Trackside perception enhancement method for port-specific railways

Abstract

The present application relates to the field of vehicle-mounted sensing technology, and specifically discloses a trackside sensing enhancement method for a port special railway, comprising: constructing multi-source sensing data according to train operation targets, loading and unloading machinery targets, operation personnel targets, trackside facility targets and environmental disturbance information in a port railway operation area; enhancing identification of key targets of the port railway in the standardized multi-source sensing data results, and constructing an enhanced target set; performing cross-source collaborative correlation and scene relationship enhancement modeling on the enhanced target set to obtain a port railway relationship enhancement graph; performing port railway scene risk identification and scene-based risk attribution according to the port railway relationship enhancement graph; and performing risk classification early warning and disposal according to a port railway risk event set. The trackside sensing enhancement method for the port special railway provided by the present application can solve the problems of limited far front sight distance, lateral intrusion blind area and lagging obstacle identification behind the shelter of vehicle-mounted sensing.

Description

Technical Field

[0001] This invention relates to the field of vehicle-mounted sensing technology, and in particular to a method for enhancing trackside sensing on a dedicated port railway. Background Technology

[0002] Port-dedicated railways undertake tasks such as loading and unloading, shunting, marshalling, and cargo transfer within the port area. They are characterized by short lines, dense tracks, numerous switches, concentrated operating equipment, and frequent overlap between railway operations and port machinery operations, making their operating environment significantly different from ordinary trunk railways. With the increasing automation and intelligence of port railways, higher demands are placed on the timeliness of train automatic or assisted driving systems in perceiving obstacles ahead, judging track passability, and handling risks. Existing automatic trains typically use onboard cameras, radar, or lidar as their primary sensing methods. However, in port scenarios, factors such as curves, switches, container yards, loading and unloading machinery, lateral crossings, and complex obstructed environments limit the ability of onboard perception to detect risks ahead, lateral intrusions, and sudden risks arising from obstructions. Furthermore, existing trackside monitoring solutions mostly employ video surveillance or multi-sensor intrusion alarms, primarily issuing alarms after a target enters a fixed area, making it difficult to directly provide forward-looking risk assessment results and action suggestions to the train's automatic driving system.

[0003] Existing technology typically employs a train obstacle detection scheme based on onboard perception. This scheme installs cameras, millimeter-wave radar, lidar, or combinations thereof at the front of the train to continuously detect the area ahead of the train in the direction of travel, collecting image information, distance information, contour information, and target motion information. Subsequently, through processing methods such as target detection, target segmentation, distance calculation, and track area matching, it identifies whether there are personnel, vehicles, foreign objects, or other obstacles that may affect train operation in the area ahead. Then, combining the spatial relationship between the target and the track centerline, track boundaries, or the train's expected path, it determines whether the target has entered the train's operating space or whether it affects train passage. When an obstacle risk is determined, it outputs deceleration, alarm, or stop commands to the train control system or driver assistance system. The advantage of this existing technology is that the processing link is relatively direct, and the perception results can be directly used for onboard control, making it suitable for real-time detection of near-field obstacles ahead of the train under ordinary track conditions. However, this solution mainly relies on the perspective of the train itself for perception, and its perception range is usually mainly the forward field of view. In port-specific railway scenarios such as curves, switches, yard obstructions, loading and unloading machinery obstructions, and sudden intrusion of lateral targets, there are problems such as insufficient detection of risks far ahead, large lateral blind spots, and delayed target recognition after obstruction.

[0004] Existing technology two typically employs a zone intrusion detection scheme based on trackside video surveillance or trackside multi-sensor fusion. This scheme deploys trackside sensing devices such as cameras, infrared thermal imaging devices, millimeter-wave radar, and lidar along the track, in switch areas, loading / unloading intersections, or on both sides of key tracks to continuously monitor the track area and adjacent areas. Subsequently, the collected images, distance, contours, thermal features, or motion information are transmitted to a backend processing unit to identify and track targets such as personnel, vehicles, loading / unloading machinery, lifting equipment, and foreign objects within the monitored area. Combined with pre-set electronic fences, clearance boundaries, or hazardous areas, the system determines whether a target has entered a preset area. When a target is detected crossing boundaries, encroaching on clearances, or occupying track area, the system outputs an alarm message and sends the result to the monitoring platform, dispatching platform, or train operation-related systems. Although this technical solution monitors the track and adjacent areas by deploying multiple sensing devices along the track and uses fixed electronic fences or preset clearances for intrusion detection, its core remains a static processing mode of "target identification - zone encroachment - alarm output." This scheme primarily triggers alarms only after an obstacle actually enters the monitoring area, lacking the ability to proactively analyze the evolution trend of potential obstacles, the dynamic threat of approaching the track, and the probability of short-term intrusion. On the other hand, trackside monitoring typically relies on fixed deployment locations and preset areas, making it difficult to dynamically adjust safety boundaries according to changes in train speed, driving task mode, switch status, and port operations. Therefore, it lacks adaptability in areas with complex obstructions, curves, lateral intrusions, and areas with dense loading and unloading operations. Furthermore, existing technologies primarily output monitoring alarms or target location information, lacking risk levels and action suggestions that can be directly applied to the train's automatic driving system, resulting in limited executability of the perception results and limited safety compensation capabilities.

[0005] Therefore, how to solve the problems of limited forward visibility, lateral intrusion blind spots, and delayed obstacle recognition after obstruction in the onboard perception of dedicated railway automatic trains in complex operating environments has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] This invention provides a method for enhancing trackside perception on dedicated port railways, aiming to at least solve one of the problems existing in related technologies, such as limited forward visibility, lateral intrusion blind spots, and delayed obstacle recognition after obstruction, in the case of onboard perception of automated driving trains on dedicated port railways in complex operating environments.

[0007] As a first aspect of the present invention, a method for enhancing trackside perception of a dedicated port railway is provided, comprising:

[0008] Multi-source sensing data is constructed based on train operation targets, loading and unloading machinery targets, personnel targets, trackside facility targets, and environmental disturbance information in the port railway operation area, and standardized multi-source sensing data results are obtained.

[0009] Enhanced identification of key port and railway targets in the standardized multi-source sensing data results is performed, and multi-dimensional features are constructed from the enhanced identified key targets to obtain an enhanced target set;

[0010] Cross-source collaborative association and scene relationship enhancement modeling are performed on the enhanced target set to obtain a port-railway relationship enhancement map;

[0011] Based on the enhanced port-railway relationship map, port-railway scenario risk identification and scenario-based risk attribution are performed to obtain a set of port-railway risk events. The set of port-railway risk events includes at least event category, event level, affected area, key relationship edge, and triggering condition.

[0012] Risk classification, early warning, and handling are carried out based on the aforementioned set of port and railway risk events.

[0013] Furthermore, based on train operation targets, loading and unloading machinery targets, personnel targets, trackside facility targets, and environmental disturbance information in the port railway operation area, multi-source sensing data is constructed to obtain standardized multi-source sensing data results, including:

[0014] A set of scenario areas is constructed based on the operational elements in the port railway operation area, wherein the operational elements in the port railway operation area include at least basic map data, track distribution data, operation boundary data, and key monitoring location data;

[0015] Acquire multi-source sensing data from multi-source sensing devices within the port railway operation area;

[0016] Time alignment of multi-source sensing data from multi-source sensing devices;

[0017] Spatial mapping and unified coordinate transformation are performed on the multi-source sensing data from multi-source sensing devices;

[0018] After time alignment and spatial mapping, the multi-source sensing data is subjected to sensing data quality screening and standardization processing to obtain standardized multi-source sensing data results.

[0019] Furthermore, the key targets of the port railway in the standardized multi-source sensing data results are enhanced and identified, and a multi-dimensional feature set is constructed from the enhanced key targets to obtain an enhanced target set, including:

[0020] The standardized multi-source sensing data results are grouped according to timestamp, region number, source device type, and spatial proximity to obtain candidate target clusters within the same observation period;

[0021] The video region, radar spot region, and their neighboring background corresponding to the candidate target cluster are enhanced and saliency-enhanced.

[0022] The enhanced saliency clusters are subjected to target category discrimination and fine-grained attribute recognition to obtain the joint discrimination results of target category and attributes;

[0023] Extract multidimensional representation information from the target that has completed target category discrimination, and construct a multidimensional feature vector of the target based on the multidimensional representation information of the target;

[0024] An enhanced target set is obtained by performing confidence integration based on the joint discrimination result of the target category and attributes and the multidimensional feature vector of the target.

[0025] Further, based on the joint discrimination result of the target category and attribute and the target multidimensional feature vector, confidence integration is performed to obtain an enhanced target set, including:

[0026] Based on the target category discrimination result, attribute recognition result, and the target multidimensional feature vector for each target, confidence integration is performed to obtain the comprehensive confidence level of target enhancement recognition;

[0027] The overall confidence level of each target is compared with a preset enhanced output threshold;

[0028] If the overall confidence level of the target is greater than the preset enhanced output threshold, then the target is written into the enhanced target set;

[0029] If the overall confidence level of the target is not greater than the preset enhanced output threshold, then the target is marked as a target to be reviewed;

[0030] After comparing the overall confidence of all targets with the preset enhanced output threshold, an enhanced target set is obtained. The enhanced target set includes at least the target identifier, target category, attribute description, global position, velocity direction, target multidimensional feature vector, and overall confidence.

