Intelligent material warehouse and intelligent tool management method based on railway safety management and control platform

Abstract

The invention discloses an intelligent material warehouse and intelligent tool management method based on a railway safety management and control platform, and relates to the technical field of railway transportation safety, and the method comprises the steps: extracting a motion track of a material tool in a multi-modal data flow, monitoring an operation safety risk event in real time through an intelligent video monitoring method, and generating a risk early warning data packet; cleaning the risk early warning data packet, and performing structured processing in combination with the operation and maintenance asset multi-dimensional data to generate a structural early warning data set; constructing a causal graph model based on the structure early warning data set, mining a risk root cause through a space-time causal discovery algorithm, and obtaining a material library tool co-processing instruction; and executing a material library tool co-processing instruction, optimizing parameters of the causal graph model, and generating a closed-loop optimization package. By constructing the causal graph model and applying the space-time causal discovery algorithm, safety management is changed from passive response to active early warning, and the intrinsic safety level and the intelligent degree of railway operation and maintenance are improved.

Description

Technical Field

[0001] This invention relates to the field of railway transportation safety technology, and in particular to a management method for intelligent material warehouses and intelligent tooling based on a railway safety management and control platform. Background Technology

[0002] With the evolution of IoT, big data, and AI technologies, asset management methods under railway safety control platforms are undergoing a profound transformation from mechanization and informatization to intelligentization. By deploying RFID, intelligent weighing containers, and machine vision technologies, smart warehouses achieve real-time perception of material inventory, environmental parameters, and inbound / outbound dynamics. Furthermore, the integration of AR glasses, professional large-scale models, and electronic fence tools enables automatic inventory counting of personnel and materials on and off tracks, remote visual monitoring of power outages and restorations, and real-time early warnings for operations exceeding permitted scope. Together, these constitute a labor safety control method for specialized scenarios such as railway power supply.

[0003] However, limitations remain in achieving deep collaboration and inherent safety control between smart material warehouses and smart tooling. On the one hand, data fusion largely remains at the platform level, failing to fundamentally break down "data silos." This results in the inability to dynamically trace the complete operational chain of "specific personnel using specific tooling to operate specific materials," making it difficult to accurately assess complex risks across asset linkages. On the other hand, intelligent analysis capabilities are still primarily rule-driven and rely on post-event alerts, lacking a forward-looking causal discovery mechanism. Existing early warnings are mostly based on threshold judgments or fixed rules. While they can achieve in-process intervention, they cannot integrate historical operation sequences, environmental parameters, and asset status changes from a spatiotemporal perspective, hindering the deepening of safety management from in-process intervention to pre-event prevention. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a management method for intelligent material warehouses and intelligent tooling based on a railway safety management and control platform to solve the problems of collaborative risk assessment and the inability of rule-driven early warning to prevent risks in advance.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a management method for intelligent material warehouses and intelligent tooling based on a railway safety management and control platform. The method includes: collecting multi-dimensional data of operation and maintenance assets and performing fusion preprocessing to obtain a multi-modal data stream; extracting the motion trajectory of materials and tooling from the multi-modal data stream and monitoring operational safety risk events in real time using intelligent video monitoring methods to generate risk warning data packets; cleaning the risk warning data packets and performing structured processing in conjunction with the multi-dimensional data of operation and maintenance assets to generate a structured warning dataset; constructing a causal graph model based on the structured warning dataset and using a spatiotemporal causal discovery algorithm to uncover the root causes of risks and obtain collaborative handling instructions for material warehouses and tooling; executing the collaborative handling instructions for material warehouses and tooling and optimizing the parameters of the causal graph model to generate a closed-loop optimization package.

[0008] As a preferred embodiment of the management method for smart material warehouses and smart tooling based on a railway safety management and control platform as described in this invention, the multi-dimensional data of the operation and maintenance assets includes material identity information, real-time status parameters of tooling, personnel operation behavior trajectory, and environmental temperature and humidity.

[0009] The preprocessing includes spatiotemporal reference alignment, cross-modal missing value compensation, and feature dimensionality reduction.

[0010] As a preferred embodiment of the management method for smart material warehouses and smart tooling based on the railway safety management platform described in this invention, the multimodal data stream includes material identity sequences, tooling status time-series data, personnel behavior trajectory fragments, and environmental parameters.

[0011] As a preferred embodiment of the management method for intelligent material warehouse and intelligent tooling based on the railway safety management and control platform described in this invention, the extraction of material tooling movement trajectory from the multimodal data stream refers to extracting the event anchor point set of material identity sequence and tooling status time sequence data, and performing alignment gating interpolation with personnel behavior trajectory fragments to obtain the material tooling movement trajectory.

[0012] As a preferred embodiment of the intelligent material warehouse and intelligent tooling management method based on the railway safety management platform described in this invention, the step of monitoring operational safety risk events in real time through intelligent video monitoring and generating risk warning data packets includes the following specific steps.

[0013] Extract the event anchor set of material identity sequence and tooling status time series data, and perform alignment gating interpolation with personnel behavior trajectory fragments to obtain the material tooling movement trajectory;

[0014] Based on the movement trajectory of materials and tooling, the trajectory gating adjudication method determines the triggering conditions of risk events in real time in the aligned video and generates a list of risk events.

[0015] By performing deep feature association between the risk event list and the multimodal data stream, and calculating the evidence support, risk warning data packages are obtained.

