080] Specifically, a block-based time series transformer architecture is designed, consisting of an input layer, a Transformer encoding layer, a decoding layer, and an output layer. Data is then fed into the model, preprocessed and chronologically arranged into time series data. A sliding window approach is used to construct input-output sample pairs. Through training and testing of the Transformer backbone network, the fluctuation amplitude of marine environmental data over time is output, and the stability or abnormal changes of the data are determined, thereby achieving a marine environmental fluctuation rating.

[0081] In step S103, spatial hotspot area clustering is performed on the high-precision navigation and positioning data in the spatial dimension to achieve regional division and pattern recognition of the maritime space.

[0082] Specifically, the density clustering algorithm DBSCAN is used to identify high-frequency maritime activity areas, such as port anchorages and channel intersections. First, the neighborhood radius eps and the minimum number of neighborhood samples minPts are determined using the K-distance curve method to adapt to the sparse and dense characteristics of ship spatial distribution. The density accessibility of ship locations is then calculated based on Euclidean distance. Continuous areas that meet the density conditions are divided into independent clusters to form hotspot area boundaries. Low-density areas are marked as noise points. The positioning data is then subjected to preliminary spatial clustering. The clustering results are then used to calculate the centers of each cluster using the Mean-Shift algorithm to identify hotspot centers. Finally, a spatial hotspot map is generated by combining spatial density features to achieve spatial hotspot rating.

[0083] In step S104, the ocean environment fluctuation is evaluated based on the predicted ocean environment time series data, and the spatial hotspot level is evaluated based on the ocean spatial hotspot map obtained by clustering. The two types of data are combined to further evaluate the ocean environment risk.

[0084] Specifically, a BP neural network optimized by the sine-cosine algorithm (SCA) is constructed. First, global parameter optimization is performed to initialize the BP network weight W and threshold b. Then, a global search is performed in the solution space through SCA: the individual position update formula is: W new =W+r1sin(r2)|r3P best -W|, where r1, r2, r3 are random numbers, P best is the historical optimal solution. The iterative minimization loss function is: L=∑(y true -y pred ) 2 +λ||W|| 2 The input layer integrates the time series prediction results and spatial hotspot intensity, and the hidden layer uses an adaptive learning rate To accelerate convergence, the output layer generates risk probability through the Sigmoid function.

[0085] In step S105, the current ocean observation data and environmental risk assessment results are introduced, and a multi-agent reinforcement learning framework is used to dynamically allocate resources for the ocean perception data.

[0086] Specifically, the feedback control problem of marine meteorological and hydrological information screening is modeled as a Markov multi-agent game process, in which observation nodes such as buoys, satellites, and unmanned ships serve as independent agents. Their state space is defined as the current sensor data quality, network load, energy consumption level, and environmental dynamic characteristics. The action space includes sampling frequency adjustment, data transmission priority selection, and anomaly detection threshold optimization. Distributed execution-centralized training is adopted, and each agent makes distributed decisions based on local observations. At the same time, the global state is aggregated through a central coordinator to perform policy gradient updates. Based on the multi-agent deep deterministic policy gradient algorithm, a joint reward function R = α×x1-β×x2-γ×x3 is designed. Among them, x1 is the information value, x2 is the communication energy consumption, and x3 is the network delay. During the training process, the experience replay pool is used to store the state transition tuple (s t , a t , r t , s t +1), and update the policy network parameters through the target network delay to solve the policy oscillation problem in non-steady-state environments.

[0087] In some embodiments, ocean sensing time series data repair is performed based on the time series data repair method provided by the present invention to obtain complete and smooth time series data, such as Figure 3 As shown, the process specifically includes the following steps S301 to S303:

[0088] Step S301: Receive oceanographic and hydrological time-series data collected by multiple sensors, including buoys, satellites, and underwater robots. Parse the data format and extract the timestamp, spatial coordinates, sensor type, and measurement values. Perform a preliminary data quality check, marking missing segments, outliers, and noise points. Outliers include points outside the physically reasonable range and points with significant abrupt changes. Noise points represent transient sensor failure data.