[0031] Furthermore, cross-source collaborative association and scene relationship enhancement modeling are performed on the enhanced target set to obtain a port-railway relationship enhancement graph, including:

[0032] Perform cross-source consistency matching and identity verification of the enhancement targets based on each target object in the enhancement target set;

[0033] For enhanced targets that have completed identity verification, perform target temporal trajectory extension and continuous state recovery;

[0034] Identify the interaction relationships between augmented targets that have formed continuous trajectories;

[0035] Establish coupling relationships between targets and regions, targets and facilities, and targets and operational status to form coupling relationship models;

[0036] An enhanced map of port-railway relationships is constructed based on the continuous trajectory of the enhanced targets, the interaction relationships between targets, and the coupling relationship model.

[0037] Furthermore, based on the enhanced port-railway relationship map, port-railway scenario risk identification and scenario-based risk attribution are performed to obtain a set of port-railway risk events, including:

[0038] The node attributes, relationship attributes, and scene coupling attributes in the port-railway relationship enhancement graph are traversed and analyzed to extract abnormal signs and form a set of risk candidate triggers. The abnormal signs include any one or more of the following: abnormal single node attributes, abnormal enhancement of relationship edges between nodes, and constraint mismatch between nodes, regions, and facilities.

[0039] Based on the enhanced port-railway relationship map, abnormal signs in the risk candidate trigger set are traced and a multi-entity risk relationship chain is formed;

[0040] Risk category identification and scenario-based risk attribution are performed based on the multi-entity risk relationship chain.

[0041] A comprehensive risk score is calculated for the multi-subject risk relationship chain that has completed risk category identification and scenario-based risk attribution, and the event level is classified according to the comprehensive risk score;

[0042] The risk events that have been classified into risk levels are formed into a risk event output set, and warning trigger conditions are generated for each risk event in the risk event output set.

[0043] Furthermore, a comprehensive risk score is calculated for the multi-entity risk relationship chain that has completed risk category identification and scenario-based risk attribution, and the event level is classified according to the comprehensive risk score, including:

[0044] A comprehensive risk score is calculated for the multi-agent risk relationship chain based on the abnormal intensity, propagation intensity, duration, scope of impact, importance of affected objects, severity of conflict, and current operational status of the scenario.

[0045] The comprehensive risk score is classified into event levels according to a preset event level range, wherein the event levels include general warning events, key warning events, serious warning events, and emergency response events.

[0046] Furthermore, risk classification, early warning, and handling are carried out based on the aforementioned set of port railway risk events, including:

[0047] The risk events in the port railway risk events are traversed, and the corresponding early warning release methods are matched according to the preset early warning rule base. The early warning release methods include at least interface pop-up alarms, sound and light alarms, push notifications from job terminals, broadcast prompts in the work area, alarms displayed on the duty center screen, and message distribution from mobile inspection terminals.

[0048] Based on the event level, risk chain structure, and scope of impact of the scenario, corresponding linkage control strategies and handling instructions are generated for risk events that match the early warning release method. The linkage control strategies include at least automatic video focusing, highlighting of electronic fences in risk areas, output of broadcast warning information, location and dispatch of patrol terminals, confirmation requests from on-duty personnel, and status linkage with external business or control systems.

[0049] Real-time tracking of risk event status during the early warning execution process, and formation of early warning execution effectiveness evaluation results based on real-time tracking results;

[0050] The risk warning strategy is adaptively adjusted based on the evaluation results of the early warning implementation effect.

[0051] Furthermore, the risk warning strategy is adaptively adjusted based on the evaluation results of the warning implementation effectiveness, including:

[0052] Statistical analysis was performed based on the evaluation results of the early warning execution effect, the results of manual confirmation, and the final initial results of the event.

[0053] Based on the statistical analysis results, the matching relationship between event types and early warning actions, linkage priority parameters, level classification thresholds, and handling recommendation rules are revised in reverse.

[0054] Furthermore, the risk classification, early warning, and handling based on the aforementioned port railway risk event set also includes:

[0055] Risk events that have completed the early warning issuance, joint response, effect collection, and feedback correction should be archived in a closed loop to form a closed-loop archive that includes the cause of the event, the relationship chain structure, the early warning action, the joint response process, the response result, the time information, the responsible position, the manual confirmation opinion, and the strategy correction record.

[0056] The method for enhancing trackside perception of dedicated port railways provided by this invention obtains standardized multi-source perception data results by constructing multi-source perception data, and constructs an enhancement target set based on the standardized multi-source perception data results. Cross-source collaborative association and scenario relationship enhancement modeling are performed on the enhancement target set to obtain a port railway relationship enhancement map. Based on the port railway relationship enhancement map, risk identification and scenario-based risk attribution are performed to obtain a port railway risk event set. Finally, risk classification, early warning and processing are performed based on the port railway risk event set. The trackside perception enhancement method for this port-dedicated railway completes the collaborative acquisition, unified access, temporal alignment, and spatial benchmark construction of multi-source perception data in port railway scenarios, solving the problem of inconsistent perception input layers. Based on unified input, it completes the enhanced identification and multi-dimensional feature construction of key targets, solving the problems of unstable target identification and incomplete expression in complex scenarios. It performs cross-source collaborative association and scene relationship enhancement modeling on enhanced targets, solving the problem of difficulty in characterizing the relationships between targets and between targets and regional facilities. Based on the relationship enhancement graph, it completes risk chain identification, event attribution, and level determination, solving the problems of difficulty in identifying risk evolution processes and difficulty in quantifying risk levels. It performs hierarchical early warning, coordinated handling, effect recovery, strategy correction, and closed-loop archiving for risk events, solving the problem of lack of coordinated response and continuous optimization mechanisms after risk is discovered. Therefore, the trackside perception enhancement method for this port-dedicated railway constructs a closed-loop perception, analysis, judgment, action, and verification mechanism, enabling the trackside perception results to directly provide the train automatic driving system with operable traffic status, risk level, and action suggestions. It also dynamically adjusts perception resources and analysis strategies according to actual risks, thereby effectively compensating for the deficiencies of onboard perception in complex port operation scenarios and improving the safety, robustness, and risk controllability of automatic train operation in the port-dedicated railway. Attached Figure Description

[0057] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the following detailed description to explain the invention, but do not constitute a limitation thereof.

[0058] Figure 1 A flowchart of the trackside sensing enhancement method for port-dedicated railways provided by the present invention.

[0059] Figure 2 The flowchart for constructing multi-source sensing data provided by the present invention.

[0060] Figure 3 The flowchart for constructing and obtaining the enhanced target set provided by the present invention.

[0061] Figure 4 The flowchart for obtaining the enhanced port-railway relationship map provided by the present invention.

[0062] Figure 5A flowchart for obtaining a set of port railway risk events provided by the present invention.

[0063] Figure 6 A flowchart for risk classification, early warning, and handling provided by this invention. Detailed Implementation

[0064] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0065] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0066] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of the invention described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0067] This embodiment provides a method for enhancing trackside perception of a dedicated port railway. Figure 1 This is a flowchart of a trackside sensing enhancement method for a port-dedicated railway according to an embodiment of the present invention, as shown below. Figure 1 As shown, it includes:

[0068] S100. Construct multi-source sensing data based on train operation targets, loading and unloading machinery targets, personnel targets, trackside facility targets, and environmental disturbance information in the port railway operation area, and obtain standardized multi-source sensing data results.

[0069] Specifically, a multi-source perception input foundation layer is constructed to address the challenges posed by train operation targets, loading and unloading machinery targets, personnel targets, trackside facility targets, and environmental disturbance information within the port railway operation area. This step does not simply aggregate data from cameras, radar, positioning units, or edge acquisition terminals. Instead, it first establishes a unified scene perception coordinate benchmark based on the port railway line distribution, track operation boundaries, loading and unloading operation surfaces, intersections, and key risk locations. Then, image streams, point streams, status streams, and location information streams from different devices undergo temporal alignment, spatial mapping, and quality filtering to form a standardized perception matrix that can be used for subsequent enhanced recognition and event reasoning. This step effectively solves the problems of inconsistent time benchmarks across multiple devices, fragmented viewpoints, severe local occlusion, and difficulty in unifying and correlating different sensor data in existing technologies.

[0070] S200. Enhance the identification of key port and railway targets in the standardized multi-source sensing data results, and construct multi-dimensional features of the enhanced key targets to obtain an enhanced target set.

[0071] In this embodiment of the invention, to address the problems existing in the standardized multi-source sensing data results output in step S100, such as mixed target categories, severe local occlusion, dense and overlapping work objects, and significant changes in lighting, smoke and dust, and vehicle scale in the port railway environment, a target enhancement recognition and multi-dimensional feature construction mechanism for complex scenarios is constructed. This step does not directly perform a conventional detection or classification of the input target. Instead, based on the regional constraint information, time synchronization relationship, and global coordinate results established in step S100, candidate targets from different source sensing units are reorganized. Subsequently, by combining visual texture, spatial contour, trajectory dynamics, regional semantics, and equipment source credibility, enhanced recognition is performed on trains, carriages, loading and unloading machinery, workers, trackside obstacles, and abnormal intrusion objects, forming a unified target feature expression vector. This step effectively solves the problems of insufficient stability of single-source recognition, unclear target boundaries under complex occlusion conditions, severe confusion of similar targets, and insufficient basis for subsequent risk assessment in existing port railway sensing schemes.