[0016] As a preferred embodiment of the management method for smart material warehouses and smart tooling based on the railway safety management platform described in this invention, the step of cleaning the risk warning data package refers to obtaining a preliminary set of warning events by calculating the similarity of multimodal features.

[0017] As a preferred embodiment of the intelligent material warehouse and intelligent tooling management method based on the railway safety management platform described in this invention, the specific steps for generating the structure early warning dataset are as follows:

[0018] Based on the initial screening and early warning event set and multi-dimensional data of operation and maintenance assets, a multi-dimensional data snapshot is constructed through entity parsing algorithm to generate associated context data blocks;

[0019] The associated context data blocks are processed into spatiotemporal feature vectors and uniformly encoded to generate a sequence of feature vectors.

[0020] Based on feature vector sequences and combined with multimodal data streams to construct entity relationship edges, node mapping is performed through an index archiving algorithm to generate a structure early warning dataset.

[0021] As a preferred embodiment of the intelligent material warehouse and intelligent tooling management method based on the railway safety management platform described in this invention, the specific steps for constructing a causal graph model based on the structural early warning dataset are as follows:

[0022] Based on the structural early warning dataset, an enhanced graph structure is generated by mapping the feature vector sequence to node attributes using the entity relationship edges as the topological basis through a graph embedding algorithm.

[0023] Based on the enhanced graph structure, the temporal dependency and spatial propagation path of node state changes are analyzed by the spatiotemporal gating partial order propagation method, and a preliminary causal graph skeleton is constructed.

[0024] The initial causal graph skeleton and entity relationship edges are constrained and optimized, and spurious related edges are eliminated through a causal pruning algorithm to generate a causal graph model.

[0025] As a preferred embodiment of the intelligent material warehouse and intelligent tooling management method based on the railway safety management platform described in this invention, the specific steps for obtaining the collaborative processing instructions for the material warehouse and tooling are as follows:

[0026] Based on the causal graph model, the causal influence strength of feature vector sequences and entity relationship edges is quantified by a spatiotemporal causal discovery algorithm to generate a time-varying causal intensity graph;

[0027] Based on time-varying causal intensity maps, multi-hop source tracing analysis is used to identify root cause nodes and key propagation paths, and to obtain key causal paths.

[0028] Key causal paths are mapped to root causes of risks, and time-series scheduling is used to obtain instructions for collaborative handling of material warehouse tooling.

[0029] As a preferred embodiment of the intelligent material warehouse and intelligent tooling management method based on the railway safety management platform described in this invention, the steps of executing the collaborative disposal instructions for the material warehouse and tooling, optimizing the parameters of the cause-effect graph model, and generating a closed-loop optimization package are as follows:

[0030] Execute material warehouse tooling collaborative processing instructions and obtain instruction execution logs by monitoring the execution status in real time;

[0031] Based on the instruction execution log, the root cause node and key propagation path are calibrated for consistency before and after execution to obtain the optimization parameters of the cause-effect graph model.

[0032] The parameters of the causal graph model are optimized to minimize the temporal dependency of node state changes, and a closed-loop optimization package is generated by selecting the optimal spatial propagation path.

[0033] The beneficial effects of this invention are as follows: By constructing a causal graph model and applying a spatiotemporal causal discovery algorithm, complex causal relationships among multiple elements such as materials, tooling, personnel, and environment are deeply mined from fused multimodal data, thereby accurately locating the collaborative root causes of risks. This achieves a qualitative leap in safety management from passive response and in-process intervention to proactive early warning and autonomous prevention, improving the inherent safety level and intelligence of railway operation and maintenance. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a flowchart illustrating the management method for smart material warehouses and smart tooling based on a railway safety management platform.

[0036] Figure 2 A flowchart for generating a structure early warning dataset.

[0037] Figure 3 A flowchart for constructing a causal graph model.

[0038] Figure 4 A flowchart for obtaining the closed-loop optimization package. Detailed Implementation

[0039] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0040] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0041] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0042] Reference Figures 1-4 This is one embodiment of the present invention, which provides a management method for a smart material warehouse and smart tooling based on a railway safety management and control platform, including the following steps:

[0043] S1: Collect multi-dimensional data of operation and maintenance assets and perform fusion preprocessing to obtain multi-modal data streams.

[0044] S1.1: Multidimensional data of operation and maintenance assets includes material identity information, real-time status parameters of tooling, personnel operation behavior trajectory and environmental temperature and humidity;

[0045] Specifically, the material identification information comes from the barcode or RFID scanning records in the inbound and outbound process and is linked to the timestamps in the inbound and outbound ledgers;

[0046] The real-time status parameters of the tooling come from the lock / unlock records, borrowing and returning records of the smart tooling cabinet, and the operating status reports from the tooling's built-in status sensors.

[0047] The personnel operation behavior trajectory is derived from the location point sequence of the personnel positioning base station and aligned with the step time record of the work ticket and the timestamp of the access control card swipe record;

[0048] The ambient temperature and humidity are sampled by temperature and humidity sensors deployed in the storage area at a fixed frequency, such as once per second (1Hz) or once every 5 seconds (0.2Hz), forming a time series.

