An edge-computing-based water traffic risk prediction method

Abstract

This invention discloses a method for predicting water traffic risks based on edge computing, comprising the following steps: acquiring water traffic data and performing data preprocessing and partitioning to obtain a subset of partitioned samples; performing state decomposition processing on the subset of partitioned samples to obtain a candidate kernel space sample set; performing Nyström kernel mean embedding on the candidate kernel space sample set to form a local risk distribution representation; establishing a sliding memory pool set to generate new state samples when water traffic data is updated; establishing a cross-partition embedding relay rule to transfer the local risk distribution representation across partitions to form a feedforward risk distribution representation; setting a frozen state and a refresh state for the Nyström kernel mean embedding to form an updated local risk distribution representation; and performing risk prediction processing to generate a risk prediction result. This invention improves the real-time performance of edge-side water traffic risk prediction.

Description

Technical Field

[0001] This invention relates to the field of water traffic safety and intelligent monitoring technology, and in particular to a water traffic risk prediction method based on edge computing. Background Technology

[0002] With the continuous growth of ship traffic and the increasing number of complex navigation scenarios such as ports, waterways, bridge areas, and winding waterways, the operational status of waterway traffic exhibits high dynamism, strong coupling, and significant regional correlation. Ships are affected not only by their own speed, course, and position changes during navigation, but also by a combination of factors including encounters with surrounding vessels, local traffic density, channel boundary constraints, and hydrological and meteorological disturbances. Therefore, timely and accurate prediction of waterway traffic risks has become a crucial issue in waterway traffic safety management. Current technologies for waterway traffic risk prediction typically rely on automatic identification systems, radar monitoring systems, video monitoring systems, and hydrological and meteorological monitoring systems to acquire operational data, and then use statistical analysis, rule-based judgment, or machine learning methods to estimate the risk status for future periods.

[0003] However, existing technologies still rely on a relatively simplistic approach to characterizing maritime traffic risks. They often rely directly on raw traffic data or simply extracted statistical features for prediction, lacking the ability to hierarchically represent multidimensional risk factors such as motion evolution, encounter relationships, environmental disturbances, and regional constraints. This makes it difficult to accurately depict the intrinsic mechanisms of risk formation and evolution in local waters. For cross-regional risk propagation issues, existing solutions typically focus on state analysis within a single region, failing to adequately consider the feedforward risk impacts caused by changes in navigation flow direction, encounter relationships, and regional constraint transmission between adjacent waters. Consequently, the prediction results do not fully reflect the risk diffusion and risk relay process.

[0004] Therefore, how to provide a method for predicting water traffic risks based on edge computing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a water traffic risk prediction method based on edge computing. This invention improves the real-time performance, accuracy of local risk characterization, and cross-regional risk transmission perception capability of water traffic risk prediction on the edge side, reduces the computational overhead and data transmission pressure caused by continuous reconstruction of edge nodes, and improves the stability and accuracy of risk prediction results in complex navigation environments.

[0006] A method for predicting water traffic risks based on edge computing according to an embodiment of the present invention includes the following steps: Acquire water traffic data within the water area covered by edge nodes and perform data preprocessing and partitioning to obtain partitioned sample subsets; The partitioned sample subset is decomposed into a state decomposition sample set, and a hierarchical kernel operator mapping is performed to obtain a candidate kernel space sample set. Calculate the risk-sensitive threshold value for the candidate kernel space sample set, filter the candidate kernel space sample set according to the risk-sensitive threshold value, form the Nyström anchor point sample set, and perform Nyström kernel mean embedding on the candidate kernel space sample set based on the Nyström anchor point sample set to form a local risk distribution representation; A sliding memory pool set is established and a subset of partitioned samples is written to it, forming new state samples when water traffic data is updated; Based on the risk functional partition set, a cross-partition embedding relay rule is established, and the local risk distribution representation is transmitted across partitions according to the cross-partition embedding relay rule to form a feedforward risk distribution representation; The Nyström kernel mean embedding is set to a frozen state and a refresh state. When the stability condition is met, the local risk distribution representation remains unchanged. When the risk triggering event is met, the embedding is reconstructed based on the newly added state samples and the sliding memory pool set to form an updated local risk distribution representation. Risk prediction processing is performed based on the updated local risk distribution representation and feedforward risk distribution representation to generate risk prediction results.

[0007] Optionally, obtaining the partitioned sample subset specifically includes: Acquire water traffic data within the water area covered by edge nodes, perform data preprocessing on the water traffic data, and form a local traffic state sample set; Simultaneously, channel topology information and regional constraint information are acquired from electronic navigation data, and zoning processing is performed on the waters covered by edge nodes to form a set of risk functional zones; The local traffic status sample set is input into the risk functional zoning set, and the zoning is performed according to the spatial location and navigation affiliation of each local traffic status sample within the water area covered by the edge node, forming a zoning sample subset.

