A coal mine terminal risk scoring method based on a Bayesian network

Abstract

This invention discloses a method for coal mine terminal risk scoring based on Bayesian networks, comprising the following steps: Step 1: Collecting coal mine terminal risk data to form a terminal risk dataset; Step 2: Extracting risk variable nodes, barrier variable nodes, and terminal scoring nodes, constructing a Bayesian network, and forming a node conditional probability table; Step 3: Performing deontization, triangulation, and clique partitioning to form a connection tree and a partition set; Step 4: Calculating the probability estimate of clique nodes and the barrier state description term; Step 5: Constructing propagation messages using an improved Shafer-Shenoy algorithm to form a penetrating residual component; Step 6: Performing state enhancement marginalization to form a partition set propagation message; Step 7: Forming the terminal residual posterior risk distribution; Step 8: Forming the coal mine terminal risk scoring result. This invention improves the ability to characterize coal mine terminal residual risk and the reliability of risk scoring.

Description

Technical Field

[0001] This invention relates to the field of coal mine safety risk assessment technology, and in particular to a coal mine terminal risk scoring method based on Bayesian networks. Background Technology

[0002] With the increasing demand for intelligent coal mine construction and refined management of terminal risks, modeling and risk scoring techniques for the correlation between coal mine terminal operating status, barrier status, and risk outcomes have received widespread attention. Existing coal mine terminal risk assessment methods mainly rely on static threshold determination, empirical rule matching, or ordinary Bayesian network inference for risk identification, but these methods generally suffer from the following problems in practical applications:

[0003] The collected terminal risk data is complex in origin, lacks a unified time base, and contains missing and abnormal status records, resulting in unstable temporal correlations between risk status, barrier status, and scoring status, which affects the accuracy of node conditional probability statistics. Traditional Bayesian networks often directly output risk results using conventional inference methods after network construction, making it difficult to reflect the constraint effect of barrier status on risk propagation paths and intensity during connection tree propagation, leading to discrepancies between terminal risk results and actual protection conditions. Existing risk scoring methods typically focus on posterior probability output or level classification, lacking residual processing and state enhancement marginalization mechanisms for messages propagated through the separation set, making it difficult to accurately characterize the residual risk distribution after barrier penetration, thus affecting the accuracy and applicability of coal mine terminal risk scoring results.

[0004] Therefore, how to provide a Bayesian network-based method for risk assessment in coal mines is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a risk scoring method for coal mine terminals based on Bayesian networks. This invention employs an improved Shafer-Shenoy algorithm, which performs residual decomposition and state enhancement marginalization processing by propagating messages around the partition set in the connection tree. This enables the construction of the residual posterior risk distribution of coal mine terminals and the generation of risk scoring results. It has the advantages of accurate risk propagation characterization, sufficient manifestation of barrier effects, and high reliability of scoring results.

[0006] A method for scoring terminal risks in coal mines based on Bayesian networks according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Collect coal mine terminal risk data, align execution time and associate terminal identifiers to form a terminal risk dataset;

[0008] Step 2: Extract risk variable nodes, barrier variable nodes, and terminal scoring nodes from the terminal risk dataset, construct a Bayesian network, and statistically analyze the node state relationships to form a node conditional probability table;

[0009] Step 3: Perform deontization, triangulation, and clique partitioning on the Bayesian network to form a connection tree, and form a partition set by the intersection variables of adjacent clique nodes;

[0010] Step 4: Calculate the probability estimate of the cluster nodes based on the node conditional probability table, and generate barrier state description items based on the state of the barrier variable nodes in the partition set;

[0011] Step 5: The improved Shafer-Shenoy algorithm is used to construct propagation messages between adjacent clique nodes. At the split set, the barrier state description term is used to perform residual decomposition on the propagation messages to form penetrating residual components.

[0012] Step 6: Perform state enhancement marginalization on the penetrating residual component within the delimiter variable domain, form a delimiter propagation message, and pass the delimiter propagation message to the adjacent clique node;

[0013] Step 7: Aggregate the propagation messages of the segmented set and the probability estimates of the cluster nodes where the terminal scoring node is located, perform combined calculations, and form the terminal residual posterior risk distribution;

[0014] Step 8: Perform score mapping and level classification on the distribution of terminal residual posterior risk to form the coal mine terminal risk score result.

[0015] Optionally, step one specifically includes:

[0016] The risk records generated by the coal mine terminal within a preset collection period are collected, and the terminal identifier, collection timestamp, status marker and numerical field are extracted from the risk records to form the original risk record set.

