A spatiotemporal intelligent marine satellite internet of things perception information screening method and device

Abstract

The application provides a kind of spatio-temporal intelligent marine satellite internet of things sensing information screening method and device, belong to data processing technical field, this method is based on convolution bidirectional long short-term memory neural network realizes marine environment time series data repair, based on block time series transformer carries out marine environment data space-time prediction and based on mean shift density clustering algorithm carries out marine hotspot area mining, the marine environment data includes sea surface temperature, salinity, wave height and other sensing data, based on BP neural network carries out marine environment risk assessment and multi-agent reinforcement learning carries out information screening feedback control, meet the complex marine environment detection scene efficient information processing demand, realize low redundancy, high-precision data acquisition transmission adaptive feedback control effect.The present application can be directly applied to marine internet of things in marine observation detection facility networking system, can provide strong support in military application and civil application, has wide and important application prospect and value.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a spatiotemporal intelligent marine satellite IoT sensing information filtering method and device, which improves the accuracy of marine information filtering and control and reduces information redundancy by meeting the high-efficiency information processing requirements of complex marine environment detection scenarios. Background Technology

[0002] With the increasing demand for marine resource development and ecological protection, marine satellite IoT has achieved all-weather data collection of multi-dimensional data such as ocean temperature, salinity, current velocity, and biological activity through a three-dimensional sensing network including low-orbit satellites, buoys, submersible moorings, underwater sonar, and satellite remote sensing. However, the marine environment is highly dynamic, wide-ranging, and complex, and the sensed data exhibits characteristics such as multi-source heterogeneity (e.g., buoy point data, sonar sequences, remote sensing images), large differences in spatiotemporal scales (from minutes to days), and significant noise interference (e.g., ocean current disturbances, equipment signal attenuation). Buoy data collected by satellites is prone to drift errors due to surface ocean currents, underwater sonar suffers from unstable resolution due to deep-sea pressure and signal scattering limitations, and while satellite remote sensing has a wide coverage area, it is easily affected by meteorological conditions such as clouds and sea fog. Traditional information filtering methods are mostly based on single sensor data or static spatiotemporal models, which makes it difficult to effectively integrate the spatiotemporal correlation of multi-source data. This results in insufficient filtering of redundant information, bias in the extraction of key features, and an inability to adapt to the real-time analysis needs of dynamic scenarios such as tidal changes and sudden storms, which seriously restricts the accuracy of marine monitoring and disaster early warning.

[0003] Current technologies for filtering marine satellite IoT sensing information face several challenges. Firstly, data from different devices exhibit significant differences in spatiotemporal resolution, data format, and noise type. Traditional fusion algorithms are prone to information conflicts or insufficient utilization of complementarity due to spatiotemporal scale mismatches. Secondly, in dynamic marine scenarios, data distribution exhibits non-equilibrium characteristics due to abrupt environmental changes, and existing methods face bottlenecks in dynamic weight allocation and cross-modal correlation mining. Furthermore, traditional technologies have a weak ability to combine marine physical mechanisms with data characteristics, making it difficult to effectively suppress environmental noise and extract implicit spatiotemporal evolution patterns through data-driven models, resulting in insufficient real-time filtering capabilities for sudden events.

[0004] To address the above problems, this invention proposes a spatiotemporal intelligent method for filtering marine satellite IoT sensing information, which has the following innovations and advantages:

[0005] A time-series data restoration method based on an improved convolutional bidirectional long short-term memory neural network (CNN-BiLSTM) is proposed. This method employs a bidirectional moving average (BiMA) model to pre-imputate missing data. The pre-imputated data is then used to construct a training dataset for the CNN-BiLSTM neural network. The fully trained CNN-BiLSTM is then used to predict the missing data, and the predicted values ​​are used to update the imputation results. Finally, the BiMA model is used to smooth the obtained imputation values, thereby improving the global consistency of the data. This approach more accurately reflects the true time-series variation patterns and provides 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 (PatchTST) divides marine environmental data such as sea surface temperature and salinity into local blocks according to time and space dimensions. Multi-scale features are captured using the PatchTST transformer, and multi-modal fusion is performed by combining multiple marine environmental feature data. A sparse attention mechanism is used to model long-range spatiotemporal dependencies, and physical constraints from ocean dynamic equations are introduced during the decoding stage to improve the predictive rationality. Finally, future spatiotemporal sequences are generated through recursive block division, achieving efficient and accurate prediction of marine environmental parameters while reducing error accumulation and maintaining computational efficiency.

[0007] A method for mining marine hotspots based on an improved mean-drift density clustering algorithm is proposed. The mean-drift algorithm (MS) is used to adaptively determine the initial cluster centers. 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 regions. Subsequently, the density clustering algorithm (DBSCAN) is combined to refine the coarse clustering results, and finally, marine hotspots with significant statistical characteristics are extracted.

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

[0009] An improved information filtering feedback control method based on multi-agent reinforcement learning is used to model the problem of multi-source information collaborative processing as a Markov game process. A distributed decision-making and centralized training framework is adopted, and key information filtering among agents is achieved through a two-level attention mechanism. Finally, it is applied to real-time data filtering, anomaly detection and collaborative control in dynamic environments, which meets the high-efficiency information processing requirements of complex marine environment detection scenarios and achieves an adaptive feedback control effect that significantly reduces transmission load while ensuring data validity.

[0010] Through these innovations, this 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 sensing. It enables spatiotemporal collaborative modeling and dynamic weight optimization of multimodal marine data, significantly improving the extraction accuracy of key features in marine environmental monitoring, the timeliness of early warning of sudden disasters, and the real-time processing capability of resource-constrained devices in edge computing scenarios, providing 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 spatiotemporally intelligent marine satellite IoT sensing information filtering method and device to eliminate or improve the defects existing in the prior art, such as weak spatiotemporal correlation of multi-source heterogeneous data, low efficiency of redundant information filtering, and poor adaptability to dynamic scenes in marine satellite IoT sensing. It realizes spatiotemporal collaborative modeling, dynamic weight optimization and lightweight intelligent filtering of multimodal marine data, thereby improving the accuracy of marine information filtering and control and reducing information redundancy.