[0049] S1.2: Preprocessing includes spatiotemporal reference alignment, cross-modal missing value compensation, and feature dimensionality reduction;

[0050] Specifically, the barcode or RFID scan records of the inbound and outbound processes in the material identification information are aligned with the timestamps of the inbound and outbound ledgers under a unified time base, and rearranged according to the storage location and material code order; for missing scan records within a unified time window, records of adjacent time windows are supplemented using a unified time base; duplicate identification fields and constant fields are deleted, and representative fields that can distinguish changes in inbound / outbound and holding relationships are retained, thereby obtaining the material identification sequence.

[0051] The real-time status parameters of the tooling include the opening and closing records of the smart tooling cabinets, the borrowing and returning records, and the operating status reports of the tooling's built-in status sensors. These are aligned under a unified time reference and organized into continuous records according to the tooling number and time sequence. Within a unified time window, intermittently reported status records are supplemented by continuing the previous valid status, and time position is supplemented when there is a consistent status in adjacent time windows. Equivalent status indication fields are merged, and representative fields that can reflect start-stop and borrowing / returning changes are retained, thus obtaining the tooling status time sequence data.

[0052] Align the location point sequence generated by the personnel positioning base station with the time records of the steps on the work ticket and the timestamps of the access control card swipe records under a unified time reference, and organize them into a continuous trajectory according to personnel identification and path order; within a unified time window, short-term positioning gaps are spliced ​​together by adjacent time window trajectory segments, while maintaining path continuity and filling in the time and position of segments consistent with the step time; delete redundant coordinate fields and duplicate step markers, and retain representative fields that can represent the intervals of stay, turning and movement, thereby obtaining personnel behavior trajectory segments.

[0053] The environmental temperature and humidity are aligned with the spatiotemporal reference. The temperature and humidity sampling sequences deployed in the storage area are aligned with the unified time reference and consistent with the corresponding storage location spatial identifiers. For occasional missing measurements within the unified time window, the sampling values ​​of adjacent time windows are used to fill in the gaps to form a continuous time series. The duplicate sampling columns and constant interval fields are deleted, and the representative fields that can reflect the trend of temperature and humidity changes are retained, thereby obtaining the environmental parameters.

[0054] The material identification sequence, tooling status time series data, personnel behavior trajectory fragments, and environmental parameters are aligned under a unified time base and connected in series into a four-channel parallel time series to obtain a multimodal data stream.

[0055] S2: Extract the material tooling movement trajectory from the multimodal data stream, and monitor operational safety risk events in real time through intelligent video monitoring methods to generate risk warning data packages.

[0056] S2.1: Extract the event anchor set of material identity sequence and tooling status time series data, and perform alignment gating interpolation with personnel behavior trajectory segments to obtain the material tooling movement trajectory;

[0057] It should be noted that in the multimodal data stream, the time points of retrieval, return and storage location change are located from the material identity sequence according to a unified time base, and the state steps of unlocking, powering on, borrowing, returning and stopping are located from the tooling status time sequence data according to a unified time base, and are merged into an event anchor set in chronological order.

[0058] Establish a one-to-one correspondence between the event anchor point set and the personnel behavior trajectory segment within the same aligned time window, based on personnel identification and timestamp. After completing the correspondence, sequentially traverse the trajectory coordinate points along the time sequence of the personnel behavior trajectory segment between each pair of adjacent event anchor points. When the material identity sequence shows that the holding relationship is established and the tooling status time sequence data shows that the active status is established (meaning that any one of the unlocking mark, power-on mark, borrowing mark, return mark, or deactivation mark is true within the time window), write the current trajectory coordinate point and the corresponding timestamp into the material tooling movement trajectory. When any gating condition is not met, pause the writing and continue to the next time point until the gating conditions are met simultaneously. When the next event anchor point is reached, end the interpolation of this segment and encapsulate it into a continuous position segment. Serialize all continuous position segments according to the event anchor point sequence and attach the material code, tooling number, and source event anchor point identifier to obtain the material tooling movement trajectory.

[0059] S2.2: Based on the movement trajectory of materials and tooling, the trajectory gating adjudication method determines the triggering conditions of risk events in real time in the aligned video and generates a list of risk events;

[0060] It should be noted that the timestamps and position coordinates in the movement trajectory of the material tooling are used to locate and align the corresponding frames in the video one by one. The monitoring area is then cropped from the corresponding frame with the position coordinates as the center to form a ROI frame sequence consistent with the time axis.

[0061] Within the ROI frame sequence, the spatial relationship between personnel behavior trajectory segments and material tooling movement trajectories is compared in chronological order. The moment when separation changes to intersection or crossing is recorded as a topology hit. Within the same aligned time window, the tooling status time sequence data and material identity sequence are read synchronously to determine whether the active status in the tooling status time sequence data is valid or whether the holding relationship in the material identity sequence has changed. If either one is valid, it is recorded as a status hit. When topology hit and status hit occur simultaneously, a risk event record is generated. The record includes the event type, occurrence time, associated trajectory identifier, trigger anchor point identifier, and ROI frame positioning information. All risk event records are accumulated in chronological order and organized into a fixed field format to obtain a risk event list.

[0062] Furthermore, the aligned video includes video data, video frame timestamps, and spatial coordinates; the video data is acquired from a fixed camera, the video frame timestamps are aligned with a unified time reference, and the spatial coordinates are mapped from pixel coordinates to library coordinates through library location calibration points; after the timestamps and coordinates are unified, the aligned video is formed.

[0063] S2.3: Perform deep feature association between the risk event list and the multimodal data stream, calculate the evidence support, and obtain the risk warning data package.