[0017] Time mapping is performed on the collection timestamps in the original risk record set according to a unified time benchmark, and the original risk record set is arranged in order according to the time mapping results to form a time-aligned record set;

[0018] The time-aligned record set is associated and merged according to the terminal identifier, and the merged records are then filled with missing information, corrected for anomalies, and standardized in terms of fields to form a standardized risk record set.

[0019] The risk records are arranged in order of collection timestamps to form a terminal risk dataset.

[0020] Optionally, step two specifically involves:

[0021] State markers and numerical fields are extracted from the terminal risk dataset. State markers and numerical fields that reflect changes in risk status are categorized to form risk status items. State markers and numerical fields that reflect changes in barrier status are categorized to form barrier status items. State markers and numerical fields that reflect changes in terminal risk scoring status are categorized to form scoring status items, thus obtaining the node extraction record set.

[0022] Risk variable nodes are established based on risk status items extracted from the record set of nodes; barrier variable nodes are established based on barrier status items extracted from the record set of nodes; and terminal scoring nodes are established based on scoring status items extracted from the record set of nodes, thus forming a node set.

[0023] The node states in the node set are paired and statistically analyzed according to the collection timestamp order to determine the state association between risk variable nodes and barrier variable nodes, the state association between risk variable nodes and terminal scoring nodes, and the state association between barrier variable nodes and terminal scoring nodes, forming a node relationship record set.

[0024] A Bayesian network is constructed by connecting risk variable nodes, barrier variable nodes, and terminal scoring nodes based on the node relationship record set. The probability distribution of risk variable nodes, barrier variable nodes, and terminal scoring nodes under conditional states is statistically analyzed based on the node relationship record set to form a node conditional probability table.

[0025] Optionally, step three specifically includes:

[0026] Extract risk variable nodes, barrier variable nodes, terminal scoring nodes, and node connection relationships from the Bayesian network. Convert the directed connections in the node connection relationships into undirected connections, and supplement undirected connections for nodes that all point to the same node to form a deontological network.

[0027] In the deontological network, nodes are selected for elimination in order of increasing number of node connections. For each node to be eliminated, connecting edges are added to the adjacent nodes to form a complete connection. After the connecting edges are added, the node to be eliminated is deleted, thus forming a triangulated network.

[0028] In the triangulated network, retrieve the node groups where there is a connecting edge between any two nodes, merge the node groups to form clique nodes, and form clique partitioning results according to the node group affiliation relationship;

[0029] Based on the clique partitioning results, calculate the intersection variables and the number of intersection variables between clique nodes. Select clique node connection relationships in descending order of the number of intersection variables. Delete clique node connection relationships that form closed loops to form a connection tree. Determine the intersection variables between adjacent clique nodes as the separator set.

[0030] Optionally, step four specifically involves:

[0031] Based on the clique nodes in the connection tree, extract the risk variable nodes, barrier variable nodes and terminal score nodes contained in the clique nodes, and extract the conditional probability records corresponding to the risk variable nodes, barrier variable nodes and terminal score nodes from the node conditional probability table to form a clique node probability record set;

[0032] The conditional probability records in the clique node probability record set are jointly permuted according to the combination of node states contained in the clique node. The probability values ​​corresponding to the same combination of node states in the joint permutation result are merged to form the clique node probability estimate.

[0033] Based on the segmentation set, extract the barrier variable nodes contained in the segmentation set, and extract the status flags and numerical fields corresponding to the barrier variable nodes from the terminal risk dataset to form a barrier status record set;

[0034] The state markers and numerical fields in the barrier state record set are matched and categorized according to the barrier variable nodes contained in the delimiter set to form barrier state description items.

[0035] Optionally, step five specifically includes:

[0036] Based on the connection relationships in the connection tree, adjacent clique node pairs are extracted, and for each pair of adjacent clique node pairs, the sending clique node, the receiving clique node, and the partition set located between the sending clique node and the receiving clique node are determined to form a propagation construction record set;

[0037] An improved Shafer-Shenoy algorithm is used to construct propagation messages for the propagation construction record set. The improved Shafer-Shenoy algorithm is based on the probability estimate of the clique node corresponding to the sending clique node and uses the partition set between the sending clique node and the receiving clique node as the propagation constraint position. First, it performs aggregation calculation on the combination of node states in the sending clique node, and then forms the combined propagation volume of the sending clique node pointing to the partition set.

[0038] For the combined propagation quantity, variable projection and state alignment are performed according to the partition set variables in the partition set, so that the node state combination in the combined propagation quantity is merged into the state combination corresponding to the partition set variable, forming the partition set constraint propagation quantity.

