Nuclear power buoy marine environment data perception system based on deep learning

Abstract

The present application belongs to the technical field of artificial intelligence and marine environment data processing, and specifically relates to a nuclear power buoy marine environment data sensing system based on deep learning. The system constructs an asynchronous observation sequence according to monitoring variables, sampling time, arrival time and quality markers, uses a deep time delay coding network to obtain dynamic time lag between variables and form a correction time slice, constructs a dynamic space-time correlation graph with monitoring variables and nuclear power buoys as nodes and dynamic time lag and space correlation as edges, generates a reconstruction candidate value of missing or abnormal data through a graph neural network and a multi-scale time sequence decoding, screens and determines the reliability according to the time continuity, variable response order and buoy propagation order, and adjusts the node feature contribution degree and edge weight of the subsequent time slice according to the reliability. The present application can reduce the data dislocation caused by asynchronous sampling and transmission delay, limit the diffusion of low-confidence reconstruction results, and make the recovered data maintain time continuity, variable correlation and space propagation consistency.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and marine environmental data processing technology, specifically a deep learning-based marine environmental data sensing system for nuclear power buoys. Background Technology

[0002] Nuclear power plant buoy marine environmental data sensing systems are typically used to continuously receive data on sea temperature, salinity, current velocity, current direction, waves, and other existing monitoring variables, and aggregate the data collected by each buoy to a shore-based processing platform according to timestamps. Existing systems mostly employ fixed sampling periods and uniform data formats. The data receiving program first performs message parsing, duplicate data removal, missing data marking, and outlier screening, and then resampling is performed at preset time intervals to arrange data from different variables and different buoys onto the same timeline. For situations where sampling periods are inconsistent or data arrival is delayed, conventional processing methods typically select the most recent observation based on the original timestamp, or generate aligned results through linear interpolation, moving averages, and time window aggregation. After time alignment, the system uses threshold rules, statistical distributions, or historical means to determine whether the data deviates from the normal range, and provides the processed sequence to the environmental state analysis model. This type of solution has a clear implementation process and can meet the routine monitoring needs under conditions of relatively complete data and stable transmission delays; however, its time alignment process usually assumes that each variable has a directly comparable state of change at the same time.

[0003] To address continuous missing data, isolated anomalies, and slow drift in buoy data, existing technologies typically employ interpolation compensation, state estimation, or predictive models to recover the data. Interpolation compensation methods calculate substitute values ​​based on valid observations before and after the missing location, state estimation methods use pre-defined state equations and observation equations to infer the current state, and predictive models generate future or reconstructed values ​​based on historical time series. In recent years, some approaches have also used deep learning models to extract the temporal dependencies of marine environmental data. Multiple monitoring variables are input into recurrent neural networks, temporal convolutional networks, or attention networks at a unified time step, and the model outputs estimates of the missing locations, using the difference between the estimated values ​​and the true values ​​of the training samples as the training basis. While these approaches can handle nonlinear variations, they typically involve resampling at a fixed time scale before model input, and most models treat the correlation between different variables as a static relationship within the same time step, failing to separately model the sequential changes in response between variables as sea state changes. Furthermore, the reconstructed results are often written back to the data sequence in single numerical form.

[0004] Graph-based multi-buoy data processing schemes represent different buoys or monitoring variables as graph nodes, establish graph edges based on spatial distance, historical correlation coefficients, or manually defined relationships, aggregate neighboring node information through graph neural networks, and combine this with time-series coding networks to predict missing data. Some schemes also include range checks, rate-of-change checks, or adjacent data consistency checks after model output to exclude results exceeding preset ranges. While these schemes can utilize cross-buoy and cross-variable information, graph edges are typically fixed before training or updated only based on current feature correlations, making it difficult to express the dynamic propagation lags between monitoring variables caused by tides, currents, and local sea conditions. Post-output rule checks are independent of the internal feature propagation within the model, and the check results cannot inversely constrain node contributions and edge weights in subsequent time slices. When the reconstructed values ​​themselves contain biases, they may still be used as subsequent prediction inputs to continue participating in feature aggregation, causing the bias to propagate continuously along the time series and graph connectivity.

[0005] When multivariate data from nuclear power plant buoys suffers from sampling asynchrony, arrival delays, and variable response lags, conventional time alignment and depth reconstruction methods cannot distinguish between apparent differences caused by time misalignment and actual changes in the marine environment. This results in reconstructed data that is numerically continuous but does not conform to the variable response order and the propagation process of adjacent buoys. The technical reason for this is that fixed resampling compresses the original asynchronous observations to a unified time point, and static correlation models assume that variable relationships remain unchanged across time periods, lacking joint processing of dynamic lags, spatial propagation order, and reconstruction reliability. When a variable experiences continuous missing measurements or anomalies, the model may generate candidate values ​​based on temporally misaligned correlation data. Subsequent processing only checks the numerical range and cannot determine whether the candidate values ​​conform to the actual evolutionary relationship. Low-reliability reconstructed values ​​will continue to enter subsequent time slices and change the prediction results of other nodes. Therefore, how to correct for multivariate dynamic lags while preserving the original sampling and arrival relationships, and how to form a closed loop of spatiotemporal correlation, consistency screening, and reliability feedback, are the main technical problems that need to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide a deep learning-based marine environmental data sensing system for nuclear power buoys, which can solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a deep learning-based nuclear power buoy marine environmental data sensing system, including a data processor and a memory connected to the data processor. The memory stores a program executed by the data processor, which enables the data processor to construct an asynchronous observation sequence based on the monitoring variables uploaded by the nuclear power buoy, the original sampling time, the data arrival time, and the quality markers.

[0008] The dynamic time delay between each monitored variable is obtained through a deep time delay coding network, and a time delay correction time slice is formed.

[0009] A dynamic spatiotemporal correlation graph is generated using the monitored variables and the nuclear power buoy as nodes and the dynamic time delay and spatial correlation as edges.

[0010] Based on the dynamic spatiotemporal correlation graph, candidate values ​​for reconstructing missing or abnormal data are generated.

[0011] The reconstruction candidate values ​​are screened and their credibility is determined based on the time continuity, the order of variable response, and the propagation order of adjacent buoys.

[0012] Adjust the node feature contribution and edge weight of subsequent time slices according to the aforementioned credibility.

[0013] Preferably, the deep time-delay coding network includes a multi-scale temporal feature extraction layer, a variable pair association layer, and a time delay estimation layer.