[0012] This invention provides a time-series data restoration method based on an improved convolutional bidirectional long short-term memory neural network. It employs a bidirectional moving average (BiMA) model to pre-imputate missing data, uses a CNN-BiLSTM model to predict the time-series data, and smooths the obtained imputation values ​​using the BiMA model to achieve smooth imputation restoration of the time-series data. The method includes the following steps:

[0013] Pre-interpolation stage: First, a two-way moving average model is used to pre-interpolate the time-series data of the marine satellite IoT sensing. An adaptive window with an initial length of m is constructed at the left end of the time series, and the mean of the m observations within this window is calculated. 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 equals m, and then calculate the mean. The left end of the augmented time series is k. Similarly, construct a window of length m at the right end of the time series, and calculate the mean of the m observations within this window. If the number of observations in the window is less than m, then expand the window to the left until the number of observations in the window equals m, and then calculate the mean. The right end of the augmented time series is k. .

[0014] Prediction Phase: A CNN-BiLSTM training dataset is constructed using pre-impregnated time-series data and a rolling prediction approach. This involves using the first n data points to predict the (n+1)th data point, where n is the length of the time series fluctuation period. The time-series data is converted into two-dimensional segments via a sliding window and input into the CNN network. Convolutional layers extract local temporal features, and pooling layers reduce dimensionality. Subsequently, the BiLSTM network receives the feature sequences output by the CNN, capturing both forward and backward long-term dependencies through bidirectional long short-term memory units, and modeling complex time-series patterns using contextual information. Finally, fully connected layers map the hidden states of the BiLSTM to the prediction results, and model training is completed through loss function optimization and backpropagation, enabling temporal prediction of missing values ​​in time-series data segments.

[0015] Iteration Phase: To ensure the model focuses more on fitting actual observations during training, thus learning more real-world time-series features, the network's loss calculation method is adjusted. If the label is the actual observation, the original loss is used; if the label is an estimated value, the loss is calculated as t times the original value. This indicates the degree of confidence in the estimated value. The mean squared error (MSE) function is used to calculate the model's loss. During the iteration process, the model with the smallest loss value is retained, and the predictions of this model are used to update the imputation results.

[0016] Smoothing stage: First, a fixed or dynamic window slides along the forward and reverse directions of the time series data, respectively, calculating the moving average of the data within each window. The forward sliding window calculates a local mean based on historical data to eliminate random noise, while the reverse sliding window uses the current point and its subsequent adjacent points to capture potential trends. By fusing the bidirectional calculation results through weighted averaging, the lag bias of the unidirectional moving average is balanced, while preserving the overall trend and local details of the data. This effectively suppresses imputation errors, improves the coherence of the time series curve, and provides a low-noise, highly stable data foundation for subsequent analysis.

[0017] The bidirectional moving average is characterized by constructing an adaptive window of initial length m at the left end of the time series, and calculating the mean of the m observations within this window. 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 equals m, and then calculate the mean. The left end of the augmented time series is k. Similarly, construct a window of length m at the right end of the time series, and calculate the mean of the m observations within this window. If the number of observations in the window is less than m, then expand the window to the left until the number of observations in the window equals m, and then calculate the mean. The right end of the augmented time series is k. .

[0018] The convolutional bidirectional long short-term memory neural network is characterized by using multi-scale convolutional kernel groups to extract local spatial features from the input data in parallel, compressing them through pooling layers to output feature maps, inputting the feature maps into a bidirectional long short-term memory network (BiLSTM), capturing long-term dependencies using forward and reverse time series respectively, generating temporal hidden state vectors, employing a gating mechanism to fuse the bidirectional hidden states, and dynamically allocating feature weights through a self-attention layer to enhance the representation of key temporal nodes, mapping the weighted features into a fully connected layer to the target dimension, generating the final prediction result through an inverse normalization module, suppressing overfitting through a Dropout layer, and optimizing network parameters using a backpropagation algorithm to minimize prediction error.

[0019] This invention provides a spatiotemporal prediction method for marine environmental data based on an improved block time series transformer (BST). The method utilizes a block-based mechanism to segment marine environmental data into local blocks according to time and spatial dimensions. It employs the channel independence mechanism of PatchTST to capture multi-scale features and combines multi-feature marine environmental data for multimodal fusion. A sparse attention mechanism is used to model long-range spatiotemporal dependencies. Physical constraints from ocean dynamics equations are introduced during the decoding stage to improve the predictive validity. Finally, future spatiotemporal sequences are generated through recursive block segmentation, achieving efficient and accurate prediction of marine environmental parameters. The method includes the following steps:

[0020] Sequence data block division: We will divide the first... A one-dimensional sequence is represented as Each sequence is independently input into the Transformer backbone network. Each input is a one-dimensional time series. First, the data is divided into patches, which can be overlapping or non-overlapping. Let P be the patch length and S be the non-overlapping region between two consecutive patches. After the patching process, a patch sequence will be generated. Where N represents the number of blocks generated. .

[0021] Transformer encoding: through trainable linear projections and learnable addition position encoding Mapping the blocks to a latent representation space of dimension D, the data input to the Transformer encoder is then represented as follows: ,in Then each attention head in multi-head attention Convert it into a query matrix Key matrix and value matrix ,in The attention output obtained using scaling products is shown in the following formula:

[0022]

[0023] After processing the attention output through a normalization layer and a feedforward network layer with residual connections, the resulting output is represented as follows: Finally, a fully connected layer with a linear head is used to obtain the prediction result, as shown in the following equation:

[0024]

[0025] Physical Constraint Decoding and Recursive Prediction: In the decoding stage, 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 from physical laws, ensuring that the predicted velocity and temperature field parameters conform to the basic laws of fluid mechanics. For example, when predicting ocean velocity field data, assume that the initial predicted velocity field output by the model decoder is... Pressure field The physical residuals were calculated using the Navier-Stokes equations, and the constrained predicted values ​​conformed to the laws of fluid dynamics. Then, physical residuals are introduced to correct the predicted values. Subsequently, a recursive block generation strategy is adopted, using the current prediction block as the input for the next time step to gradually generate the future spatiotemporal sequence.

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

[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, wherein each individual time series represents a different channel encapsulating a specific signal.

[0028] The sparse attention mechanism is characterized by selective attention, focusing only on and calculating the relationships between elements that have a significant impact on model performance, while ignoring elements 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.