[0064] It should be noted that, assuming the risk event list has been generated, the material identity sequence, tooling status time series data, personnel behavior trajectory fragments, and environmental parameter window records are located one by one in the multimodal data stream according to the occurrence time, associated trajectory identifier, and trigger anchor point identifier in the risk event list, and the field-level one-to-one connection is completed; for each risk event record, the number of consecutive frames in the ROI frame sequence that change from separation to intersection or crossing is counted as the topology hit count, the number of times the active status in the tooling status time series data appears within the event time window is counted as the active hit count, the number of times the holding relationship in the material identity sequence changes within the event time window is counted as the holding change count, and the overlap time between the event time window and the environmental parameter indicating the workable range (adjusted according to the season or warehouse scenario, such as temperature [−10,40]℃, relative humidity [0.2,0.8]) is counted as the environmental consistency time, and the time series consistency results with the personnel behavior trajectory fragments within the event time window are recorded;

[0065] The expression for calculating the evidence support is:

[0066] ;

[0067] in, For the degree of support of evidence, For consecutive intersecting frames, Total number of frames in the time window. For active hit rate, For changes in shareholding percentage, For the sake of environmental consistency, To observe the coverage percentage, This represents the percentage of cases with inconsistent timing.

[0068] After completing the counting and duration statistics, the event type, occurrence time, associated trajectory identifier, trigger anchor point identifier, ROI frame location information and evidence support are assembled into a fixed format record, summarized and packaged in chronological order, and a risk warning data package is generated.

[0069] Ideally, material identification sequences, tooling status time-series data, personnel behavior trajectory fragments, and environmental parameters are unified under event fingerprints for field-level integration, enabling traceable aggregation of multi-source evidence; and the recorded content is output in a structured manner as a risk warning data package, supporting direct invocation for subsequent causal analysis and handling linkage.

[0070] S3: Clean the risk warning data packets and perform structured processing in conjunction with multi-dimensional data of operation and maintenance assets to generate a structured warning dataset.

[0071] S3.1: Clean the risk warning data package and obtain the initial screening warning event set by calculating the multimodal feature similarity;

[0072] It should be noted that an event fingerprint is generated by splicing the event type, occurrence time, associated trajectory identifier, trigger anchor point identifier, and ROI frame positioning information in the risk warning data packet. Duplicate event fingerprints are merged one by one within the same aligned time window, and the event with the lowest evidence support is removed. Based on the occurrence time and associated trajectory identifier in the event fingerprint, the material identity sequence, tooling status time series data, personnel behavior trajectory fragments, and environmental parameters are located one by one in the multimodal data stream and the field-level one-to-one docking is completed.

[0073] The expression for calculating multimodal feature similarity is:

[0074] ;

[0075] in, For multimodal feature similarity, Indicating consistency of event type, To ensure consistency between the associated trajectory identifiers, To trigger the anchor point identifier consistency indication, The overlap ratio between the occurrence time window and the ROI time window. It is the ratio of environmental consistency duration to the event time window duration. This is the timing consistency ratio. It is the hyperbolic tangent function.

[0076] Records that meet the selection criteria based on multimodal feature similarity (event type consistency, associated trajectory identification consistency, trigger anchor point identification consistency, time window overlap is true, and at least one of the following is true in active state or holding relationship change, and spatial relationship changes from separation to intersection or crossing is true) are aggregated to obtain the initial screening warning event set.

[0077] A better approach is to ensure consistency of origin and full-chain traceability by deduplicating event fingerprints and connecting them with fields; and to jointly constrain spatial intersection ratio, time window overlap, environmental consistency, temporal consistency and active holding changes, thereby reducing false alarms and false negatives, and forming a set of preliminary screening and early warning events with low noise, strong correlation and usable for subsequent entity parsing and causal analysis.

[0078] S3.2: Based on the initial screening and early warning event set and multi-dimensional data of operation and maintenance assets, construct a multi-dimensional data snapshot through entity parsing algorithm to generate associated context data blocks;

[0079] It should be noted that, based on the occurrence time window, material code, tooling number, personnel identifier and storage location identifier in the initial screening and early warning event set as search keys, the entry and exit records of material identity information, the status change records of tooling real-time status parameters, the location sequence of personnel operation behavior trajectory and the time record of operation steps, and the same window sampling sequence of environmental temperature and humidity are located in the multi-dimensional data of operation and maintenance assets.

[0080] By material code, inbound and outbound records and holding relationships are concatenated chronologically to form a continuous segment of material identification information; by tooling number, unlocking, powering on, borrowing, returning, and operating status reporting are concatenated chronologically to form a continuous segment of tooling real-time status parameters; by personnel identification, location sequence and operation step time records are spliced ​​chronologically to form a continuous segment of personnel operation behavior trajectory; by storage location identification, temperature sampling and humidity sampling are combined chronologically to form a continuous segment of environmental temperature and humidity.

[0081] Alignment is performed one by one around the occurrence time window, and the start and end times and timestamp formats are standardized. The consistency of material codes, tooling numbers, personnel identification and storage location identification is checked item by item. Continuous segments of material identity information, continuous segments of tooling real-time status parameters, continuous segments of personnel operation behavior trajectories and continuous segments of environmental temperature and humidity are encapsulated according to fixed field position order. These are summarized to form a multi-dimensional data snapshot organized by event fingerprints. The data is then integrated according to the event sequence to generate associated context data blocks.