[0072] S300. Perform cross-source collaborative association and scene relationship enhancement modeling on the enhanced target set to obtain a port-railway relationship enhancement map;

[0073] In this embodiment of the invention, a cross-source target collaborative association and scene relationship enhancement mechanism is constructed to address the spatial proximity, operational coupling, operational dependence, and risk transmission relationships among trains, carriages, loading and unloading machinery, trackside facilities, workers, and abnormal encroachments in port railway scenarios. This step is not merely a simple matching of the enhanced target set output by the S200; instead, it first confirms consistency and cross-frame association of targets from different sensing sources, observation times, and perspectives based on their global coordinates, temporal continuity, category attributes, regional semantics, and motion trends. Subsequently, it further identifies the relative distance relationships, following relationships, intersection relationships, encroachment relationships, operational coupling relationships, and facility dependence relationships between targets, constructing a relationship enhancement graph for complex port railway operation scenarios. This step effectively solves the problems in existing technologies where it is difficult to understand the interaction relationships between a single target and its surrounding objects after identification, difficult to stably track cross-device targets, and difficult to pre-judge abnormal events based on relationship chains.

[0074] S400. Based on the enhanced port-railway relationship graph, identify port-railway scenario risks and attribute them to specific scenarios to obtain a set of port-railway risk events. The set of port-railway risk events includes at least event category, event level, affected area, key relationship edge, and triggering condition.

[0075] In this embodiment of the invention, a risk chain identification and event level determination mechanism oriented towards scenario relationship graphs is constructed to address the complex risk relationships that may arise between train operation, loading and unloading machinery operations, personnel activities, trackside obstacle intrusion, and equipment status fluctuations in port railway operation areas. This step does not isolate the abnormal state of a single target; instead, it uses the enhanced port railway relationship graph output in step S300 as input to comprehensively analyze the coupling strength, propagation direction, and duration between entity nodes, relationship edges, regional constraints, and facility status. It identifies the associated chains evolving from local anomalies to scenario-level risks and provides risk scores, risk categories, and event levels. This step effectively solves the problems of existing technologies that can only detect single-point anomalies, cannot identify multi-entity collaborative risks, struggle to characterize risk evolution processes, and lack sufficient basis for early warning level classification.

[0076] S500. Based on the aforementioned set of port and railway risk events, conduct risk classification, early warning, and handling.

[0077] In this embodiment of the invention, a closed-loop feedback mechanism for coordinated early warning execution and handling is constructed based on the event categories, event levels, affected areas, key relationship edges, and triggering conditions included in the port railway risk event set output in step S400. This step does not simply provide alarm prompts for risk events, but rather generates differentiated early warning actions, coordinated control actions, and handling suggestions for different types of risk events based on the event level and the scope of impact of the scenario, combined with the on-site operation rules, equipment linkage strategies, job response mechanisms, and historical handling experience of the port railway. Simultaneously, after the early warning is executed, changes in event status, response effects, and handling completion status are continuously collected to provide feedback and correction to the early warning strategy, and the complete process results are archived in the event knowledge base. This step effectively solves the problem in existing technologies of "only alarming after anomaly detection, without tracking, correction, or accumulation," thereby improving the timeliness of risk response, the consistency of execution, and the continuous optimization capability of system operation in port railway scenarios.

[0078] In summary, the trackside perception enhancement method for port-dedicated railways provided by this invention obtains standardized multi-source perception data results by constructing multi-source perception data, and constructs an enhancement target set based on the standardized multi-source perception data results. Cross-source collaborative association and scenario relationship enhancement modeling are performed on the enhancement target set to obtain a port railway relationship enhancement map. Based on the port railway relationship enhancement map, risk identification and scenario-based risk attribution are performed to obtain a port railway risk event set. Finally, risk classification, early warning and processing are performed based on the port railway risk event set. The trackside perception enhancement method for this port-dedicated railway completes the collaborative acquisition, unified access, temporal alignment, and spatial benchmark construction of multi-source perception data in port railway scenarios, solving the problem of inconsistent perception input layers. Based on unified input, it completes the enhanced identification and multi-dimensional feature construction of key targets, solving the problems of unstable target identification and incomplete expression in complex scenarios. It performs cross-source collaborative association and scene relationship enhancement modeling on enhanced targets, solving the problem of difficulty in characterizing the relationships between targets and between targets and regional facilities. Based on the relationship enhancement graph, it completes risk chain identification, event attribution, and level determination, solving the problems of difficulty in identifying risk evolution processes and difficulty in quantifying risk levels. It performs hierarchical early warning, coordinated handling, effect recovery, strategy correction, and closed-loop archiving for risk events, solving the problem of lack of coordinated response and continuous optimization mechanisms after risk is discovered. Therefore, the trackside perception enhancement method for this port-dedicated railway constructs a closed-loop perception, analysis, judgment, action, and verification mechanism, enabling the trackside perception results to directly provide the train automatic driving system with operable traffic status, risk level, and action suggestions. It also dynamically adjusts perception resources and analysis strategies according to actual risks, thereby effectively compensating for the deficiencies of onboard perception in complex port operation scenarios and improving the safety, robustness, and risk controllability of automatic train operation in the port-dedicated railway.

[0079] It should be noted that the train operation target, loading and unloading machinery target, operator target, and trackside facility target in step S100 are identified and extracted in step S200, then the trajectory and interrelationship are established in step S300, the intrusion and conflict risk are judged in step S400, and the warning and handling instructions are generated in step S500. The environmental disturbance information in step S100 is mainly used to correct the identification placement confidence, association reliability, and risk level.

[0080] In this embodiment of the invention, multi-source sensing data is constructed based on train operation targets, loading and unloading machinery targets, personnel targets, trackside facility targets, and environmental disturbance information in the port railway operation area, to obtain standardized multi-source sensing data results, such as... Figure 2 As shown, it includes:

[0081] S110. Construct a scene area set based on the operation elements in the port railway operation area, wherein the operation elements in the port railway operation area include at least basic map data, track distribution data, operation boundary data, and key monitoring location data;

[0082] Specifically, basic map data, track distribution data, operational boundary data, and key monitoring location data of the target port area's railway operation area are acquired. The train passage area, loading and unloading operation area, personnel activity area, equipment parking area, and restricted area are then divided into layers, forming a set of scene areas A = {A1, A2, ..., An}. Each area Ai corresponds to a unique area number, boundary coordinates, risk level, and permitted behavior type, used to limit the legal activity range of different sensing targets. The following area constraint expression can be used to establish the area baseline values:

[0083] Ri(x,y) = wi1Bi(x,y) + wi2Ti(x,y) + wi3Oi(x,y) + wi4Si(x,y),

[0084] Where: Ri(x,y) represents the regional constraint response value of coordinate point (x,y) in the i-th region; Bi(x,y) represents the boundary response value of whether the point is within the regional boundary buffer zone; Ti(x,y) represents the constraint value of the track passage attribute corresponding to the point; Oi(x,y) represents the constraint value of the loading and unloading operation attribute corresponding to the point; Si(x,y) represents the constraint value of the safety risk attribute corresponding to the point; wi1, wi2, wi3, and wi4 represent the weight coefficients of each attribute component.

[0085] It should be noted that when Ri(x,y) is greater than the preset threshold, the point is determined to be a valid monitoring point in the key constraint area.

[0086] It should be understood that the set of scene regions constructed in the embodiments of the present invention is the basis for the unified spatial constraints of subsequent steps.

[0087] S120. Obtain multi-source sensing data from multi-source sensing devices within the port railway operation area;

[0088] Specifically, video acquisition equipment, radar sensing equipment, trackside status acquisition units, edge processing nodes, and positioning and timing units are deployed within the port railway operation area. Each device is assigned a unique device identifier, and a mapping table between the device and the area is established. Among them, the video acquisition equipment is used to acquire visible light image sequences, the radar sensing equipment is used to acquire moving target point information, the trackside status acquisition unit is used to acquire level crossing status, signal status, track occupancy status, or equipment start / stop status, and the positioning and timing unit is used to provide unified time identifiers and location reference information to each sensing node.

[0089] To ensure consistency in subsequent processing, the original output of any device ek is denoted as:

[0090] Dk(t) = {Vk(t), Pk(t), Ck(t), Mk},

[0091] Where Dk(t) represents the raw data packet of the k-th sensing device at time t; Vk(t) represents the sensing master data output by the device; when the device is a video device, Vk(t) is an image frame; when the device is a radar device, Vk(t) is a set of target points; Pk(t) represents the location or installation attitude parameters of the device; Ck(t) represents the device status parameters; Mk represents the static metadata of the device, including device type, installation area and acquisition direction.

[0092] S130. Time alignment of the multi-source sensing data from the multi-source sensing devices;

[0093] Specifically, timestamps are extracted from data packets from various sensing devices, and a global time axis is constructed based on a unified timing signal. Data arriving at different times are merged into time windows to form joint sensing slices within the same observation period. For device data that is not strictly synchronized, compensation, interpolation, or discarding is performed according to the magnitude of the time deviation to reduce correlation errors caused by acquisition delays and transmission jitter.

[0094] The time-aligned joint sensing slice can be represented as:

[0095] Gt(m) = Union(D1(t1), D2(t2), ..., Dq(tq)), satisfying |tk - tm| <= delta_t.

[0096] Where: Gt(m) represents the joint sensing slice corresponding to the m-th global observation time; Union(.) represents the aggregation operation of multi-source data within the same time window; D1 to Dq represent the data packets of the q sensing devices participating in the fusion; t1 to tq represent the original sampling time of each device; tm represents the global observation reference time; delta_t represents the allowed time registration window threshold.

[0097] S140. Perform spatial mapping and unified coordinate transformation on the multi-source sensing data of the multi-source sensing devices;

[0098] Specifically, based on the installation location, orientation parameters, internal and external parameter calibration results, and regional coordinate reference of each device, the data from the video plane coordinate system, radar local coordinate system, and trackside device coordinate system are uniformly mapped to the global coordinate system of the port railway scene to form the position expression result of the target under the same spatial reference; for video perception targets, the center of their target box, the center of their bottom edge, or the projection position of key points are extracted; for radar point targets, their polar coordinates or Cartesian coordinates are extracted; and then uniformly mapped to a global position vector.