[0029] This invention provides a method for mining marine hotspot regions based on an improved mean-drift density clustering algorithm. Based on the density gradient direction iteration mechanism of the mean-drift algorithm, the neighborhood radius parameter of DBSCAN is dynamically adjusted to capture high-density regions. Furthermore, DBSCAN is combined to refine the coarse clustering results, ultimately extracting marine hotspot regions with significant statistical characteristics. The method includes the following steps:

[0030] DBSCAN single-stage clustering: Randomly select a point as the current observation point, and then calculate the value of the current point based on the given neighborhood radius eps and the minimum number of points MinPts. - Neighborhood. If the If the number of points in its neighborhood is at least MinPts, then the current point is marked as a core point, and the points in its neighborhood are added to 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 their... - All points within the neighborhood. For each newly discovered core point, if it is already in the core point candidate pool, it is ignored; otherwise, it is added to the core point pool, and this process is repeated recursively. - This expansion process is repeated for points within the neighborhood. Through this process, all density-reachable core points can be discovered and grouped into the same cluster. Boundary points are assigned to clusters of core points connected to them.

[0031] Mean-Shift quadratic clustering: Randomly select a point as the center point, find all points within the bandwidth of the center point, denoted as set M, and consider these points to belong to cluster C, i.e., the density function calculation region in the first iteration is denoted as cluster C. At the same time, increment the access frequency of these points belonging to cluster C by 1; calculate the vector from the center point to each element in set M, and 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 labeled and classified.

[0032] The mean shift algorithm is characterized in that, for a given d-dimensional space... n sample points For any point x, its Mean-Shift vector is: ,in This represents data points in the dataset whose distance from x is less than the radius h of the sphere, i.e.: Introducing a Gaussian kernel function to shift the mean allows the center point of the distance calculation to have a larger weight, reflecting the characteristic that the shorter the distance, the larger the weight. The improved shift mean is: , where x is the center point; Let n be the number of points within the range. To find the negative of the derivative of the kernel function. The drift process involves calculating the drift vector and iteratively updating the position of the sphere's center, using the following formula: .

[0033] The density-based clustering algorithm is characterized by selecting any sample point M from the research objects; if it is a core object, then sample point M is taken as the center. Draw a circle with a radius of 10 ... - The sample points in the neighborhood are also included in cluster X, and the process is repeated recursively until cluster X can no longer be expanded.

[0034] This invention provides a marine environmental risk assessment method based on an improved sine and cosine algorithm-optimized BP neural network. The method constructs a BP network structure optimized using the sine and cosine algorithms, globally optimizes the initial weights and thresholds of the BP neural network, and enhances the algorithm's convergence 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, achieving a high-precision and robust dynamic risk assessment of the marine environment. The method includes the following steps:

[0035] Determine the BP neural network topology: Select a three-layer BP neural network structure consisting of an input layer, hidden layers, and an output layer. Set the number of input and output nodes based on the number of sample feature values. Determine the optimal number of hidden layer nodes using empirical formulas and an increment / decrement method. Based on the determined number of input, output, and hidden layer nodes, calculate the initial weights and thresholds of the BP neural network; this represents the spatial search dimension for each individual node 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 fitness value of the individual is calculated by the fitness function, which 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 position of the individual is gradually updated, and the optimal value is found and assigned to the BP network as the initial weight threshold.

[0037] Risk prediction: After the SCA algorithm finds the optimal value and assigns it as the weight threshold of the BP network, the SCA-BP model is constructed through network training. The model is then applied to risk level prediction, and the prediction performance is measured by an accuracy evaluation metric.

[0038] The sine and cosine algorithm is characterized by assuming a population size of N, a search space of n dimensions, and the position of the i-th individual represented as: First, randomly generate the positions of N individuals in the n-dimensional solution space. The fitness value of an individual is calculated using an objective function, and the fitness value of the i-th individual is expressed as: Record the position of the best individual in the population. The j-th dimension of the i-th individual in the population is updated as follows:

[0039]

[0040] in This represents the current optimal individual position. To control parameters, determine the area for the next position.

[0041] The BP network is characterized by comprising an input layer, hidden layers, and an output layer. It includes a forward propagation phase, where the signal travels from the input layer through the hidden layers and finally reaches the output layer. It also includes a backward propagation phase, where the signal travels from the output layer to the hidden layers and finally to the input layer, with the weights and biases adjusted sequentially from the hidden layers to the output layer and from the input layer to the hidden layers. During forward propagation, the signal travels from the input layer to the hidden layers... 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, and d is the number of inputs. For input data, Let be the activation function. During backpropagation, the network parameters are adjusted by calculating the error between the output layer and the desired value, thereby reducing the error. The formula is: Where E is the error, The output layer outputs the results. This is the actual value.

[0042] The Sigmoid activation function is characterized by introducing nonlinearity into the model to prevent the neural network from being unable to solve linearly inseparable problems due to linear mapping. The original cascaded features are reconstructed to obtain a cascaded fused enhanced feature map. The feature value input is set to x, and e is a natural constant. The Sigmoid activation layer maps the feature values ​​to the range [0, 1], Sigmoid(x) = 1 / (1 + e^(-x)). The feature values ​​obtained in this way can be regarded as the importance parameters corresponding to each feature channel. Finally, the obtained parameters are combined with the original feature map through an inner product to obtain an enhanced representation of the features.

[0043] This invention provides an improved information filtering feedback control method based on multi-agent reinforcement learning. It employs a multi-agent reinforcement learning algorithm, deploying node agents at each sensing node. The marine meteorological and hydrological information filtering feedback control problem is described as a Markov intelligent game process. Node agents adopt differentiated data collection based on risk levels. A decision network is obtained through distributed decision-making and centralized training, enabling the transmission of full information for key data and the sending of statistical summaries of daily data. This completes the adaptive feedback control of the entire information filtering process. The method includes the following steps:

[0044] Multi-agent system architecture deployment: A node agent is deployed at each sensing node to handle local data collection and initial screening; a coordinating agent is set up in areas with the same risk level to achieve cross-node collaboration; and an optimization agent is set up globally to dynamically adjust the network-wide screening strategy. The node agent directly receives the preset regional risk level and generates data priorities based on the device status.

