High-speed rail platform safety determination method and system based on mixed precision inference

Abstract

The application relates to the technical field of intelligent reasoning and safety judgment, and discloses a high-speed rail platform safety judgment method and system based on mixed-precision reasoning, which comprises the following steps: acquiring multi-source semantic observation records and generating a safety observation element set; constructing a platform safety relation graph; determining relation conflict density, closed residual error, cross-source divergence degree and reasoning difficulty level; generating a mixed-precision bit width scheduling table; performing graph relation reasoning to obtain an initial safety judgment vector and an initial judgment boundary quantity; in step 6, a final judgment vector is determined; and in step 7, a locked safety level is determined and a safety judgment package is output. The application realizes mixed-precision safety judgment and locked output driven by multi-source semantic relation of a high-speed rail platform.

Description

Technical Field

[0001] This invention belongs to the field of intelligent reasoning and safety determination technology, specifically relating to a method and system for determining the safety of high-speed railway platforms based on hybrid precision reasoning. Background Technology

[0002] In high-speed rail station scenarios, passenger flow, area occupancy, train arrival processes, equipment status, and alarm handling often change simultaneously. Safety judgments are not only affected by the status of a single object but also by the relationships between objects, areas, processes, and handling procedures. Existing station safety handling methods mostly rely on single-point detection results or fixed rule-based judgments, typically processing data from different sources separately before summarizing. This makes it difficult to incorporate the correlation conflicts between multi-source semantic information, the closure of handling procedures, and cross-source consistency into a single analysis chain. Especially when deployed at the station edge, computational resources and response latency are constrained. If a uniform precision and depth of inference is consistently used, computational redundancy can occur in low-complexity scenarios, while judgments become unstable in high-complexity scenarios. Furthermore, when different source records give inconsistent statuses to the same object, or when there are breaks in the link between alarms, responses, and confirmations, existing methods often lack a unified means of characterizing relational structures, leading to significant fluctuations in judgment results. Summary of the Invention

[0003] This invention provides a method and system for determining the safety of high-speed railway platforms based on hybrid precision reasoning, which solves the technical problems in the background art.

[0004] This invention provides a method for determining the safety of high-speed railway platforms based on mixed-precision reasoning, comprising the following steps:

[0005] Step 1: Obtain the multi-source semantic observation records for the current judgment period, standardize and encapsulate the multi-source semantic observation records, and determine the observation weights according to the preset source mapping table to obtain a set of safe observation elements;

[0006] Step 2: Construct a platform safety relationship diagram based on the set of safety observation elements and the preset relationship rule table;

[0007] Step 3: Based on the platform safety relationship diagram, the preset conflict matrix, and the preset closed link template, determine the relationship conflict density, closure residual, and cross-source divergence degree, and determine the reasoning difficulty level;

[0008] Step 4: Based on the inference difficulty level, pre-stored layer importance, and quantization sensitivity, generate a mixed precision bit-width scheduling table;

[0009] Step 5: Based on the platform safety relationship diagram and the mixed precision bit-width scheduling table, perform graph relationship reasoning to obtain the initial safety judgment vector and the initial judgment boundary quantity;

[0010] Step 6: Determine the final decision vector based on the initial decision boundary quantity, the relationship conflict density, and the closure residual; if the preset conditions are not met, extract the conflict kernel graph based on the platform safety relationship graph and perform high-precision deep reasoning to obtain the final decision vector;

[0011] Step 7: Based on the final decision vector, the relation conflict density, the closure residual, and the initial decision boundary quantity, determine the locking security level and output a security decision packet.

[0012] This invention also provides a high-speed railway platform safety determination system based on hybrid precision inference, comprising:

[0013] The security observation element generation module is used to obtain multi-source semantic observation records for the current judgment period, standardize and encapsulate the multi-source semantic observation records, and determine the observation weights according to a preset source mapping table to obtain a security observation element set.

[0014] The platform relationship diagram construction module is used to construct a platform safety relationship diagram based on the set of safety observation elements and the preset relationship rule table.

[0015] The relationship index determination module is used to determine the relationship conflict density, closure residual and cross-source divergence degree based on the platform safety relationship diagram, the preset conflict matrix and the preset closed link template, and to determine the reasoning difficulty level.

[0016] The bit-width scheduling generation module is used to generate a mixed-precision bit-width scheduling table based on the inference difficulty level, pre-stored layer importance, and quantization sensitivity.

[0017] The graph relationship reasoning module is used to perform graph relationship reasoning based on the platform safety relationship graph and the mixed precision bit-width scheduling table to obtain the initial safety judgment vector and the initial judgment boundary quantity.

[0018] The conflict kernel deep reasoning module is used to determine the final decision vector based on the initial decision boundary quantity, the relationship conflict density, and the closure residual; when the preset conditions are not met, it extracts the conflict kernel graph based on the platform safety relationship graph and performs high-precision deep reasoning to obtain the final decision vector.

[0019] The security level locking module is used to determine the locking security level based on the final decision vector, the relation conflict density, the closure residual, and the initial decision boundary quantity, and output a security decision packet.

[0020] The beneficial effects of this invention are as follows: Based on the multi-source semantic observation results of high-speed railway platforms, this invention first constructs a platform safety relationship graph, then determines the inference difficulty level by combining relationship conflict density, closure residuals, and cross-source divergence, and generates a mixed-precision bit-width scheduling table accordingly, allocating different inference bit widths for different network layers. This not only compresses invalid computational overhead under the condition of limited computing power at the edge, but also concentrates high-precision computation on key layers and conflict kernel graphs. For cases where the boundaries of the initial judgment result are unclear or the handling link is not closed, this invention further performs local high-precision deep inference, thereby improving the judgment stability in complex scenarios. At the same time, this invention also combines a continuous periodic locking mechanism and a safety judgment packet output method to ensure that the safety level result is consistent with the conflict chain, involved area, involved object, and unclosed link, facilitating subsequent linkage handling and result traceability. Attached Figure Description

[0021] Figure 1 This is a flowchart of the high-speed rail platform safety determination method based on hybrid precision reasoning of the present invention. Detailed Implementation

[0022] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0023] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0024] like Figure 1 As shown, the high-speed rail platform safety determination method based on hybrid precision inference includes the following steps:

[0025] Step 1: Obtain the multi-source semantic observation records for the current judgment period, standardize and encapsulate the multi-source semantic observation records, and determine the observation weights according to the preset source mapping table to obtain a set of safe observation elements;

[0026] Step 2: Construct a platform safety relationship diagram based on the set of safety observation elements and the preset relationship rule table;

[0027] Step 3: Based on the platform safety relationship diagram, the preset conflict matrix, and the preset closed link template, determine the relationship conflict density, closure residual, and cross-source divergence degree, and determine the reasoning difficulty level;

[0028] Step 4: Based on the inference difficulty level, pre-stored layer importance, and quantization sensitivity, generate a mixed precision bit-width scheduling table;

[0029] Step 5: Based on the platform safety relationship diagram and the mixed precision bit-width scheduling table, perform graph relationship reasoning to obtain the initial safety judgment vector and the initial judgment boundary quantity;

[0030] Step 6: Determine the final decision vector based on the initial decision boundary quantity, the relationship conflict density, and the closure residual; if the preset conditions are not met, extract the conflict kernel graph based on the platform safety relationship graph and perform high-precision deep reasoning to obtain the final decision vector;

[0031] Step 7: Based on the final decision vector, the relation conflict density, the closure residual, and the initial decision boundary quantity, determine the locking security level and output a security decision packet.