[0008] Optionally, obtaining the candidate kernel space sample set specifically includes: According to the sampling time sequence and partition affiliation of each sample in the partitioned sample subset, the state feature information of each sample is extracted and organized to form a one-to-one corresponding original state component. The original state components are decomposed into motion evolution state components, encounter relationship state components, environmental disturbance state components, and regional constraint state components. They are then combined according to the correspondence of each sample to form state decomposition samples. The state decomposition samples corresponding to the partition sample subsets are collected to form a state decomposition sample set. Each state decomposition sample in the state decomposition sample set is input into the hierarchical kernel mapping process. Motion layer kernel mapping is performed on the motion evolution state component, relationship layer kernel mapping is performed on the encounter relationship state component, disturbance layer kernel mapping is performed on the environmental disturbance state component, and constraint layer kernel mapping is performed on the regional constraint state component, thus forming a hierarchical mapping result corresponding to each state decomposition sample. According to the corresponding order of each state decomposition sample in the partition sample subset, the association combination processing and unified representation processing are performed on each hierarchical mapping result. The processed hierarchical mapping result is used as the candidate kernel space sample, and the candidate kernel space samples corresponding to the same partition sample subset are aggregated to obtain the candidate kernel space sample set.

[0009] Optionally, the local risk distribution representation is obtained by specifically including: Read the candidate kernel space sample set, extract the risk association information corresponding to each candidate kernel space sample in the corresponding order of each candidate kernel space sample in the partition sample subset, calculate the risk sensitivity threshold value for each candidate kernel space sample, and associate each risk sensitivity threshold value with the corresponding candidate kernel space sample to form a gated association sample sequence; Anchor point screening is performed on the gated associated sample sequence. The samples are sorted from largest to smallest according to their risk-sensitive threshold values. Candidate kernel space samples whose risk-sensitive threshold values ​​meet the preset conditions are selected as Nyström anchor point samples. Unselected candidate kernel space samples are used as non-anchor point samples. The risk-sensitive threshold values ​​corresponding to each Nyström anchor point sample are retained as anchor point gating coefficients, forming a Nyström anchor point sample set, a non-anchor point sample set, and an anchor point gating coefficient sequence. The Nyström anchor sample set, non-anchor sample set, and anchor gating coefficient sequence are used as inputs. Nyström kernel mean embedding is performed. Anchor main embedding is performed on the Nyström anchor sample set. An internal kernel association structure is constructed based on the kernel similarity relationship between each Nyström anchor sample within the Nyström anchor sample set. At the same time, the kernel association strength corresponding to each Nyström anchor sample within the anchor point is gated and modulated according to the anchor gating coefficient sequence to obtain the gated and modulated anchor kernel association structure. Anchor kernel representations corresponding to each Nyström anchor sample are generated based on the gated and modulated anchor kernel association structure, and mean aggregation is performed on each anchor kernel representation to form the anchor main embedding result. Anchor association projection is performed on the non-anchor sample set. The non-anchor sample set is projected to the embedding space corresponding to the anchor main embedding result according to the kernel association relationship with the Nyström anchor sample set. The projection result is then gated and corrected in combination with the anchor gating coefficient sequence to form the non-anchor compensated embedding result. Gated fusion processing is performed on the anchor point main embedding results and the non-anchor point compensated embedding results. The high-risk representation components in the anchor point main embedding results are enhanced according to each risk-sensitive threshold value, and the low-risk background components in the non-anchor point compensated embedding results are constrained to form a local risk distribution representation.

[0010] Optionally, obtaining the newly added state samples specifically includes: Read the partition sample subset within the edge node, establish a sliding memory pool corresponding to each partition sample subset, and write each sample in each partition sample subset into the corresponding sliding memory pool according to the sampling time sequence. When water traffic data is updated at the edge node, data preprocessing is performed on the updated water traffic data, and partition mapping is performed according to the partition mapping relationship to obtain the updated data corresponding to the partition. The updated data corresponding to the partition is written to the corresponding sliding memory pool, and the sliding update process is performed on the samples in the corresponding sliding memory pool according to the sample writing time. The samples within the preset time window are retained, and the samples outside the preset time window are removed. At the same time, the corresponding samples newly written after the sliding update process are extracted from each sliding memory pool to form the new state samples.

[0011] Optionally, the formation of the feedforward risk distribution representation specifically includes: Read each risk functional zone in the risk functional zone set, extract the navigation flow direction relationship, encounter relationship evolution relationship and regional constraint transmission relationship between each risk functional zone, and determine the transmission direction, transmission sequence and transmission object between each risk functional zone based on the navigation flow direction relationship, encounter relationship evolution relationship and regional constraint transmission relationship, forming cross-zone embedded relay rules; According to the cross-partition embedding relay rule, the local risk distribution representation in each risk functional partition is matched with the upstream partition and the current partition. The local risk distribution representation that meets the cross-partition transmission direction is determined as the embedding information to be transmitted, and the embedding information to be transmitted is transmitted to the corresponding current risk functional partition. The embedded information to be transmitted is processed by relay adjustment, and the directional consistency adjustment is performed based on the navigation flow relationship, the risk evolution correlation adjustment is performed based on the encounter relationship evolution relationship, and the constraint impact adjustment is performed based on the regional constraint transmission relationship, thus forming feedforward embedded information; The feedforward embedded information and the local risk distribution representation corresponding to the current risk functional partition are subjected to feedforward fusion processing to form a feedforward risk distribution representation.