[0045] Risk level-driven dynamic screening rules: Node agents perform differentiated data collection based on risk level: only raw data of key events exceeding the threshold are transmitted in high-risk areas, while compressed sensing technology is used to reduce redundancy in medium- and low-risk areas; at the same time, priority scores are calculated for each data point, and full information is uploaded for high-scoring data, while statistical summaries are sent for low-scoring data.

[0046] Markov Intelligent Game Modeling: The problem of feedback control for screening marine meteorological and hydrological information is modeled as a partially observable Markov game (POSG), defining the state vector. ,in, For environmental observation data, Encoding historical information, To control the target constraints, each agent corresponds to a region or a type of sensor node, and local actions... Indicates whether to filter data for this node. The reward function is designed as a multi-objective reward. ,in Adjust dynamically according to task requirements.

[0047] Multi-agent reinforcement learning: A knowledge distillation and transfer mechanism is proposed, and a teacher-student network architecture is designed. After centralized training of the teacher network, the global policy is distilled to the local policy networks of each agent through KL divergence loss. A priority experience replay (PER) mechanism is adopted, based on TD error. Experience samples are sampled in a tiered manner, with high-value transitions prioritized for replay. A centralized training-distributed execution (CTDE) framework is adopted, with a central coordinator fusing global state during the training phase. With joint actions The QMIX algorithm is used to decompose the global Q value into individual agents.

[0048] Feedback control and dynamic optimization: The agent dynamically adjusts observation weights based on the importance of regions; the inner loop adjusts data filtering actions in real time based on the MARL decision network; the outer loop dynamically updates the reward function weights based on the system state. This allows for automatic adjustment of the data collection frequency and transmission priority based on real-time changes in risk levels, ensuring that high-value information can be promptly and effectively aggregated in the data center, which is beneficial for data analysis and risk prediction and prevention.

[0049] The Markov game process is characterized by describing the cooperative and competitive relationships among multiple agents. Let's assume that in a Markov game involving n agents, each agent i's observation of the environmental state S is... Choose to use action The strategy is described as follows The next shift in the environment Depending on the actions of all agents, each agent gains a reward based on the environmental state, i.e., its own actions. Total revenue for each agent , Let be the discount rate, t be any time, and T be the termination time; the ultimate goal is to find that the expected reward of each agent under the connection strategy satisfies the condition that the strategy chosen by any agent is optimal given that the strategies of other agents are determined.

[0050] The multi-agent reinforcement learning is characterized by establishing a centralized critic network for each perceptual node's node agent. This network can acquire global information, including the global state and the actions of all agents, and provide the corresponding value functions. Each agent's actor network only needs to make decisions based on local observation information. This enables distributed control of multiple agents, allowing perception nodes in high-risk areas to transmit key information in a timely manner, while perception 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 as follows:

[0052] This invention provides a spatiotemporally intelligent method and device for filtering marine satellite IoT sensing information, belonging to the field of data processing technology. The method is based on an improved convolutional bidirectional long short-term memory neural network for time-series data restoration, providing a more reliable foundation for subsequent data analysis and prediction. A spatiotemporal prediction method for marine environmental data based on an improved block time series transformer captures multi-scale features, achieving efficient and accurate prediction of marine environmental parameters. A method for mining marine hotspot areas based on an improved mean-shift density clustering algorithm extracts marine hotspot areas with significant statistical characteristics using density clustering and mean-shift algorithms. A method for marine environmental risk assessment based on an improved sine and cosine algorithm-optimized BP neural network constructs a BP network structure optimized by sine and cosine algorithms to achieve high-precision and robust dynamic risk assessment of the marine environment. An information filtering feedback control method based on improved multi-agent reinforcement learning employs a distributed decision-making and centralized training framework to achieve low-redundancy and high-precision adaptive feedback control.

[0053] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0054] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0055] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. In the drawings:

[0056] Figure 1 This is a schematic diagram illustrating the steps of a spatiotemporal intelligent marine satellite IoT sensing information filtering method in one embodiment of the present invention.

[0057] Figure 2 This is a flowchart illustrating a spatiotemporal intelligent marine satellite IoT sensing information filtering method according to an embodiment of the present invention.

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

[0059] Figure 4This is a flowchart illustrating a spatiotemporal prediction method for marine environmental data in one embodiment of the present invention.

[0060] Figure 5 This is a flowchart illustrating a method for identifying marine hotspot areas according to an embodiment of the present invention.

[0061] Figure 6 This is a flowchart illustrating a marine environmental risk assessment method according to an embodiment of the present invention.

[0062] Figure 7 This is a flowchart illustrating an information filtering feedback control method according to an embodiment of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

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

[0065] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence 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" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0067] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0068] To address the contradiction between the high redundancy transmission of multi-source heterogeneous data from marine observation and sensing, and the resource allocation issues in sensor networks, as well as the insufficient real-time filtering accuracy of key information in dynamic spatiotemporal environments, this invention provides a spatiotemporally intelligent marine satellite IoT sensing information filtering method and device. This method utilizes a convolutional bidirectional long short-term memory neural network for marine environmental time-series data restoration, a block-based time series transformer for marine environmental data time-series prediction, and a mean-shift-based density clustering algorithm for marine hotspot area mining. The marine environmental data includes sea surface temperature, salinity, and wave height sensing data. It employs a backpropagation neural network for marine environmental risk assessment and multi-agent reinforcement learning for information filtering feedback control, meeting the high-efficiency information processing requirements of complex marine environmental detection scenarios and achieving low-redundancy, high-precision adaptive feedback control. Figure 1 As shown, the method includes the following steps S101~S105:

[0069] Step S101: Receive user requests and impute missing values ​​in the time-series data of marine observation and sensing data collected by multi-source sensors. Pre-impute the missing data to generate a preliminary complete sequence; then, use a neural network model to mine spatiotemporal features, iteratively predict and correct the pre-impute results, and perform secondary smoothing to eliminate local noise interference and improve the global consistency of the data.

[0070] Step S102: Perform time series prediction on marine environmental data in the time dimension to achieve accurate assessment of the impact of multiple factors on the marine environment. A block time series transformer is used to cut multidimensional data such as sea surface temperature and salinity into local blocks according to the spatiotemporal dimension. Long-range spatiotemporal dependencies are modeled through a sparse attention mechanism, and the future spatiotemporal sequences of each data are generated by combining the ocean dynamics equations to constrain the decoding process.