[0039] Perform state matching between the partition set constraint propagation quantity and the barrier state description item corresponding to the partition set, and perform residual decomposition on each state combination in the partition set constraint propagation quantity according to the state matching result. Separate the propagation part that is reduced by the barrier state, and determine the propagation part that penetrates the barrier state as the penetration residual component.

[0040] Optionally, step six specifically includes:

[0041] Based on the propagation construction record set, the penetration residual component, sending clique node, receiving clique node and delimiter variable are extracted, and then aggregated according to the correspondence between sending clique node, receiving clique node and delimiter variable to form the edge processing record set;

[0042] For the penetrating residual components in the marginalized record set, variable retention and variable reduction are performed according to the separator set variable, so that the node state combination in the penetrating residual component is merged into the state combination corresponding to the separator set variable, forming the marginalized candidate record set.

[0043] State enhancement marginalization is performed on the node state combination in the marginalization candidate record set. Merging calculation is performed on the propagation volume with the same value of the separator set variable state, and the merge calculation result is organized into the message volume corresponding to the value of the separator set variable state to form the separator set propagation message.

[0044] The propagation relationship of the partition set propagation message is determined based on the receiving group node and partition set variables in the propagation construction record set, and the partition set propagation message is propagated to the receiving group node according to the propagation relationship.

[0045] Optionally, step seven specifically includes:

[0046] Locate the clique node where the terminal scoring node is located, and collect the partitioned propagation messages transmitted to the clique node where the terminal scoring node is located and the probability estimate of the clique node corresponding to the clique node where the terminal scoring node is located, to form a combined processing record set;

[0047] For the segmented set propagation messages and the probability estimates of the clique nodes in the combined processing record set, state matching and state alignment are performed according to the node state combination in the clique node where the terminal scoring node is located, so that the message volume in the segmented set propagation message and the probability value in the clique node probability estimate correspond to the same node state combination, forming a combined candidate record set.

[0048] Perform combination calculations on the message volume and probability values ​​in the candidate record set according to the node state combination, determine the terminal posterior probability value corresponding to the node state combination, and form the terminal posterior record set.

[0049] Merging and probability normalization are performed on the terminal scoring node state values ​​corresponding to the terminal posterior probability values ​​in the terminal posterior record set to form a posterior probability set corresponding to each state value of the terminal scoring node. The set is then arranged according to the state values ​​of the terminal scoring nodes to form the terminal residual posterior risk distribution.

[0050] Optionally, step eight specifically includes:

[0051] Extract the posterior probability set corresponding to each state value of the terminal scoring node from the terminal residual posterior risk distribution, and establish a correspondence between the state values ​​of the terminal scoring node and the posterior probability set to form a score mapping record set;

[0052] According to the risk level order of the terminal scoring node status values, a status score is assigned to each status value of the terminal scoring node, and the status score is matched with the corresponding posterior probability set to form a scoring calculation record set;

[0053] A weighted convergence calculation is performed based on the state scores and posterior probability set in the scoring calculation record set to form the coal mine terminal risk score, and the risk level is determined based on the score range corresponding to the coal mine terminal risk score.

[0054] The coal mine terminal risk score is correlated and organized with the risk level to form the coal mine terminal risk score result.

[0055] The beneficial effects of this invention are:

[0056] Compared with existing risk assessment methods that rely on static threshold judgment, empirical rule matching, or direct inference using ordinary Bayesian networks, this invention constructs a Bayesian network consisting of risk variable nodes, barrier variable nodes, and terminal scoring nodes by performing time alignment and terminal identifier association on coal mine terminal risk data. Furthermore, it forms a connection tree and a partition set, enabling the correlation between terminal risk status, barrier status, and scoring status to be characterized within a unified probabilistic inference framework. This improves the accuracy of node state relationship statistics and node conditional probability table construction, and reduces risk assessment bias caused by complex data sources, inconsistent time bases, and missing or abnormal state records.

[0057] Meanwhile, this invention introduces an improved Shafer-Shenoy algorithm during the connection tree propagation process. At the delimiter set, it utilizes barrier state description terms to perform residual decomposition on the propagated message and performs state enhancement marginalization on the penetration residual component. This allows the risk propagation process to reflect the constraint effect of barrier states on the propagation path and intensity, ultimately forming a terminal residual posterior risk distribution and outputting the coal mine terminal risk score result accordingly. Compared to technical solutions that only output general posterior probabilities or simple level determinations, this invention can more accurately characterize the residual risk level after the penetration barrier effect, making the risk score result more consistent with the actual protection conditions. It has the advantages of accurate risk propagation characterization, sufficient representation of barrier effect, and high reliability of the score result. Attached Figure Description

[0058] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0059] Figure 1 This is a flowchart of a coal mine terminal risk scoring method based on Bayesian networks proposed in this invention.