[0012] Optionally, obtaining the updated local risk distribution representation specifically includes: Within the edge node, set the frozen state and refresh state for the Nyström kernel mean embedding corresponding to the current partition, configure the current local risk distribution representation as the preserved result corresponding to the frozen state, and configure the newly added state samples and the sliding memory pool set as the reconstruction input corresponding to the refresh state; Based on the comparison of newly added state samples with existing samples in the sliding memory pool, the degree of change and risk of local traffic state are judged, and the switching state between the Nyström kernel mean and the frozen state is determined according to the judgment result. When the local traffic state meets the stability condition, the Nyström kernel mean embedding is kept in a frozen state. The preserved result corresponding to the frozen state is called as the local risk distribution representation of the current partition, and the current local risk distribution representation is used as the preserved result under the frozen state. When the local traffic state meets the risk triggering event, the Nyström kernel mean embedding is switched to the refresh state. Triggering samples corresponding to the risk triggering event are extracted from the newly added state samples, and existing samples that match the triggering samples in terms of partition affiliation, sampling time sequence and risk association are extracted from the sliding memory pool set. The triggering samples and the extracted existing samples are aggregated to form the embedding reconstruction samples. The embedded reconstruction samples are subjected to state decomposition processing to form reconstructed state decomposition samples. Hierarchical kernel operator mapping is performed on the reconstructed state decomposition samples to form reconstructed candidate kernel space samples. Risk-sensitive gate values ​​are calculated on the reconstructed candidate kernel space samples. Reconstructed candidate kernel space samples are selected according to the risk-sensitive gate values ​​to form reconstructed Nyström anchor point samples. Based on the reconstructed Nyström anchor samples, Nyström kernel mean embedding is performed on the reconstructed candidate kernel space samples to form an updated local risk distribution representation.

[0013] Optionally, obtaining the risk prediction result specifically includes: The local risk characterization information in the updated local risk distribution representation and the cross-regional risk transmission information in the feedforward risk distribution representation are extracted and aligned according to the corresponding region and the corresponding prediction time window to form risk prediction input information; The risk prediction input information is subjected to correlation and fusion processing, which correlates and jointly represents the local risk characterization information with the cross-regional risk transmission information to form a prediction fusion representation. The risk prediction results are generated by calculating the risk evolution trend, risk transmission trend, and risk clustering trend in the prediction fusion representation according to the prediction time window.

[0014] The beneficial effects of this invention are: This invention preprocesses, partitions, decomposes, and maps hierarchical kernel operators to water traffic data within the water area covered by edge nodes. This results in candidate kernel space samples that better reflect local navigation conditions, improving the ability to characterize local risk states in complex water areas and thus enhancing the accuracy and stability of risk prediction results.

[0015] This invention calculates risk-sensitive threshold values ​​and filters Nyström anchor point samples, and combines Nyström kernel mean embedding to form a local risk distribution representation. This can reduce the computational burden under the condition of limited computing resources on the edge side, while enhancing the characterization of high-risk state samples and improving the real-time performance and local identification capability of risk prediction.

[0016] This invention establishes a sliding memory pool set and combines frozen and refresh states to control the update process of the local risk distribution representation. This can reduce redundant calculations when the local traffic conditions are stable and complete the embedding reconstruction in a timely manner when risk-triggered events occur, thus balancing edge-side computational efficiency and emergency risk response capabilities.

[0017] This invention establishes cross-regional embedded relay rules to form a feedforward risk distribution representation, and combines it with the updated local risk distribution representation to perform risk prediction processing. This enhances the ability to perceive cross-regional risk propagation and continuous evolution risks, and improves the prediction effect of collision, aggregation, conflict, yaw and navigation restriction risks. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a water traffic risk prediction method based on edge computing proposed in this invention; Figure 2 This is a schematic diagram of the Nyström kernel mean embedding algorithm for a water traffic risk prediction method based on edge computing proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-2 A method for predicting water traffic risks based on edge computing includes the following steps: Acquire water traffic data within the water area covered by edge nodes and perform data preprocessing and partitioning to obtain partitioned sample subsets; The partitioned sample subset is decomposed into a state decomposition sample set, and a hierarchical kernel operator mapping is performed to obtain a candidate kernel space sample set. Calculate the risk-sensitive threshold value for the candidate kernel space sample set, filter the candidate kernel space sample set according to the risk-sensitive threshold value, form the Nyström anchor point sample set, and perform Nyström kernel mean embedding on the candidate kernel space sample set based on the Nyström anchor point sample set to form a local risk distribution representation; A sliding memory pool set is established and a subset of partitioned samples is written to it, forming new state samples when water traffic data is updated; Based on the risk functional partition set, a cross-partition embedding relay rule is established, and the local risk distribution representation is transmitted across partitions according to the cross-partition embedding relay rule to form a feedforward risk distribution representation; The Nyström kernel mean embedding is set to a frozen state and a refresh state. When the stability condition is met, the local risk distribution representation remains unchanged. When the risk triggering event is met, the embedding is reconstructed based on the newly added state samples and the sliding memory pool set to form an updated local risk distribution representation. Risk prediction processing is performed based on the updated local risk distribution representation and feedforward risk distribution representation to generate risk prediction results.