[0014] The multi-scale time series feature extraction layer encodes the short-period fluctuations, tidal period changes, and long-term trends of each monitored variable. The variable pair correlation layer forms a joint feature sequence for any two monitored variables within their overlapping effective observation intervals. The time delay estimation layer generates a time delay distribution based on the direction of change, local phase difference, and rate of change difference in the joint feature sequence, and performs a differentiable offset on the feature time index of the corresponding monitored variable according to the time delay distribution to form a time delay correction time slice that retains the original timestamp mapping relationship.

[0015] Preferably, the dynamic spatiotemporal correlation graph includes variable layer nodes, buoy layer nodes, and cross-layer edges connecting the two layers of nodes.

[0016] The variable layer nodes carry the temporal characteristics of different monitoring variables within the same nuclear power buoy, and the buoy layer nodes carry the aggregation characteristics of different nuclear power buoys in the same time slice.

[0017] The initial weights of the cross-layer edges are determined by historical stable associations, spatial proximity relationships between nuclear power buoys, and dynamic time delays, and are modified by the graph attention network based on the consistency of local changes in the current time slice.

[0018] The graph attention network retains the edge state of the previous time slice in each time slice, and encodes the newly added edges, persistent edges, and decaying edges respectively before inputting them into the reconstruction network.

[0019] Preferably, the reconstruction candidate value is jointly determined by the data-driven branch and the consistency constraint branch.

[0020] The data-driven branch extracts neighborhood aggregation features from the dynamic spatiotemporal correlation graph through a graph neural network, and generates multiple candidate sequences corresponding to missing or abnormal positions through a multi-scale temporal decoder.

[0021] The consistency constraint branch calculates the time continuity deviation, variable response order deviation, adjacent buoy propagation order deviation, and historical stable interval deviation for each candidate sequence, and constructs a consistency vector based on each deviation.

[0022] The data processor jointly encodes the consistency vector and the uncertainty of the candidate sequence, and outputs the reconstruction candidate value and the confidence level corresponding to the reconstruction candidate value.

[0023] Preferably, the variable pair association layer establishes a validity mask matrix and a multi-window association matrix for each variable pair. The validity mask matrix records the data availability status, quality mark, and arrival delay status of the variable pair at each time index. The multi-window association matrix records the short window change synchronization degree, the medium window phase migration degree, and the long window trend consistency degree, respectively.

[0024] The time delay estimation layer performs gated fusion of the validity mask matrix and the multi-window correlation matrix to generate multiple candidate time delays and corresponding weights. It performs cross-variable feature shift only on candidate time delays that maintain the same direction in continuous time slices, and retains the remaining candidate time delays as low-weight time delay components in the time delay distribution.

[0025] Preferably, the dynamic spatiotemporal correlation graph is divided into multiple local subgraphs according to the monitoring variable category and the spatial neighborhood of the nuclear power buoy, and propagation edges connecting different local subgraphs are set.

[0026] The node features within each local subgraph are formed by splicing together the current observation features, time delay correction features, and historical reconstruction confidence. The edge features of the propagation edge are formed by splicing together the spatial distance code, dynamic time delay code, and historical propagation order code.

[0027] The graph attention network calculates the attention within the local subgraph and the attention propagated across the subgraph, respectively, and performs gating normalization on the attention coefficients based on the confidence level of the corresponding nodes. The normalized local aggregation results and the cross-subgraph propagation results are then input into the reconstruction network.

[0028] Preferably, the multi-scale temporal decoder generates candidate sequences based on short-period context, tidal period context, and long-term trend context for the same missing position or the abnormal position, respectively. The consistency constraint branch forms a corresponding consistency vector for each candidate sequence and maps each consistency vector to a candidate selection weight.

[0029] The data processor performs weighted fusion of the candidate sequences according to the candidate selection weights, and generates uncertainty based on the dispersion between each candidate sequence and the characteristic perturbation response of the graph neural network. The uncertainty and the weighted fusion result are stored as reconstruction records for subsequent time slice calls.

[0030] Preferably, when subsequent real observation data arrives, the data processor matches the real observation data with the corresponding reconstructed record according to the original sampling time to generate a reconstruction deviation sequence and a time delay deviation sequence.

[0031] The decoding parameters associated with the corresponding monitoring variables in the data-driven branch are corrected according to the reconstructed deviation sequence, the time delay distribution parameters of the deep time delay coding network are corrected according to the time delay deviation sequence, and the changes in edge weights before and after correction are written into the edge state memory.

[0032] When the deviation directions corresponding to multiple consecutive real observation segments are consistent, the state of the continuous edge in the dynamic spatiotemporal correlation graph is updated; otherwise, the edge state before correction is restored.

[0033] Preferably, the graph attention network establishes a credible feature channel and a weak evidence feature channel within the local subgraph. Node features with a credibility not lower than the credibility quantile of the corresponding time slice enter the credible feature channel, while the remaining node features, along with the quality label, enter the weak evidence feature channel.

[0034] The trusted feature channel updates the node state using the neighborhood aggregation result, and the weak evidence feature channel updates the node state using the restricted aggregation result with residual preservation. The node states of the two channels are then passed to the cross-subgraph propagation attention.

[0035] The cross-subgraph propagation attention is gating the propagation edges based on the product encoding of the credibility of each node on the propagation path.

[0036] Preferably, the edge state memory establishes a hierarchical index according to the monitoring variable pairs, nuclear power buoy pairs and sea state characteristic intervals. Each index entry records the edge weight version, the corresponding time delay distribution, the reconstruction deviation statistics and the actual observation coverage.

[0037] Before updating the persistent edge state, the data processor selects historical index items similar to the current time slice from the same sea state feature interval for playback, and compares the consistency of the currently updated reconstruction result with the actual observations corresponding to the historical index items.

[0038] When the time continuity deviation of the reconstructed result after the update or the deviation of the variable response order is greater than the corresponding deviation before the correction, the update of the current edge weight version and the corresponding time delay distribution is cancelled, and the deviation sequence is written into the weak evidence record of the subsequent update.