[0060] Figure 2 This is a schematic diagram of the connection tree construction for a Bayesian network-based coal mine terminal risk scoring method proposed in this invention.

[0061] Figure 3 This is a schematic diagram of the segmented set propagation message processing in a coal mine terminal risk scoring method based on Bayesian networks proposed in this invention. Detailed Implementation

[0062] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0063] refer to Figures 1-3 A method for scoring terminal risks in coal mines based on Bayesian networks includes the following steps:

[0064] Step 1: Collect coal mine terminal risk data, align execution time and associate terminal identifiers to form a terminal risk dataset;

[0065] Step 2: Extract risk variable nodes, barrier variable nodes, and terminal scoring nodes from the terminal risk dataset, construct a Bayesian network, and statistically analyze the node state relationships to form a node conditional probability table;

[0066] Step 3: Perform deontization, triangulation, and clique partitioning on the Bayesian network to form a connection tree, and form a partition set by the intersection variables of adjacent clique nodes;

[0067] Step 4: Calculate the probability estimate of the cluster nodes based on the node conditional probability table, and generate barrier state description items based on the state of the barrier variable nodes in the partition set;

[0068] Step 5: The improved Shafer-Shenoy algorithm is used to construct propagation messages between adjacent clique nodes. At the split set, the barrier state description term is used to perform residual decomposition on the propagation messages to form penetrating residual components.

[0069] Step 6: Perform state enhancement marginalization on the penetrating residual component within the delimiter variable domain, form a delimiter propagation message, and pass the delimiter propagation message to the adjacent clique node;

[0070] Step 7: Aggregate the propagation messages of the segmented set and the probability estimates of the cluster nodes where the terminal scoring node is located, perform combined calculations, and form the terminal residual posterior risk distribution;

[0071] Step 8: Perform score mapping and level classification on the distribution of terminal residual posterior risk to form the coal mine terminal risk score result.

[0072] In this embodiment, step one specifically includes:

[0073] The system collects risk records generated by coal mine terminals within a preset collection period and extracts terminal identifiers, collection timestamps, status markers, and numerical fields from these risk records. The preset collection period is a continuous time interval determined before the start of collection, which includes the start time and the end time. During collection, the risk records generated between the start time and the end time are used as the extraction range. The terminal identifiers, collection timestamps, status markers, and numerical fields are then organized into the same record to form the original risk record set.

[0074] Time mapping is performed on the collection timestamps in the original risk record set according to a unified time benchmark. The unified time benchmark is a unified timing standard used when aligning all collection timestamps in the original risk record set. During time mapping, the collection timestamps corresponding to each record are converted to the time position under the unified timing standard, and the original risk record set is arranged in the order of execution according to the time mapping result, so that the records in the original risk record set are arranged continuously according to the time sequence, forming a time-aligned record set.

[0075] The time-aligned record set is associated and merged according to the terminal identifier, so that records with the same terminal identifier are grouped together. Then, missing completion, anomaly correction and field normalization are performed on the associated and merged records. Missing completion is used to fill in missing fields, anomaly correction is used to correct fields that exceed the normal value range or deviate significantly from the change pattern of adjacent records, and field normalization is used to unify the record format of status markers and numerical fields to form a normalized risk record set.

[0076] The risk record set is arranged in the order of its collection timestamps, so that the records in the risk record set are arranged in a continuous time sequence, forming a terminal risk dataset.

[0077] In this embodiment, step two specifically includes:

[0078] State markers and numerical fields are extracted from the terminal risk dataset. The state markers and numerical fields are classified and organized according to the meaning of the records. The state markers and numerical fields used to characterize the changes in the risk status of the coal mine terminal are classified into the risk status item. The state markers and numerical fields used to characterize the changes in the safety barrier status of the coal mine terminal are classified into the barrier status item. The state markers and numerical fields used to characterize the changes in the risk score status of the coal mine terminal are classified into the score status item. The risk status item, barrier status item and score status item are organized into the same group of records to obtain the node extraction record set.

[0079] Risk variable nodes are established based on risk status items extracted from the record set. When establishing them, the status category corresponding to the risk status item is used as the range of node status values, and the record affiliation corresponding to the risk status item is used as the source of node identification. Barrier variable nodes are established based on barrier status items extracted from the record set. When establishing them, the status category corresponding to the barrier status item is used as the range of node status values, and the record affiliation corresponding to the barrier status item is used as the source of node identification. Terminal scoring nodes are established based on scoring status items extracted from the record set. When establishing them, the status category corresponding to the scoring status item is used as the range of node status values, and the record affiliation corresponding to the scoring status item is used as the source of node identification, thus forming a node set.