[0071] Step S103: Perform spatial hotspot clustering 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 according to density, forming hotspot region boundaries. Simultaneously, low-density areas are marked as noise points to achieve spatial hotspot rating.

[0072] Step S104: Based on the real-time nature of marine environmental data transmission and the different needs for acquiring specific environmental data elements, a marine environmental risk assessment is conducted using the results of time series data prediction and spatial hotspot clustering. Through a neural network structure, fluctuations in time series data and spatial hotspots are input into the risk assessment model as influencing factors to identify abnormal marine weather and risks such as ship collisions and accidents.

[0073] Step S105: Combining marine observation and sensing data with marine risk assessment results, key impact information categories are screened. Based on a multi-agent reinforcement learning framework, key node data is screened to achieve low-redundancy, high-precision adaptive feedback control.

[0074] In this invention, the marine satellite sensing data consists of marine environmental time-series data and high-precision navigation and positioning data at sea.

[0075] like Figure 2 The diagram shown is a flowchart of a spatiotemporal intelligent marine satellite IoT sensing information filtering method and device.

[0076] In step S101, the marine satellite sensing data is preprocessed to fill data gaps caused by various anomalies, and the marine satellite sensing data is standardized in a time sequence.

[0077] Specifically, observational data from marine observation and monitoring platforms often contains anomalies, errors, and useless information due to factors such as severe weather, sensor signal inaccuracies, or equipment malfunctions. First, anomalies that significantly deviate from the normal range are removed based on physical thresholds and statistical methods, and systematic errors caused by severe weather or equipment failures are identified by combining contextual information. Second, interpolation algorithms are used to fill data gaps caused by signal inaccuracies or missing data, ensuring the continuity of the time series. Simultaneously, moving average techniques are used to smooth noisy data. Finally, the data is standardized or normalized to form a spatiotemporally continuous and quality-controllable standardized dataset, providing reliable input for marine environmental data prediction and spatial clustering.

[0078] In step S102, time series prediction is performed on marine environmental data in the time dimension to achieve accurate assessment of the impact of multiple factors on the marine environment.

[0079] Specifically, a block-based time series transformer structure is designed, including an input layer, a Transformer encoding layer, a decoding layer, and an output layer. Data is then input; the preprocessed data is arranged in chronological order to form the time series data input model, and input-output sample pairs are constructed using a sliding window method. Through training and testing the Transformer backbone network, the fluctuation amplitude of marine environmental data over time is output to determine the stability or abnormal changes in the data, thus achieving a marine environmental fluctuation rating.

[0080] In step S103, spatial hotspots are clustered in the high-precision navigation and positioning data in the spatial dimension to achieve regional division and pattern recognition of the maritime space.

[0081] Specifically, the DBSCAN density clustering algorithm is used to identify high-frequency activity areas at sea, such as port anchorages and channel intersections. First, the neighborhood radius eps and minimum neighborhood sample size minPts are determined using the K-distance curve method to accommodate the sparsity and density characteristics of ship spatial distribution. Then, the density reachability of ship locations is calculated based on Euclidean distance, dividing continuous areas that meet the density conditions into independent clusters to form hotspot boundaries. Simultaneously, low-density areas are marked as noise points. Preliminary spatial clustering of the location data is performed. The Mean-Shift algorithm is then used to calculate the center of each cluster, aiming to identify hotspot centers. Finally, a spatial hotspot map is generated by combining spatial density features to achieve spatial hotspot rating.

[0082] In step S104, marine environmental fluctuations are assessed based on the predicted marine environmental time series data, and the spatial hotspot level is assessed based on the marine spatial hotspot map obtained by clustering. The marine environmental risks are further assessed by combining these two types of data.

[0083] Specifically, a BP neural network optimized by the Sine and Cosine Algorithm (SCA) is constructed. First, global parameter optimization is performed, initializing the BP network weights W and the threshold b. Then, a global search is conducted in the solution space using SCA: the individual position update formula is: ,in It is a random number. This is the historical optimal solution. The iterative minimization loss function is: The input layer integrates temporal prediction results and spatial hotspot intensity, while the hidden layer employs an adaptive learning rate. To accelerate convergence, the output layer generates risk probabilities using the sigmoid function. .

[0084] In step S105, 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 ocean perception data.

[0085] Specifically, the marine meteorological and hydrological information screening feedback control problem is modeled as a Markov multi-agent game process, where observation nodes such as buoys, satellites, and unmanned vessels are treated as independent agents. Their state space is defined by the current sensor data quality, network load, energy consumption level, and environmental dynamics, while their action space includes sampling frequency adjustment, data transmission priority selection, and anomaly detection threshold optimization. A distributed execution-centralized training approach is adopted, with each agent making distributed decisions based on local observations, while a central coordinator aggregates the global state for policy gradient updates. A joint reward function is designed based on a multi-agent deep deterministic policy gradient algorithm. .in For information value, For communication energy consumption, Due to network latency. During training, an experience replay pool is used to store state transition tuples. Furthermore, the policy network parameters are updated by delaying the target network update to address the policy oscillation problem under unsteady conditions.

[0086] In some embodiments, ocean sensing time-series data is repaired 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 specific steps include S301~S303:

[0087] Step S301: Receive marine meteorological and hydrological time-series data collected by multiple sources of sensors, including buoys, satellites, and underwater robots. Parse the data format and extract timestamps, spatial coordinates, sensor types, and measurement values. Perform preliminary data quality checks, marking missing segments, outliers, and noise points. Outliers include points exceeding physically reasonable ranges and points with significant abrupt changes. Noise points are transient fault data from sensors.

[0088] Step S302: For the missing data segment, construct a bidirectional moving average interpolation model. Construct a sliding window on the time series data and select the mean within the window as the interpolation value. That is, construct an adaptive window with an initial length of m at the left end of the time series, and calculate the mean of the m observations within this window. 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 equals m, and then calculate the mean. The left end of the augmented time series is k. Similarly, construct a window of length m at the right end of the time series, and calculate the mean of the m observations within this window. If the number of observations in the window is less than m, then expand the window to the left until the number of observations in the window equals m, and then calculate the mean. The right end of the augmented time series is k. .