[0021] In this embodiment, obtaining the partitioned sample subset specifically includes: Acquire water traffic data within the water area covered by edge nodes, and perform data preprocessing on the water traffic data. Data preprocessing includes time alignment, spatial unification, target association, anomaly removal, and state completion processing to form a local traffic state sample set. Simultaneously, channel topology information and regional constraint information are acquired from electronic navigation data, and zoning processing is performed on the waters covered by edge nodes to form a set of risk functional zones. Electronic navigation data is digital navigation information data that characterizes the channel structure, navigation boundaries and regional constraint relationships within the waters covered by edge nodes. The local traffic status sample set is input into the risk functional zoning set, and the zoning is performed according to the spatial location and navigation affiliation of each local traffic status sample within the water area covered by the edge node, forming a zoning sample subset.

[0022] In this embodiment, obtaining the candidate kernel space sample set specifically includes: According to the sampling time sequence and partition affiliation of each sample in the partitioned sample subset, the state feature information of each sample is extracted and organized to form a one-to-one corresponding original state component. The state feature information includes ship position change information, speed and heading change information, ship encounter relationship information, environmental disturbance information, and regional constraint information. The original state components are decomposed into motion evolution state components, encounter relationship state components, environmental disturbance state components, and regional constraint state components. They are then combined according to the correspondence of each sample to form state decomposition samples. The state decomposition samples corresponding to the partition sample subsets are collected to form a state decomposition sample set. Each state decomposition sample in the state decomposition sample set is input into the hierarchical kernel mapping process. Motion layer kernel mapping is performed on the motion evolution state component, relationship layer kernel mapping is performed on the encounter relationship state component, disturbance layer kernel mapping is performed on the environmental disturbance state component, and constraint layer kernel mapping is performed on the regional constraint state component, thus forming a hierarchical mapping result corresponding to each state decomposition sample. According to the corresponding order of each state decomposition sample in the partition sample subset, the association combination processing and unified representation processing are performed on each hierarchical mapping result. The processed hierarchical mapping result is used as the candidate kernel space sample, and the candidate kernel space samples corresponding to the same partition sample subset are aggregated to obtain the candidate kernel space sample set.

[0023] In this embodiment, obtaining the representation of local risk distribution specifically includes: Read the candidate kernel space sample set, extract the risk association information corresponding to each candidate kernel space sample in the corresponding order of each candidate kernel space sample in the partition sample subset, calculate the risk sensitivity threshold value for each candidate kernel space sample, and associate each risk sensitivity threshold value with the corresponding candidate kernel space sample to form a gated association sample sequence; Anchor point screening is performed on the gated associated sample sequence. The samples are sorted from largest to smallest according to their risk-sensitive threshold values. Candidate kernel space samples whose risk-sensitive threshold values ​​meet the preset conditions are selected as Nyström anchor point samples. Unselected candidate kernel space samples are used as non-anchor point samples. The risk-sensitive threshold values ​​corresponding to each Nyström anchor point sample are retained as anchor point gating coefficients, forming a Nyström anchor point sample set, a non-anchor point sample set, and an anchor point gating coefficient sequence. The Nyström anchor sample set, non-anchor sample set, and anchor gating coefficient sequence are used as inputs. Nyström kernel mean embedding is performed. Anchor main embedding is performed on the Nyström anchor sample set. An internal kernel association structure is constructed based on the kernel similarity relationship between each Nyström anchor sample within the Nyström anchor sample set. At the same time, the kernel association strength corresponding to each Nyström anchor sample within the anchor point is gated and modulated according to the anchor gating coefficient sequence to obtain the gated and modulated anchor kernel association structure. Anchor kernel representations corresponding to each Nyström anchor sample are generated based on the gated and modulated anchor kernel association structure, and mean aggregation is performed on each anchor kernel representation to form the anchor main embedding result. Anchor association projection is performed on the non-anchor sample set. The non-anchor sample set is projected to the embedding space corresponding to the anchor main embedding result according to the kernel association relationship with the Nyström anchor sample set. The projection result is then gated and corrected in combination with the anchor gating coefficient sequence to form the non-anchor compensated embedding result. Gated fusion processing is performed on the anchor point main embedding results and the non-anchor point compensated embedding results. The high-risk representation components in the anchor point main embedding results are enhanced according to each risk-sensitive threshold value, and the low-risk background components in the non-anchor point compensated embedding results are constrained to form a local risk distribution representation.

[0024] This invention introduces a risk-sensitive gating value to drive the screening of Nyström anchor points, and performs gating modulation and fusion of anchor point kernel correlation and non-anchor point projection results during the Nyström kernel mean embedding process. This enhances the characterization ability of high-risk state samples, suppresses low-risk background interference, and improves the accuracy of local risk distribution representation in depicting the formation, aggregation, and evolution of complex water traffic risks, thereby improving the accuracy, stability, and relevance of risk prediction results.

[0025] In this embodiment, obtaining the new state samples specifically includes: Read the partition sample subset within the edge node, establish a sliding memory pool corresponding to each partition sample subset, and write each sample in each partition sample subset into the corresponding sliding memory pool according to the sampling time sequence. When water traffic data is updated at the edge node, data preprocessing is performed on the updated water traffic data, and partition mapping is performed according to the partition mapping relationship to obtain the updated data corresponding to the partition. The updated data corresponding to the partition is written to the corresponding sliding memory pool, and the sliding update process is performed on the samples in the corresponding sliding memory pool according to the sample writing time. The samples within the preset time window are retained, and the samples outside the preset time window are removed. At the same time, the corresponding samples newly written after the sliding update process are extracted from each sliding memory pool to form the new state samples.