[0082] Furthermore, the entity parsing algorithm uses a unified association standard based on material code, tooling number, personnel identifier, and storage location identifier to eliminate conflicts such as duplicate codes for the same name and multiple codes for one item; it binds inbound and outbound records, status change records, trajectory points, and temperature and humidity samples one by one within the same time window to reconstruct continuous context segments; and it outputs traceable multidimensional data snapshots to reduce the probability of false matching and missed matching, thereby improving the accuracy and stability of subsequent spatiotemporal feature vector sequence generation and causal graph construction.

[0083] S3.3: Perform spatiotemporal feature vectorization on the associated context data block and encode it uniformly to generate a feature vector sequence;

[0084] It should be noted that the data is divided into continuous equal-length time windows based on the unified start and end times and timestamps in the associated context data block, and the values ​​of the preceding timestamps at the gaps are filled in to fill the gaps; the material identity information is expanded into a pick-up flag, a return flag, and a holding relationship flag within each time window; the real-time status parameters of the tooling are expanded into an unlock flag, a power-on flag, a borrowing flag, a return flag, a deactivation flag, and a corresponding duration within each time window; the personnel operation behavior trajectory is expanded into a position coordinate sequence, whether displacement occurred, whether a stop occurred, and whether a turn occurred within each time window in chronological order; the environmental temperature and humidity are expanded into a temperature sequence and a humidity sequence within each time window.

[0085] Event fingerprints, material codes, tooling numbers, personnel identifiers, and storage location identifiers are stably numbered and encoded according to fixed column positions. Time-related fields are serialized according to a unified timestamp format. The expanded columns of each channel are concatenated in a fixed order. For time window columns that are insufficient in length (the number of records within the agreed time window steps is less than the target number of steps), the previous value is retained and filled. For example, if the target time window contains 5 time steps but the personnel operation behavior trajectory column only has 3 records, the previous value of the last record is retained and filled twice to make it 5 records. Time window columns that exceed the length are truncated to the same length according to the time sequence. The encoding results of the time windows are concatenated in the time sequence and integrated according to the event fingerprint to generate a feature vector sequence.

[0086] S3.4: Based on the feature vector sequence and combined with multimodal data stream, entity relationship edges are constructed, and node mapping is performed through index archiving algorithm to generate a structure early warning dataset.

[0087] It should be noted that event nodes, material nodes, tooling numbers, personnel identifiers, storage location identifiers, and occurrence time windows are created based on event fingerprints, material codes, tooling numbers, personnel identifiers, storage location identifiers, and occurrence time windows in the feature vector sequence. Entity relationship edges are generated based on the aligned material identity sequences, tooling status time-series data, personnel behavior trajectory fragments, and environmental parameters in the multimodal data stream. The relationship types of entity relationship edges include event participation relationships (event nodes connect to material nodes, tooling nodes, personnel nodes, and environment nodes), holding relationships (material nodes and personnel nodes are bound within a time window), borrowing and returning relationships (tooling nodes and personnel nodes are bound within a time window), trajectory sharing window relationships (personnel nodes and material nodes or tooling nodes are spatially adjacent within the same time window), and temporal sequence relationships (event nodes in the previous time window point to event nodes in the next time window).

[0088] Based on entity relationship edges, an index is established using a combination of event fingerprint, associated trajectory identifier, and occurrence time window as the primary key. An inverted index is established using material code, tooling number, personnel identifier, storage location identifier, and ROI frame location information. Nodes and entity relationship edges are mapped according to the primary key index and stored in partitions according to line, station, storage location, and time window to generate a structural early warning dataset.

[0089] S4: Construct a causal graph model based on the structural early warning dataset, and use a spatiotemporal causal discovery algorithm to uncover the root causes of risks and obtain collaborative disposal instructions for material warehouse tooling.

[0090] S4.1: Based on the structural early warning dataset, an enhanced graph structure is generated by mapping the feature vector sequence to node attributes using the entity relationship edges as the topological basis through a graph embedding algorithm.

[0091] It should be noted that, based on the entity relationship edges in the structural early warning dataset, the corresponding records are located one by one within the same occurrence time window, using the fixed column position in the feature vector sequence as the basis.

[0092] Write the material identification information, including the retrieval mark, return mark, and holding relationship mark, into the material node attributes; write the tooling real-time status parameters, including the unlock mark, power-on mark, borrow mark, return mark, deactivation mark, and corresponding duration, into the tooling node attributes; write the personnel operation behavior trajectory's location coordinate sequence, whether displacement occurred, whether pause occurred, and whether turning occurred into the personnel node attributes; write the ambient temperature and humidity sequences into the environment node attributes, and record the event type, occurrence time, ROI frame location information, and trigger anchor point identifier in the event node;

[0093] The entity relationship edges are connected to the corresponding nodes according to the event participation relationship, holding relationship, borrowing and returning relationship, trajectory co-window relationship and time sequence relationship, and the timestamp field and primary key information are retained to generate an enhanced graph structure.

[0094] S4.2: Based on the enhanced graph structure, the temporal dependency and spatial propagation path of node state changes are analyzed by the spatiotemporal gating partial order propagation method, and a preliminary causal graph skeleton is constructed.