[0032] In one embodiment of the present invention, before constructing the platform safety relationship graph, the system first uniformly organizes the multi-source semantic observation records within the current judgment period to obtain a set of safety observation elements. Here, multi-source semantic observation records refer to the structured semantic results output by the front-end acquisition device or front-end processing unit, and do not directly use the original image frames or original underlying signals. A safety observation element refers to a standardized semantic unit after field merging of a single multi-source semantic observation record. Through the above processing, subsequent relationship modeling, difficulty determination, and hierarchical reasoning all revolve around the same data object, maintaining consistency in data entry points before and after.

[0033] In step 11, the system first acquires the multi-source semantic observation records output by the front-end acquisition device, and then filters out records belonging to the current judgment period according to the record timestamp, obtaining a subset of multi-source semantic observation records. The multi-source semantic observation records include object passage records, area dwell records, boundary crossing records, train process status records, equipment status records, alarm trigger records, manual confirmation records, and response completion records. The record timestamp refers to the time attribution information corresponding to a single record; the current judgment period refers to a single safety judgment time window preset by the system. Through the above processing, the range of records entering subsequent calculations in this period is fixed, and data from other periods will not be mixed in later.

[0034] It should be noted that if an alarm trigger record belongs to the current judgment period, while the corresponding confirmation record falls into the next judgment period, then the two will enter the processing chain of their respective periods and will not be cross-merged in this step.

[0035] In step 12, the system extracts object identifiers, region identifiers, action tags, role tags, judgment periods, and source categories from a subset of multi-source semantic observation records, and performs standardized encapsulation according to a unified field order to obtain security observation elements. The object identifier refers to the attribution marker of the observed object; the region identifier refers to the unified number of the spatial region within the platform; the action tag refers to the semantic action type corresponding to the record; the role tag refers to the object's identity category in the platform scene; and the source category refers to the type of pre-collection device or pre-processing path from which the record originates. In other words, records from different sources are written into data units with the same structure in this step, eliminating the need for separate field format processing later. For example, an object access record and a device status record may have different content formats before entering this step; after encapsulation, they are both converted into security observation elements with the same field order and can directly participate in the same relational reasoning process.

[0036] In step 13, the system unifies the object identifier, region identifier, action label, role label, and source category of the security observers. Then, based on the preset source mapping table, it retrieves the source credibility coefficient corresponding to the source category of each security observer and determines the source credibility coefficient as the observation weight. The observation weight is then associated with the corresponding security observer to obtain a set of security observers. The preset source mapping table refers to a pre-established correspondence table between source categories and source credibility coefficients; the source credibility coefficient refers to the quantified result of the credibility of different source categories in the current system; and the observation weight refers to the weight value bound to a single security observer. Through the above processing, the system first merges synonymous expressions into a unified definition, and then converts source differences into computable weight information. This ensures that nodes and edges in the subsequent graph structure have a consistent semantic basis and that different source results maintain a distinguishable contribution in subsequent reasoning.

[0037] In summary, this process first defines the period range, then standardizes the field structure, and subsequently normalizes the identifiers and assigns weights to the sources, thus obtaining a set of security observation elements. This not only organizes the multi-source inputs from high-speed rail platforms into unified semantic units, but also provides a stable data foundation for subsequent relational graph construction, relational conflict analysis, and hierarchical reasoning control.

[0038] In one embodiment of the present invention, after obtaining the set of security observation elements, the system continues to construct a platform security relationship graph based on the set of security observation elements and a preset relationship rule table. Here, the platform security relationship graph refers to data objects that organize the relationships between objects, areas, processes, equipment, alarms, and actions into a unified graph structure; the preset relationship rule table refers to a pre-defined set of rules for establishing node relationships. Since the previous part has already organized the multi-source semantic observation records into a set of security observation elements, this part does not return to the original record level for repeated processing, but directly completes node extraction, node merging, relationship edge construction, and edge weight determination around the security observation element set. This not only transforms scattered semantic units into a unified relationship structure but also provides direct input for subsequent relationship conflict analysis and hierarchical reasoning.

[0039] In step 21, the system first extracts node candidates from the safety observation element set. Specifically, it extracts object node candidates based on object identifiers, region node candidates based on region identifiers, process node candidates based on action tags and role tags, device node candidates based on action tags and source categories, and alarm node candidates and handling node candidates based on action tags, region identifiers, and their respective judgment cycles, thus obtaining a node candidate sequence. Here, node candidates refer to the original node entries that have not yet been merged; process nodes refer to nodes corresponding to train operation, platform operation, or scene process states; and handling nodes refer to nodes corresponding to handling steps such as manual confirmation and response completion. In other words, this step first breaks down the key fields in the safety observation elements and then maps them into different categories of node candidates according to the semantics of the fields.

[0040] For example, if the same boundary crosses a security observer, it can simultaneously contribute both object node candidates and region node candidates; if an alarm triggers a security observer, it can form an alarm node candidate. After this processing, subsequent mapping no longer relies on comparing each original security observer one by one, but instead directly expands on the sequence of node candidates.

[0041] In step 22, the system merges the candidate node sequences according to node categories to obtain a node set and further determines the observation weights corresponding to the nodes. Specifically, object node candidates are merged by object identifier, region node candidates by region identifier, process node candidates by action label and role label, equipment node candidates by action label and source category, and alarm node candidates and handling node candidates by action label, region identifier, and judgment cycle. After merging, the system sums the observation weights merged into the same node and divides by the number of merges to obtain the node's corresponding observation weight. Here, the node's corresponding observation weight refers to the weight value of the merged node in the current graph structure; the number of merges refers to the number of security observation elements merged into that node. This processing ensures that the same object, region, or event segment will not be duplicated in the graph, and the observation weights of each source semantic unit will naturally be passed to the node layer.

[0042] For example, if two security observers describe similar actions of the same object within the same area, but from different source categories, they will first be merged into the same node. Then, the observation weights of the two observers will jointly form the node's corresponding observation weight. This reduces duplicate nodes in the graph structure and ensures that the node weights remain consistent with the source credibility information from the previous step.