[0026] In this embodiment, the formation of the feedforward risk distribution representation specifically includes: Read each risk functional zone in the risk functional zone set, extract the navigation flow direction relationship, encounter relationship evolution relationship and regional constraint transmission relationship between each risk functional zone, and determine the transmission direction, transmission sequence and transmission object between each risk functional zone based on the navigation flow direction relationship, encounter relationship evolution relationship and regional constraint transmission relationship, forming cross-zone embedded relay rules; According to the cross-partition embedding relay rule, the local risk distribution representation in each risk functional partition is matched with the upstream partition and the current partition. The local risk distribution representation that meets the cross-partition transmission direction is determined as the embedding information to be transmitted, and the embedding information to be transmitted is transmitted to the corresponding current risk functional partition. The embedded information to be transmitted is processed by relay adjustment, and the directional consistency adjustment is performed based on the navigation flow relationship, the risk evolution correlation adjustment is performed based on the encounter relationship evolution relationship, and the constraint impact adjustment is performed based on the regional constraint transmission relationship, thus forming feedforward embedded information; The feedforward embedded information and the local risk distribution representation corresponding to the current risk functional partition are subjected to feedforward fusion processing to form a feedforward risk distribution representation.

[0027] This invention establishes cross-regional embedded relay rules and combines navigation flow direction relationships, encounter relationship evolution relationships, and regional constraint transmission relationships to perform cross-regional transmission, adjustment, and fusion of local risk distribution representations. This enhances the perception of risk propagation, risk relay, and forward impact of risks in adjacent waters, improves the representation effect of feedforward risk distribution representations on continuous and diffuse risks, and thus enhances the comprehensiveness, accuracy, and foresight of water traffic risk prediction in complex navigation environments.

[0028] In this embodiment, obtaining the updated local risk distribution representation specifically includes: Within the edge node, the Nyström kernel mean embedding corresponding to the current partition is set to a frozen state and a refresh state. The current local risk distribution representation is configured as the preserved result corresponding to the frozen state, and the newly added state samples and the sliding memory pool set are configured as the reconstruction input corresponding to the refresh state. The frozen state is the state in which the local risk distribution representation remains unchanged and the Nyström kernel mean embedding reconstruction is stopped when the stability condition is met. The refresh state is the state in which the Nyström kernel mean embedding reconstruction is re-executed based on the newly added state samples and the sliding memory pool set when the risk trigger event is met. Based on the comparison of newly added state samples with existing samples in the sliding memory pool, the degree of change and risk of local traffic state are judged, and the switching state between the Nyström kernel mean and the frozen state is determined according to the judgment result. When the local traffic state meets the stability condition, the Nyström kernel mean embedding is kept in a frozen state. The preserved result corresponding to the frozen state is called as the local risk distribution representation of the current partition, and the current local risk distribution representation is used as the preserved result under the frozen state. When the local traffic state meets the risk triggering event, the Nyström kernel mean embedding is switched to the refresh state. Triggering samples corresponding to the risk triggering event are extracted from the newly added state samples, and existing samples that match the triggering samples in terms of partition affiliation, sampling time sequence and risk association are extracted from the sliding memory pool set. The triggering samples and the extracted existing samples are aggregated to form the embedding reconstruction samples. The embedded reconstruction samples are subjected to state decomposition processing to form reconstructed state decomposition samples. Hierarchical kernel operator mapping is performed on the reconstructed state decomposition samples to form reconstructed candidate kernel space samples. Risk-sensitive gate values ​​are calculated on the reconstructed candidate kernel space samples. Reconstructed candidate kernel space samples are selected according to the risk-sensitive gate values ​​to form reconstructed Nyström anchor point samples. Based on the reconstructed Nyström anchor samples, Nyström kernel mean embedding is performed on the reconstructed candidate kernel space samples to form an updated local risk distribution representation.

[0029] This invention sets frozen and refresh states for Nyström kernel mean embedding within edge nodes, and combines newly added state samples with a sliding memory pool set to determine local traffic state changes. Under stable conditions, the local risk distribution representation remains unchanged, while under risk-triggered events, the local risk distribution representation is reconstructed. This reduces unnecessary redundant calculations at the edge, lowers the computational power consumption caused by continuous embedding updates, and improves the timeliness of response to sudden risks, risk clusters, and risk evolution changes. It also enhances the dynamic updating capability of local risk representation and the real-time performance, accuracy, and stability of water traffic risk prediction results.

[0030] In this embodiment, obtaining the risk prediction results specifically includes: The local risk characterization information in the updated local risk distribution representation and the cross-regional risk transmission information in the feedforward risk distribution representation are extracted and aligned according to the corresponding region and the corresponding prediction time window to form risk prediction input information; The risk prediction input information is subjected to correlation and fusion processing, which correlates and jointly represents the local risk characterization information with the cross-regional risk transmission information to form a prediction fusion representation. The risk evolution trend, risk transmission trend, and risk clustering trend in the prediction fusion representation are calculated according to the prediction time window to generate risk prediction results. The risk prediction results are prediction results that reflect the changes in the water traffic operation status and risk development trend of the corresponding risk functional zone within the prediction time window. They are used to characterize the degree of risk of collision, clustering, conflict, deviation, or navigation restriction of ships in the corresponding water zone.