[0099] In this embodiment of the invention, the spatial mapping result can be written as:

[0100] Qg = Tk * Ql + bk,

[0101] Where Qg represents the global coordinate vector after target mapping; Ql represents the original position vector of the target in the device's local coordinate system; Tk represents the spatial transformation matrix corresponding to the k-th device; and bk represents the translation compensation vector corresponding to the k-th device. Through the above transformation, the same target observed by different devices can be compared and correlated in a unified scene coordinate system.

[0102] S150. Perform sensor data quality screening and standardization on the multi-source sensing data after time alignment and spatial mapping to obtain standardized multi-source sensing data results.

[0103] Specifically, a quality assessment is performed on the joint sensing data after time alignment and spatial mapping, removing data containing severe blur, overexposure, frame loss, abnormal drift, invalid points, duplicate reports, or data exceeding the area boundary. A standardized output result is generated based on target category, location, velocity, orientation, source device, and confidence level. This standardized output result serves as input to subsequent step S200 for target enhancement recognition, trackside behavior understanding, or cross-device collaborative correlation analysis.

[0104] In this embodiment of the invention, the standardized sensing unit can be represented as:

[0105] Us = {id, cls, pos, vel, dir, src, conf, reg, ts},

[0106] Wherein, id represents the unique identifier of the target; cls represents the target category; pos represents the global location of the target; vel represents the target velocity information; dir represents the target direction information; src represents the set of devices from which the target originates; conf represents the target credibility; reg represents the region number to which the target belongs; and ts represents the standard timestamp corresponding to the target.

[0107] It should be understood that step S110 is mainly responsible for delineating the area and setting rules, while subsequent steps S120 to S150 are for accessing, aligning, mapping, and filtering data according to the delineated area benchmark.

[0108] In this embodiment of the invention, key targets of ports and railways in the standardized multi-source sensing data results are enhanced and identified, and multi-dimensional features are constructed from the enhanced key targets to obtain an enhanced target set, such as... Figure 3 As shown, it includes:

[0109] S210. The standardized multi-source sensing data results are grouped according to timestamp, region number, source device type and spatial proximity to obtain candidate target clusters within the same observation period;

[0110] Specifically, the standardized sensing unit set Us output in step S100 is grouped according to timestamp, region number, source device type and spatial proximity to form candidate target clusters within the same observation period; for sensing units from different devices but located in close spatial positions and meeting the synchronization requirements in time, a candidate merging relationship is established to reduce the repeated representation of the same target under multiple source inputs.

[0111] The construction result of the candidate target cluster can be represented as:

[0112] Hj(t) = Sum(r=1 to nr) ar * Phi(Usr, t, reg, pos),

[0113] Where Hj(t) represents the response value of the j-th candidate target cluster constructed at time t; Usr represents the r-th standardized sensing unit participating in the construction of the j-th candidate target cluster; Phi(.) represents the joint merge function for temporal consistency, regional consistency, and spatial proximity; ar represents the contribution weight of the r-th sensing unit in the construction of the candidate cluster; nr represents the number of sensing units contained in the candidate target cluster; reg represents the region number to which the target belongs; and pos represents the global position parameter of the target.

[0114] It should be noted that when Hj(t) is greater than the candidate clustering threshold, the corresponding sensing unit combination is determined to belong to the same candidate target.

[0115] S220. Enhance and saliency the video target region and its neighborhood background, radar target traces and their neighboring traces corresponding to the candidate target cluster;

[0116] Specifically, the video target region and its neighborhood background, radar target traces and their neighboring traces are subjected to enhancement and saliency processing to highlight the train edges, personnel outlines, equipment components, intrusion obstacle outlines and motion direction information, and suppress recognition interference caused by backlight, rain and fog, dust, uneven night lighting or reflection on the vehicle surface; the enhancement and saliency processing includes at least local structure response extraction, background suppression, adaptive contrast enhancement and boundary continuity compensation.

[0117] In this embodiment of the invention, the enhanced target response map can be represented as:

[0118] Ej(u,v) = b1Gj(u,v) + b2Lj(u,v) + b3Mj(u,v) - b4Nj(u,v),

[0119] Where: Ej(u,v) represents the enhanced response value of the candidate target at image coordinates (u,v); Gj(u,v) represents the local gradient structure response value; Lj(u,v) represents the local brightness contrast enhancement value; Mj(u,v) represents the motion consistency enhancement value; Nj(u,v) represents the background noise suppression term; b1, b2, b3, and b4 represent the weight coefficients of each enhancement component. It should be understood that the above processing improves the discriminability of key target areas in complex port environment.

[0120] S230. Perform target category discrimination and fine-grained attribute recognition on the candidate target cluster after the enhanced saliency processing to obtain the joint discrimination result of target category and attribute;

[0121] Specifically, the candidate target clusters after enhanced saliency processing are classified to identify whether the target belongs to the train body, carriage unit, hoisting machinery, traction equipment, workers, trackside facilities, yard obstacles or abnormal encroachments. On this basis, fine-grained attribute information is further extracted, including the target length scale, shape type, operating status, orientation, track occupation attribute, whether it is in operation, and its relative relationship with the track boundary.

[0122] The joint discrimination result of target category and attribute can be written as:

[0123] Cj = argmaxc [ qc1Fc_tex(c) + qc2Fc_geo(c) + qc3Fc_dyn(c) + qc4Fc_reg(c) + qc5*Fc_src(c) ],

[0124] Where: Cj represents the final category discrimination result of the j-th candidate target; argmaxc represents selecting the category with the largest comprehensive score among all candidate categories c as the output category; Fc_tex(c) represents the texture feature matching score of the target in category c; Fc_geo(c) represents the geometric contour matching score of the target in category c; Fc_dyn(c) represents the dynamic behavior matching score of the target in category c; Fc_reg(c) represents the matching score of the target in category c for consistency with the region semantic rules; Fc_src(c) represents the matching score of the target in category c corresponding to the source device credibility; qc1, qc2, qc3, qc4, and qc5 represent the fusion weights of each discrimination component. Through this joint discrimination, the coarse classification that only relies on image appearance can be further extended into a robust classification that fuses trajectory, region semantics, and source reliability.

[0125] S240. Extract multidimensional representation information from the target that has completed the target category discrimination, and construct a multidimensional feature vector of the target based on the multidimensional representation information of the target;

[0126] Specifically, for targets that have completed category discrimination, their multidimensional representation information is extracted and organized. The multidimensional representation information includes at least appearance texture features, geometric morphology features, spatial location features, velocity direction features, regional semantic features, source credibility features, and temporal change features. These are then uniformly constructed into a target feature vector that can be used for subsequent association analysis. Different types of targets can have different feature weight configurations. For example, for train targets, the length continuity and track alignment consistency are emphasized; for personnel targets, the contour changes and movement patterns are emphasized; and for intruding obstacles, the static occupancy and boundary intrusion depth are emphasized.

[0127] In this embodiment of the invention, the target multidimensional feature vector can be defined as:

[0128] Zj = [ wt1Tj , wt2Gj , wt3Pj , wt4Vj , wt5Rj , wt6Sj , wt7*Yj ],

[0129] Where Zj represents the unified multidimensional feature vector of the j-th target; Tj represents the texture appearance feature component of the target; Gj represents the geometric shape feature component of the target; Pj represents the spatial position feature component of the target; Vj represents the velocity and direction feature component of the target; Rj represents the regional semantic feature component of the target; Sj represents the source device credibility feature component of the target; Yj represents the temporal variation feature component of the target; and wt1 to wt7 represent the weight parameters corresponding to each feature component. This unified feature vector representation provides a unified input basis for subsequent cross-source matching and event inference.

[0130] S250. Based on the joint discrimination result of the target category and attribute and the target multidimensional feature vector, confidence integration is performed to obtain an enhanced target set.

[0131] In this embodiment of the invention, for each candidate target that has completed category discrimination and multidimensional feature construction, its category discrimination result, fine-grained attributes, target multidimensional feature vector Zj, global position, velocity direction, and source device information are obtained. The overall recognition confidence of the target is calculated by comprehensively considering the target's category credibility, regional rule consistency, temporal continuity, spatial rationality, multi-source observation consistency, and multidimensional feature stability. Specifically, multidimensional feature stability is used to determine whether the texture, geometry, position, velocity, regional semantics, source credibility, and temporal variation components in the target's multidimensional feature vector Zj are complete, and whether the feature components remain continuous at adjacent observation times. This avoids directly outputting targets with missing or abnormally changing features to subsequent steps.

[0132] Specifically, based on the joint discrimination result of the target category and attribute and the target multidimensional feature vector, confidence integration is performed to obtain an enhanced target set, including:

[0133] (1) Based on the target category discrimination result, attribute recognition result and the target multidimensional feature vector of each target, confidence integration is performed to obtain the comprehensive confidence of target enhancement recognition;

[0134] (2) Compare the overall confidence level of each target with the preset enhanced output threshold;

[0135] (3) If the overall confidence level of the target is greater than the preset enhanced output threshold, then the target is written into the enhanced target set;

[0136] (4) If the overall confidence level of the target is not greater than the preset enhanced output threshold, then the target is marked as a target to be reviewed;

[0137] (5) After comparing the overall confidence of all targets with the preset enhanced output threshold, an enhanced target set is obtained. The enhanced target set includes at least the target identifier, target category, attribute description, global position, velocity direction, target multidimensional feature vector and overall confidence.