[0080] The node states in the node set are paired and statistically analyzed according to the collection timestamp order. During the statistics, the risk variable node state corresponding to the same collection timestamp is matched with the barrier variable node state, the risk variable node state corresponding to the same collection timestamp is matched with the terminal scoring node state, and the barrier variable node state corresponding to the same collection timestamp is matched with the terminal scoring node state. Continuous statistics are performed on the changes in node states between adjacent collection timestamps to determine the state association relationships between risk variable nodes and barrier variable nodes, between risk variable nodes and terminal scoring nodes, and between barrier variable nodes and terminal scoring nodes, forming a node relationship record set.

[0081] Based on the node relationship record set, risk variable nodes, barrier variable nodes, and terminal scoring nodes are connected. When connecting, the connection relationship between nodes is determined according to the state association direction in the node relationship record set. A Bayesian network is constructed, and the probability distribution of risk variable nodes, barrier variable nodes, and terminal scoring nodes under conditional states is statistically analyzed based on the node relationship record set. During the statistical analysis, the number of records in which nodes appear under each combination of conditional states is merged, and the probability conversion is performed on the merged results to form a node conditional probability table.

[0082] In this embodiment, step three specifically includes:

[0083] Extract risk variable nodes, barrier variable nodes, terminal score nodes, and node connection relationships from the Bayesian network. Remove the direction label from each directed connection in the node connection relationship and retain the connection endpoints to obtain undirected connections. Add undirected connections between nodes that all point to the same node, so that the original parent nodes form direct associations and form a deontological network.

[0084] In the deontological network, the number of node connections corresponding to each node is counted. Nodes to be eliminated are selected in order of increasing number of node connections. The connection status of each pair of adjacent nodes corresponding to each elimination node is checked. For adjacent nodes that do not have connecting edges, connecting edges are added to make all nodes adjacent to the elimination node fully connected. After the connecting edges are added, the elimination node is deleted to form a triangulated network.

[0085] In a triangulated network, a node group is retrieved where there is a connecting edge between any two nodes. During the retrieval, the node group is expanded along the connecting edge from any node and the pairwise connection relationship within the node group is verified. Node groups that meet the pairwise connection conditions are merged to form a clique node, and the clique partitioning result is formed according to the node group affiliation relationship.

[0086] Based on the clique partitioning results, the intersection variables and the number of intersection variables between clique nodes are calculated. During the calculation, the common variables in any two clique nodes are extracted and counted. Clique node connection relationships are selected in descending order of the number of intersection variables. Clique node connection relationships that form closed loops are deleted to ensure that the intersection variables are continuously distributed in the connection path, forming a connection tree. The intersection variables between adjacent clique nodes are determined as the separator set.

[0087] In this embodiment, step four specifically includes:

[0088] Based on the clique nodes in the connection tree, extract the risk variable nodes, barrier variable nodes and terminal score nodes contained in the clique nodes, and retrieve the conditional probability records corresponding to the risk variable nodes, barrier variable nodes and terminal score nodes from the node conditional probability table. During the retrieval, establish the corresponding relationship according to the node identifier of each node in the clique node, and group the conditional probability records belonging to the same clique node into the same record unit to form a clique node probability record set.

[0089] The conditional probability records in the clique node probability record set are jointly arranged according to the node state combinations contained in the clique node. During the joint arrangement, the state values ​​of each node in the clique node are used as the arrangement dimension. All state combinations formed by different node state values ​​are matched with conditional probability records one by one. The probability values ​​corresponding to the same node state combination in the joint arrangement result are merged and calculated. During the merge calculation, the conditional probability records corresponding to the same node state combination are merged and sorted to obtain the corresponding probability values, forming the clique node probability estimate.

[0090] Based on the segment set, extract the barrier variable nodes contained in the segment set, and extract the status markers and numerical fields corresponding to the barrier variable nodes from the terminal risk dataset. During extraction, retrieve the corresponding records in the terminal risk dataset according to the barrier variable node identifier in the segment set, and collect and organize the retrieved status markers and numerical fields according to the barrier variable nodes to form a barrier status record set.

[0091] The state markers and numerical fields in the barrier state record set are matched and categorized according to the barrier variable nodes contained in the separator set. During state matching, the state markers and numerical fields in the barrier state record set are matched with the barrier variable nodes in the separator set one by one. During state categorization, the matching results are classified and organized according to the state value and numerical range to form barrier state description items.