[0089] Step S302: For the missing data segment, a two-way moving average interpolation model is constructed, a sliding window is constructed on the time series data, and the mean value in the window is selected as the interpolation value. That is, an adaptive window with an initial length of m is constructed at the left end of the time series, and the mean x of the m observations in this window is calculated. l If the number of observations in the window is less than m, then expand the window to the right until the number of observations in the window is equal to m, and calculate the mean x l The left end of the expanded time series is k x l Similarly, a window of length m is constructed at the right end of the time series, and the mean x of the m observations in this window is calculated. r If the number of observations in the window is less than m, expand the window to the left until the number of observations in the window is equal to m, and calculate the mean x r The right end of the expanded time series is k x r.

[0090] Step S303: Use the pre-interpolated data and adopt a rolling prediction method to construct a training data set for CNN-BiLSTM, that is, use the first n data to predict the n+1th data, and the length of n is set to the length of the time series fluctuation period.

[0091] Step S304: Adjust the network loss calculation method. If the label is an actual observation, the original loss is used. If the label is an estimate, the loss is multiplied by t, where t∈(0,1) represents the confidence level in the estimate. The mean squared error (MSE) function is used to calculate the model loss. During the iteration process, the model with the minimum loss value is retained and the prediction value of this model is used to update the interpolation result.

[0092] Step S305: To avoid large fluctuations in the interpolation results obtained by CNN-BiLSTM, BiMA is used to smooth the interpolated values obtained. The smoothed results can be adjusted again.

[0093] like Figure 4 As shown, the spatiotemporal prediction method of ocean environment data based on the improved block time series converter of the present invention specifically includes the following steps S401 to S404:

[0094] Step S401: Express the i-th one-dimensional sequence as Each sequence is independently input into the Transformer backbone network. (i) First, it is divided into blocks. These blocks can be overlapping or non-overlapping. The block length is recorded as P, and the non-overlapping area between two consecutive blocks is recorded as S. After the block process, a block sequence will be generated. Among them, N represents the number of blocks generated,

[0095] Step S402: Through the trainable linear projection W P ∈R D×P , and a learnable additive positional encoding W pos ∈R D ×N , mapping the blocks to a potential representation space of dimension D, the data input to the Transformer encoder is represented as

[0096]

[0097] in

[0098] Then each attention head h=1,...,H of the multi-head attention transforms it into a query matrix Bond Matrix and the value matrix in The attention output is obtained using the scaled product as follows:

[0099]

[0100] After the attention output is processed by the normalization layer and the feedforward network layer with residual connection, the generated result is represented as z (i) ∈R D×N , and finally a fully connected layer with a linear head is used to obtain the prediction result, as shown in the following formula:

[0101]

[0102] Step S403: In the decoding stage, the ocean dynamics equations are introduced as physical constraints. Through residual connection or loss function design, the model prediction results are jointly optimized with the residuals after decoupling the physical laws to ensure that the predicted flow velocity field, temperature field and other parameters conform to the basic laws of fluid mechanics.

[0103] For example, when predicting ocean velocity field data, it is assumed that the initial predicted velocity field output by the model decoder is u p , pressure field p p The physical residuals are calculated using the Navier-Stokes equations, constraining the predicted values to conform to the laws of fluid mechanics:

[0104]

[0105] Then introduce the physical residual to correct the predicted value, u final =u p -αR m , p final =p p -βR c Subsequently, a recursive block generation strategy is adopted to take the current prediction block as the input of the next time step and gradually generate future spatiotemporal sequences.

[0106] like Figure 5 As shown, the ocean hotspot area mining method based on the improved mean shift density clustering algorithm provided by the present invention specifically includes the following steps S501 to S504:

[0107] Step S501: Determine the neighborhood radius (eps) and the minimum number of neighborhood samples (minPts) by using the K-distance curve method to adapt to the sparse and dense characteristics of the navigation data spatial distribution.

[0108] Step S502: Perform initial clustering based on DBSCAN, randomly select a point as the current inspection point, and then calculate the ε-neighborhood of the current point based on the given neighborhood radius eps and the minimum number of points MinPts. If the number of points in the v-neighborhood is not less than MinPts, the current point is marked as a core point, and the points in these neighborhoods are placed in the candidate pool of core points. Otherwise, the point is marked as a noise point. For each core point, all points in its eps-neighborhood (including other core points) are connected to form a cluster. For core points, the algorithm will examine all points in its ε-neighborhood. For each newly discovered core point, if the point is already in the core point candidate pool, it is ignored; if not, it is added to the core point pool, and this expansion process is recursively repeated for the points in its ε-neighborhood. Through this process, all density-reachable core points can be found and summarized into the same cluster. Boundary points are assigned to the cluster of the core points connected to them.