[0138] In this embodiment of the invention, the overall confidence level of target enhancement recognition can be expressed as:

[0139] Qj = d1*Sc_j + d2*Sr_j + d3*St_j + d4*Sp_j + d5*Sm_j + d6*Sz_j,

[0140] Where Qj represents the comprehensive recognition confidence of the j-th target; Sc_j represents the target category discrimination confidence score; Sr_j represents the consistency score between the target and the rules of its region; St_j represents the temporal continuity score of the target under continuous observation time; Sp_j represents the spatial rationality score of the target's position, scale, and motion state; Sm_j represents the multi-source consistency score of the observation results of the target by different sensing devices; d1, d2, d3, d4, d5, and d6 represent the weight coefficients corresponding to each confidence component. When Qj is greater than the preset enhancement output threshold, the target is added to the enhanced target set; when Qj is not greater than the preset enhancement output threshold, it is marked as a target to be verified, and its multi-dimensional feature vector Zj is retained for subsequent multi-source observation results to supplement and verify.

[0141] Specifically, the final set of enhanced targets can be represented as:

[0142] Eout = { oe1, oe2, ..., oem},

[0143] Where oej represents the j-th enhanced recognition target, which includes at least the target identifier, target category, attribute description, global position, velocity direction, multi-dimensional feature vector, and comprehensive recognition confidence.

[0144] More specifically, each enhanced recognition target can be represented as:

[0145] oej ={idj, clsj, attrj, posj, velj, dirj, srcj, Zj, Qj, regj, tsj},

[0146] Wherein, idj represents the target identifier; clsj represents the target category; attrj represents the target fine-grained attribute; posj represents the target global position; velj represents the target velocity; dirj represents the target direction of motion; srcj represents the set of target source devices; Zj represents the target multidimensional feature vector constructed in step S240; Qj represents the comprehensive identification confidence of the j-th target; regj represents the region number to which the target belongs; and tsj represents the standard timestamp corresponding to the target.

[0147] It should be noted that in this embodiment of the invention, the enhanced target set Eout is used as the input to step S300. Step S310 calculates the feature vector similarity Simz(a,b) based on the multidimensional feature vectors Za and Zb carried by different targets, and combines this with position, velocity, category, and time information to determine whether targets from different devices or different observation times belong to the same real entity.

[0148] In this embodiment of the invention, cross-source collaborative association and scene relationship enhancement modeling are performed on the enhanced target set to obtain a port-railway relationship enhancement map, such as... Figure 4 As shown, it includes:

[0149] S310. Perform cross-source consistency matching and identity verification of the enhanced targets based on each target object in the enhanced target set;

[0150] Specifically, for each target object in the enhanced target set Eout output in step S200, a candidate matching set between different sensing sources is established based on global location proximity, category compatibility, velocity direction consistency, feature vector similarity, and observation time continuity. For target pairs that satisfy the consistency constraint, they are determined to be repeated observations of the same real entity under different devices or different time slices, and are assigned a unified entity number. The cross-source same entity confirmation score can be expressed as:

[0151] Mj(a,b) = p1Simz(a,b) + p2Simp(a,b) + p3Simv(a,b) + p4Simc(a,b) + p5*Simt(a,b),

[0152] Where Mj(a,b) represents the same-entity matching score between target a and target b; Simz(a,b) and Simp(a,b) represent the global positional proximity between target a and target b; Simv(a,b) represents the velocity direction consistency between target a and target b; Simc(a,b) represents the class compatibility between target a and target b; Simt(a,b) represents the temporal continuity consistency between target a and target b; and p1, p2, p3, p4, and p5 represent the weight coefficients of each matching component. When Mj(a,b) is greater than the same-entity matching threshold, target a and target b are merged into the same entity object.

[0153] S320. Extend the target's temporal trajectory and restore its continuous state for the enhanced target that has completed the same body confirmation.

[0154] Specifically, after the entity object has been confirmed as a single entity, its trajectory is concatenated across time periods. Based on its position changes, velocity changes, direction changes, and region migration results during continuous observation periods, an entity-level temporal trajectory is generated. For target interruption segments caused by occlusion, illumination changes, equipment perspective switching, or short-term loss of detection, continuous state recovery is performed based on trajectory extension rules and adjacent state prediction results to improve the integrity of the target trajectory. The trajectory extension state of target entity e can be written as:

[0155] Te(t) = q1Pe(t) + q2Ve(t-1) + q3Ae(t-1) + q4Ke(t),

[0156] Where Te(t) represents the trajectory extension state vector of entity e at time t; Pe(t) represents the observation position vector of entity e at time t; Ve(t-1) represents the velocity vector of entity e at the previous time; Ae(t-1) represents the change in direction or acceleration trend of entity e at the previous time; Ke(t) represents the trajectory correction term of entity e under the semantic constraints of the current region; q1, q2, q3, and q4 represent the fusion coefficients of each state component. Through this calculation, the target maintains a relatively stable continuous trajectory expression even under short-term interruptions or viewpoint switching.

[0157] S330. Identify the interaction relationships between enhanced targets that have formed continuous trajectories;

[0158] Specifically, for entities that have formed continuous trajectories, based on target category, relative position, direction of movement, relative distance, region affiliation, and behavioral attributes, it determines whether there are interactive relationships between different entities, such as parallel relationships, following relationships, proximity relationships, intersection relationships, yielding relationships, loading and unloading coordination relationships, personnel approaching equipment relationships, personnel intruding into the track area relationships, and obstacle encroachment relationships, and assigns a corresponding relationship label to each type of relationship. The relationship strength between entity e1 and entity e2 can be expressed as:

[0159] Re(e1,e2) = s1De(e1,e2) + s2Be(e1,e2) + s3He(e1,e2) + s4Ge(e1,e2) +s5*Ue(e1,e2),

[0160] Where Re(e1,e2) represents the strength of the comprehensive relationship between entities e1 and e2; De(e1,e2) represents the distance coupling response value between the two entities; Be(e1,e2) represents the motion direction coordination response value between the two entities; He(e1,e2) represents the historical co-occurrence response value between the two entities; Ge(e1,e2) represents the interaction constraint response value between the two entities in their respective region rules; Ue(e1,e2) represents the semantic association response value between the two entity categories; s1, s2, s3, s4, and s5 represent the weight coefficients of each relation component.

[0161] It should be noted that when Re(e1,e2) is greater than the threshold for determining the corresponding relation type, the entity pair (e1,e2) is assigned the corresponding relation label.

[0162] S340. Construct coupling relationships between targets and regions, targets and facilities, and targets and operational status to form coupling relationship models;

[0163] Specifically, based on the identification of relationships between targets, the coupling relationships between targets and track boundaries, level crossing facilities, loading and unloading equipment, work platforms, warning lines, signaling devices, clearance zones, and restricted areas are further analyzed to form a multi-layered scene relationship network of "entity-region-facility-state". For the same entity, its region, current facility proximity status, whether it is within the work chain, whether it has entered the restriction boundary, and whether it has triggered scene constraint rules are recorded simultaneously. The coupling relationship value between entity e and scene element g can be defined as:

[0164] Cg(e,g) = m1Lg(e,g) + m2Ng(e,g) + m3Wg(e,g) + m4Xg(e,g),

[0165] Where: Cg(e,g) represents the coupling relationship value between entity e and scene element g; Lg(e,g) represents the spatial adjacency response value between entity and scene element; Ng(e,g) represents the functional dependency response value between entity and scene element; Wg(e,g) represents the state association response value of entity to scene element in the current work process; Xg(e,g) represents the regional rule conflict response value between entity and scene element; m1, m2, m3, and m4 represent the weight coefficients of each coupling component. Through this relationship modeling, the target can be expanded from a simple "location object" to "work entity with scene context meaning".

[0166] S350. Construct a port-railway relationship enhancement map based on the continuous trajectory of the enhanced target, the interaction relationship between targets, and the coupling relationship model.

[0167] Specifically, the entity objects, continuous trajectories, inter-target relationships, target-region relationships, target-facility relationships, and operational status relationships obtained in steps S310 to S340 are uniformly organized to construct an enhanced relationship graph in the port railway scenario. In the enhanced relationship graph, nodes are used to represent trains, carriages, personnel, machinery, obstacles, facilities, and regional objects, while edges are used to represent relationships such as following, approaching, intrusion, loading and unloading coordination, boundary conflict, facility dependence, and risk transmission. Both nodes and edges carry timestamps, confidence levels, regional labels, and status attributes, which are used as the input basis for risk judgment and linkage early warning in step S400.

[0168] Relationship enhancement graphs can be abstractly represented as:

[0169] Gport(t) = {Vt, Et, At}

[0170] Wherein, Gport(t) represents the port-railway relationship enhancement graph at time t; Vt represents the set of nodes in the graph, with nodes corresponding to entity objects, region objects, and facility objects; Et represents the set of relationship edges in the graph, with edges corresponding to various target interaction relationships and scene coupling relationships; At represents the set of attributes of nodes and edges, including time attributes, category attributes, state attributes, risk attributes, and confidence attributes.

[0171] Furthermore, the overall credibility of any relation edge ek can be expressed as:

[0172] Fk = n1Osrc(k) + n2Otime(k) + n3Ospace(k) + n4Ologic(k),

[0173] Where Fk represents the overall credibility of relation edge ek; Osrc(k) represents the support strength of relation edge derived from the consistency of multi-source observations; Otime(k) represents the continuous stability of relation edge in the time dimension; Ospace(k) represents the positional rationality of relation edge in the spatial dimension; Ologic(k) represents the consistency between relation edge and scene rule logic; n1, n2, n3, and n4 represent the weight coefficients of each credibility component. Only relation edges with Fk greater than the relation output threshold are retained for the formal graph output.