[0089] Step S303: Construct the training dataset of CNN-BiLSTM using pre-impregnated data and rolling prediction, that is, use the first n data to predict the (n+1)th data, and set the length of n to the length of the time series fluctuation period.

[0090] Step S304: Adjust the way the network loss is calculated. If the label is the actual observation value, calculate the original loss; if the label is the estimated value, calculate the loss as t times the original loss, where t∈(0,1) represents the degree of confidence in the estimated value. Use the mean squared error function (MSE) to calculate the model loss. During the iteration process, retain the model corresponding to the minimum loss value and use the prediction value of this model to update the imputation results.

[0091] Step S305: To avoid large fluctuations in the interpolation results obtained by CNN-BiLSTM, BiMA is used to smooth the obtained interpolation values. The smoothed results can be adjusted again.

[0092] like Figure 4 As shown, the spatiotemporal prediction method for marine environmental data based on the improved block time series transformer of the present invention specifically includes the following steps S401~S404:

[0093] Step S401: Represent the i-th one-dimensional sequence as Each sequence is independently input into the Transformer backbone network. Each input is a one-dimensional time series. First, the data is divided into blocks, which can be overlapping or non-overlapping. Let P be the block length and S be the non-overlapping region between two consecutive blocks. After the block division process, a block sequence will be generated. Where N represents the number of blocks generated. .

[0094] Step S402: Through trainable linear projection and learnable addition position encoding Map the blocks to dimensions. In the latent representation space, the data input to the Transformer encoder is represented as

[0095]

[0096] in .

[0097] Then each attention head of multi-head attention Convert it into a query matrix Key matrix and value matrix ,in , The attention output obtained using scaling products is shown in the following formula:

[0098]

[0099] After processing the attention output through a normalization layer and a feedforward network layer with residual connections, the resulting output is represented as follows: Finally, a fully connected layer with a linear head is used to obtain the prediction result, as shown in the following equation:

[0100]

[0101] 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 and the residuals after decoupling from the physical laws are jointly optimized to ensure that the predicted velocity field and temperature field parameters conform to the basic laws of fluid mechanics.

[0102] For example, when predicting ocean current field data, suppose the initial predicted current field output by the model decoder is... Pressure field The physical residuals were calculated using the Navier-Stokes equations, and the constrained predicted values ​​conformed to the laws of fluid dynamics.

[0103]

[0104] Then, physical residuals are introduced to correct the predicted values. Subsequently, a recursive block generation strategy is adopted, using the current prediction block as the input for the next time step to gradually generate the future spatiotemporal sequence.

[0105] like Figure 5 As shown, the marine hotspot region mining method based on an improved mean-shift density clustering algorithm provided by this invention specifically includes the following steps S501~S504:

[0106] Step S501: Determine the neighborhood radius (eps) and minimum number of neighborhood samples (minPts) using the K-distance curve method to adapt to the sparsity and density characteristics of the spatial distribution of navigation data.

[0107] Step S502: Perform initial clustering based on DBSCAN, randomly select a point as the current observation point, and then calculate the current point's clustering based on the given neighborhood radius eps and minimum number of points MinPts. - Neighborhood. If the If the number of points in its neighborhood is at least MinPts, then the current point is marked as a core point, and the points in its neighborhood are added to 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 their... - All points within the neighborhood. For each newly discovered core point, if it is already in the core point candidate pool, it is ignored; otherwise, it is added to the core point pool, and this process is repeated recursively. - This expansion process is repeated for points within the neighborhood. Through this process, all density-reachable core points can be discovered and grouped into the same cluster. Boundary points are assigned to clusters of core points connected to them.

[0108] Step S503: Perform secondary clustering based on Mean-Shift. Randomly select a point as the center point, find all points within the bandwidth of the center point, denoted as set M, and consider these points to belong to cluster C, i.e., the density function calculation region in the first iteration is denoted as cluster C. At the same time, increment the access frequency of these points belonging to cluster C 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 magnitude of the offset vector meets the set threshold requirement, record the center point at this time, and continue to repeat until all points are labeled and classified.

[0109] This invention also provides a marine environmental risk assessment method based on an improved sine and cosine algorithm-optimized BP neural network, such as... Figure 6 As shown, the specific steps include S601~S603:

[0110] Step S601: Select a three-layer BP neural network structure consisting of an input layer, hidden layers, and an output layer. Set the number of input and output nodes based on the number of sample feature values. Determine the optimal number of hidden layer nodes using empirical formulas and an increment / decrement method. Based on the determined number of input, output, and hidden layer nodes, calculate the initial weights and thresholds of the BP neural network; this represents the spatial search dimension for each individual node in the sine / cosine algorithm.

[0111] Step S602: Map the individuals in the SCA algorithm to the initial weight threshold of the BP neural network. The individual calculates its fitness value through the fitness function, which 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 position of the individual is gradually updated, and the optimal value is found and assigned to the BP network as the initial weight threshold.

[0112] Step S603: After the SCA algorithm assigns the optimal value to the weight threshold of the BP network, the SCA-BP model is constructed through network training. The model is then applied to risk level prediction, and its prediction performance is measured using evaluation metrics such as accuracy.

[0113] like Figure 7 As shown, the information filtering feedback control method based on improved multi-agent reinforcement learning provided by the present invention specifically includes the following steps S701~S705:

[0114] Step S701: Deploy node agents for each sensing node to handle local data collection and initial screening; set up coordinating agents in areas with the same risk level to achieve cross-node collaboration; and globally set up optimization agents to dynamically adjust the network-wide screening strategy. Node agents directly receive preset regional risk levels and generate data priorities based on device status.

[0115] Step S702: Risk level-driven dynamic screening rules: The node agent performs differentiated collection according to the risk level: only the original data of key events exceeding the threshold is transmitted in high-risk areas, and compressed sensing technology is used to reduce redundancy in medium and low-risk areas; at the same time, a priority score is calculated for each data, and full information is uploaded for high-scoring data, while statistical summary is sent for low-scoring data.

[0116] 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.

[0117]

[0118] in, For environmental observation data, Encoding historical information, To control the target constraints.

[0119] Each agent corresponds to a region or a type of sensor node, and local actions... Indicates whether to filter data for this node. The reward function is designed as a multi-objective reward.

[0120]

[0121] in Adjust dynamically according to task requirements.