[0095] It should be noted that a temporal partial order is established from early to late according to the occurrence time of the enhanced graph structure, and entity relationship edges are used as propagable adjacencies; within each time window, the attribute change markers of event nodes, material nodes, tooling nodes, personnel nodes, and environmental nodes are scanned (material identity information: retrieval marker, return marker, holding relationship marker; tooling real-time status parameters: unlock marker, power-on marker, borrow marker, return marker, and deactivation marker; whether personnel operation behavior trajectory displacement, dwell time, and turning have occurred; environmental temperature and humidity sequences corresponding to prohibited or permitted intervals).

[0096] Provided that the time partial order is valid and the entity relationship edges are connected, the trigger node of the earlier time window is pointed to the response node of the later time window to form a directed edge only when the active related flags (unlock flag, power-on flag, borrow flag, return flag, and deactivation flag) of the tooling real-time status parameters are true, or the holding relationship flag of the material identity information changes to true in two adjacent time windows, or the gate condition of the ambient temperature and humidity falling into the prohibited or allowed range is true.

[0097] After propagation through all time windows is complete, all directed edges and their corresponding nodes are aggregated to generate a preliminary causal graph skeleton.

[0098] S4.3: Constrain and optimize the initial causal graph skeleton and entity relationship edges, and eliminate spurious related edges through the causal pruning algorithm to generate a causal graph model;

[0099] It should be noted that, based on the preliminary cause-effect graph skeleton and entity relationship edges, for each directed edge in the preliminary cause-effect graph skeleton, according to the time sequence, for the event node, material node, tooling node, personnel node and environment node where the start and end points are located, the corresponding relationship is queried in the set of entity relationship edges that are consistent with the time window of the directed edge.

[0100] Verify that the start time stamp of the connection is strictly earlier than the end time stamp, or that the ROI frame location information frame number is earlier than the end frame number at the same time stamp; verify that the active related markers, holding relationship change markers, or environmental permitted and prohibited zone markers are consistent with the connection direction; immediately delete any directed connection that fails the verification and record the reason for deletion.

[0101] Using the same event type and the same entity combination as units, if the remaining directed edges show a situation where the starting trigger appears multiple times in consecutive time windows but the ending response does not appear in the corresponding time windows, the corresponding directed edges will be deleted; if the ending response appears within the time window but the starting trigger does not appear in the same window or the previous window, the corresponding directed edges will be deleted, thus obtaining a causal graph model composed of the checked node set and the retained edges.

[0102] Furthermore, by eliminating or shrinking edges with no cause and no effect, inconsistent time sequence with gating conditions, low frequency of occurrence, and low temporal consistency, pseudo-correlated edges and redundant back edges are removed; the sparsity and stability of the causal graph model are improved, making key causal paths more focused, root cause location more accurate, and action mapping more verifiable, while also improving the generalization reliability under different operational scenarios.

[0103] A superior cause-effect graph model connects events, materials, tools, personnel, and the environment into traceable directed links based on chronological order and gating relationships, making the deduction from alarms to root causes transparent and verifiable. Under a unified data standard, it prunes spurious directed links and quantifies and verifies propagation delays and triggering conditions, thereby reducing false alarms and missed detections. Key causal paths can be directly mapped to handling actions, supporting the coordinated implementation of pre-event prevention and in-event intervention. Execution logs can also reverse-calibrate link parameters, forming a continuously converging closed-loop management capability.

[0104] S4.4: Based on the causal graph model, the causal influence strength of feature vector sequences and entity relationship edges is quantified through a spatiotemporal causal discovery algorithm to generate a time-varying causal intensity graph;

[0105] It should be noted that, for each directed edge in the causal graph model, the participating entities and corresponding occurrence time windows are located on the entity relationship edge. The source node time sequence markers and response node time sequence markers consistent with the occurrence time windows are extracted from the feature vector sequence and aligned to a unified time axis. Quantitative indicators are statistically analyzed window by window on the aligned time axis, including the successive hit count (the number of hits with the source node first and the response node second), the consecutive hit duration (the length of time that the source node and response node are consecutively established), and the gating consistency ratio (the proportion of active related markers, holding relationship change markers, environmental allowance and prohibition interval markers, and the proportion of edges with consistent directions). When all of these conditions are met, the time window is marked as "valid" as true. For example, the successive hit count is higher than once, the consecutive hit duration is not less than one time window, and the gating consistency ratio is not less than 0.7. The difference between the source time and the response time is also recorded. The quantitative indicators, valid markers, and time differences of each time window are summarized in chronological order according to the directed edge number to form the causal influence intensity, generating a time-varying causal intensity map.

[0106] S4.5: Based on the time-varying causal intensity map, identify root cause nodes and key propagation paths through multi-hop source tracing analysis to obtain key causal paths;

[0107] It should be noted that alarm nodes are extracted from the time-varying causal intensity graph, valid markers are selected in chronological order, and the time window where the source time is earlier than the response time is used as the tracing starting point. Based on the tracing starting point, a multi-hop search is performed along the directed edges, and only the predecessor edge with the largest causal influence strength and consistent gating conditions is retained in each hop. During the continuous hop process, the source node, response node, source time, response time and causal influence strength are recorded for each hop, and the search stops when there is no predecessor node or the predecessor edge no longer satisfies the valid marker.

[0108] Each path from the alarm node back to the stop node is organized into a candidate propagation path set in chronological order. The paths are sorted according to the continuity of causal influence intensity and the coverage length of the time window. The path with the longest coverage length and true continuity of causal influence intensity is selected as the critical propagation path. The starting node of the critical propagation path is identified as the root cause node, thus generating the critical causal path.