[0092] In this embodiment, step five specifically includes:

[0093] Based on the connection relationships in the connection tree, adjacent clique node pairs are extracted, and for each pair of adjacent clique node pairs, the sending clique node, the receiving clique node, and the partition set located between the sending clique node and the receiving clique node are determined to form a propagation construction record set. During extraction, the connection edges between clique nodes are first searched one by one in the connection tree, and then the clique nodes corresponding to both ends of each connection edge are determined as a pair of adjacent clique node pairs. The intersection variables corresponding to the connection edges are determined as the partition set. Then, the sending direction, receiving direction, and partition set are organized into the same record to form a propagation construction record set.

[0094] An improved Shafer-Shenoy algorithm is used to construct propagation messages for the propagation construction record set. The improved Shafer-Shenoy algorithm is based on the probability estimate of the clique node corresponding to the sending clique node and uses the partition set between the sending clique node and the receiving clique node as the propagation constraint position. First, the aggregation calculation is performed on the node state combination in the sending clique node, and then the combined propagation volume of the sending clique node pointing to the partition set is formed. During the aggregation calculation, the probability estimate of the clique node corresponding to the sending clique node is expanded item by item according to the node state combination, and the node state combination in the sending clique node that is not related to the partition set is merged, so that the remaining node state combination is concentrated in the partition set association range, forming the combined propagation volume of the sending clique node pointing to the partition set.

[0095] For the combined propagation quantity, variable projection and state alignment are performed according to the partition set variables in the partition set, so that the node state combinations in the combined propagation quantity are merged into the state combinations corresponding to the partition set variables, forming the partition set constraint propagation quantity. When projecting variables, the variable terms in the combined propagation quantity that correspond to the partition set variables are retained and irrelevant variable terms are eliminated. When aligning states, the state combinations in the combined propagation quantity are matched one-to-one and ordered according to the state values ​​of the partition set variables, so that the propagation quantities corresponding to the state values ​​of the same partition set variables are arranged to the same state position, forming the partition set constraint propagation quantity.

[0096] The propagation quantity of the partition set constraint is matched with the barrier state description item corresponding to the partition set. Based on the state matching result, the residual decomposition is performed on each state combination in the propagation quantity of the partition set constraint. The propagation part that is reduced by the barrier state is separated, and the propagation part that penetrates the barrier state is determined as the penetration residual component. During state matching, each state combination in the propagation quantity of the partition set constraint is matched with each state description item in the barrier state description item according to the state value of the partition set variable. During residual decomposition, the propagation quantity in each state combination is separated according to the barrier state reduction relationship, so that the propagation part that is reduced by the barrier state is stripped from the corresponding state combination, and the propagation part that penetrates the barrier state is retained in the corresponding state combination, forming the penetration residual component.

[0097] Compared to the unmodified Shafer-Shenoy algorithm, this invention makes targeted adjustments in four main aspects: propagation path construction, propagation volume formation method, partition set constraint processing, and barrier state intervention mechanism. The unmodified Shafer-Shenoy algorithm typically only performs conventional combination and marginalization processing based on the local probability information of the sending cluster node and adjacent messages, forming ordinary propagation messages oriented towards the partition set. This is insufficient to reflect the actual attenuation process of coal mine terminal risks under the action of safety barriers. In contrast, this invention first clarifies the correspondence between sending cluster nodes, receiving cluster nodes, and partition sets in the connection tree, ensuring that each propagation path has clear structural constraints. Then, during the propagation message construction stage, it performs aggregation calculations on the node state combinations in the sending cluster nodes, merging state combinations unrelated to the partition set. This allows the propagation volume to undergo centralized processing oriented towards the partition set's association range before entering the partition set. Subsequently, through variable projection and state alignment, the combined propagation volume is accurately merged into the state combinations corresponding to the partition set variables, forming... The propagation quantity with the characteristics of the partition set constraint is further matched with the barrier state description item by item, and residual decomposition is performed on each state combination. The propagation part reduced by the barrier state is stripped off, and only the propagation part that penetrates the barrier state is retained as the penetration residual component. This allows the barrier state to directly affect the risk propagation intensity and propagation range. After this improvement, the problem of the propagation quantity being too high due to ignoring the barrier reduction effect in the risk propagation process in the unimproved algorithm can be effectively avoided. It can improve the consistency between the propagation message and the actual working conditions, improve the authenticity of the terminal residual posterior risk distribution, improve the responsiveness of the risk scoring results to changes in the barrier state, and enhance the pertinence and stability of the connection tree propagation process. It can reasonably suppress high-risk states when the barrier is effective and fully retain the risk penetration characteristics when the barrier fails. In the end, the risk propagation characterization is closer to the actual operating conditions, the barrier effect is more fully reflected, and the terminal scoring results and risk level judgment are more consistent with the actual on-site conditions.