[0122] Step S704: Propose a knowledge distillation transfer mechanism and design a teacher-student network architecture. After centralized training of the teacher network, the global policy is distilled to the local policy networks of each agent through KL divergence loss. A Priority Experience Replay (PER) mechanism is adopted, based on TD error. Experience samples are sampled in a tiered manner, with high-value transitions prioritized for replay. A centralized training-distributed execution (CTDE) framework is adopted, with a central coordinator fusing global state during the training phase. With joint actions The QMIX algorithm is used to decompose the global Q value into individual agents.

[0123] Step S705: The agent dynamically adjusts the observation weights based on the importance of the region; the inner loop adjusts the data filtering actions in real time based on the MARL decision network; the outer loop dynamically updates the reward function weights based on the system state. This allows for automatic adjustment of the data collection frequency and transmission priority based on real-time changes in risk levels, ensuring that high-value information can be promptly and effectively aggregated in the data center, which is beneficial for data analysis and risk prediction and prevention.

[0124] The present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of a spatiotemporally intelligent marine satellite IoT sensing information filtering method and a spatiotemporally intelligent marine satellite IoT sensing information filtering system.

[0125] Corresponding to the above method, the present invention also provides an apparatus comprising a computer device, the computer device including a processor and a memory, the memory storing computer instructions, the processor executing the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the apparatus performs the steps of the method as described above.

[0126] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned edge computing server deployment method. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0127] In summary, this invention provides a spatiotemporally intelligent method and device for filtering marine satellite IoT sensing information, belonging to the field of data processing technology. This method utilizes a convolutional bidirectional long short-term memory neural network to repair marine environmental time-series data, a block-based time series transformer for spatiotemporal prediction of marine environmental data, and a density clustering algorithm based on mean shift for mining marine hotspot areas. The marine environmental data includes sensing data such as sea surface temperature, salinity, and wave height. It employs a backpropagation neural network for marine environmental risk assessment and multi-agent reinforcement learning for information filtering and feedback control, meeting the high-efficiency information processing needs of complex marine environmental monitoring scenarios and achieving low-redundancy, high-precision adaptive feedback control for data acquisition and transmission. This invention can be directly applied to the networking system of marine observation and monitoring facilities in marine satellite IoT, providing strong support for both military and civilian applications, and possesses broad and significant application prospects and value.

[0128] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0129] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0130] In this 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 in place of features of other embodiments.

[0131] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for filtering marine IoT sensing information based on spatiotemporal intelligence, characterized in that, The method includes the following steps: The first step is to repair time-series data based on an improved convolutional bidirectional long short-term memory neural network. The missing data is pre-imputed and repaired based on bidirectional moving average. The time-series data is predicted using a CNN-BiLSTM model. The obtained imputed and repaired values ​​are smoothed using a BiMA model to achieve smooth imputation and repair of time-series data. The second step involves spatiotemporal prediction of marine environmental data based on an improved block time series transformer. This step utilizes a block mechanism to segment marine environmental data into local blocks according to time and space dimensions, employs the channel independence mechanism of PatchTST to capture multi-scale features, and combines multi-feature marine environmental data for multimodal fusion. A sparse attention mechanism is used to model long-range spatiotemporal dependencies, and physical constraints from marine dynamic equations are introduced during the decoding stage to improve the rationality of the prediction. Finally, future spatiotemporal sequences are generated through recursive block segmentation to achieve efficient prediction of marine environmental parameters. The third step is to mine marine hotspot regions based on the improved density clustering algorithm with mean drift. Based on the density gradient direction iteration mechanism of the mean drift algorithm, the neighborhood radius parameter of DBSCAN is dynamically corrected to capture high-density regions. DBSCAN is further combined to refine the coarse clustering results, and finally marine hotspot regions with significant statistical characteristics are extracted. The fourth step is to optimize the marine environmental risk assessment of the BP neural network based on the improved sine and cosine algorithm. The BP network structure is optimized by using the sine and cosine algorithm, and the initial weights and thresholds of the BP neural network are globally optimized. The algorithm convergence is enhanced by combining an adaptive parameter adjustment strategy. Finally, real-time monitoring data is input, and the risk probability distribution is output through the Sigmoid activation function to output multi-dimensional risk assessment results. The fifth step involves information filtering feedback control based on improved multi-agent reinforcement learning. A multi-agent reinforcement learning algorithm is used to deploy node agents for each sensing node. The problem of marine meteorological and hydrological information filtering feedback control is described as a Markov intelligent game process. The node agents adopt differentiated collection based on risk level. The decision network is obtained through distributed decision-making and centralized training. Combined with feedback control, abnormal data filtering and dynamic optimization of transmission load are achieved.

2. The marine IoT sensing information filtering method based on spatiotemporal intelligence according to claim 1, characterized in that, A time-series data restoration method based on an improved convolutional bidirectional long short-term memory neural network is proposed. The method uses a bidirectional moving average (BiMA) model to pre-imputate and restore missing data, and a CNN-BiLSTM model to predict time-series data. The obtained imputation and restoration values ​​are smoothed by the BiMA model to achieve smooth imputation and restoration of time-series data. The bidirectional moving average constructs an adaptive window of initial length m at the left end of the time series, and calculates the mean of the m observations within this window. 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 equals m, and then calculate the mean. The left end of the augmented time series is k. Similarly, construct a window of length m at the right end of the time series, and calculate the mean of the m observations within this window. If the number of observations in the window is less than m, then expand the window to the left until the number of observations in the window equals m, and then calculate the mean. The right end of the augmented time series is k. ; The convolutional bidirectional long short-term memory neural network utilizes multi-scale convolutional kernel groups to extract local spatial features from the input data in parallel. After compression by pooling layers, the output feature map is input into the bidirectional long short-term memory network BiLSTM, which captures long-term dependencies using forward and backward time series respectively, generating temporal hidden state vectors. A gating mechanism is used to fuse the bidirectional hidden states, and a self-attention layer is used to dynamically allocate feature weights to enhance the representation of key temporal nodes. The weighted features are input into the fully connected layer and mapped to the target dimension. The final prediction result is generated by the inverse normalization module. At the same time, the dropout layer is used to suppress overfitting, and the backpropagation algorithm is combined to optimize the network parameters to minimize the prediction error.