[0039] Compared with existing technologies, the advantages of this invention are as follows: 1. By preserving the original sampling time, data arrival time, and quality markers to construct asynchronous observation sequences, and utilizing a deep time-delay coding network to obtain dynamic time delays from variable change direction, local phase difference, and change rate difference, time misalignment caused by sampling asynchrony and transmission delay can be eliminated before the data enters the reconstruction process. The time-delay-corrected multivariate time slices are used to establish a spatiotemporal correlation graph containing variable nodes, buoy nodes, and dynamic correlation edges, ensuring that candidate values ​​at missing or anomalous positions are simultaneously constrained by variable relationships within the same buoy, spatial relationships between different buoys, and propagation order. After the data-driven branch generates multiple candidate sequences, the consistency constraint branch filters them according to temporal continuity, variable response order, propagation order of adjacent buoys, and historical stable intervals, avoiding selection of reconstruction results solely based on numerical similarity. This reduces the likelihood of time-misaligned data being mistakenly treated as real anomalies in reconstruction, ensuring that the recovered data simultaneously satisfies temporal continuity, variable correlation, and spatial propagation consistency, solving the problem of distorted evolutionary relationships in data obtained by conventional methods.

[0040] 2. The system generates a confidence level for each reconstruction result and adjusts the node feature contribution and graph edge weights of subsequent time slices based on the confidence level. This allows low-confidence reconstruction values ​​to participate in subsequent calculations as weak evidence, limiting their spread along the time series and graph connectivity. The dynamic spatiotemporal correlation graph saves edge states according to newly added edges, persistent edges, and decaying edges. Confidential feature channels and weak evidence feature channels are set within local subgraphs, enabling the separate aggregation of real observation features and uncertain reconstruction features, reducing the interference of low-quality data on adjacent variables and other buoy node states. When subsequent real observations arrive, the system matches the reconstruction records according to the original sampling time, corrects decoding parameters, time delay distribution, and persistent edge states using reconstruction bias and time delay bias, and undoes updates that cause consistency degradation through edge state memory and historical index replay. The feedback process allows time delay estimation, correlation relationships, and reconstruction models to be gradually corrected based on real observations, while retaining recoverable historical edge weight versions, reducing error accumulation and correlation structure shifts under continuous missing data conditions. Attached Figure Description

[0041] Figure 1 This is a flowchart of the overall processing of a deep learning-based marine environmental data sensing system for nuclear power buoys.

[0042] Figure 2 This is a flowchart for dynamic time delay estimation and time slice correction of asynchronous observation data.

[0043] Figure 3 A flowchart for constructing dynamic spatiotemporal correlation graphs and propagating trusted features.

[0044] Figure 4This is a flowchart of multi-scale candidate reconstruction, consistency screening, and feedback correction. Detailed Implementation

[0045] In one embodiment, please refer to Figure 1 The deep learning-based marine environmental data sensing system for nuclear power buoys includes a data processor and a memory connected to the data processor. The memory stores a program executed by the data processor. The program enables the data processor to construct an asynchronous observation sequence based on the monitoring variables uploaded by the nuclear power buoys, the original sampling time, the data arrival time, and the quality markers. It obtains the dynamic time delay between each monitoring variable through a deep time delay coding network and forms a time delay correction time slice. It generates a dynamic spatiotemporal correlation graph with the monitoring variables and nuclear power buoys as nodes and the dynamic time delay and spatial correlation as edges. Based on the dynamic spatiotemporal correlation graph, it generates candidate values ​​for reconstructing missing or abnormal data. It filters the candidate values ​​for reconstruction based on the temporal continuity, the variable response order, and the propagation order of adjacent buoys and determines their credibility. It adjusts the node feature contribution and edge weight of subsequent time slices according to the credibility. The system thus forms a closed-loop processing link from the original asynchronous data access, time delay correction, relationship modeling, candidate reconstruction, consistency screening to credibility feedback.

[0046] In this embodiment, the monitored variables are the marine environmental variables corresponding to the existing acquisition channels of the nuclear power buoy. The data processor does not require different monitored variables to have the same sampling period, nor does it force the data to be compressed to a uniform time during the receiving phase. Instead, it establishes a data tuple for each observation record, consisting of a buoy identifier, a variable identifier, the original sampling time, the data arrival time, the observed value, a quality marker, and a message sequence marker. The original sampling time represents the time when the data is formed, the data arrival time represents the time when the data enters the processing end, the quality marker is used to record the completeness, missing, verification anomaly, duplication, delay, and pending confirmation status, and the message sequence marker is used to identify out-of-order data packets. The data processor establishes a primary index according to the original sampling time, an auxiliary index according to the data arrival time, and retains the difference between the two to characterize the transmission delay.

[0047] Specifically, asynchronous observation sequences are stored separately according to variable identifiers. Within each variable sequence, they are sorted according to the original sampling time. Different variable sequences are linked through overlapping time windows. The left boundary of the overlapping time window is the most recent common valid observation time before the time slice to be processed, and the right boundary is the most recent common valid observation time received after the time slice to be processed. When a variable has consecutive missing measurements in the overlapping time window, the data processor retains the missing measurement position and generates a missing measurement mask, without replacing the original missing position with the interpolation result. When the data arrival order of a variable is inconsistent with the original sampling order, it is rearranged according to the original sampling time and the misordered state is written into the quality mark, so that the subsequent model can distinguish between numerical mutations caused by environmental changes and apparent mutations caused by transmission misorders.

[0048] Please refer to Figure 2 The deep time delay coding network receives valid observation segments and corresponding masks of each variable within the same overlapping time window. For each variable, it extracts short-period variation, tidal period variation, and long-term trend. Short-period variation is extracted by local convolutional kernels to extract the amplitude and direction changes between consecutive adjacent observations. Tidal period variation is jointly extracted by dilated convolution and periodic position coding. Long-term trend is extracted by a time-series coding layer with gated state memory. The three types of features maintain the original sampling index in the time dimension and are concatenated in the feature dimension. During the training phase, the network uses complete segments to construct random occlusion and random delay samples, enabling the time delay estimation to simultaneously address natural asynchrony, transmission delay, and local missing data.

[0049] For the dynamic time delays of variables i and j at time index t, the candidate time delay distribution is calculated using the following formula:

[0050] Where i and j represent the monitoring variable indices, t represents the time index, and d and r represent the candidate time delays. Represents the set of allowed candidate time delays. This represents the joint matching score between the current feature of variable i and variable j after offset by the candidate time lag d. Represents the normalized weight of the candidate time delay d, when When taking a discrete set consisting of negative time delay, zero time delay, and positive time delay, the position with the largest weight is used to represent the order of the main responses, and the remaining positions are retained as secondary time delay components.

[0051] In a preferred embodiment, the joint matching score is calculated by the similarity of change direction, the local phase difference, the rate of change difference, and the effective observation coverage. The similarity of change direction is obtained based on the consistency of the signs of the differences between adjacent observations. The local phase difference is obtained based on the difference between the peak and valley positions of the two variables within the sliding window. The rate of change difference is obtained based on the distance between the standardized difference sequences. The effective observation coverage is obtained based on the proportion of the two variables having effective values ​​at the candidate offset positions. The candidate time delay set is not limited by fixed physical constants, but is determined by the available history length and observation density of the current processing window. Candidate positions lacking common effective observations are masked out.