[0031] Example 1: To verify the feasibility of the present invention in practice, it was applied to a complex inland waterway navigation scenario consisting of a main channel, tributary channels, bridge-restricted waterways, confluence areas, and anchorage transition areas. The vessel types in this scenario include bulk carriers, container ships, engineering vessels, and small inspection vessels. The navigation conditions exhibit significant multi-source interference characteristics. On the one hand, the continuous passage of large-tonnage vessels in the main channel easily leads to encounters with significant speed differences in the bridge and confluence areas. On the other hand, the entry of vessels from tributary channels into the main channel causes a sudden increase in local traffic density. Combined with changes in wind speed, water flow deviation, visibility fluctuations, and boundary constraints, this can easily induce deviations, conflicts, localized congestion, and collision risks. Existing technologies in this scenario typically employ a centralized platform to aggregate data from the entire domain and perform unified calculations. This approach suffers from issues such as large data upload volumes, high processing latency, and insufficient response to sudden changes in local risk states. In particular, when local waterway traffic conditions alternate between stable and sudden events, the fixed-period full recalculation method leads to a waste of edge-side resources. Furthermore, simply making predictions based on the original state data is insufficient to effectively characterize cross-regional risk propagation and the forward impact of adjacent regions. This is precisely the technical problem that this invention aims to solve.

[0032] In this scenario, edge nodes are deployed at each navigation monitoring unit, ensuring that each edge node covers a corresponding water area and continuously receives data from the automatic identification system, shore-based radar tracking data, video recognition results, hydrological monitoring data, and meteorological monitoring data. The edge nodes first perform time alignment, spatial unification, target association, anomaly removal, and state completion processing on the received water traffic data to form a local traffic state sample set. Then, combining this with the channel topology, navigation boundary relationships, and regional constraints in the electronic waterway data, the covered water area is divided into several risk functional zones.

[0033] After forming a subset of samples within each zone, the state feature information reflecting changes in position, speed and heading, ship encounter relationships, environmental disturbances, and regional constraints is extracted and organized. This information is further decomposed into motion evolution state components, encounter relationship state components, environmental disturbance state components, and regional constraint state components. Then, a hierarchical kernel operator mapping is performed on each of these state components to obtain a candidate kernel space sample set. Unlike the traditional method of directly inputting raw samples into the risk prediction model, this invention first completes state decomposition and hierarchical mapping at the edge nodes, thereby transforming the factors contributing to complex maritime traffic risks into a uniformly processed kernel space representation, laying the foundation for subsequent construction of local risk distribution representations.

[0034] After the candidate kernel space sample set is formed, this invention continues to calculate the risk-sensitive gate value corresponding to each candidate kernel space sample. The risk-sensitive gate value not only reflects the risk level of the candidate kernel space sample in the current partition, but also reflects the importance of the sample in the subsequent Nyström kernel mean embedding. By screening Nyström anchor point samples according to the risk-sensitive gate value and retaining the risk-sensitive gate value as gating adjustment information, the embedding process can prioritize high-risk state samples, while suppressing the interference of low-risk background samples on the local risk distribution representation. After obtaining the local risk distribution representation, this invention does not employ a fixed-period full recalculation. Instead, it establishes a sliding memory pool within the edge nodes, corresponding one-to-one with the sample subsets of each partition. When water traffic data is updated, the processed updated data is written into the corresponding sliding memory pool, forming new state samples. When the new state samples do not show significant changes compared to existing samples in the sliding memory pool, the Nyström kernel mean embedding remains frozen, and the current local risk distribution representation is directly reused. When the degree of change in local traffic state, risk, or local density corresponding to the new state samples exceeds a set threshold, the system switches to a refresh state. Samples matching the risk-triggered event are selected from the new state samples and the sliding memory pool, and embedding reconstruction is performed, thereby forming the updated local risk distribution representation. This reduces the amount of repetitive computation in a stable state while ensuring responsiveness during sudden changes in local risks.

[0035] To verify the effectiveness of this invention, a comparative test was conducted in the same complex general aviation scenario, comparing the traditional centralized full recalculation scheme, the traditional edge node fixed-period update scheme, and the scheme of this invention. The traditional centralized full recalculation scheme uploads all monitoring data to the central platform for unified processing; the traditional edge node fixed-period update scheme completes data processing at the edge, but performs embedding reconstruction on all samples in each fixed period. The specific comparison data is shown in Table 1: Table 1. Performance Comparison Results of Different Risk Prediction Schemes in Complex General Aviation Scenarios

[0036] As shown in Table 1, in complex navigation scenarios, the proposed solution outperforms traditional centralized full recalculation and traditional edge fixed-period update solutions in terms of average single-risk prediction latency, average early warning time, overall prediction accuracy, high-risk event recall rate, pre-congestion identification rate, and yaw risk identification accuracy. Specifically, the average single-risk prediction latency is reduced to 1.4 seconds, the average early warning time is increased to 68 seconds, and the overall prediction accuracy reaches 93.8%. At the same time, the average processor utilization rate is reduced to 52.8%, and the data upload volume per unit time window is reduced to 22MB. This indicates that the proposed solution can effectively reduce the computational burden and data transmission pressure on the edge side while improving the real-time performance, accuracy, and foresight of risk prediction.