[0043] In step 23, the system establishes an edge set based on the node set and the preset relationship rule table, and determines the relationship establishment indicator for each edge. Specifically, a region affiliation edge is established between object nodes and region nodes; a process dependency edge is established between process nodes and object nodes / region nodes; a device support edge is established between device nodes and region nodes / process nodes; and a handling closure edge is established between alarm nodes and handling nodes. Simultaneously, time-consistent edges and conflict-rejecting edges are also established. Here, the relationship establishment indicator refers to the judgment result of whether a preset relationship is established; when the preset relationship is established, the relationship establishment indicator is 1; when the preset relationship is not established, the relationship establishment indicator is 0. A time-consistent edge is an edge established when nodes satisfy a preset sequential relationship; a conflict-rejecting edge is an edge established when there is a mutually exclusive relationship between nodes.

[0044] It should be noted that this step does not involve making indiscriminate connections to all nodes, but rather establishes edges strictly according to a pre-defined relationship rule table, ensuring that each edge has a clear origin and a clear relational meaning. As a result, subsequent relationship conflict analysis deals not with a loose set of nodes, but with a complete set of edges that possess attributes of belonging, dependence, support, closure, and conflict.

[0045] In step 24, the system continues to determine the edge weights of each edge based on the relationship establishment indicator and the observation weights of the nodes at both ends of each edge, thus constructing a platform safety relationship graph. Specifically, the observation weights of the nodes at both ends of each edge are first added together and then divided by two to obtain the average observation weight. Then, the average observation weight is multiplied by the relationship establishment indicator to obtain the edge weight. Here, the edge weight refers to the relationship strength of an edge in the current decision period; the average observation weight is the average of the weights of the nodes at both ends of the edge. If the relationship establishment indicator is zero, the corresponding edge weight is zero; if the relationship establishment indicator is valid, the edge weight changes with the weights of the nodes at both ends. In other words, this step further transmits the weight information from the node layer to the edge layer, so that the edge not only indicates whether a relationship exists but also the validity of that relationship in the current graph. This preserves the differences in node weights from the previous step and allows subsequent relationship reasoning to directly utilize the edge weights in the calculation.

[0046] In summary, this part first decomposes the set of safety observation elements into node candidates, then merges the node candidates into a node set, and establishes an edge set and edge weights under the constraints of a preset relational rule table, ultimately forming a platform safety relational graph. After this processing, the object activities, area states, process states, equipment states, and alarm handling links in the platform are no longer separate semantic fragments, but are organized into a unified relational structure. Subsequent calculations, whether of relation conflict density or hierarchical reasoning based on the graph structure, can directly revolve around the platform safety relational graph, with unified data entry points between steps and a smoother processing flow.

[0047] In one embodiment of the present invention, after constructing the platform safety relationship graph, the system continues to calculate the relationship conflict density, closure residual, and cross-source divergence degree around the graph, and determines the reasoning difficulty level based on these calculations. The preset closure link template refers to a template pre-defined for the link structure and sequence that alarm nodes, response nodes, and confirmation nodes should satisfy. Since the platform safety relationship graph has already been obtained in the previous part, this part does not re-extract nodes or reconstruct edges, but directly performs relationship analysis on the existing graph structure and the set of safety observation elements to obtain the difficulty control results required for subsequent mixed-precision bit-width scheduling.

[0048] In step 31, the system extracts the state labels of the nodes at both ends of each edge based on the platform safety relationship graph and retrieves the corresponding conflict output values ​​from the preset conflict matrix. Here, the state label refers to the unified state expression of the action label and role label corresponding to the node in the current graph structure; the preset conflict matrix is ​​a two-dimensional correspondence table pre-established for the state label combinations of the nodes at both ends of the relationship edge in the platform safety relationship graph. The row index of the preset conflict matrix corresponds to the state label of one end node, and the column index corresponds to the state label of the other end node. Each matrix element represents the conflict output value of the corresponding state label combination. The conflict output value is the lookup result of the two state label combinations in the preset conflict matrix; it takes a value greater than zero when the corresponding preset relationship is true (conflict), and zero when it is false. Subsequently, the system multiplies the edge weight of each edge by the corresponding conflict output value edge by edge to obtain the local conflict quantity of each edge. Then, it sums all the local conflict quantities and divides by the sum of all edge weights to obtain the relationship conflict density. If the sum of all edge weights is zero, the relationship conflict density is directly determined to be zero. In other words, this step doesn't simply count the number of conflicting edges, but simultaneously considers whether an edge is conflicting and its weight within the current graph. This preserves local conflicts while also consolidating the degree of conflict within the graph into a unified metric. For example, if two edges are both conflicting, but one edge has a higher weight, then that edge has a greater impact on the relationship conflict density.

[0049] In step 32, based on the platform safety relationship diagram and the preset closed link template, the system extracts the links that should be closed, consisting of alarm nodes, response nodes, and confirmation nodes, and counts the total number of links that should be closed and the number of unclosed links. Unclosed links here refer to links that lack response nodes, lack confirmation nodes, or whose node sequence does not meet the closure requirements. Subsequently, the system divides the number of unclosed links by the total number of links that should be closed to obtain the closure residual. If the total number of links that should be closed is zero, the closure residual is directly determined to be zero.

[0050] It's important to note that this step focuses not on whether a single alarm node occurs, but on whether a complete response chain is established after an alarm. This is because safety assessments at high-speed rail platforms depend not only on the presence of anomalies but also on the continuity of subsequent responses. For example, if an alarm is triggered, and one link has already responded and acknowledged while the other remains in the alarm stage, the latter will be included in the unclosed link statistics.

[0051] In step 33, the system, based on the set of security observers, filters out pairs of security observers with the same object identifier but different source categories, and counts the number of pairs where both action and role labels are consistent, as well as the total number of security observer pairs. Then, the number of pairs where both action and role labels are consistent is divided by the total number of security observer pairs to obtain the consistency ratio. This consistency ratio is then subtracted from one to obtain the cross-source divergence degree. Here, the consistency ratio represents the degree to which different sources give consistent judgments about the same object, while the cross-source divergence degree represents the degree of inconsistency between different sources. If the total number of security observer pairs is zero, the consistency ratio is determined to be zero, and the cross-source divergence degree is determined to be one. After this processing, the system can reflect both the conflict within the relationship graph and whether there are significant divergences between multi-source inputs. For example, if the same object is recorded by two sources simultaneously, and their action and role labels are consistent, the consistency ratio will increase; if their states are inconsistent, the cross-source divergence degree will increase.