[0052] Preferably, the time delay correction does not directly move the original observations to the new timestamp, but rather performs a differentiable offset on the encoded feature time index while preserving the original timestamp mapping relationship. The correction feature of variable j relative to variable i is obtained by weighting multiple offset features according to the candidate time delay weights. A positive time delay corresponds to reading the feature of variable j from an earlier index, a negative time delay corresponds to reading the feature of variable j from a later index, and a zero time delay corresponds to reading the feature of the same index. When the offset position falls between discrete indices, a linear combination of two adjacent feature vectors is used, so that the time delay parameter can be updated in reverse through reconstruction error and consistency deviation, while avoiding changes to the original data record.

[0053] The time delay correction feature of variable j relative to variable i at time index t is formed by the following formula:

[0054] in, This indicates the time delay correction characteristic of variable j relative to variable i. This represents the original encoded feature sequence of variable j. This indicates a differentiable interpolation operation performed at a given time position, where td represents the read position. When the candidate time delay weights are concentrated in positive values, the correction features are mainly taken from earlier observations. When the weights are distributed across multiple positions, the correction features retain multiple possible response paths.

[0055] In another embodiment, the deep time delay coding network includes a multi-scale time series feature extraction layer, a variable pair association layer, and a time delay estimation layer. The multi-scale time series feature extraction layer encodes the short-period fluctuations, tidal period changes, and long-term trends of each monitored variable. The variable pair association layer forms a joint feature sequence for any two monitored variables within overlapping effective observation intervals. The time delay estimation layer generates a time delay distribution based on the direction of change, local phase difference, and rate of change difference in the joint feature sequence, and performs a differentiable offset on the feature time index of the corresponding monitored variable according to the time delay distribution to form a time delay correction time slice that retains the original timestamp mapping relationship. The joint feature sequence also carries quality mark encoding and transmission delay encoding to prevent the model from treating low-quality observations as evidence of stable time delays.

[0056] Furthermore, the variable pair correlation layer establishes an effectiveness mask matrix and a multi-window correlation matrix for each variable pair. The effectiveness mask matrix records the data availability status, quality label, and arrival delay status of the variable pair at each time index. The multi-window correlation matrix records the short window change synchronization degree, the medium window phase migration degree, and the long window trend consistency degree, respectively. The time delay estimation layer performs gated fusion of the effectiveness mask matrix and the multi-window correlation matrix to generate multiple candidate time delays and their corresponding weights. Only the candidate time delays that maintain the same direction in continuous time slices are subjected to cross-variable feature shift, while the remaining candidate time delays are retained as low-weight time delay components in the time delay distribution.

[0057] Specifically, each element of the validity mask matrix consists of data availability bits, quality confidence bits, and arrival delay level codes. Data availability bits are used to distinguish between real observations and missing locations, quality confidence bits are used to distinguish between data that has passed verification and data that is yet to be confirmed, and arrival delay levels are discretely encoded according to the recent arrival delay distribution of the same variable. In the multi-window correlation matrix, short windows cover local continuous changes, medium windows cover at least one major periodic segment, and long windows cover the historical range that can characterize trend reversals. Correlation features are calculated for each window at a common valid location, and windows with insufficient common valid locations are marked as not eligible to participate in the major time delay determination.

[0058] During gated fusion, the data processor concatenates the validity mask matrix encoding result with the multi-window correlation matrix encoding result and inputs it into the gated network. The gated network outputs the retention coefficient of each candidate time delay. The retention coefficient is multiplied element-wise with the candidate time delay distribution and then renormalized. The directional consistency in continuous time slices is determined by comparing the sign and change magnitude of the main candidate time delays. When the current main candidate time delay has the same direction as the main candidate time delay of the previous time slice and does not cross multiple discrete positions, it is used for differentiable feature migration. When the direction is frequently reversed or there are insufficient common valid positions, it is reduced to a low-weight time delay component instead of being directly deleted, so that it can be re-evaluated after subsequent real observations arrive.

[0059] In one embodiment, please refer to Figure 3 The data processor generates a dynamic spatiotemporal correlation graph using monitored variables and nuclear power buoys as nodes and dynamic time delays and spatial correlations as edges. The dynamic spatiotemporal correlation graph includes variable layer nodes, buoy layer nodes, and cross-layer edges connecting the two layers. Variable layer nodes carry the temporal characteristics of different monitored variables within the same nuclear power buoy, while buoy layer nodes carry the aggregation characteristics of different nuclear power buoys in the same time slice. The initial weights of the cross-layer edges are jointly determined by historical stable correlations, spatial proximity relationships between nuclear power buoys, and dynamic time delays, and are corrected by the graph attention network based on the consistency of local changes in the current time slice. The graph attention network retains the edge states of the previous time slice in each time slice and encodes the newly added edges, persistent edges, and decaying edges respectively before inputting them into the reconstruction network.

[0060] In this embodiment, variable layer nodes are uniquely determined by buoy identifiers and variable identifiers. Node features include current observation code, time delay correction code, quality mark code, and historical reconstruction confidence. Buoy layer nodes are obtained by aggregating variable layer nodes within the same buoy through masked aggregation. During aggregation, only real observations or high-confidence reconstruction records are used as the main input, while low-confidence records participate through independent channels. Cross-layer edges include variable association edges within the same buoy, spatial propagation edges of the same type of variables between different buoys, and belonging edges between variable layer nodes and their respective buoy layer nodes. The graph structure is updated in each time delay correction time slice, but the identity and historical state of the edges remain continuous.

[0061] The update weight of any edge at time index t is calculated using the following formula:

[0062] Where a and b represent graph node indices, This represents the current edge weight from node a to node b. This indicates the edge weight at the previous time. This represents the edge state preservation coefficient. This represents a function that maps the input to a predetermined weight range. Indicates historical stable association encoding, This indicates a joint encoding of spatial proximity and dynamic time delay. This represents the consistency correction amount for the current local change. When an edge continuously receives consistent observation support, the coefficient is kept to a larger value, and when an edge is triggered by a new relationship, the coefficient is kept to a smaller value.