[0109] S4.6: Map key causal paths to root causes of risks and obtain material warehouse tooling collaborative handling instructions through time-series scheduling.

[0110] It should be noted that, according to the critical causal path, the event type, occurrence time, material code, tooling number, personnel identifier, storage location identifier, and trigger anchor point identifier are read edge by edge, and a detailed table of "node-edge-time window" is generated; the root cause node that is triggered earliest within the effective time window at the starting point of the critical causal path is marked as "root cause of risk" in the detailed table; based on the fixed correspondence between entity relationship edge type and event type, each directed edge is mapped to a disposal action item for materials, tooling, personnel, and storage location, and the start and end times and preconditions are given. For example, when the tooling is powered on, a power-off order is issued for the tooling and personnel are notified to evacuate; when materials are held by unauthorized personnel, an audible and visual alarm is triggered for the storage location and the tooling is locked for borrowing or returning; when environmental parameters exceed the allowable range and there is live operation, the live operation is suspended and personnel are prompted to evacuate.

[0111] The disposal action items are topologically sorted according to the time sequence of the critical cause-effect path, items with overlapping times of the same tooling number or the same storage location identifier are merged, and items with pre-constraints are ordered in execution order. The disposal action items ordered in time sequence are then summarized to generate a material warehouse tooling collaborative disposal instruction.

[0112] S5: Execute the material warehouse tooling collaborative processing instructions and optimize the parameters of the cause-effect graph model to generate a closed-loop optimization package.

[0113] S5.1: Execute the collaborative processing instructions for the material storage tooling and obtain the instruction execution log by monitoring the execution status in real time;

[0114] It should be noted that the disposal action items in the material warehouse tooling collaborative disposal instruction are triggered one by one in the order of execution time (based on the action object identifier and action content in the disposal action item, and executed at the start time of the plan). After the disposal action item is triggered, the start time and action object identifier of the disposal action item are recorded immediately. At the same time, it is monitored whether there is a change in the material identity sequence that is consistent with the time window of the disposal action item; it is monitored whether the status is started, stopped or locked in the tooling status time sequence data; it is monitored whether the arrival or departure time appears in the personnel operation behavior trajectory; it is monitored whether the temperature and humidity are within the allowable range in the environmental parameters, and the corresponding frame number is recorded in the ROI frame positioning information of the aligned video.

[0115] Summarize the start time, end time, action object identifier, action content, execution result, trigger anchor point identifier, ROI frame positioning information, material identity sequence change record, tooling status time sequence data change record, personnel operation behavior trajectory arrival and departure time, and environmental parameter range hit status according to the action item number, and generate an instruction execution log.

[0116] S5.2: Based on the instruction execution log, perform consistency calibration on the root cause node and key propagation path before and after execution, and obtain the optimization parameters of the cause-effect graph model;

[0117] It should be noted that, based on the generated instruction execution log and the root cause node and critical propagation path, a "pre-execution prediction chain and post-execution observation chain" are established for each connection according to the connection order of the critical propagation path; the start time, end time, action object identifier, execution result, ROI frame positioning information, material identity sequence change record, tooling status time sequence data change record, personnel operation behavior trajectory arrival and departure time, and environmental parameter range hit status in the instruction execution log are compared one by one with the source node time, response node time, and trigger anchor point identifier in the critical propagation path;

[0118] If the source node does not trigger a corresponding response node after completing the action, it is marked as "cause without effect" and the difference between the source time and the response time is recorded as the basis for latency correction. If a response node appears but the preceding edge is not established, it is marked as "effect without cause" and the trigger anchor point identifier and occurrence time are recorded as the basis for tightening the gating time window. After all edges are marked, the number of times each edge appears in the instruction execution log coverage, the length of the continuous effective time window, and the number of times the ROI frame sequence number is satisfied are counted. The ratio of the number of times the ROI frame sequence number is satisfied to the number of times each edge appears in the instruction execution log coverage is used as the time consistency index.

[0119] The "delay correction basis, gating time window tightening basis and time consistency index" are summarized by edge numbering to generate causal graph model optimization parameters.

[0120] S5.3: Optimize parameters based on the cause-effect graph model, minimize the temporal dependency of node state changes, and generate a closed-loop optimization package by selecting the optimal spatial propagation path.

[0121] It should be noted that, based on the causal graph model, the parameters are optimized by applying the latency correction criteria and the gating time window tightening criteria to the corresponding directed edges, updating the allowable difference between the source time and the response time, and tightening the effective interval of the trigger anchor point; the parallel directed edges of the same response node are filtered based on the time consistency index, and the edge with the lowest occurrence frequency in the instruction execution log coverage is deleted; then, in each group of candidate propagation paths, the spatial propagation path with the shortest difference between the source time and the response time and the longest continuous effective time window coverage is selected first to generate a closed-loop optimization package.

[0122] In summary, this invention, by constructing a causal graph model and employing a spatiotemporal causal discovery algorithm, deeply mines the complex causal relationships among multiple elements such as materials, tooling, personnel, and environment from fused multimodal data, thereby accurately locating the collaborative root causes of risks. This achieves a qualitative leap in safety management, moving from passive response and in-process intervention to proactive early warning and autonomous prevention, thus improving the inherent safety level and intelligence of railway operation and maintenance.