[0109] Step S503: Perform secondary clustering based on Mean-Shift. Randomly select a point as the center point, find all points within the bandwidth from the center point, record them as set M, and consider these points to belong to cluster C. That is, the density function calculation area in the first iteration is recorded as cluster C. At the same time, the access frequency of these points belonging to cluster C is increased by 1; calculate the vector from the center point to each element in set M, add these vectors to obtain the offset vector Shift; move the center point along the direction of the offset vector by a distance of ||Shift||; repeat the steps until the size of the offset vector meets the set threshold requirement, record the center point at this time, and continue to repeat until all points are marked and classified.

[0110] The present invention also provides a marine environmental risk assessment method based on an improved sine-cosine algorithm to optimize the BP neural network, such as Figure 6 As shown, the process specifically includes the following steps S601 to S603:

[0111] Step S601: Select a three-layer BP neural network structure consisting of an input layer, a hidden layer, and an output layer. The number of input and output nodes of the network is set based on the number of sample eigenvalues. The optimal number of hidden layer nodes is determined using an empirical formula and the addition and subtraction method. Based on the determined number of input, output, and hidden layer nodes, the initial weights and thresholds of the BP neural network are calculated, which are the spatial search dimensions of each individual in the sine-cosine algorithm.

[0112] Step S602: Map the individuals in the SCA algorithm to the initial weight threshold of the BP neural network. The fitness value of the individual is calculated through the fitness function. The fitness function is the mean square error function in the standard BP neural network training algorithm. By continuously comparing the fitness values of individuals at different positions, the positions of the individuals are gradually updated, and the optimal value is found and assigned to the BP network as the initial weight threshold.

[0113] Step S603: After the SCA algorithm assigns the optimal value found as the weight threshold of the BP network, the SCA-BP model is constructed after network training. The model is applied to risk level prediction, and the accuracy and other evaluation indicators are used to measure the prediction effect of the model.

[0114] like Figure 7 As shown, the information screening feedback control method based on improved multi-agent reinforcement learning provided by the present invention specifically includes the following steps S701 to S705:

[0115] Step S701: Deploy a node agent for each sensor node, responsible for local data collection and preliminary screening. A coordination agent is set up in areas with the same risk level to achieve cross-node collaboration. A global optimization agent is set up to dynamically adjust the network-wide screening strategy. The node agent directly receives the preset regional risk level and generates data priority based on the device status.

[0116] Step S702: Dynamic screening rules driven by risk level: Node intelligent agents perform differentiated collection according to risk level: high-risk areas only transmit the original data of critical events that exceed the threshold, and medium and low-risk areas use compressed sensing technology to reduce redundancy; at the same time, a priority score is calculated for each data, and the full information is uploaded for high-scoring data, and a statistical summary is sent for low-scoring data.

[0117] Step S703: Model the marine meteorological and hydrological information screening feedback control problem as a partially observable Markov game (POSG), and define the state vector

[0118] s t =[E t , H t , C t ]

[0119] Among them, E t is the environmental observation data, H t To encode historical information, C t To control the target constraints.

[0120] Each agent corresponds to a region or a type of sensor node, and local actions Indicates whether to filter the node data. The reward function is designed as a multi-objective reward

[0121] rt =αr accuracy +βr efficiency -γr cost

[0122] Among them, α, β, and γ are dynamically adjusted according to task requirements.

[0123] Step S704: Propose a knowledge distillation transfer mechanism and design a teacher-student network architecture. After the teacher network is trained centrally, the global strategy is distilled to the local strategy network of each agent through KL divergence loss. The priority experience replay (PER) mechanism is used. Based on the TD error δ t =|Q target -Q pred |The experience samples are graded and high-value transitions are replayed first. The centralized training-distributed execution (CTDE) framework is adopted. During the training phase, the central coordinator integrates the global state S t With joint action A t , the QMIX algorithm is used to decompose the global Q value to each agent.

[0124] Step S705: The intelligent agent dynamically adjusts the observation weight according to the importance of the region. The inner loop adjusts the data screening action in real time based on the MARL decision network. The outer loop dynamically updates the reward function weights α, β, and γ according to the system status, and automatically adjusts the collection frequency and the priority of data transmission according to the real-time risk level changes, ensuring that high-value information can be timely and effectively aggregated to the data center, which is conducive to data analysis and risk prediction and prevention.