[0063] Historical stable association encoding is obtained from training data and confirmed real observation segments. The data processor records the association direction, typical time delay, and duration of common change for each variable pair and buoy pair, and writes the records in the stable interval into the edge state memory. The joint encoding of spatial proximity and dynamic time delay does not judge the propagation relationship based solely on geographical distance, but maps the spatial proximity order, historical propagation order, and current dynamic time delay between buoys into edge features. The current local change consistency correction is calculated by the node change direction, change amplitude level, and confidence difference, avoiding the formation of high-weight propagation edges due to short-term anomalies of a single node.

[0064] New edges, persistent edges, and decaying edges are determined jointly based on the current edge weight and the historical edge state. New edges represent connections that did not form a stable association before but have shown consistent changes in the current consecutive time slices. Persistent edges represent connections where the current relationship is consistent with the historical relationship and is continuously supported by reliable observations. Decaying edges represent connections where the historical relationship is still retained but the current support strength decreases. The three types of edges are reconstructed using different state encoding inputs. The state encoding does not directly determine the candidate value, but is used to control the mixing ratio of historical information and current information in message passing, so that dynamic relationship changes can be expressed and the graph structure is not completely reset due to a single time slice change.

[0065] In a preferred embodiment, the dynamic spatiotemporal correlation graph is divided into multiple local subgraphs according to the monitoring variable category and the spatial neighborhood of the nuclear power buoy, and propagation edges connecting different local subgraphs are set. The node features in each local subgraph are formed by splicing the current observation features, time delay correction features and historical reconstruction confidence. The edge features of the propagation edges are formed by splicing the spatial distance code, dynamic time delay code and historical propagation order code. The graph attention network calculates the attention within the local subgraph and the cross-subgraph propagation attention respectively, and performs gating normalization on the attention coefficients according to the confidence of the corresponding nodes. Then, the normalized local aggregation results and the cross-subgraph propagation results are input into the reconstruction network.

[0066] The division of local subgraphs is based on maintaining semantic consistency of monitored variables and continuity of spatial propagation. Closely related variable nodes within the same buoy are assigned to the variable local subgraph, while similar variable nodes in adjacent buoys are assigned to the spatial local subgraph. The variable local subgraph is used to describe the order of variable response at the same location, while the spatial local subgraph is used to describe the propagation order of environmental changes between adjacent buoys. Propagation edges connecting different local subgraphs only participate in the main aggregation when the nodes at both ends have common valid observations or time delay relationships confirmed by continuous time slices. Other propagation edges retain structural identifiers but participate in a weak evidence manner.

[0067] The graph attention network normalizes the attention within the local subgraph and the cross-subgraph propagation attention separately. The attention within the local subgraph is calculated based on the matching degree of node features and the order of variable responses, while the cross-subgraph propagation attention is calculated based on the features of the propagation edges and the historical propagation order. The node credibility is used as a gating factor and multiplied into the unnormalized attention value. Nodes with low credibility are not completely deleted, but their direct contribution to neighboring nodes is reduced. The local aggregation result and the cross-subgraph propagation result are combined through a gating fusion layer. The gating fusion layer determines the mixing ratio of the two types of results based on the missing state, abnormal state, and number of available neighbors of the target node.

[0068] Furthermore, the graph attention network establishes a credible feature channel and a weak evidence feature channel within the local subgraph. Node features with a credibility not lower than the credibility quantile of the corresponding time slice enter the credible feature channel, while the remaining node features, along with the quality label, enter the weak evidence feature channel. The credible feature channel updates the node state using the neighborhood aggregation result, while the weak evidence feature channel updates the node state using the restricted aggregation result with residual preservation. The node states of the two channels are then passed to the cross-subgraph propagation attention, which gates the propagation edges based on the product encoding of the credibility of each node along the propagation path.

[0069] The trusted gating code for propagation path q at time index t is calculated using the following formula:

[0070] in, Represents the set of nodes along the propagation path. This indicates the credibility of node a at time index t. The trusted gating code represents the propagation path. When the trustworthiness of each node on the path is high, the product remains large. When the trustworthiness of any intermediate node is low, the product decreases accordingly. This code restricts the spread of uncertain information along multi-hop paths by controlling the propagation of attention across subgraphs.

[0071] The credibility quantile is calculated from the set of nodes with credibility records in the current time slice. No fixed absolute threshold is set to adapt to the credibility distribution under different sea conditions and different data integrity. The credibility feature channel performs regular weighted aggregation on neighbor messages, while the weak evidence feature channel merges the aggregated messages with the original state of the nodes according to a limited ratio and retains the quality label and missing mask to prevent weak evidence from being misinterpreted as real observations after propagation in the multi-layer graph. The two channels remain independent during the cross-subgraph propagation stage and are only merged through conditional gating before entering the reconstructed network.

[0072] In one embodiment, please refer to Figure 4 The reconstructed candidate values ​​are jointly determined by the data-driven branch and the consistency constraint branch. The data-driven branch extracts neighborhood aggregation features from the dynamic spatiotemporal correlation graph through a graph neural network and generates multiple candidate sequences corresponding to missing or abnormal positions through a multi-scale time series decoder. The consistency constraint branch calculates the temporal continuity deviation, variable response order deviation, adjacent buoy propagation order deviation, and historical stable interval deviation for each candidate sequence. Based on each deviation, a consistency vector is constructed. The data processor jointly encodes the consistency vector and the uncertainty of the candidate sequence, and outputs the reconstructed candidate values ​​and the corresponding confidence levels.

[0073] The data-driven branch receives the target node's own historical encoding, local subgraph aggregation results, cross-subgraph propagation results, and target location mask. The graph neural network distinguishes between real observation messages, credible reconstruction messages, and weak evidence messages in each layer of message passing, and appends the message type encoding to the neighborhood features. The multi-scale temporal decoder sets short-cycle decoding paths, tidal cycle decoding paths, and long-term trend decoding paths. The three paths share the static identifier of the target node but use different receptive ranges. Each path outputs a candidate sequence with the same length as the target missing interval, avoiding the use of a single time scale to average and reconstruct abrupt and slow processes.

[0074] The consistency constraint branch does not directly delete candidate sequences using fixed-range rules. Instead, it writes each candidate sequence back to a temporary copy and recalculates its connection with the preceding and following real observations. The temporal continuity bias is obtained from the difference in amplitude and direction of change between the candidate sequence boundary and the adjacent real observation. The variable response order bias is obtained from the time delay deviation between the candidate change start point and the related variable change start point. The adjacent buoy propagation order bias is obtained from the deviation between the candidate change start point and the spatial propagation edge record. The historical stable interval bias is obtained from the distance between the candidate feature and the historical real segment within the same sea state feature interval.