[0037] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting water traffic risks based on edge computing, characterized in that, Includes the following steps: Acquire water traffic data within the water area covered by edge nodes and perform data preprocessing and partitioning to obtain partitioned sample subsets; The partitioned sample subset is decomposed into a state decomposition sample set, and a hierarchical kernel operator mapping is performed to obtain a candidate kernel space sample set. Calculate the risk-sensitive threshold value for the candidate kernel space sample set, filter the candidate kernel space sample set according to the risk-sensitive threshold value, form the Nyström anchor point sample set, and perform Nyström kernel mean embedding on the candidate kernel space sample set based on the Nyström anchor point sample set to form a local risk distribution representation; A sliding memory pool set is established and a subset of partitioned samples is written to it, forming new state samples when water traffic data is updated; Based on the risk functional partition set, a cross-partition embedding relay rule is established, and the local risk distribution representation is transmitted across partitions according to the cross-partition embedding relay rule to form a feedforward risk distribution representation; The Nyström kernel mean embedding is set to a frozen state and a refresh state. When the stability condition is met, the local risk distribution representation remains unchanged. When the risk triggering event is met, the embedding is reconstructed based on the newly added state samples and the sliding memory pool set to form an updated local risk distribution representation. Risk prediction processing is performed based on the updated local risk distribution representation and feedforward risk distribution representation to generate risk prediction results.

2. The water traffic risk prediction method based on edge computing according to claim 1, characterized in that, The specific steps involved in obtaining the partitioned sample subset are as follows: Acquire water traffic data within the water area covered by edge nodes, perform data preprocessing on the water traffic data, and form a local traffic state sample set; Simultaneously, channel topology information and regional constraint information are acquired from electronic navigation data, and zoning processing is performed on the waters covered by edge nodes to form a set of risk functional zones; The local traffic status sample set is input into the risk functional zoning set, and the zoning is performed according to the spatial location and navigation affiliation of each local traffic status sample within the water area covered by the edge node, forming a zoning sample subset.

3. The water traffic risk prediction method based on edge computing according to claim 1, characterized in that, The specific steps involved in obtaining the candidate kernel space sample set are as follows: According to the sampling time sequence and partition affiliation of each sample in the partitioned sample subset, the state feature information of each sample is extracted and organized to form a one-to-one corresponding original state component. The original state components are decomposed into motion evolution state components, encounter relationship state components, environmental disturbance state components, and regional constraint state components. They are then combined according to the correspondence of each sample to form state decomposition samples. The state decomposition samples corresponding to the partition sample subsets are collected to form a state decomposition sample set. Each state decomposition sample in the state decomposition sample set is input into the hierarchical kernel mapping process. Motion layer kernel mapping is performed on the motion evolution state component, relationship layer kernel mapping is performed on the encounter relationship state component, disturbance layer kernel mapping is performed on the environmental disturbance state component, and constraint layer kernel mapping is performed on the regional constraint state component, thus forming a hierarchical mapping result corresponding to each state decomposition sample. According to the corresponding order of each state decomposition sample in the partition sample subset, the association combination processing and unified representation processing are performed on each hierarchical mapping result. The processed hierarchical mapping result is used as the candidate kernel space sample, and the candidate kernel space samples corresponding to the same partition sample subset are aggregated to obtain the candidate kernel space sample set.

4. The water traffic risk prediction method based on edge computing according to claim 1, characterized in that, The local risk distribution representation is obtained specifically by: Read the candidate kernel space sample set, extract the risk association information corresponding to each candidate kernel space sample in the corresponding order of each candidate kernel space sample in the partition sample subset, calculate the risk sensitivity threshold value for each candidate kernel space sample, and associate each risk sensitivity threshold value with the corresponding candidate kernel space sample to form a gated association sample sequence; Anchor point screening is performed on the gated associated sample sequence. The samples are sorted from largest to smallest according to their risk-sensitive threshold values. Candidate kernel space samples whose risk-sensitive threshold values ​​meet the preset conditions are selected as Nyström anchor point samples. Unselected candidate kernel space samples are used as non-anchor point samples. The risk-sensitive threshold values ​​corresponding to each Nyström anchor point sample are retained as anchor point gating coefficients, forming a Nyström anchor point sample set, a non-anchor point sample set, and an anchor point gating coefficient sequence. The Nyström anchor sample set, non-anchor sample set, and anchor gating coefficient sequence are used as inputs. Nyström kernel mean embedding is performed. Anchor main embedding is performed on the Nyström anchor sample set. An internal kernel association structure is constructed based on the kernel similarity relationship between each Nyström anchor sample within the Nyström anchor sample set. At the same time, the kernel association strength corresponding to each Nyström anchor sample within the anchor point is gated and modulated according to the anchor gating coefficient sequence to obtain the gated and modulated anchor kernel association structure. Anchor kernel representations corresponding to each Nyström anchor sample are generated based on the gated and modulated anchor kernel association structure, and mean aggregation is performed on each anchor kernel representation to form the anchor main embedding result. Anchor association projection is performed on the non-anchor sample set. The non-anchor sample set is projected to the embedding space corresponding to the anchor main embedding result according to the kernel association relationship with the Nyström anchor sample set. The projection result is then gated and corrected in combination with the anchor gating coefficient sequence to form the non-anchor compensated embedding result. Gated fusion processing is performed on the anchor point main embedding results and the non-anchor point compensated embedding results. The high-risk representation components in the anchor point main embedding results are enhanced according to each risk-sensitive threshold value, and the low-risk background components in the non-anchor point compensated embedding results are constrained to form a local risk distribution representation.