[0125] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of a spatiotemporal intelligent ocean satellite Internet of Things perception information screening method and a spatiotemporal intelligent ocean satellite Internet of Things perception information screening system.

[0126] Corresponding to the above method, the present invention also provides a device, which includes a computer device, the computer device includes a processor and a memory, the memory stores computer instructions, the processor is used to execute the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the device implements the steps of the method described above.

[0127] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned edge computing server deployment method. The computer-readable storage medium may be a tangible storage medium, such as a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a floppy disk, a hard disk, a removable storage disk, a CD-ROM, or any other form of storage medium known in the art.

[0128] In summary, the present invention provides a spatiotemporal intelligent ocean satellite Internet of Things perception information screening method and device, which belongs to the field of data processing technology. The method realizes ocean environment time series data repair based on convolutional bidirectional long short-term memory neural network, performs spatiotemporal prediction of ocean environment data based on block time series transformer, and mines ocean hotspot areas based on density clustering algorithm with mean shift. The ocean environment data includes perception data such as sea surface temperature, salinity, and wave height. The BP neural network is used to perform ocean environment risk assessment and multi-agent reinforcement learning to perform information screening feedback control, which meets the efficient information processing requirements of complex ocean environment detection scenarios and realizes low redundancy, high-precision data acquisition and transmission adaptive feedback control effects. The invention can be directly applied to the ocean observation and detection facility networking system in the ocean satellite Internet of Things, and can provide strong support in both military and civilian applications. It has broad and important application prospects and value.

[0129] It should be understood by those skilled in the art that the various exemplary components, systems and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software or a combination of the two. Whether it is specifically performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention. When implemented in hardware, it can be, for example, an electronic circuit, an application specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc. When implemented in software, the elements of the present invention are programs or code segments that are used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link via a data signal carried in a carrier.

[0130] It should be understood that the present invention is not limited to the specific configurations and processes described above and illustrated in the figures. For the sake of brevity, a detailed description of known methods is omitted. In the above embodiments, several specific steps are described and illustrated as examples. However, the method of the present invention is not limited to the specific steps described and illustrated. Those skilled in the art may make various changes, modifications, and additions, or change the order of the steps after understanding the spirit of the present invention.

[0131] In the present invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or replace features of other embodiments.

[0132] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations to the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A method for screening marine IoT perception information based on spatiotemporal intelligence, characterized in that: The method comprises the following steps: The first step is to repair time series data based on an improved convolutional bidirectional long short-term memory neural network. Based on a joint repair method of bidirectional sliding average and neural network, the missing data is pre-interpolated through a sliding window to construct an initial dataset. The interpolation results are dynamically updated using a spatiotemporal joint prediction model, and the repair accuracy is optimized based on time series consistency constraints. The second step is to predict the spatiotemporal ocean environmental data based on the improved block time series transformer. The ocean environmental data is modeled in spatiotemporal blocks. The block spatiotemporal transformer is used to capture long-range dependencies. The prediction results are constrained by integrating prior knowledge of ocean dynamics. The recursive block generation mechanism is used to reduce error accumulation. The third step is to mine ocean hotspots based on an improved mean-shift density clustering algorithm. This adaptive density clustering method uses a sliding window to iteratively calculate local statistical features, combined with a gradient-optimized clustering strategy to extract hotspots with significant spatiotemporal correlations. The fourth step is to optimize the marine environmental risk assessment of the BP neural network based on the improved sine-cosine algorithm, enhance the neural network model with an intelligent optimization algorithm, implement the marine environmental risk probability distribution modeling through a dynamic parameter adjustment mechanism, and output multi-dimensional risk assessment results; The fifth step is to use the improved multi-agent reinforcement learning information screening feedback control and the multi-agent collaborative decision-making framework to screen key spatiotemporal features through the attention mechanism, and combine feedback control to achieve abnormal data filtering and dynamic optimization of transmission load.