[0052] In step 34, the system calculates and sums the weighted results based on preset weight coefficients corresponding to relation conflict density, closure residuals, and cross-source divergence, respectively, to obtain a reasoning difficulty score. Then, the reasoning difficulty score is further categorized according to a first preset difficulty threshold and a second preset difficulty threshold: a reasoning difficulty score below the first preset difficulty threshold is classified as low-difficulty reasoning; a reasoning difficulty score greater than or equal to the first preset difficulty threshold but lower than the second preset difficulty threshold is classified as medium-difficulty reasoning; and a reasoning difficulty score greater than or equal to the second preset difficulty threshold is classified as high-difficulty reasoning.

[0053] In other words, this step unifies and compresses the three indicators from different sources and with different meanings into a single level result that can directly participate in subsequent bit-width scheduling. This preserves the three types of information—relationship conflict, closed link, and cross-source consistency—while also bringing them to a unified inference control entry point.

[0054] Through the above steps, the system first extracts relationship conflict information from the platform security relationship graph, then extracts handling integrity information from the closed link, and simultaneously extracts cross-source consistency information from the security observation set. Finally, these three results are merged into an inference difficulty level. In this way, subsequent mixed-precision bit-width scheduling is not set independently of the scenario state, but is directly built upon the current platform relationship structure, handling link state, and multi-source consistency state. The steps are well-connected, and the data objects remain consistent. Subsequent steps only need to read the inference difficulty level to continue completing hierarchical inference control.

[0055] In one embodiment of the present invention, after determining the inference difficulty level, the system continues to generate a mixed-precision bit-width scheduling table based on the inference difficulty level and the layer importance and quantization sensitivity of each layer in the security decision graph network. Here, layer importance refers to a preset measure of the contribution of a certain network layer to the current security decision result; quantization sensitivity refers to a preset measure of the impact of a certain network layer on decision stability after the inference bit width is reduced; high bit-width reserved layers refer to network layers that need to maintain a high inference bit width during the current bit-width scheduling process; and the mixed-precision bit-width scheduling table refers to a structured result recording the layer number and corresponding inference bit width of each network layer.

[0056] In step 41, the system first reads the inference difficulty level and simultaneously reads the pre-stored layer importance and quantization sensitivity of each layer in the security decision graph network. Pre-storage here means that these parameters have been set before the start of the current decision cycle and saved one by one according to the network layer number. Subsequently, the system identifies network layers with a layer importance greater than or equal to a preset layer importance threshold, or a quantization sensitivity greater than or equal to a preset quantization sensitivity threshold, as high-bit-width reserved layers. In other words, this step first selects network layers that have a significant impact on the result or are sensitive to bit-width changes, preventing them from entering the subsequent low-bit-width allocation range. If a layer contributes significantly to the result, it will be reserved as a high-bit-width reserved layer even if its quantization sensitivity is low; conversely, if a layer, although not a major contributing layer, is prone to decision fluctuations if its bit-width is lowered, it will also be classified as a high-bit-width reserved layer. In this way, the network layers that truly need protection are fixed first, and the subsequent bit-width allocation will not lower the bit-width of all network layers simultaneously.

[0057] In step 42, based on the already determined high-bit-width reserved layers, the system allocates specific inference bit widths to each network layer. For high-bit-width reserved layers, the system directly allocates a 16-bit inference bit width; for network layers not determined as high-bit-width reserved layers, the allocation is based on the inference difficulty level. When the inference difficulty level is low, a 4-bit inference bit width is allocated; when the inference difficulty level is medium or high, an 8-bit inference bit width is allocated. Here, the high-bit-width reserved layers always maintain high precision, while other network layers switch between 4-bit and 8-bit inference widths according to the difficulty changes in the current decision cycle.

[0058] It's important to note that this step doesn't simply assign values ​​according to a uniform rule across the entire network. Instead, it first considers the network layer's own attributes, then the inference difficulty level of the current period, with both factors jointly determining the final bit width. For example, at a low inference difficulty level, a typical network layer might be assigned a four-bit inference width. However, if the same layer corresponds to a medium inference difficulty level in another period, it will be assigned an eight-bit inference width. This approach incorporates the difficulty differences resulting from changes in the current platform state into the network inference process while also stabilizing critical and sensitive layers, preventing arbitrary fluctuations across different periods.

[0059] In step 43, the system organizes the inference bit widths corresponding to each layer according to the layer sequence number of each layer in the security decision graph network, and records the layer sequence number and corresponding inference bit width of each layer to form a mixed-precision bit-width scheduling table. Here, the layer sequence number refers to the fixed sequential number of the network layer in the security decision graph network; recording the layer sequence number and corresponding inference bit width indicates that each layer has a unique position and a unique bit width value in the scheduling table. After this step, the mixed-precision bit-width scheduling table becomes the direct basis for the subsequent graph relation inference stage. Subsequent self-updating weights and relation update weights no longer require separate determination of which bit width to use, but instead simply retrieve the corresponding result from the scheduling table according to the layer sequence number. This ensures that the weight conversion caliber of each layer is consistent and that the entire inference process maintains a uniform bit width configuration within the same decision cycle.

[0060] Through the steps described above, this section effectively transcribes the inference difficulty level into the inference bit width of each network layer. The system first reads inference difficulty level, layer importance, and quantization sensitivity to determine which layers must maintain a high bit width. Then, it allocates four- or eight-bit inference bit widths to the remaining network layers according to the difficulty level of the current cycle. Finally, it organizes the allocation results of all layers into a mixed-precision bit width scheduling table. In this way, subsequent graph relation inference no longer needs to temporarily determine which layer should use which bit width; instead, it executes layer by layer directly according to the scheduling table.

[0061] In one embodiment of the present invention, after generating the mixed-precision bit-width scheduling table, the system continues to perform graph relation reasoning based on the platform safety relationship graph and the mixed-precision bit-width scheduling table to obtain the initial safety judgment vector and the initial judgment boundary quantity. Here, graph relation reasoning refers to the reasoning process of updating the representation of each node layer by layer around the nodes, edges, and edge weights in the platform safety relationship graph; the initial safety judgment vector refers to the vectorized output result corresponding to each safety level in the current judgment period; and the initial judgment boundary quantity refers to the difference between the largest and second-largest components in the initial safety judgment vector.

[0062] In step 51, the system first determines the initial representation of each node based on the node category, the observation weight corresponding to the node, the edges connected to each node, and the edge weights corresponding to those edges in the platform safety relationship graph. The node categories here include object nodes, region nodes, process nodes, equipment nodes, alarm nodes, and handling nodes; the observation weight corresponding to the node refers to the node weight formed after the node merging in the previous step; the edges connected to each node and the edge weights corresponding to those edges represent the association range and association strength of that node in the current relationship graph. In other words, this step first transforms the discrete graph structure into a node representation that can directly participate in subsequent layer-by-layer updates. After this processing, the subsequent graph relationship reasoning no longer deals with scattered nodes and edges, but with initial node representations that have a unified expression form. For example, even among object nodes, different object nodes will have different initial representations due to different associated regions and associated edge weights.