[0174] In this embodiment of the invention, port railway scenario risk identification and scenario-based risk attribution are performed based on the enhanced port-railway relationship map to obtain a set of port-railway risk events, such as... Figure 5 As shown, it includes:

[0175] S410. Perform a traversal analysis on the node attributes, relationship attributes, and scene coupling attributes in the port railway relationship enhancement graph to extract abnormal signs and form a set of risk candidate triggers. The abnormal signs include any one or more of the following: abnormal single node attributes, abnormal enhancement of relationship edges between nodes, and constraint mismatch between nodes, regions, and facilities.

[0176] Specifically, the node attributes, relationship attributes, and scene coupling attributes in the port-railway relationship enhancement graph Gport(t) output in step S300 are traversed and analyzed to identify whether there are abnormal signs such as target category conflict, regional rule conflict, boundary intrusion, trajectory mutation, abnormal relationship enhancement, abnormal linkage of facility status, or inconsistent operation process, and the abnormal signs are mapped to a set of risk candidate triggers; wherein, the abnormal signs can be from the abnormality of single node attribute, the abnormal enhancement of the relationship edge between nodes, or the constraint mismatch between nodes, regions, and facilities.

[0177] The abnormal response value of any abnormal symptom unit zr can be expressed as:

[0178] Ar(zr) = u1Ir(zr) + u2Jr(zr) + u3Kr(zr) + u4Lr(zr) + u5*Mr(zr),

[0179] Where Ar(zr) represents the comprehensive anomaly response value of the anomaly symptom unit zr; Ir(zr) represents the entity attribute deviation response value corresponding to the anomaly symptom; Jr(zr) represents the relation strength imbalance response value corresponding to the anomaly symptom; Kr(zr) represents the regional rule conflict response value corresponding to the anomaly symptom; Lr(zr) represents the facility status coupling anomaly response value corresponding to the anomaly symptom; Mr(zr) represents the time-series persistence anomaly response value corresponding to the anomaly symptom; and u1, u2, u3, u4, and u5 represent the weighting coefficients of each anomaly component.

[0180] It should be noted that when Ar(zr) is greater than the abnormal trigger threshold, the abnormal symptom unit is written into the risk candidate trigger set Zrisk.

[0181] S420. Based on the enhanced port-railway relationship map, trace the abnormal signs in the risk candidate trigger set and form a multi-subject risk relationship chain;

[0182] Specifically, for each anomalous symptom in the Zrisk set of risk candidate triggers, its formation path is traced forward and its potential impact path is deduced backward based on the connection relationship between its corresponding entity, relation edge, regional object and facility object, thereby constructing a risk relationship chain containing "source anomaly - intermediate coupling - end effect". For multiple anomalous symptoms that overlap in time and space, cross entities or share facilities in the same period, they are merged at the chain level to identify compound risk events.

[0183] The chain propagation strength of any risk relationship chain cl can be expressed as:

[0184] Pc(cl) = v1Sc(cl) + v2Tc(cl) + v3Rc(cl) + v4Ec(cl) + v5*Dc(cl),

[0185] Wherein, Pc(cl) represents the chain propagation strength of the risk relationship chain cl; Sc(cl) represents the strength value of the initial anomaly source of the risk relationship chain; Tc(cl) represents the continuous expansion strength of the risk chain in the time dimension; Rc(cl) represents the coupling diffusion strength of the risk chain in the relationship graph; Ec(cl) represents the comprehensive strength of the number and importance of entities involved in the risk chain; Dc(cl) represents the potential impact strength of the risk chain on train operation, personnel safety, loading and unloading operations, or facility safety; v1, v2, v3, v4, and v5 represent the weight coefficients of each chain propagation component. Through this calculation, isolated anomalies can be further enhanced into risk relationship chain expressions with propagation logic.

[0186] S430. Based on the multi-subject risk relationship chain, identify risk categories and perform scenario-based risk attribution;

[0187] Specifically, based on the risk relationship chain constructed in step S420, the risk is categorized and determined to belong to personnel intrusion risk, equipment intrusion risk, train approach conflict risk, loading and unloading operation mismatch risk, trackside obstacle threat risk, facility status linkage imbalance risk, or compound operation conflict risk. After completing the category identification, the causes of the risk are further attributed to the scenario, clarifying the dominant source, key relationship edge, core area and main affected object of the risk, so as to form an interpretable risk judgment result.

[0188] The attribution score of any risk chain cl under category g can be expressed as:

[0189] Yg(cl) = r1Hg_attr(g) + r2Hg_rel(g) + r3Hg_reg(g) + r4Hg_fac(g) + r5*Hg_trd(g),

[0190] Wherein, Yg(cl) represents the attribution score of risk chain cl belonging to risk category g; Hg_attr(g) represents the matching score of risk chain to category g at the entity attribute level; Hg_rel(g) represents the matching score of risk chain to category g at the relation structure level; Hg_reg(g) represents the matching score of risk chain to category g at the regional rule level; Hg_fac(g) represents the matching score of risk chain to category g at the facility status level; Hg_trd(g) represents the matching score of risk chain to category g at the evolution trend level; r1, r2, r3, r4, and r5 represent the weight coefficients of each attribution component. The risk category with the largest Yg(cl) is selected as the final category identification result of the risk chain.

[0191] S440. Calculate the comprehensive risk score for the multi-subject risk relationship chain that has completed risk category identification and scenario-based risk attribution, and classify the event level according to the comprehensive risk score;

[0192] In this embodiment of the invention, for risk relationship chains that have completed category identification and scenario-based attribution, a comprehensive risk score is calculated by combining the anomaly intensity, propagation intensity, duration, scope of impact, importance of affected objects, severity of conflict, and current operational status of the scenario. Based on a preset level range, risk events are classified into general warning events, key warning events, severe warning events, or emergency response events. Among them, for risk relationship chains involving train operation channels, personnel life safety, critical loading and unloading facilities, or multi-area linkage anomalies, the sensitivity of their event level is increased.

[0193] Specifically, a comprehensive risk score is calculated for the multi-entity risk relationship chain that has completed risk category identification and scenario-based risk attribution, and the event level is classified according to the comprehensive risk score, including:

[0194] (1) Calculate the comprehensive risk score for the multi-subject risk relationship chain based on the abnormal intensity, propagation intensity, duration, scope of impact, importance of the affected objects, severity of conflict and current operational status of the scenario;

[0195] (2) The comprehensive risk score is divided into event levels according to the preset event level range, wherein the event levels include general warning events, key warning events, serious warning events and emergency response events.

[0196] Specifically, the comprehensive risk score of any risk event ev can be expressed as:

[0197] W(ev) = k1Ae(ev) + k2Pe(ev) + k3Te(ev) + k4Ie(ev) + k5Ce(ev) + k6Ne(ev),

[0198] Wherein, W(ev) represents the comprehensive risk score of risk event ev; Ae(ev) represents the aggregation intensity of the abnormal symptoms corresponding to the risk event; Pe(ev) represents the diffusion intensity of the risk propagation chain corresponding to the risk event; Te(ev) represents the duration or acceleration intensity of the evolution of the risk event; Ie(ev) represents the intensity of the impact range of the risk event; Ce(ev) represents the severity of the conflict between the risk event and the critical operating area or critical facility; Ne(ev) represents the importance intensity of the target involved in the risk event; k1, k2, k3, k4, k5, and k6 represent the weight coefficients of each scoring component.

[0199] Furthermore, the event level (ev) can be divided according to the range of values ​​for W (ev):

[0200] level(ev) =L1, when 0 <= W(ev) < theta1; level(ev) =L2, when theta1 <= W(ev)< theta2; level(ev) =L3, when theta2 <= W(ev) < theta3; level(ev) =L4, when W(ev) >= theta3;

[0201] Where L1 represents a general warning event; L2 represents a key warning event; L3 represents a severe warning event; L4 represents an emergency response event; theta1, theta2, and theta3 represent risk level classification thresholds, and theta1 must be satisfied. <theta2 < theta3。

[0202] S450. Form a risk event output set from the risk events that have been classified into risk levels, and generate early warning trigger conditions for each risk event in the risk event output set.

[0203] Specifically, risk events that have been classified into risk events are organized into a risk event output set, and a corresponding event identifier, event category, dominant source, risk chain structure, event level, affected area, involved entities, key relationship edges, handling priority and early warning triggering conditions are generated for each risk event. For risk events that reach the level of key early warning or above, a description of linkage early warning conditions is further generated, which is used as the direct input for alarm release, linkage control and handling closed loop in step S500.

[0204] The risk event output unit can be represented as:

[0205] Oev = { eid, etype, esrc, echain, escore, elevel, earea, eobj, erel,eprio, etrig, ets},

[0206] Wherein, eid represents the unique identifier of the risk event; etype represents the risk event category; esrc represents the dominant anomaly source of the risk event; echain represents the risk relationship chain corresponding to the risk event; escore represents the comprehensive risk score; elevel represents the event level; earea represents the risk impact area; eobj represents the set of involved entity objects; erel represents the set of key relationship edges; eprio represents the handling priority; etrig represents the description of the warning triggering conditions; and ets represents the event output timestamp.

[0207] The final set of risk events is denoted as:

[0208] Erisk(t) = { Oev1, Oev2, ..., Oevh},

[0209] Where Erisk(t) represents the set of port railway risk events output at time t; Oev1 to Oevh represent h risk event units that meet the output conditions. The risk event set Erisk(t) serves as the input basis for early warning issuance, linkage response, and result closed-loop recording in step S500.