2. The method for screening marine IoT perception information based on spatiotemporal intelligence according to claim 1 is characterized in that: A time series data repair method based on an improved convolutional bidirectional long short-term memory neural network uses a bidirectional moving average (BiMA) model to pre-interpolate and repair missing data, a CNN-BiLSTM model is used to predict time series data, and the interpolation and repair values obtained are smoothed using the BiMA model to achieve smooth interpolation and repair of time series data. The two-way sliding average constructs an adaptive window with an initial length of m at the left end of the time series and calculates the mean x of the m observations in this window. l If the number of observations in the window is less than m, then expand the window to the right until the number of observations in the window is equal to m, and calculate the mean x l The left end of the expanded time series is k x l Similarly, a window of length m is constructed at the right end of the time series, and the mean x of the m observations in this window is calculated. r If the number of observations in the window is less than m, expand the window to the left until the number of observations in the window is equal to m, and calculate the mean x r The right end of the expanded time series is k x r ; The convolutional bidirectional long short-term memory neural network uses a multi-scale convolution kernel group to extract local spatial features of the input data in parallel, outputs a feature map after compression through a pooling layer, inputs the feature map into a bidirectional long short-term memory network BiLSTM, captures long-term dependencies in forward and reverse time series, generates a temporal latent state vector, uses a gating mechanism to fuse the bidirectional latent states, and dynamically allocates feature weights through a self-attention layer to enhance the representation of key temporal nodes. The weighted feature input is mapped into a target dimension through a fully connected layer, and the final prediction result is generated through a denormalization module. At the same time, overfitting is suppressed through a Dropout layer, and the network parameters are optimized in combination with a backpropagation algorithm to minimize the prediction error.

3. The method for screening marine IoT perception information based on spatiotemporal intelligence according to claim 1 is characterized in that: A spatiotemporal prediction method for ocean environmental data based on an improved block time series transformer uses a block mechanism to cut ocean environmental data into local blocks according to time and space dimensions. The channel independence mechanism of PatchTST is used to capture multi-scale features, and multimodal fusion is performed by combining multi-feature ocean environmental data. A sparse attention mechanism is used to model long-range spatiotemporal dependencies. Physical constraints such as ocean dynamics equations are introduced in the decoding stage to improve the rationality of prediction. Finally, future spatiotemporal series are generated through recursive block partitioning to achieve efficient and accurate prediction of ocean environmental parameters. The block mechanism divides the input sequence into smaller subsequences, and then feeds these subsequences into the encoder in the Transformer backbone structure to calculate the self-attention between these blocks as a whole, thereby reducing the spatial and temporal complexity of the model and enabling the extraction of meaningful temporal relationships using longer input sequences. The channel independence mechanism, the PatchTST model applies attention weights to each channel separately, and the multivariate time series input data is regarded as a multi-channel signal, where each individual time series represents a different channel that encapsulates a specific signal; The sparse attention mechanism performs selective attention, focusing on and calculating only the relationships between elements that have a greater impact on model performance, while ignoring those with less impact. It retains key attention heads through dynamic pruning and maps high-dimensional features to low-dimensional buckets through local hashing, significantly reducing the complexity of the attention mechanism.

4. The method for screening marine IoT perception information based on spatiotemporal intelligence according to claim 1, characterized in that: A method for mining ocean hotspots based on an improved mean-shift density clustering algorithm. Based on the density gradient directional iteration mechanism of the mean-shift algorithm, the neighborhood radius parameter of DBSCAN is dynamically modified to capture high-density areas. This method is then combined with DBSCAN to refine the coarse clustering results, ultimately extracting ocean hotspots with significant statistical characteristics. The mean shift algorithm, for a given d-dimensional space R d n sample points x in i (i=1, 2...n), for any point x, its Mean-Shift vector is: Among them S h It represents the data point whose distance from the point in the data set to x is less than the radius of the sphere h, that is: S h (x) = {y:(yx i ) T (yx i ) <h 2 The introduction of the drift mean of the Gaussian kernel function can make the center point of the calculated distance have a greater weight, reflecting the characteristic that the shorter the distance, the greater the weight. The improved drift mean is: Where x is the center point; x i is a point within the range; n is the number of points within the range, g(x) is the negative derivative of the kernel function; the drift process is to iteratively update the position of the center of the sphere by calculating the drift vector, and the update formula is: x′=x+M h ; The density-based clustering algorithm selects any sample point M in the research object. If it is a core object, a circle is drawn with the sample point M as the center and ε as the radius. This circle forms a cluster X. Then, all points in the cluster are traversed to find the core object. If the sample point P is the object being sought, the sample points in P's ε-neighborhood are also included in the cluster X. The algorithm is recursively executed in sequence until the cluster X can no longer be expanded.