[0098] In this embodiment, step six specifically includes:

[0099] Based on the propagation construction record set, the penetration residual components, sending clique nodes, receiving clique nodes, and partition set variables are extracted and grouped according to the correspondence between sending clique nodes, receiving clique nodes, and partition set variables to form a marginalization processing record set. During extraction, each record in the propagation construction record set is searched item by item. The penetration residual components that point to the same sending clique node and correspond to the same partition set variable are grouped into the same record unit. Within the same record unit, the one-to-one correspondence between sending clique nodes, receiving clique nodes, and partition set variables is maintained to form a marginalization processing record set.

[0100] For the penetrating residual components in the marginalization processing record set, variable retention and variable elimination are performed according to the delimiter variable. This merges the node state combinations in the penetrating residual components into the state combinations corresponding to the delimiter variable, forming a marginalization candidate record set. When retaining variables, the variable items in the penetrating residual components that correspond to the delimiter variable are retained. When eliminating variables, the variable items in the penetrating residual components that do not belong to the delimiter variable are eliminated. The node state combinations after eliminating variable items are rearranged according to the state values ​​of the delimiter variable, so that the node state combinations corresponding to the same delimiter variable state value are merged into the same state position, forming a marginalization candidate record set.

[0101] State enhancement marginalization is performed on the node state combinations in the marginalization candidate record set. The propagation quantities with the same state value of the separator set variable are merged and the merged calculation results are organized into the message quantity corresponding to the state value of the separator set variable, forming the separator set propagation message. During state enhancement marginalization, the node state combinations in the marginalization candidate record set are searched item by item according to the state value of the separator set variable. The propagation quantities with the same state value of the separator set variable are aggregated into the same calculation unit. Then, the propagation quantities in the same calculation unit are merged and the merged calculation results are established with the corresponding state value of the separator set variable, forming the separator set propagation message.

[0102] Based on the receiving group nodes and partition set variables in the propagation construction record set, the transmission relationship of partition set propagation messages is determined, and the partition set propagation messages are transmitted to the receiving group nodes according to the transmission relationship. When determining the transmission relationship, the state values ​​of partition set variables in the partition set propagation messages are matched with the corresponding variable state positions in the receiving group nodes. Then, the matched message quantities are arranged into the transmission path corresponding to the receiving group nodes, thus completing the transmission of partition set propagation messages to the receiving group nodes.

[0103] In this embodiment, step seven specifically includes:

[0104] Locate the clique node where the terminal scoring node is located, and collect the partition set propagation messages transmitted to the clique node where the terminal scoring node is located and the probability estimate of the clique node corresponding to the clique node where the terminal scoring node is located to form a combined processing record set. During the location process, first search the connection tree for clique nodes containing the terminal scoring node, then extract the partition set propagation messages pointing to the clique node where the terminal scoring node is located item by item, and organize the extracted partition set propagation messages and the probability estimate of the clique node corresponding to the clique node where the terminal scoring node is located into the same record unit to form a combined processing record set.

[0105] For the segmented set propagation messages and clique node probability estimates in the combined processing record set, state matching and state alignment are performed according to the node state combinations in the clique node where the terminal scoring node is located. This ensures that the message volume in the segmented set propagation message and the probability value in the clique node probability estimate are mapped to the same node state combination, forming a combined candidate record set. During state matching, the segmented set variable state values ​​in the segmented set propagation message are matched item by item with the corresponding node state values ​​in the clique node where the terminal scoring node is located. During state alignment, the corresponding message volume and probability values ​​are rearranged according to the same node state combination, so that the message volume and probability values ​​corresponding to the same node state combination are aggregated to the same state position, forming a combined candidate record set.

[0106] The message volume and probability value in the combined candidate record set are combined and calculated according to the node state combination to determine the terminal posterior probability value corresponding to the node state combination and form the terminal posterior record set. During the combined calculation, the message volume and probability value corresponding to the same node state combination are jointly calculated so that the propagation impact reflected by the propagation message of the separated set and the probability relationship reflected by the probability estimate of the cluster node are converged into the same node state combination. The joint calculation result is determined as the terminal posterior probability value corresponding to the node state combination and forms the terminal posterior record set.

[0107] Based on the terminal posterior probability values ​​in the terminal posterior record set, merging and probability normalization are performed on the terminal scoring node state values ​​to form a posterior probability set corresponding to each state value of the terminal scoring node. The set is then arranged according to the terminal scoring node state values ​​to form the terminal residual posterior risk distribution. During merging, terminal posterior probability values ​​with the same terminal scoring node state values ​​are grouped into the same state category. During probability normalization, the terminal posterior probability values ​​corresponding to each state category are normalized to form a unified probability distribution. Finally, the set is arranged according to the order of the terminal scoring node state values ​​to form the terminal residual posterior risk distribution.