[0123] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A management method of a smart warehouse and a smart tool based on a railway safety management and control platform, characterized in that: The application relates to a method for realizing safety risk early warning of material warehouse and tooling. The method comprises the following steps: collecting and fusing multi-dimensional data of operation and maintenance assets to obtain a multi-modal data stream; extracting a movement track of a material tooling in the multi-modal data stream, and monitoring a work safety risk event in real time through an intelligent video monitoring method to generate a risk early warning data package; cleaning the risk early warning data package, and combining the multi-dimensional data of operation and maintenance assets to perform structured processing to generate a structured early warning data set; constructing a causal diagram model based on the structured early warning data set, and mining a risk root cause through a spatio-temporal causal discovery algorithm to obtain a material warehouse tooling collaborative disposal instruction; 2. The management method of the smart warehouse and smart tool based on the railway safety management and control platform according to claim 1, characterized in that: executing the material warehouse tooling collaborative disposal instruction, and optimizing parameters of the causal diagram model to generate a closed-loop optimization package. The multi-dimensional data of operation and maintenance assets comprises material identity information, real-time state parameters of tooling, personnel operation behavior tracks and environmental temperature and humidity. 3.The management method of the smart warehouse and smart tool based on the railway safety management and control platform according to claim 2, characterized in that: The preprocessing comprises time-space benchmark alignment, cross-modal missing value compensation and feature dimension reduction.

4. The management method of the smart warehouse and smart tool based on the railway safety management and control platform according to claim 3, characterized in that: The multi-modal data stream comprises a material identity sequence, tooling state time series data, personnel behavior track segments and environmental parameters.

5. The management method of the smart warehouse and smart tool based on the railway safety management and control platform according to claim 4, characterized in that: The extraction of the material tooling movement track in the multi-modal data stream refers to the extraction of an event anchor point set of the material identity sequence and the tooling state time series data, and the alignment gate control interpolation with the personnel behavior track segments to obtain the material tooling movement track. The method for monitoring the work safety risk event in real time through the intelligent video monitoring method to generate the risk early warning data package comprises the following steps: extracting the event anchor point set of the material identity sequence and the tooling state time series data, and performing alignment gate control interpolation with the personnel behavior track segments to obtain the material tooling movement track; based on the material tooling movement track, a track gate control decision method judges a risk event triggering situation in the aligned video in real time to generate a risk event list; 6. The management method of the smart warehouse and smart tool based on the railway safety management and control platform according to claim 5, characterized in that: deep feature association is performed on the risk event list and the multi-modal data stream, and evidence support degree is calculated to obtain the risk early warning data package.

7. The management method of the smart warehouse and smart tool based on the railway safety management and control platform according to claim 6, characterized in that: The cleaning of the risk early warning data package refers to the calculation of multi-modal feature similarity to obtain a preliminary screening early warning event set. The generation of the structured early warning data set comprises the following steps: based on the preliminary screening early warning event set and the multi-dimensional data of operation and maintenance assets, a multi-dimensional data snapshot is constructed through an entity analysis algorithm to generate an associated context data block; the associated context data block is subjected to time-space feature vectorization processing and unified coding to generate a feature vector sequence; 8. The management method of the smart warehouse and smart tool based on the railway safety management and control platform according to claim 7, characterized in that: based on the feature vector sequence and in combination with the multi-modal data stream, an entity relationship edge is constructed, node mapping is performed through an index archiving algorithm, and a structured early warning data set is generated. The construction of the causal diagram model based on the structured early warning data set comprises the following steps: based on the structured early warning data set, a graph embedding algorithm is used to map the feature vector sequence to node attributes based on the entity relationship edge as a topological basis to generate an enhanced graph structure; according to the enhanced graph structure, a spatio-temporal gate control partial order propagation method is used to analyze time sequence dependency and space propagation path of node state changes to construct a preliminary causal diagram skeleton; 9.The management method of the smart warehouse and smart tool based on the railway safety management and control platform according to claim 8, characterized in that: the preliminary causal diagram skeleton is subjected to constraint optimization with the entity relationship edge, and a causal pruning algorithm is used to eliminate pseudo-related edges to generate a causal diagram model. The obtaining of the material warehouse tooling collaborative disposal instruction comprises the following steps: Based on the causal graph model, the time-varying causal strength graph is generated by quantifying the causal influence strength of the feature vector sequence and the entity relationship edge through the spatiotemporal causal discovery algorithm; Based on the time-varying causal strength graph, the root cause node and the key propagation path are identified through multi-hop tracing analysis to obtain the key causal path; The key causal path is mapped to the root cause of the risk, and the warehouse tool collaborative disposal instruction is obtained through timing scheduling.

10. The management method of the smart warehouse and smart tool based on the railway safety management and control platform according to claim 9, characterized in that: The warehouse tool collaborative disposal instruction is executed, and the parameters of the causal graph model are optimized to generate a closed-loop optimization package, and the specific steps are as follows, The warehouse tool collaborative disposal instruction is executed, and the instruction execution log is obtained by monitoring the execution state in real time; Based on the instruction execution log, the consistency calibration before and after the execution of the root cause node and the key propagation path is performed to obtain the optimization parameters of the causal graph model; According to the optimization parameters of the causal graph model, the timing dependency of the node state change is minimized, and the closed-loop optimization package is generated by screening the optimal spatial propagation path.

Images