[0075] The consensus cost of candidate sequence k is calculated using the following formula:

[0076] in, This represents the consistency cost of the candidate sequence at time index t. Indicates continuous time deviation. Indicates the deviation in the order of variable responses. This indicates the deviation in the propagation order of adjacent buoys. Indicates the deviation from the historical stable range. to This represents the non-negative weight determined by the effective observation coverage. When spatially adjacent data is missing, the corresponding spatial bias weight is reduced to 0 and then normalized by the remaining weights.

[0077] In a preferred embodiment, the multi-scale time-series decoder generates candidate sequences based on short-period context, tidal period context, and long-term trend context for the same missing or abnormal position. The consistency constraint branch forms a corresponding consistency vector for each candidate sequence and maps each consistency vector to a candidate selection weight. The data processor performs weighted fusion of the candidate sequences according to the candidate selection weight. At the same time, it generates uncertainty based on the degree of dispersion between each candidate sequence and the characteristic perturbation response of the graph neural network. The uncertainty and the weighted fusion result are stored as a reconstruction record for subsequent time slice calls.

[0078] Candidate selection weights are determined by both consistency cost and candidate uncertainty. Candidates with lower consistency cost and lower uncertainty receive higher weights. The dispersion between candidate sequences is calculated by the weighted variance of multiple candidate values ​​at the same position. Feature perturbation response is obtained by applying a small occlusion to the input neighborhood features while keeping the quality label unchanged and repeatedly decoding. When a slight occlusion causes a large change in the output, the corresponding candidate uncertainty increases. In addition to saving the fusion value and uncertainty, the reconstruction record also saves the candidate index, candidate selection weight, edge state version used, and original timestamp mapping involved in the fusion, so as to perform corresponding backtracking after the actual observation arrives.

[0079] Reconstructed value and its credibility Determine using the following formula:

[0080] in, Represents the set of candidate sequences. This indicates that the candidate selection weights are equal to 1. This represents the value of variable i in the candidate sequence k. The confidence level represents the uncertainty formed by the candidate dispersion and characteristic perturbation response. When both the candidate cost and uncertainty are small, the confidence level is close to 1. When the divergence between candidates increases or the consistency cost increases, the confidence level decreases.

[0081] In this embodiment, the reconstruction confidence directly participates in the calculation of node features and edge weights in the next time slice. The confidence of the real observation node is determined by the quality label and data verification results. The confidence of the reconstruction node is obtained by the aforementioned joint calculation. The node feature contribution is gated by confidence. The consistency correction amount of the current local change in the edge weight update reads the confidence of the nodes at both ends. Low confidence reconstruction records continue to be retained in the sequence and graph structure, but will not be propagated with the same message strength as the real observation, thereby maintaining the continuous representation of the missing interval and limiting the diffusion of uncertain information.

[0082] In another embodiment, when subsequent real observation data arrives, the data processor matches the real observation data with the corresponding reconstructed record according to the original sampling time to generate a reconstruction deviation sequence and a time delay deviation sequence. The decoding parameters associated with the corresponding monitoring variables in the data-driven branch are corrected according to the reconstruction deviation sequence, and the time delay distribution parameters of the deep time delay coding network are corrected according to the time delay deviation sequence. The changes in edge weights before and after correction are written into the edge state memory. When the deviation directions corresponding to multiple consecutive real observation segments are consistent, the continuous edge state in the dynamic spatiotemporal correlation graph is updated; otherwise, the edge state before correction is restored.

[0083] Specifically, the matching process uses the original sampling time as the primary key and variable identifiers and buoy identifiers as joint constraints. When a late real observation covers the reconstructed location, the reconstructed record is not directly deleted. Instead, the real value, reconstructed value, candidate components, confidence level, and edge state version are combined to form a correction sample. The reconstructed deviation sequence is formed by subtracting the reconstructed value from the real value. The time delay deviation sequence is formed by comparing the starting point of variable change in the real observation segment with the main position of the current time delay distribution. Decoding parameter correction only applies to the decoding path related to the corresponding variable and time scale. Time delay distribution parameter correction only applies to the variable pairs participating in this reconstruction. Parameters not participating in the reconstruction remain unchanged.

[0084] The parameter correction adopts a restricted incremental approach. The data processor saves a snapshot of the parameters before each correction and the referenced data range. When the reconstruction deviation is consistent in direction in consecutive segments, the corresponding decoding parameters are updated in the direction that reduces the deviation. When the time delay deviation points to the same candidate position in consecutive segments, the time delay score of that position is increased and the score of the conflict position is decreased. If the deviation direction changes frequently with segments, the persistent edge state is not updated and the current sample is written into the weak evidence record to avoid the continuous shift of the association structure caused by a single late data or local anomaly.

[0085] Furthermore, the edge state memory establishes a hierarchical index based on monitoring variable pairs, nuclear power buoy pairs, and sea state characteristic intervals. Each index entry records the edge weight version, the corresponding time delay distribution, the reconstruction deviation statistic, and the actual observation coverage. Before updating the persistent edge state, the data processor selects historical index entries similar to the current time slice from the same sea state characteristic interval for playback. The updated reconstruction result is compared with the actual observations corresponding to the historical index entries. When the time continuity deviation or variable response order deviation of the updated reconstruction result is greater than the corresponding deviation before correction, the update of the current edge weight version and the corresponding time delay distribution is canceled, and the deviation sequence is written into the weak evidence record of subsequent updates.

[0086] The sea state characteristic interval is determined by the multivariate trend code, periodic code, and spatial change code of the current time slice. The first layer of the hierarchical index is located according to the monitoring variable pairs, the second layer is located according to the nuclear power buoy pairs, and the third layer is located according to the sea state characteristic interval. The edge weight version and time delay distribution in the same index item have a common version identifier. The reconstructed deviation statistics record the direction, duration, and corresponding candidate sequence. The actual observation coverage record is used to verify the time interval of the version. The selection of historical index items is based on the distance between the current sea state code and the historical sea state code, as well as the completeness of the actual observation coverage.

[0087] During historical replay, the data processor fixes the model parameters other than the current parameters to be updated, and re-inputs the selected historical real segments into the time delay correction, graph construction, and link reconstruction processes to obtain candidate sequences of the pre-correction and post-correction versions, respectively. Then, it compares the time continuity deviation and variable response order deviation according to the same consistency cost calculation method. If the post-correction version reduces the deviation on the current segment but expands the deviation on similar historical segments, it is not registered as a persistent edge state. Instead, the edge weight version and time delay distribution before correction are restored, and the current correction sample is retained for re-judgment when more real observations are available later.