[0063] In step 52, the system reads the corresponding inference bit width according to the layer number of each layer in the mixed precision bit width scheduling table, and converts the self-update weight and relation update weight of that layer into the quantization format corresponding to the inference bit width. Here, the self-update weight refers to the weight used by a node when updating in this layer based only on its own representation in the previous layer; the relation update weight refers to the weight used by a node when absorbing relation information of neighboring nodes in this layer.

[0064] It should be noted that this step does not recalculate the bit width; it simply applies the mixed-precision bit-width scheduling table generated in the previous part to the weight quantization layer by layer. This ensures that different layers maintain the already determined bit-width configuration and that the subsequent layer-by-layer inference process remains consistent with the difficulty level of the current cycle.

[0065] In step 53, the system uses the initial representation of each node as the first layer input. Based on the quantized self-update weight, the quantized relation update weight, and the edge weight of each edge, it calculates the self-update result of each node and the aggregation result of adjacent nodes layer by layer. After adding the self-update result and the aggregation result of adjacent nodes, it performs activation processing to obtain the updated representation of each node in this layer. The updated representation of each node in this layer is then used as the input of the next layer.

[0066] The formula for updating each node is as follows: ,in, Represents a node In the The update representation of the layer, Indicates the preset activation function. Indicates the first Layer self-updating weights Represents a node In the Layer representation, Represents a node The set of adjacent nodes, Represents a node With nodes The edge weight between them Indicates the first Layer Relationship Types The corresponding relationship is updated with weights. Indicates adjacent nodes In the Layer representation.

[0067] In other words, the system first uses the first Layer self-updating weights for nodes In the The layer representation is self-updated and then the nodes are processed. Each of the adjacent nodes reads its value in the 1st... The layer representation is then aggregated by combining the edge weights of the corresponding edges and the relation update weights corresponding to the relation types. The self-update results are then merged with the adjacent node aggregation results, and finally processed by a preset activation function to obtain the node. In the The updated representation of each node in this layer continues to serve as input for the next layer. After all network layers have been updated, the initial security decision vector is read out based on the updated representation of each node in the final layer.

[0068] The adjacency node aggregation result here refers to the summary result of the relational information absorbed by the current node from its connected adjacent nodes; the activation process refers to the operation of a preset activation function on the update result to form a new node representation.

[0069] In other words, this step does not yield a decision all at once, but rather allows the node representation to be progressively transmitted and updated along the network hierarchy. In this way, the edge weight information in the platform security relationship graph and the bit width configuration given in the previous section will work together in the same inference link. For example, if a node is connected to multiple high-weight adjacent nodes, the aggregation result of its adjacent nodes will be more fully incorporated into the current layer's updated representation.

[0070] In step 54, the system extracts the maximum component from all components of the initial security decision vector and the second largest component from the remaining components. The maximum component is then subtracted from the second largest component to obtain the initial decision boundary quantity. Here, the maximum component represents the security level output with the highest degree of support in the current initial security decision vector, and the second largest component represents the other security level output with the highest degree of support besides the maximum component. This is straightforward: if the difference between the maximum and second largest components is large, it indicates that the current initial decision result is more clearly distinguishable; if they are close, it indicates that the current result still needs further correction. Subsequent steps, based on this initial decision boundary quantity and combined with the relational conflict density and closure residuals, determine whether to proceed to high-precision deep inference using the conflict kernel graph.

[0071] Through the above steps, this section completes the transformation from the platform safety relationship diagram and the mixed-precision bit-width scheduling table to the initial safety judgment result. The system first converts the graph structure information into initial node representations, then implements the hierarchical bit-width configuration into the weight quantization of each layer, then updates the node representations layer by layer, and finally reads out the initial safety judgment vector and further obtains the initial judgment boundary quantity. This not only preserves the relationship structure information formed earlier, but also truly implements the bit-width scheduling result of the current period into the inference execution stage. Subsequent steps then continue to perform local high-precision corrections around the initial judgment boundary quantity and the initial safety judgment vector, resulting in a smoother transition between steps.

[0072] In one embodiment of the present invention, after obtaining the initial safety judgment vector, initial judgment boundary quantity, relationship conflict density, and closure residual, the system continues to perform the final judgment vector determination process. This part no longer reconstructs the platform safety relationship graph or recalculates the reasoning difficulty level. Instead, it verifies the initial judgment result of the current cycle based on the question of whether to enter a local high-precision re-judgment. Here, the initial judgment boundary quantity refers to the difference between the largest and second-largest components in the initial safety judgment vector; the direct output judgment result refers to the system's judgment result based on preset conditions regarding whether to directly adopt the initial safety judgment vector; the conflict kernel graph refers to the local relationship subgraph extracted from the platform safety relationship graph; and the final judgment vector refers to the unique judgment result directly called upon in subsequent safety level calculations of this cycle.

[0073] In step 61, the system, based on the initial decision boundary quantity, relationship conflict density, and closure residual, determines whether the following conditions are simultaneously met: the initial decision boundary quantity is greater than or equal to a preset decision boundary quantity threshold, the relationship conflict density is less than or equal to a preset conflict density threshold, and the closure residual is zero. This results in a direct output decision result. The decision logic here does not simply consider whether the initial safe decision vector is concentrated; it also considers whether there are still significant conflicts in the current relationship structure and whether the handling link has been closed. If all three conditions are met simultaneously, it indicates that the initial decision result of the current cycle has a basis for direct output; if any condition is not met, further local high-precision re-decision is required. This processing prevents the system from pulling all cycles into the same depth of computation, instead performing a clear entry screening first.

[0074] In step 62, when the direct output judgment result is not satisfied, the system extracts the nodes at both ends of the edges with local conflict quantities greater than zero, nodes located in unclosed links, nodes sharing object identifiers, area identifiers, or process states with unclosed links, and their associated edges from the platform safety relationship graph to obtain a conflict kernel graph. Here, shared object identifiers, area identifiers, or process states indicate that although the node is not directly located in an unclosed link, it has a clear propagation association with the unclosed link and therefore still needs to be included in the subsequent processing scope. In other words, this step does not recalculate the entire platform safety relationship graph, but rather extracts the local structures that are truly related to the current instability judgment. This preserves the conflict-concentrated areas and avoids subsequent high-bit-width calculations falling on irrelevant nodes.