5. The method for screening marine IoT perception information based on spatiotemporal intelligence according to claim 1 is characterized in that: A marine environmental risk assessment method based on an improved sine-cosine algorithm to optimize BP neural network was developed. The sine-cosine algorithm was used to optimize the BP network structure, and the initial weights and thresholds of the BP neural network were globally optimized. An adaptive parameter adjustment strategy was combined to enhance the algorithm convergence. Finally, real-time monitoring data was input, and the risk probability distribution was output through the Sigmoid activation function, achieving a highly accurate and robust dynamic marine environmental risk assessment. The sine-cosine algorithm assumes that the population size is N, the search space is n-dimensional, and the position of the i-th individual is expressed as: x i =(x i1 , x i2 ,…,x in ), i=1,2,…,N; first, randomly generate N individual positions x1,x2,…,x N , the fitness value of the individual is calculated by the objective function, and the fitness value of the i-th individual is expressed as Record the optimal individual position x in the population * :x * =arcsinf(x i ), i=1,2,…,N, the jth dimension of the i-th individual in the population is updated as follows: in is the current optimal individual position; r1 is the control parameter that determines the area of the next position (movement direction); The BP network consists of an input layer, a hidden layer, and an output layer. It has a forward propagation phase for signals, which goes from the input layer through the hidden layer and finally reaches the output layer. It has a backward propagation phase for errors, which goes from the output layer to the hidden layer and finally to the input layer, and adjusts the weights and biases from the hidden layer to the output layer, and from the input layer to the hidden layer in turn. In the forward propagation, the weights and biases from the input layer to the hidden layer are From the hidden layer to the output layer is Where v and w are the weights from the input layer to the hidden layer and from the hidden layer to the output layer, d is the number of inputs, and x is the number of i is the input data, θ h is the activation function; in back propagation, the network parameters are adjusted by calculating the error between the output layer and the expected value, so that the error becomes smaller. The formula is Where E is the error, y k Output layer output result, T k is the true value; The Sigmoid activation function introduces nonlinearity into the model to prevent the neural network from being unable to solve the linear inseparability problem due to linear mapping; the original cascade features are reconstructed to obtain an enhanced feature map of cascade fusion, and the eigenvalue input is set to x, e is a natural constant, and the Sigmoid activation layer maps the value of the feature to between [0, 1], Sigmoid(x)=1 / (1+e^(-x)); the eigenvalue obtained in this way can be regarded as the importance parameter corresponding to each feature channel, and finally the obtained parameter is applied to the original feature map by the inner product method to obtain an enhanced representation of the feature.

6. The method for screening marine IoT perception information based on spatiotemporal intelligence according to claim 1 is characterized in that: An improved information screening feedback control method based on multi-agent reinforcement learning employs a multi-agent reinforcement learning algorithm, deploying a node agent for each sensing node. The problem of marine meteorological and hydrological information screening feedback control is described as a Markov intelligent game process. Node agents adopt differentiated collection methods based on risk levels. A decision network is obtained through distributed decision-making and centralized training. This method achieves full transmission of critical data and statistical summaries of daily data, completing adaptive feedback control of the entire information screening process. The Markov game process describes the relationship between cooperation and competition among multiple agents. Assume that in a Markov game process involving n agents, each agent i observes the environment state S as O i , choose to take action A i The strategy is described as π i :O i ×A i , the next transformation of the environment is T:S×A1×…A n Depends on the actions of all agents. Each agent obtains a benefit r based on the state of the environment, that is, its own actions. i :S×A i , the total benefit of each agent γ is the discount rate, t is any time, and T is the termination time. The final goal is to find the expected reward of each agent under the connection strategy so that the strategy selected by any agent is the best when the strategies of other agents are determined. The multi-agent reinforcement learning establishes a centralized critic network for each node agent of the perception node, which can obtain global information, including the global state and the actions of all agents, and give the corresponding value function Q i (x,a1,…,a n ); the actor network of each intelligent agent only needs to make decisions based on local observation information, which can realize distributed control of multiple intelligent agents, enable perception nodes in high-risk areas to transmit key information in a timely manner, and perception nodes in low-risk areas to collect comprehensive information, reduce data redundancy, and significantly reduce transmission load while ensuring data validity.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images