[0088] In one application implementation, a monitoring variable of a nuclear power buoy is missing within a continuous time slice, while the associated variables of the same buoy and the same type of variables of adjacent buoys remain available. The system confirms that the missing data is not due to message out-of-order based on the original sampling time and the data arrival time. The available variables are input into a deep time delay coding network to obtain the current dynamic time delay. Based on the time delay correction characteristics, a variable local subgraph and a spatial local subgraph are established. The data-driven branch generates short-period candidates, tidal period candidates, and long-term trend candidates, respectively. The consistency constraint branch combines the actual observations before and after the missing interval, the response order of the associated variables, and the propagation order of adjacent buoys to form candidate selection weights. The fused reconstructed value and confidence level are written into the corresponding original sampling time.

[0089] When the aforementioned missing data arrives late, the system reads the corresponding reconstruction record based on the original sampling time, compares the true value with the fused reconstruction value, and backtracks the candidate sequence and edge state version that participated in the fusion. If the continuous true segments indicate that the short-period candidates are consistent on the boundary connection and the main position of the time delay is consistent with the true response order, then the parameters related to the short-period decoding path of the variable are corrected and the continuous edge state is updated. If subsequent historical playback finds that the update increases the deviation of the variable response order under the same sea state, then the original version is restored and the current deviation sequence is included in the weak evidence record.

[0090] In another application implementation, if an isolated mutation occurs in a certain monitoring variable and the data quality is marked as pending confirmation, the system retains the observation in the asynchronous observation sequence and sends it to the weak evidence feature channel in the dynamic spatiotemporal correlation graph. During the time delay estimation stage, the main candidate time delay is not changed by the isolated mutation alone. During the reconstruction stage, a candidate sequence is generated based on the previous and subsequent real observations, the correlation variables of the same buoy, and the similar variables of adjacent buoys. If the isolated mutation is inconsistent with the variable response order and spatial propagation order, the consistency cost of its corresponding candidate increases. The fusion result is dominated by the other candidates. The original abnormal observation, the fusion reconstruction value, and the confidence level are saved simultaneously without overlapping each other.

[0091] For cases of continuous slow drift, the system records the gradually widening trend difference between the target variable and the associated variable in the long-window trend consistency. The dynamic time delay coding network retains the original main time delay and reduces the observation credibility of the trend inconsistency segment. The target node in the dynamic graph gradually moves from the credible feature channel to the weak evidence feature channel. The multi-scale time series decoder uses the tidal cycle context and long-term trend context to generate candidate sequences. The consistency constraint branch limits the drift observation to directly become the reconstruction benchmark through the deviation of the historical stable interval. After the subsequent real verification data arrives, it is determined whether to correct the decoding parameters and edge states of the corresponding variables.

[0092] During program execution, the memory stores the original asynchronous observation area, the time delay correction feature area, the dynamic graph state area, the candidate reconstruction area, and the correction playback area. The original asynchronous observation area only appends the original records and quality markers and is not overwritten by the model output. The time delay correction feature area stores the candidate time delay distribution and original timestamp mapping of each variable pair. The dynamic graph state area stores the node state, edge type, and version identifier. The candidate reconstruction area stores the candidate sequence, selection weight, uncertainty, and confidence level. The correction playback area stores late true observations, bias sequences, parameter snapshots, and historical index items. This ensures that each reconstruction and update can be traced back to the original sampling time.

[0093] The model training data consists of historical real observation segments. During training, unequal-length missing intervals, arrival delay perturbations, and local quality label changes are constructed from complete segments. Unperturbed real values ​​are retained as supervision data. The training objectives include the deviation between the reconstructed values ​​and the real values, the deviation between the candidate time delay and the segment response order, and the consistency cost of the candidate sequence. During training, the multi-scale temporal feature extraction layer first obtains a stable time representation, and then the time delay estimation layer, graph attention network, and multi-scale temporal decoder are jointly updated. The side-state memory is updated in the segment order during the training phase to maintain its consistency with the continuous operation process.

[0094] During the deployment and operation phase, the system triggers incremental processing based on the new observations received. The data processor only recalculates the time delay distribution and local subgraph for overlapping time windows affected by the new records, without recalculating historical time slices that have been sealed and do not contain late data. When a late real observation falls into a reconstructed time slice, the corresponding parameter snapshot and edge state version are read to perform local correction. The correction results are verified by historical playback and then written to the new version. The old version is retained until the new version completes continuous segment verification, so that asynchronous data processing, dynamic graph updates and model correction maintain a consistent time index and version relationship.

[0095] In this embodiment, all numerical calculations are performed while preserving the missing mask, quality markers, and original timestamp mappings. The deep learning network only processes encoded features and candidate reconstructions without altering the original observation records. Dynamic time delays are used to determine comparable feature positions between variables. Dynamic spatiotemporal correlation graphs are used to limit information transmission paths across variables and buoys. Consistency vectors are used to express the deviations between candidate sequences and temporal continuity, variable response order, propagation order of adjacent buoys, and historical stable intervals. Credibility is used to control the contribution of subsequent nodes, edge weights, and propagation path gating. Late real observations are used to correct decoding parameters, time delay distribution, and edge state versions, thereby completing the continuous perception, reliable reconstruction, and retrospective updating of marine environmental data from nuclear power buoys.

Claims

1. A deep learning-based marine environmental data sensing system for nuclear power buoys, characterized in that, It includes a data processor and a memory connected to the data processor, the memory storing a program executed by the data processor, the program instructing the data processor to: construct an asynchronous observation sequence based on the monitoring variables uploaded by the nuclear power buoy, the original sampling time, the data arrival time and the mass marker; The dynamic time delay between each monitored variable is obtained through a deep time delay coding network, and a time delay correction time slice is formed. A dynamic spatiotemporal correlation graph is generated using the monitored variables and the nuclear power buoy as nodes and the dynamic time delay and spatial correlation as edges. Based on the dynamic spatiotemporal correlation graph, candidate values ​​for reconstructing missing or abnormal data are generated. The reconstruction candidate values ​​are screened and their credibility is determined based on the time continuity, variable response order, and propagation order of adjacent buoys; Adjust the node feature contribution and edge weight of subsequent time slices according to the aforementioned credibility.