[0075] In step 63, the system performs high-precision deep inference based on the conflict kernel graph and the mixed-precision bit-width scheduling table. It reads the set of adjacent nodes for each node, combines the deep update result corresponding to each adjacent node with its corresponding conflict focusing weight, and sums the combined results to obtain the deep security decision vector. Here, the conflict focusing weight refers to the degree of focusing assigned to different adjacency relationships within the conflict kernel graph; the deep update result refers to the update result formed after adjacent nodes participate in local deep inference under high bit-width constraints.

[0076] It should be noted that this step performs re-judgment within the conflict kernel map, not by re-pushing the entire map. This allows us to continue utilizing the previously formed mixed-precision bit-width scheduling table, while also enabling high-bit-width resources to be more concentrated on the truly disputed local areas in the current cycle.

[0077] In step 64, the system determines the final decision vector based on the direct output decision result. When the direct output decision result is satisfied, the initial security decision vector is determined as the final decision vector; when the direct output decision result is not satisfied, the deep security decision vector is determined as the final decision vector. In this way, regardless of whether the current cycle enters the high-precision deep inference of the conflict kernel graph, it will converge to a unique final decision vector. Subsequent security risk value calculation, security level determination, and security decision packet generation are all carried out around this final decision vector.

[0078] From the overall processing sequence, this part first determines whether the initial results can be directly used, then extracts the conflict kernel graph and performs high-precision deep inference when necessary, and finally outputs the final decision vector. In this way, the results of the initial graph relationship inference can be retained, and when the relationship conflict density is high, the closure residual is not cleared to zero, or the initial decision boundary quantity is insufficient, the processing focus can be narrowed to the local conflict area in a timely manner. The steps are naturally connected and the data objects are kept consistent.

[0079] In one embodiment of the present invention, after obtaining the final decision vector, the system further combines the relation conflict density, closure residual, and initial decision boundary quantity to uniformly converge the output results of the current decision cycle, determine the locking security level, and generate a security decision package. Here, the final decision vector refers to the final vectorized decision result formed after direct output judgment or high-precision deep inference of the conflict kernel map; the locking security level refers to the final level result determined under the joint constraints of the current cycle risk state and the previous cycle level state; and the security decision package refers to the structured set of output results obtained by organizing the locking security level around the current cycle.

[0080] In step 71, the system first extracts the component corresponding to the security category from the final judgment vector, and then calculates the weighted results and sums them according to the preset fusion coefficients corresponding to the component corresponding to the security category minus one, the relation conflict density, the closure residual and the initial judgment boundary quantity, respectively, to obtain the security risk value.

[0081] The security risk value can be expressed as: ,in, Indicates the first The security risk value corresponding to each judgment period Indicates the first The security category component in the final decision vector of each decision cycle. Indicates the first The relationship conflict density of each judgment period Indicates the first The closed residual of each judgment period, Indicates the first The initial decision boundary quantity for each decision cycle. This represents the preset fusion coefficient corresponding to the component corresponding to the security category minus one. This represents the preset fusion coefficient corresponding to the relation conflict density. This represents the preset fusion coefficient corresponding to the closed residual. This represents the preset fusion coefficient corresponding to one minus the initial judgment boundary value. The security risk value represents the comprehensive risk level resulting from the combined effects of the final judgment result, relationship conflict state, closed link state, and initial judgment boundary state in the current judgment cycle.

[0082] In step 72, the system determines the original security level based on the security risk value and the first, second, and third preset security level thresholds. The original security level includes a security level, a concern level, an intervention level, and an emergency level. These three preset security level thresholds are arranged from low to high, and the system determines the original security level based on the range in which the security risk value falls. In this way, the continuous risk quantity obtained in the previous step is converted into discrete level results, allowing subsequent continuous periodic processing to proceed around a unified level caliber, without the need for direct cross-period comparisons of continuous values.

[0083] In step 73, the system continues to determine the locking security level based on the original security level, the closure residual, the previous cycle's locking security level, the current cycle's continuous low-risk maintenance count, and the preset continuous release threshold. Specifically, when the closure residual is greater than zero, the higher of the original security level and the previous cycle's locking security level is used; when the closure residual is equal to zero and the current cycle's continuous low-risk maintenance count is greater than or equal to the preset continuous release threshold, the original security level is used; when the closure residual is equal to zero and the current cycle's continuous low-risk maintenance count is less than the preset continuous release threshold, the previous cycle's locking security level is used. Here, the current cycle's continuous low-risk maintenance count refers to the number of consecutive low-risk cycles recorded by the system; the preset continuous release threshold refers to the number of consecutive low-risk cycles required before the locking level is allowed to fall back. This direct approach means that the system considers both the current cycle's risk status and the previous cycle's locking results and stability, thus preventing short-term fluctuations from directly lowering the final output level.

[0084] In step 74, the system generates a security determination package based on the locked security level, the dominant conflict chain corresponding to the final determination vector, the set of involved areas, the set of involved objects, the set of unclosed links, the mixed-precision bit-width scheduling table, and the inference path identifier. Here, the dominant conflict chain refers to the conflict propagation link that plays a major role in the formation of the final determination in the current cycle; the inference path identifier refers to the result mark of adopting the direct output path or the high-precision deep inference path in the current cycle. At this step, the system outputs not only a single level, but also organizes the level result and the key supporting information for forming this level together. This is not only convenient for subsequent display and recording, but also convenient for subsequent linkage processing to directly carry out around the same object.

[0085] Generally speaking, in this part, the final determination vector and structural indicators are first converged into a security risk value, then the security risk value is divided into the original security level, then the closed residual and continuous cycle conditions are introduced to determine the locked security level, and finally the locked security level is encapsulated together with the dominant conflict chain, the set of involved areas, the set of involved objects, the set of unclosed links, the mixed-precision bit-width scheduling table, and the inference path identifier into a security determination package. This can not only unify the results formed in the previous multiple steps into one output entry, but also ensure that the processing logic of the high-speed rail platform security determination results is consistent at both the current cycle and the continuous cycle levels.

[0086] The present invention also provides a high-speed rail platform security determination system based on mixed-precision inference, including:

[0087] A security observation element generation module, configured to obtain multi-source semantic observation records of the current determination cycle, perform standardized encapsulation on the multi-source semantic observation records, and determine observation weights according to a preset source mapping table to obtain a set of security observation elements;

[0088] A platform relationship graph construction module, configured to construct a platform security relationship graph based on the set of security observation elements and a preset relationship rule table;

[0089] A relationship index determination module, configured to determine relationship conflict density, closed residual, and cross-source divergence degree based on the platform security relationship graph, a preset conflict matrix, and a preset closed link template, and determine the inference difficulty level;

[0090] A bit-width scheduling generation module, configured to generate a mixed-precision bit-width scheduling table based on the inference difficulty level, the pre-stored layer importance, and quantization sensitivity;

[0091] A graph relationship reasoning module, configured to perform graph relationship reasoning based on the platform security relationship graph and the mixed-precision bit-width scheduling table to obtain an initial security determination vector and an initial determination boundary quantity;

[0092] The conflict kernel deep reasoning module is used to determine the final decision vector based on the initial decision boundary quantity, the relationship conflict density, and the closure residual; when the preset conditions are not met, it extracts the conflict kernel graph based on the platform safety relationship graph and performs high-precision deep reasoning to obtain the final decision vector.