5. The water traffic risk prediction method based on edge computing according to claim 1, characterized in that, The acquisition of the newly added state samples specifically includes: Read the partition sample subset within the edge node, establish a sliding memory pool corresponding to each partition sample subset, and write each sample in each partition sample subset into the corresponding sliding memory pool according to the sampling time sequence. When water traffic data is updated at the edge node, data preprocessing is performed on the updated water traffic data, and partition mapping is performed according to the partition mapping relationship to obtain the updated data corresponding to the partition. The updated data corresponding to the partition is written to the corresponding sliding memory pool, and the sliding update process is performed on the samples in the corresponding sliding memory pool according to the sample writing time. The samples within the preset time window are retained, and the samples outside the preset time window are removed. At the same time, the corresponding samples newly written after the sliding update process are extracted from each sliding memory pool to form the new state samples.

6. The water traffic risk prediction method based on edge computing according to claim 1, characterized in that, The formation of the feedforward risk distribution representation specifically includes: Read each risk functional zone in the risk functional zone set, extract the navigation flow direction relationship, encounter relationship evolution relationship and regional constraint transmission relationship between each risk functional zone, and determine the transmission direction, transmission sequence and transmission object between each risk functional zone based on the navigation flow direction relationship, encounter relationship evolution relationship and regional constraint transmission relationship, forming cross-zone embedded relay rules; According to the cross-partition embedding relay rule, the local risk distribution representation in each risk functional partition is matched with the upstream partition and the current partition. The local risk distribution representation that meets the cross-partition transmission direction is determined as the embedding information to be transmitted, and the embedding information to be transmitted is transmitted to the corresponding current risk functional partition. The embedded information to be transmitted is processed by relay adjustment, and the directional consistency adjustment is performed based on the navigation flow relationship, the risk evolution correlation adjustment is performed based on the encounter relationship evolution relationship, and the constraint impact adjustment is performed based on the regional constraint transmission relationship, thus forming feedforward embedded information; The feedforward embedded information and the local risk distribution representation corresponding to the current risk functional partition are subjected to feedforward fusion processing to form a feedforward risk distribution representation.

7. The water traffic risk prediction method based on edge computing according to claim 1, characterized in that, The updated local risk distribution representation is obtained specifically through: Within the edge node, set the frozen state and refresh state for the Nyström kernel mean embedding corresponding to the current partition, configure the current local risk distribution representation as the preserved result corresponding to the frozen state, and configure the newly added state samples and the sliding memory pool set as the reconstruction input corresponding to the refresh state; Based on the comparison of newly added state samples with existing samples in the sliding memory pool, the degree of change and risk of local traffic state are judged, and the switching state between the Nyström kernel mean and the frozen state is determined according to the judgment result. When the local traffic state meets the stability condition, the Nyström kernel mean embedding is kept in a frozen state. The preserved result corresponding to the frozen state is called as the local risk distribution representation of the current partition, and the current local risk distribution representation is used as the preserved result under the frozen state. When the local traffic state meets the risk triggering event, the Nyström kernel mean embedding is switched to the refresh state. Triggering samples corresponding to the risk triggering event are extracted from the newly added state samples, and existing samples that match the triggering samples in terms of partition affiliation, sampling time sequence and risk association are extracted from the sliding memory pool set. The triggering samples and the extracted existing samples are aggregated to form the embedding reconstruction samples. The embedded reconstruction samples are subjected to state decomposition processing to form reconstructed state decomposition samples. Hierarchical kernel operator mapping is performed on the reconstructed state decomposition samples to form reconstructed candidate kernel space samples. Risk-sensitive gate values ​​are calculated on the reconstructed candidate kernel space samples. Reconstructed candidate kernel space samples are selected according to the risk-sensitive gate values ​​to form reconstructed Nyström anchor point samples. Based on the reconstructed Nyström anchor samples, Nyström kernel mean embedding is performed on the reconstructed candidate kernel space samples to form an updated local risk distribution representation.

8. The water traffic risk prediction method based on edge computing according to claim 1, characterized in that, The risk prediction results are obtained specifically through: The local risk characterization information in the updated local risk distribution representation and the cross-regional risk transmission information in the feedforward risk distribution representation are extracted and aligned according to the corresponding region and the corresponding prediction time window to form risk prediction input information; The risk prediction input information is subjected to correlation and fusion processing, which correlates and jointly represents the local risk characterization information with the cross-regional risk transmission information to form a prediction fusion representation. The risk prediction results are generated by calculating the risk evolution trend, risk transmission trend, and risk clustering trend in the prediction fusion representation according to the prediction time window.

Images