2. The deep learning-based marine environmental data sensing system for nuclear power buoys according to claim 1, characterized in that, The deep time-delay coding network includes a multi-scale temporal feature extraction layer, a variable pair association layer, and a time delay estimation layer; The multi-scale time series feature extraction layer encodes the short-period fluctuations, tidal period changes, and long-term trends of each monitored variable. The variable pair correlation layer forms a joint feature sequence for any two monitored variables within their overlapping effective observation intervals. The time delay estimation layer generates a time delay distribution based on the direction of change, local phase difference, and rate of change difference in the joint feature sequence, and performs a differentiable offset on the feature time index of the corresponding monitored variable according to the time delay distribution to form a time delay correction time slice that retains the original timestamp mapping relationship.

3. The deep learning-based marine environmental data sensing system for nuclear power buoys according to claim 2, characterized in that, The dynamic spatiotemporal correlation graph includes variable layer nodes, buoy layer nodes, and cross-layer edges connecting nodes in two layers. The variable layer nodes carry the temporal characteristics of different monitoring variables within the same nuclear power buoy, and the buoy layer nodes carry the aggregation characteristics of different nuclear power buoys in the same time slice; The initial weights of the cross-layer edges are determined by historical stable associations, spatial proximity relationships between nuclear power buoys, and dynamic time delays, and are modified by the graph attention network based on the consistency of local changes in the current time slice. The graph attention network retains the edge state of the previous time slice in each time slice, and encodes the newly added edges, persistent edges, and decaying edges respectively before inputting them into the reconstruction network.

4. The deep learning-based marine environmental data sensing system for nuclear power buoys according to claim 3, characterized in that, The reconstruction candidate values ​​are jointly determined by the data-driven branch and the consistency constraint branch; The data-driven branch extracts neighborhood aggregation features from the dynamic spatiotemporal correlation graph through a graph neural network, and generates multiple candidate sequences corresponding to missing or abnormal positions through a multi-scale temporal decoder. The consistency constraint branch calculates the time continuity deviation, variable response order deviation, adjacent buoy propagation order deviation, and historical stable interval deviation for each candidate sequence, and constructs a consistency vector based on each deviation. The data processor jointly encodes the consistency vector and the uncertainty of the candidate sequence, and outputs the reconstruction candidate value and the confidence level corresponding to the reconstruction candidate value.

5. The deep learning-based marine environmental data sensing system for nuclear power buoys according to claim 2, characterized in that, The variable pair association layer establishes a validity mask matrix and a multi-window association matrix for each variable pair. The validity mask matrix records the data availability status, quality mark, and arrival delay status of the variable pair at each time index. The multi-window association matrix records the short window change synchronization degree, the medium window phase migration degree, and the long window trend consistency degree, respectively. The time delay estimation layer performs gated fusion of the validity mask matrix and the multi-window correlation matrix to generate multiple candidate time delays and corresponding weights. It performs cross-variable feature shift only on candidate time delays that maintain the same direction in continuous time slices, and retains the remaining candidate time delays as low-weight time delay components in the time delay distribution.

6. The deep learning-based marine environmental data sensing system for nuclear power buoys according to claim 3, characterized in that, The dynamic spatiotemporal correlation graph is divided into multiple local subgraphs according to the monitoring variable category and the spatial neighborhood of the nuclear power buoy, and propagation edges connecting different local subgraphs are set. The node features within each local subgraph are formed by splicing together the current observation features, time delay correction features, and historical reconstruction confidence. The edge features of the propagation edge are formed by splicing together the spatial distance code, dynamic time delay code, and historical propagation order code. The graph attention network calculates the attention within the local subgraph and the attention propagated across the subgraph, respectively, and performs gating normalization on the attention coefficients based on the confidence level of the corresponding nodes. The normalized local aggregation results and the cross-subgraph propagation results are then input into the reconstruction network.

7. The deep learning-based marine environmental data sensing system for nuclear power buoys according to claim 4, characterized in that, The multi-scale temporal decoder generates candidate sequences based on short-period context, tidal period context, and long-term trend context for the same missing position or the abnormal position, respectively. The consistency constraint branch forms a corresponding consistency vector for each candidate sequence and maps each consistency vector to a candidate selection weight. The data processor performs weighted fusion of the candidate sequences according to the candidate selection weights, and generates uncertainty based on the dispersion between each candidate sequence and the characteristic perturbation response of the graph neural network. The uncertainty and the weighted fusion result are stored as reconstruction records for subsequent time slice calls.

8. The deep learning-based marine environmental data sensing system for nuclear power buoys according to claim 7, characterized in that, When subsequent real observation data arrives, the data processor matches the real observation data with the corresponding reconstructed record according to the original sampling time to generate a reconstruction deviation sequence and a time delay deviation sequence. The decoding parameters associated with the corresponding monitoring variables in the data-driven branch are corrected according to the reconstructed deviation sequence, the time delay distribution parameters of the deep time delay coding network are corrected according to the time delay deviation sequence, and the changes in edge weights before and after correction are written into the edge state memory. When the deviation directions corresponding to multiple consecutive real observation segments are consistent, the state of the continuous edge in the dynamic spatiotemporal correlation graph is updated; otherwise, the edge state before correction is restored.

9. The deep learning-based marine environmental data sensing system for nuclear power buoys according to claim 6, characterized in that, The graph attention network establishes a credible feature channel and a weak evidence feature channel within the local subgraph. Node features with a credibility not lower than the credibility quantile of the corresponding time slice enter the credible feature channel, while the remaining node features, along with the quality label, enter the weak evidence feature channel. The trusted feature channel updates the node state using the neighborhood aggregation result, and the weak evidence feature channel updates the node state using the restricted aggregation result with residual preservation. The node states of the two channels are then passed to the cross-subgraph propagation attention. The cross-subgraph propagation attention is gating the propagation edges based on the product encoding of the credibility of each node on the propagation path.

10. The deep learning-based marine environmental data sensing system for nuclear power buoys according to claim 8, characterized in that, The edge state memory establishes a hierarchical index according to the monitoring variable pairs, nuclear power buoy pairs and sea state characteristic intervals. Each index entry records the edge weight version, the corresponding time delay distribution, the reconstruction deviation statistics and the actual observation coverage. Before updating the persistent edge state, the data processor selects historical index items similar to the current time slice from the same sea state feature interval and plays them back, and compares the consistency of the currently updated reconstruction result with the real observations corresponding to the historical index items. When the time continuity deviation of the reconstructed result after the update or the deviation of the variable response order is greater than the corresponding deviation before the correction, the update of the current edge weight version and the corresponding time delay distribution is cancelled, and the deviation sequence is written into the weak evidence record of the subsequent update.

Images