[0093] The security level locking module is used to determine the locking security level based on the final decision vector, the relation conflict density, the closure residual, and the initial decision boundary quantity, and output a security decision packet.

[0094] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0095] The embodiments of the present invention have been described above, but the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of the present embodiments, all of which are within the protection scope of the present embodiments.

Claims

1. A method for determining the safety of high-speed railway platforms based on hybrid precision reasoning, characterized in that: Includes the following steps: Step 1: Obtain the multi-source semantic observation records for the current judgment period, standardize and encapsulate the multi-source semantic observation records, and determine the observation weights according to the preset source mapping table to obtain a set of safe observation elements; Step 2: Construct a platform safety relationship diagram based on the set of safety observation elements and the preset relationship rule table; Step 3: Based on the platform safety relationship diagram, the preset conflict matrix, and the preset closed link template, determine the relationship conflict density, closure residual, and cross-source divergence degree, and determine the reasoning difficulty level, including: Step 31: Based on the platform safety relationship diagram, extract the status labels of the nodes at both ends of each side and look up the corresponding conflict output values ​​in the preset conflict matrix; multiply the edge weight of each side by the corresponding conflict output value to obtain the local conflict quantity; sum the local conflict quantities and divide by the sum of the edge weights of each side to obtain the relationship conflict density. Step 32: Based on the platform safety relationship diagram and the preset closed link template, extract the links that should be closed, which consist of alarm nodes, response nodes and confirmation nodes. Count the total number of links that should be closed and the number of unclosed links that are missing response nodes, missing confirmation nodes or whose node order does not meet the closure requirements. Divide the number of unclosed links by the total number of links that should be closed to obtain the closure residual. Step 33: Based on the set of security observers, filter out pairs of security observers with the same object identifier but different source categories, and count the number of pairs with the same action tag and role tag as well as the total number of security observer pairs; divide the number of pairs with the same action tag and role tag by the total number of security observer pairs to obtain the consistency ratio; subtract the consistency ratio from one to obtain the cross-source divergence degree. Step 34: Based on the preset weight coefficients corresponding to relation conflict density, closure residual and cross-source divergence degree respectively, calculate the weighted results and sum them to obtain the reasoning difficulty score. Divide the reasoning difficulty score according to the first preset difficulty threshold and the second preset difficulty threshold. When the reasoning difficulty score is lower than the first preset difficulty threshold, it is a low difficulty reasoning level. When the reasoning difficulty score is greater than or equal to the first preset difficulty threshold and lower than the second preset difficulty threshold, it is a medium difficulty reasoning level. When the reasoning difficulty score is greater than or equal to the second preset difficulty threshold, it is a high difficulty reasoning level. Step 4: Based on the inference difficulty level, pre-stored layer importance, and quantization sensitivity, generate a mixed precision bit-width scheduling table; Step 5: Based on the platform safety relationship diagram and the mixed precision bit-width scheduling table, perform graph relationship reasoning to obtain the initial safety judgment vector and the initial judgment boundary quantity; Step 6: Determine the final decision vector based on the initial decision boundary quantity, the relationship conflict density, and the closure residual; if the preset conditions are not met, extract the conflict kernel graph based on the platform safety relationship graph and perform high-precision deep reasoning to obtain the final decision vector; Step 7: Based on the final decision vector, the relation conflict density, the closure residual, and the initial decision boundary quantity, determine the locking security level and output a security decision packet.

2. The high-speed railway platform safety determination method based on hybrid precision reasoning according to claim 1, characterized in that, Obtain multi-source semantic observation records for the current judgment period, standardize and encapsulate the multi-source semantic observation records, and determine the observation weights according to a preset source mapping table to obtain a set of secure observation elements, including: Step 11: Obtain the multi-source semantic observation records output by the front-end acquisition device, and filter out the multi-source semantic observation record subset corresponding to the current judgment period according to the record time marker. The multi-source semantic observation records include object passage records, area stay records, boundary crossing records, train process status records, equipment status records, alarm trigger records, manual confirmation records, and response completion records. Step 12: Extract object identifier, region identifier, action tag, role tag, judgment period and source category for each record in the multi-source semantic observation record subset, and encapsulate them in a standardized manner according to a unified field order to obtain security observation elements; Step 13: Unify the object identifier, region identifier, action label, role label, and source category of the security observation element. According to the preset source mapping table, retrieve the source credibility coefficient corresponding to the source category of each security observation element, determine the source credibility coefficient as the observation weight, and associate the observation weight with the corresponding security observation element to obtain the security observation element set.

3. The high-speed railway platform safety determination method based on hybrid precision reasoning according to claim 1, characterized in that, Based on the aforementioned set of security observation elements and the preset relationship rule table, a platform security relationship diagram is constructed, including: Step 21: Extract object node candidates and area node candidates based on the object identifier and area identifier in the security observation element set; extract process node candidates, equipment node candidates, alarm node candidates and handling node candidates based on action tag, role tag, judgment period and source category to obtain a node candidate sequence. Step 22: Merge the candidate node sequences according to node category. Specifically, candidate object nodes are merged by object identifier, candidate region nodes are merged by region identifier, candidate process nodes are merged by action label and role label, candidate equipment nodes are merged by action label and source category, and candidate alarm nodes and candidate handling nodes are merged by action label, region identifier and judgment cycle to obtain a node set. Sum the observation weights of each merged node and divide by the number of merges to obtain the observation weight corresponding to the node. Step 23: Based on the node set and the preset relationship rule table, establish a region affiliation edge between the object node and the region node, establish a process dependency edge between the process node and the object node and the region node, establish a device support edge between the device node and the region node and the process node, establish a handling closure edge between the alarm node and the handling node, and establish a time sequence consistent edge and a conflict exclusion edge to obtain the edge set, and determine the relationship establishment indicator corresponding to each edge. Step 24: Based on the relationship establishment indicator of each edge and the observation weight of the node at both ends of each edge, add the observation weights of the node at both ends of each edge and divide by two to obtain the average observation weight; multiply the average observation weight of the node by the relationship establishment indicator to obtain the edge weight of each edge, and construct the platform safety relationship graph.