[0210] In this embodiment of the invention, risk classification, early warning, and handling are performed based on the set of port railway risk events, such as... Figure 6 As shown, it includes:

[0211] S510. Traverse the risk events in the port railway risk events and match the corresponding early warning release method according to the preset early warning rule library. The early warning release method includes at least interface pop-up alarm, sound and light alarm, post terminal push, work area broadcast prompt, duty center alarm screen and mobile inspection terminal message distribution.

[0212] Specifically, each risk event unit in the risk event set Erisk(t) output in step S400 is traversed to read its event category, event level, affected area, involved objects, and handling priority. Based on the pre-established "event type-level-area-action" mapping rule library, the corresponding early warning release method is matched. The early warning release method includes at least interface pop-up alarm, sound and light alarm, push notification from the post terminal, broadcast prompt in the work area, alarm display in the duty center, and message distribution from the mobile inspection terminal. For general early warning events, a prompt release is performed; for key early warning events, enhanced reminders and post confirmation are performed; and for serious early warning events and emergency response events, multi-channel synchronous release and rapid response command issuance are performed.

[0213] The matching score of the warning action for any risk event ev can be expressed as:

[0214] Ma(ev,aj) = c1Fa_type(ev,aj) + c2Fa_level(ev,aj) + c3Fa_area(ev,aj) +c4Fa_obj(ev,aj) + c5*Fa_hist(ev,aj),

[0215] Wherein, Ma(ev,aj) represents the matching score between risk event ev and candidate warning action aj; Fa_type(ev,aj) represents the matching response value between event category and warning action; Fa_level(ev,aj) represents the matching response value between event level and warning action; Fa_area(ev,aj) represents the matching response value between event impact area and warning action; Fa_obj(ev,aj) represents the matching response value between event-related object and warning action; Fa_hist(ev,aj) represents the action adaptation response value obtained based on historical handling experience; c1, c2, c3, c4, and c5 represent the weight coefficients of each matching component.

[0216] Select the candidate warning action aj with the largest Ma(ev,aj) as the main warning action for the risk event and write it into the warning release task unit.

[0217] S520. Based on the event level, risk chain structure and scope of impact of the scenario, generate corresponding linkage control strategies and handling instructions for risk events that have completed the matching of the early warning release method. The linkage control strategies include at least automatic video focusing, high-brightness of electronic fences in risk areas, output of broadcast warning information, location and dispatch of patrol terminals, confirmation requests from on-duty personnel, and status linkage with external business systems or control systems.

[0218] Specifically, for risk events that have completed the matching of early warning actions, corresponding linkage control strategies and handling instructions are generated based on their event level, risk chain structure, and scope of impact. Among them, the linkage control strategies include at least automatic video focusing, highlighting of electronic fences in risk areas, output of broadcast warning information, location dispatch of inspection terminals, confirmation requests from on-duty personnel, and status linkage with external business systems or control systems. For risk events that meet the conditions for high-level linkage, suggestions for equipment deceleration, suggestions for passage restriction, suggestions for work suspension, manual verification instructions, or zoning control instructions can be further generated to improve the efficiency of on-site risk control.

[0219] The priority value for coordinated handling of any risk event (ev) can be expressed as:

[0220] Ua(ev) = d1Ga(ev) + d2Ha(ev) + d3Ia(ev) + d4Ja(ev) + d5*Ka(ev),

[0221] Wherein, Ua(ev) represents the priority value of the coordinated handling of risk event ev; Ga(ev) represents the contribution value of event level to coordination priority; Ha(ev) represents the contribution value of risk chain spread trend to coordination priority; Ia(ev) represents the contribution value of the criticality of the affected area to coordination priority; Ja(ev) represents the contribution value of the importance of the involved object to coordination priority; Ka(ev) represents the correction value of the current scenario running status to coordination priority; d1, d2, d3, d4, and d5 represent the weight coefficients of each priority component.

[0222] The execution order of the coordinated handling tasks is determined based on the size of Ua(ev), and a corresponding set of handling instructions is generated.

[0223] S530. Track the status of risk events during the early warning execution process in real time, and generate an evaluation result of the early warning execution effect based on the real-time tracking results;

[0224] Specifically, after the early warning action and the coordinated response action are executed, the status of the scene, changes in target relationships, personnel confirmation results, response progress, and changes in the scope of the event's impact are continuously collected to track and determine whether the risk event has been mitigated, escalated, transferred, or resolved. The status tracking can be based on the continuous operation results from steps S100 to S400 to recalculate the abnormal signs, relationship edge strength, risk chain propagation strength, and risk score changes corresponding to the event, so as to form an evaluation result of the early warning execution effect.

[0225] The evaluation value of the handling effect of any risk event ev at the time of recovery tr can be expressed as:

[0226] Eb(ev,tr) = e1Rb(ev,tr) + e2Sb(ev,tr) + e3Tb(ev,tr) + e4Vb(ev,tr) -e5*Wb(ev,tr),

[0227] Wherein, Eb(ev,tr) represents the evaluation value of the handling effect of risk event ev at the time of recovery tr; Rb(ev,tr) represents the response value of the degree of risk chain weakening; Sb(ev,tr) represents the response value of the degree of abnormal symptom reduction; Tb(ev,tr) represents the response value of the degree of on-site condition recovery; Vb(ev,tr) represents the response value of the confirmation of the effectiveness of manual handling; Wb(ev,tr) represents the response value of event re-propagation or re-triggering; e1, e2, e3, e4, and e5 represent the weight coefficients of each evaluation component. When Eb(ev,tr) is greater than the effective threshold, the risk event is determined to have entered a controlled or resolved state; otherwise, the warning or upgraded handling strategy is maintained.

[0228] S540. Adaptive feedback correction of the risk warning strategy based on the evaluation results of the warning implementation effect.

[0229] Specifically, statistical analysis is performed on the evaluation results of the handling effect, the manual confirmation results, and the final handling results of the event obtained in step S530. The matching relationship between event type and early warning action, linkage priority parameters, level classification thresholds, and handling suggestion rules are then corrected in reverse to improve the accuracy of identification and the adaptability of handling of similar risk events in the future.

[0230] Specifically, the risk warning strategy is adaptively adjusted based on the evaluation results of the warning implementation effectiveness, including:

[0231] (1) Statistical analysis was performed based on the evaluation results of the early warning execution effect, the results of manual confirmation, and the final initial results of the event;

[0232] (2) Based on the statistical analysis results, reverse the matching relationship between event types and early warning actions, linkage priority parameters, level classification thresholds and handling suggestion rules.

[0233] It should be noted that for situations with frequent false alarms, ineffective handling, or delayed response, the correction strength of the corresponding parameters in the rule base will be increased; for situations with good handling results and accurate confirmation, the priority matching weight of the corresponding strategy will be increased.

[0234] The correction result of any policy parameter ps after the nth round of feedback can be expressed as:

[0235] Ps(n+1) = lam1Ps(n) + lam2Cb(n) + lam3Db(n) - lam4Fb(n),

[0236] Wherein, Ps(n+1) represents the updated value of the strategy parameter ps after the (n+1)th round of feedback; Ps(n) represents the original value of the strategy parameter ps at the nth round of feedback; Cb(n) represents the positive correction amount formed by the successful disposal sample; Db(n) represents the credibility enhancement amount formed by the manual confirmation of the valid result; Fb(n) represents the negative correction amount formed by false alarms, missed alarms or disposal failures; lam1, lam2, lam3, and lam4 represent the correction coefficients of each feedback component.

[0237] Through the above methods, the system can gradually develop early warning and response strategies that are more adapted to the actual conditions of port railway sites during continuous operation.

[0238] In this embodiment of the invention, risk classification, early warning, and handling based on the port railway risk event set further includes:

[0239] Risk events that have completed the early warning issuance, joint response, effect collection, and feedback correction should be archived in a closed loop to form a closed-loop archive that includes the cause of the event, the relationship chain structure, the early warning action, the joint response process, the response result, the time information, the responsible position, the manual confirmation opinion, and the strategy correction record.

[0240] Specifically, for risk events that have completed the early warning issuance, coordinated response, effect collection, and feedback correction, a complete closed-loop file is formed, including the event cause, relationship chain structure, early warning action, coordinated process, response result, time information, responsible position, manual confirmation opinion, and strategy correction record. This file is then stored in the risk knowledge base as a basis for rapid identification of similar events, rule revision, and management traceability. At the same time, risk events are precipitated into structured knowledge units to support long-term risk profiling analysis, hotspot area statistics, and typical event review in port and railway scenarios.

[0241] The completeness score of any closed-loop knowledge unit kv can be expressed as:

[0242] Qk(kv) = f1Ak(kv) + f2Bk(kv) + f3Ck(kv) + f4Dk(kv) + f5*Ek(kv),

[0243] Wherein, Qk(kv) represents the completeness score of the closed-loop knowledge unit kv; Ak(kv) represents the completeness of the description of the event cause and risk chain; Bk(kv) represents the completeness of the early warning action and linkage process record; Ck(kv) represents the completeness of the handling result and status recovery record; Dk(kv) represents the completeness of the manual confirmation and responsibility information record; Ek(kv) represents the completeness of the feedback correction and strategy revision record; f1, f2, f3, f4, and f5 represent the weight coefficients of each completeness component. When Qk(kv) is greater than the archiving threshold, the closed-loop knowledge unit is written into the risk knowledge base; otherwise, it is marked as an archive to be supplemented.

[0244] The final closed-loop event set can be represented as:

[0245] Eclose = { cv1, cv2, ..., cvz},

[0246] Here, Eclose represents the set of risk events that have completed closed-loop archiving; cv1 to cvz represent z risk event archive units that have completed closed-loop archiving. The closed-loop event set Eclose can serve as a data asset for system operation and management, and also provide basic support for subsequent model optimization, rule updates, and operational analysis.