3. The marine IoT sensing information filtering method based on spatiotemporal intelligence according to claim 1, characterized in that, A spatiotemporal prediction method for marine environmental data based on an improved block time series transformer (BST) is proposed. This method uses a block mechanism to divide marine environmental data into local blocks according to the temporal and spatial dimensions. It uses the channel independence mechanism of PatchTST to capture multi-scale features and performs multimodal fusion by combining multi-feature marine environmental data. It uses a sparse attention mechanism to model long-range spatiotemporal dependencies and introduces physical constraints of marine dynamic equations in the decoding stage to improve the rationality of prediction. Finally, it generates future spatiotemporal sequences through recursive block division, achieving efficient and accurate prediction of marine environmental parameters. The block-sharing mechanism divides the input sequence into smaller sub-sequences, and then inputs these sub-sequences into the encoder in the Transformer backbone to calculate the self-attention between these blocks as a whole. This reduces the space and time complexity of the model, 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. Multivariate time series input data is treated as a multi-channel signal, where each individual time series represents a different channel encapsulating a specific signal. The sparse attention mechanism selectively focuses on and computes only the relationships between elements that have a significant impact on model performance, while ignoring those elements with a smaller 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 marine IoT sensing information filtering method based on spatiotemporal intelligence according to claim 1, characterized in that, A method for mining marine hotspot regions based on an improved mean-drift density clustering algorithm is proposed. The density gradient direction iteration mechanism of the mean-drift algorithm is used to dynamically adjust the neighborhood radius parameter of DBSCAN to capture high-density regions. Furthermore, DBSCAN is combined to refine the coarse clustering results, and finally, marine hotspot regions with significant statistical characteristics are extracted. The mean shift algorithm, for a given d-dimensional space n sample points For any point x, its Mean-Shift vector is: ,in This represents data points in the dataset whose distance from x is less than the radius h of the sphere, i.e.: Introducing a Gaussian kernel function to shift the mean gives a greater weight to the center point of the distance in the calculation, reflecting the characteristic that the shorter the distance, the greater the weight. The improved shift mean is: , where x is the center point; Let n be the number of points within the range. To find the negative of the derivative of the kernel function; the drift process involves calculating the drift vector and iteratively updating the position of the sphere's center, using the following update formula: ; Density-based clustering algorithms select any sample point M from the research objects. If it is a core object, then sample point M is used as the center. Draw a circle with a radius of 10 ... - The sample points in the neighborhood are also included in cluster X, and the process is repeated recursively until cluster X can no longer be expanded.

5. The marine IoT sensing information filtering method based on spatiotemporal intelligence according to claim 1, characterized in that, A marine environmental risk assessment method based on an improved sine and cosine algorithm-optimized BP neural network is proposed. The BP network structure is optimized using the sine and cosine algorithm, and the initial weights and thresholds of the BP neural network are globally optimized. An adaptive parameter adjustment strategy is combined to enhance the convergence of the algorithm. Finally, real-time monitoring data is input, and the risk probability distribution is output through the Sigmoid activation function to achieve a high-precision and robust dynamic risk assessment of the marine environment. The sine and cosine algorithm assumes a population size of N, a search space of n dimensions, and the position of the i-th individual is represented as: First, randomly generate the positions of N individuals in the n-dimensional solution space. The fitness value of an individual is calculated using an objective function, and the fitness value of the i-th individual is expressed as: Record the position of the best individual in the population. The j-th dimension of the i-th individual in the population is updated as follows: ,in This represents the current optimal individual position. To control parameters, determine the area for the next position; The BP network consists of an input layer, hidden layers, and an output layer. It has a forward propagation phase, where the signal travels from the input layer through the hidden layers and finally reaches the output layer; and a backward propagation phase, where the signal travels from the output layer to the hidden layers and finally to the input layer, adjusting the weights and biases from the hidden layers to the output layer and from the input layer to the hidden layers in sequence. During forward propagation, the signal travels from the input layer to the hidden layers... 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, and d is the number of inputs. For input data, The activation function is used; during backpropagation, the network parameters are adjusted by calculating the error between the output layer and the expected value, thereby reducing the error. The formula is as follows: Where E is the error, The output layer outputs the results. The actual value; The Sigmoid activation function introduces nonlinearity into the model to prevent the neural network from being unable to solve linearly inseparable problems due to linear mapping. The original cascaded features are reconstructed to obtain a cascaded fusion enhanced feature map. The feature value input is set as x, and e is the natural constant. The Sigmoid activation layer maps the feature value to the range [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. Finally, the obtained parameter is combined with the original feature map through inner product to obtain the enhanced representation of the feature.

6. The marine IoT sensing information filtering method based on spatiotemporal intelligence according to claim 1, characterized in that, An improved information screening feedback control method based on multi-agent reinforcement learning is adopted. The algorithm of multi-agent reinforcement learning is used to deploy node agents for each sensing node. The problem of marine meteorological and hydrological information screening feedback control is described as a Markov intelligent game process. The node agents adopt differentiated collection according to the risk level. The decision network is obtained through distributed decision-making and centralized training. The key data transmission of full information is realized, and the daily data transmission of statistical summaries is completed, thus completing the adaptive feedback control of the entire information screening process. The Markov game process describes the cooperative and competitive relationships among multiple agents. Let's assume that in a Markov game involving n agents, each agent i's observation of the environmental state S is... Choose to use action The strategy is described as follows The next shift in the environment Depending on the actions of all agents, each agent gains a reward based on the environmental state, i.e., its own actions. Total revenue for each agent , Let be the discount rate, t be any time, and T be the termination time; the ultimate goal is to find that the expected reward of each agent under the connection strategy satisfies the condition that the strategy chosen by any agent is optimal given that the strategies of other agents are determined. The multi-agent reinforcement learning establishes a centralized critic network for each perceptual node's agent. This network acquires global information, including the global state and the actions of all agents, and provides the corresponding value function. Each agent's actor network only needs to make decisions based on local observation information, which enables distributed control of multiple agents. This allows perception nodes in high-risk areas to transmit key information in a timely manner, while perception nodes in low-risk areas collect and synthesize information, reducing data redundancy and significantly reducing 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 the processor, it implements the steps of the method as described in any one of claims 1 to 6.

Images