4. The high-speed railway platform safety determination method based on hybrid precision reasoning according to claim 1, characterized in that, Based on the inference difficulty level, pre-stored layer importance, and quantization sensitivity, a mixed-precision bit-width scheduling table is generated, including: Step 41: Read the inference difficulty level and the layer importance and quantization sensitivity of each layer of the pre-stored security decision graph network, and determine the network layers with layer importance greater than or equal to the preset layer importance threshold or quantization sensitivity greater than or equal to the preset quantization sensitivity threshold as high bit width reserved layers. Step 42: Allocate 16 bits of inference width to the high bit width reserved layer; for network layers that are not determined to be high bit width reserved layers, allocate bit width according to the inference difficulty level. Specifically, allocate 4 bits of inference width when the inference difficulty level is low, and allocate 8 bits of inference width when the inference difficulty level is medium or high. Step 43: Organize the inference bit widths corresponding to each layer according to the layer number of each layer in the security decision graph network, and record the layer number and corresponding inference bit width of each layer to obtain the mixed precision bit width scheduling table.

5. The high-speed railway platform safety determination method based on hybrid precision reasoning according to claim 1, characterized in that, Based on the platform safety relationship diagram and the mixed-precision bit-width scheduling table execution graph relationship reasoning, the initial safety decision vector and initial decision boundary quantities are obtained, including: Step 51: Based on the node category, the observation weight corresponding to the node, the edge connected to each node and the edge weight corresponding to the edge in the platform safety relationship graph, determine the initial representation of each node; Step 52: Read the corresponding inference bit width according to the layer number of each layer in the mixed precision bit width scheduling table, and convert the self-update weight and relation update weight of the layer into the quantization format corresponding to the inference bit width; Step 53: Using the initial representation of each node as the input of the first layer, calculate the self-update result and the adjacent node aggregation result of each node layer by layer based on the quantized self-update weight, the quantized relation update weight and the edge weight of each edge. Add the self-update result and the adjacent node aggregation result and perform activation processing to obtain the updated representation of each node in this layer. Use the updated representation of each node in this layer as the input of the next layer. Read out the initial security decision vector based on the updated representation of each node in the final layer. Step 54: Extract the largest component from all components of the initial security decision vector, and extract the second largest component from the remaining components excluding the largest component. Subtract the second largest component from the largest component to obtain the initial decision boundary quantity.

6. The high-speed railway platform safety determination method based on hybrid precision reasoning according to claim 1, characterized in that, The final decision vector is determined based on the initial decision boundary quantity, the relationship conflict density, and the closure residual. When the preset conditions are not met, a conflict kernel graph is extracted based on the platform safety relationship graph, and high-precision deep reasoning is performed to obtain the final judgment vector, including: Step 61: Based on the initial decision boundary quantity, the relationship conflict density, and the closure residual, determine whether the initial decision boundary quantity is greater than or equal to a preset decision boundary quantity threshold, the relationship conflict density is less than or equal to a preset conflict density threshold, and the closure residual is zero at the same time, and obtain the direct output decision result. Step 62: When the direct output judgment result is not satisfied, extract the nodes at both ends of the edge with local conflict volume greater than zero, the nodes located in the unclosed link, the nodes that share the object identifier, area identifier or process status with the unclosed link and their associated edges from the platform safety relationship graph to obtain the conflict kernel graph; Step 63: Perform high-precision deep inference based on the conflict kernel graph and the mixed precision bit-width scheduling table, read the set of adjacent nodes of each node, combine the deep update result corresponding to each adjacent node with the corresponding conflict focusing weight, and sum the combination results to obtain the deep security judgment vector. Step 64: Based on the direct output determination result, if the direct output determination result is satisfied, the initial security determination vector is determined as the final determination vector; if the direct output determination result is not satisfied, the deep security determination vector is determined as the final determination vector.

7. The high-speed railway platform safety determination method based on hybrid precision reasoning according to claim 1, characterized in that, Based on the final decision vector, the relation conflict density, the closure residual, and the initial decision boundary quantity, the locking security level is determined, and a security decision packet is output, including: Step 71: Extract the component corresponding to the safety category in the final judgment vector. Based on the preset fusion coefficients corresponding to the component corresponding to the safety category minus one, the relation conflict density, the closure residual, and the initial judgment boundary quantity, calculate the weighted results and sum them to obtain the safety risk value. Step 72: Based on the safety risk value and the first preset safety level threshold, the second preset safety level threshold, and the third preset safety level threshold, determine the original safety level. The original safety level includes: safety level, attention level, intervention level, and emergency level. Step 73: Determine the locking security level based on the original security level, closure residual, previous cycle locking security level, current cycle continuous low-risk maintenance count, and preset continuous release threshold. Specifically, when the closure residual is greater than zero, the higher of the original security level and the previous cycle locking security level is used; when the closure residual is equal to zero and the current cycle continuous low-risk maintenance count is greater than or equal to the preset continuous release threshold, the original security level is used; when the closure residual is equal to zero and the current cycle continuous low-risk maintenance count is less than the preset continuous release threshold, the previous cycle locking security level is used. Step 74: Based on the locked security level, the dominant conflict chain corresponding to the final decision vector, the set of involved regions, the set of involved objects, the set of unclosed links, the mixed precision bit-width scheduling table, and the inference path identifier, generate a security decision package.

8. A high-speed railway platform safety determination system based on hybrid precision inference, characterized in that, The high-speed railway platform safety determination method based on hybrid precision reasoning as described in any one of claims 1-7 includes: The security observation element generation module is used to obtain multi-source semantic observation records for the current judgment period, standardize and encapsulate the multi-source semantic observation records, and determine the observation weights according to a preset source mapping table to obtain a security observation element set. The platform relationship diagram construction module is used to construct a platform safety relationship diagram based on the set of safety observation elements and the preset relationship rule table. The relationship index determination module is used to determine the relationship conflict density, closure residual and cross-source divergence degree based on the platform safety relationship diagram, the preset conflict matrix and the preset closed link template, and to determine the reasoning difficulty level. The bit-width scheduling generation module is used to generate a mixed-precision bit-width scheduling table based on the inference difficulty level, pre-stored layer importance, and quantization sensitivity. The graph relationship reasoning module is used to perform graph relationship reasoning based on the platform safety relationship graph and the mixed precision bit-width scheduling table to obtain the initial safety judgment vector and the initial judgment boundary quantity. The conflict kernel deep reasoning module is used to determine the final decision vector based on the initial decision boundary quantity, the relationship conflict density, and the closure residual; when the preset conditions are not met, it extracts the conflict kernel graph based on the platform safety relationship graph and performs high-precision deep reasoning to obtain the final decision vector. The security level locking module is used to determine the locking security level based on the final decision vector, the relation conflict density, the closure residual, and the initial decision boundary quantity, and output a security decision packet.

Images