[0247] In summary, the trackside perception enhancement method for port-dedicated railways provided by this invention, compared with the prior art, firstly solves the problems of inconsistent time references, difficulty in unifying spatial locations, and discrete data representation among multi-source heterogeneous perception information by uniformly accessing, aligning, spatially mapping, and standardizing video data, radar data, trackside status data, and positioning and timing data in the port railway scenario. This improves the consistency, fusionability, and reliability of subsequent processing of perception inputs in complex operating environments.

[0248] Furthermore, this invention enhances the recognition of targets such as trains, carriages, loading and unloading machinery, workers, trackside facilities, and encroaching obstacles by performing enhanced recognition, multi-dimensional feature construction, and cross-source collaborative association. This constructs an enhanced graph of relationships between targets and between targets and regions and facilities, thereby improving the system's capabilities from traditional single-target detection to multi-subject relationship understanding for port railway scenarios. This enhances the accuracy and stability of target recognition and scene semantic understanding in complex occlusion, dynamic intersection, and strong interference environments.

[0249] Furthermore, based on the enhanced relationship graph, this invention further identifies risk chains and determines event levels. Combined with tiered early warning, coordinated response, effect feedback, strategy correction, and knowledge archiving mechanisms, it forms a complete technical chain from risk discovery and risk assessment to risk response and closed-loop optimization. This can effectively improve the early warning foresight, timely response, and long-term adaptive optimization capabilities of complex risk events in port and railway scenarios, and has strong engineering application value.

[0250] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for enhancing trackside sensing on a dedicated port railway, characterized in that, include: Multi-source sensing data is constructed based on train operation targets, loading and unloading machinery targets, personnel targets, trackside facility targets, and environmental disturbance information in the port railway operation area, and standardized multi-source sensing data results are obtained. Enhanced identification of key port and railway targets in the standardized multi-source sensing data results is performed, and multi-dimensional features are constructed from the enhanced identified key targets to obtain an enhanced target set; Cross-source collaborative association and scene relationship enhancement modeling are performed on the enhanced target set to obtain a port-railway relationship enhancement map; Based on the enhanced port-railway relationship map, port-railway scenario risk identification and scenario-based risk attribution are performed to obtain a set of port-railway risk events. The set of port-railway risk events includes at least event category, event level, affected area, key relationship edge, and triggering condition. Risk classification, early warning, and handling are carried out based on the aforementioned set of port and railway risk events.

2. The method for enhancing trackside sensing of a dedicated port railway according to claim 1, characterized in that, Multi-source sensing data is constructed based on train operation targets, loading and unloading machinery targets, personnel targets, trackside facility targets, and environmental disturbance information in the port railway operation area, resulting in standardized multi-source sensing data results, including: A set of scenario areas is constructed based on the operational elements in the port railway operation area, wherein the operational elements in the port railway operation area include at least basic map data, track distribution data, operation boundary data, and key monitoring location data; Acquire multi-source sensing data from multi-source sensing devices within the port railway operation area; Time alignment of multi-source sensing data from multi-source sensing devices; Spatial mapping and unified coordinate transformation are performed on the multi-source sensing data from multi-source sensing devices; After time alignment and spatial mapping, the multi-source sensing data is subjected to sensing data quality screening and standardization processing to obtain standardized multi-source sensing data results.

3. The method for enhancing trackside sensing of a port-dedicated railway according to claim 1, characterized in that, Enhanced identification of key port and railway targets in the standardized multi-source sensing data results is performed, and a multi-dimensional feature set is constructed from the enhanced key targets to obtain an enhanced target set, including: The standardized multi-source sensing data results are grouped according to timestamp, region number, source device type, and spatial proximity to obtain candidate target clusters within the same observation period; The video region, radar spot region, and their neighboring background corresponding to the candidate target cluster are enhanced and saliency-enhanced. The enhanced saliency clusters are subjected to target category discrimination and fine-grained attribute recognition to obtain the joint discrimination results of target category and attributes; Extract multidimensional representation information from the target that has completed target category discrimination, and construct a multidimensional feature vector of the target based on the multidimensional representation information of the target; An enhanced target set is obtained by performing confidence integration based on the joint discrimination result of the target category and attributes and the multidimensional feature vector of the target.

4. The method for enhancing trackside perception of a dedicated port railway according to claim 3, characterized in that, Based on the joint discrimination result of the target category and attribute, and the multidimensional feature vector of the target, confidence integration is performed to obtain an enhanced target set, including: Based on the target category discrimination result, attribute recognition result, and the target multidimensional feature vector for each target, confidence integration is performed to obtain the comprehensive confidence level of target enhancement recognition; The overall confidence level of each target is compared with a preset enhanced output threshold; If the overall confidence level of the target is greater than the preset enhanced output threshold, then the target is written into the enhanced target set; If the overall confidence level of the target is not greater than the preset enhanced output threshold, then the target is marked as a target to be reviewed; After comparing the overall confidence of all targets with the preset enhanced output threshold, an enhanced target set is obtained. The enhanced target set includes at least the target identifier, target category, attribute description, global position, velocity direction, target multidimensional feature vector, and overall confidence.

5. The method for enhancing trackside perception of a dedicated port railway according to claim 1, characterized in that, Cross-source collaborative association and scene relationship enhancement modeling are performed on the enhanced target set to obtain a port-railway relationship enhancement map, including: Perform cross-source consistency matching and identity verification of the enhancement targets based on each target object in the enhancement target set; For enhanced targets that have completed identity verification, perform target temporal trajectory extension and continuous state recovery; Identify the interaction relationships between augmented targets that have formed continuous trajectories; Establish coupling relationships between targets and regions, targets and facilities, and targets and operational status to form coupling relationship models; An enhanced map of port-railway relationships is constructed based on the continuous trajectory of the enhanced targets, the interaction relationships between targets, and the coupling relationship model.

6. The method for enhancing trackside perception of a dedicated port railway according to claim 1, characterized in that, Based on the enhanced port-railway relationship map, port-railway scenario risk identification and scenario-based risk attribution are performed to obtain a set of port-railway risk events, including: The node attributes, relationship attributes, and scene coupling attributes in the port-railway relationship enhancement graph are traversed and analyzed to extract abnormal signs and form a set of risk candidate triggers. The abnormal signs include any one or more of the following: abnormal single node attributes, abnormal enhancement of relationship edges between nodes, and constraint mismatch between nodes, regions, and facilities. Based on the enhanced port-railway relationship map, abnormal signs in the risk candidate trigger set are traced and a multi-entity risk relationship chain is formed; Risk category identification and scenario-based risk attribution are performed based on the multi-entity risk relationship chain. A comprehensive risk score is calculated for the multi-subject risk relationship chain that has completed risk category identification and scenario-based risk attribution, and the event level is classified according to the comprehensive risk score; The risk events that have been classified into risk levels are formed into a risk event output set, and warning trigger conditions are generated for each risk event in the risk event output set.

7. The method for enhancing trackside sensing of a dedicated port railway according to claim 6, characterized in that, A comprehensive risk score is calculated for the multi-agent risk relationship chain that has completed risk category identification and scenario-based risk attribution, and the event level is classified according to the comprehensive risk score, including: A comprehensive risk score is calculated for the multi-agent risk relationship chain based on the abnormal intensity, propagation intensity, duration, scope of impact, importance of affected objects, severity of conflict, and current operational status of the scenario. The comprehensive risk score is classified into event levels according to a preset event level range, wherein the event levels include general warning events, key warning events, serious warning events, and emergency response events.

8. The method for enhancing trackside sensing of a port-dedicated railway according to claim 1, characterized in that, Risk classification, early warning, and handling are carried out based on the aforementioned set of port railway risk events, including: The risk events in the port railway risk events are traversed, and the corresponding early warning release methods are matched according to the preset early warning rule base. The early warning release methods include at least interface pop-up alarms, sound and light alarms, push notifications from job terminals, broadcast prompts in the work area, alarms displayed on the duty center screen, and message distribution from mobile inspection terminals. Based on the event level, risk chain structure, and scope of impact of the scenario, corresponding linkage control strategies and handling instructions are generated for risk events that match the early warning release method. The linkage control strategies include at least automatic video focusing, highlighting of electronic fences in risk areas, output of broadcast warning information, location and dispatch of patrol terminals, confirmation requests from on-duty personnel, and status linkage with external business or control systems. Real-time tracking of risk event status during the early warning execution process, and formation of early warning execution effectiveness evaluation results based on real-time tracking results; The risk warning strategy is adaptively adjusted based on the evaluation results of the early warning implementation effect.

9. The method for enhancing trackside perception of a port-dedicated railway according to claim 8, characterized in that, Based on the evaluation results of the early warning implementation effect, the risk early warning strategy is adaptively adjusted through feedback, including: Statistical analysis was performed based on the evaluation results of the early warning execution effect, the results of manual confirmation, and the final initial results of the event. Based on the statistical analysis results, the matching relationship between event types and early warning actions, linkage priority parameters, level classification thresholds, and handling recommendation rules are revised in reverse.

10. The method for enhancing trackside perception of a port-dedicated railway according to claim 8, characterized in that, Risk classification, early warning, and handling based on the aforementioned set of port and railway risk events also include: Risk events that have completed the early warning issuance, joint response, effect collection, and feedback correction should be archived in a closed loop to form a closed-loop archive that includes the cause of the event, the relationship chain structure, the early warning action, the joint response process, the response result, the time information, the responsible position, the manual confirmation opinion, and the strategy correction record.

Images