[0108] In this embodiment, step eight specifically includes:

[0109] Extract the posterior probability set corresponding to each state value of the terminal scoring node from the terminal residual posterior risk distribution, and establish a correspondence between the state values ​​of the terminal scoring node and the posterior probability set to form a score mapping record set. During extraction, retrieve the corresponding probability data item by item according to the arrangement position of the state values ​​of the terminal scoring node in the terminal residual posterior risk distribution, and then organize each state value and the corresponding probability data into the same record to form a score mapping record set.

[0110] According to the risk level order of the terminal scoring node status values, a status score is assigned to each status value of the terminal scoring node, and the status score is matched with the corresponding posterior probability set to form a scoring calculation record set. When assigning status scores, the scores are configured in order of risk level from low to high. Then, the status score corresponding to each status value is matched with the posterior probability set in the score mapping record set item by item to form a scoring calculation record set.

[0111] Based on the state scores and posterior probability sets in the scoring calculation record set, a weighted convergence calculation is performed to form the coal mine terminal risk score. The risk level is determined based on the score range corresponding to the coal mine terminal risk score. During the weighted convergence calculation, the state score corresponding to each state value is multiplied with the posterior probability set, and the product results are summed to form the coal mine terminal risk score. When determining the risk level, the coal mine terminal risk score is compared with the score range segment by segment, and the score range in which the coal mine terminal risk score falls is determined as the risk level.

[0112] The coal mine terminal risk score value is correlated and organized with the risk level to form the coal mine terminal risk score result. During the correlation and organization, the coal mine terminal risk score value and the risk level are collected into the same result record, so that the coal mine terminal risk score value and the risk level form a one-to-one correspondence, thus forming the coal mine terminal risk score result.

[0113] Example 1: To verify the feasibility of this invention in practice, it was applied to a coal mine terminal risk scoring scenario in an underground working area. In this scenario, the coal mine terminal continuously collects risk records, which are then compiled into a terminal risk dataset. This dataset contains records reflecting changes in risk status, barrier status, and terminal risk score status. Existing methods in this scenario generally employ fixed threshold judgment or ordinary Bayesian network inference to identify risks. While these methods can output risk levels, inconsistencies in the collection rhythm between different risk records, unstable correlations of status changes, and the difficulty in integrating the effect of barrier status into the propagation process lead to discrepancies between the terminal risk score and the actual residual risk on-site. Especially when the risk status has increased but the barrier status has already reduced its effect, existing methods often still give an overly high score; conversely, when the risk status appears stable but the barrier status has weakened or even failed, existing methods tend to give an underly high score. Therefore, they are unable to accurately reflect the residual risk level of the coal mine terminal under actual operating conditions.

[0114] In this embodiment, time alignment and terminal identifier association are first performed on the coal mine terminal risk data to form a terminal risk dataset. Then, risk variable nodes, barrier variable nodes, and terminal scoring nodes are extracted from the terminal risk dataset to construct a Bayesian network and form a node conditional probability table. Subsequently, deontologicalization, triangulation, and clique partitioning are performed on the Bayesian network to form a connection tree and a partition set. Based on the node conditional probability table, the probability estimates of clique nodes are calculated, and barrier state description terms are generated based on the states of barrier variable nodes in the partition set. During the risk propagation stage, instead of using ordinary probability propagation, an improved Shafer-Shenoy algorithm is used to construct propagation messages between adjacent clique nodes. At the partition set, the barrier state description terms are used to perform residual decomposition on the propagation messages to form penetrating residual components. Then, state enhancement marginalization is performed on the penetrating residual components within the partition set variable domain to form partition set propagation messages, which are then passed on to adjacent clique nodes. After receiving the partition set propagation messages, the clique nodes containing the terminal scoring nodes participate in the combined calculation with the clique node probability estimates to obtain the terminal residual posterior risk distribution. Based on the terminal residual posterior risk distribution, score mapping and level classification are performed, ultimately forming the coal mine terminal risk scoring result. Through this process, the terminal output is transformed from a general risk result to a residual risk result that takes into account the barrier reduction effect, thus making the scoring result more consistent with the actual situation on site.

[0115] To demonstrate the effectiveness of this invention in application, statistical analysis was performed on 1200 continuously collected terminal samples, with on-site verification results used as a control benchmark. The samples included those with increased risk status and no barrier activation, those with increased risk status and activated barrier, and those with moderate risk status but barrier failure. Traditional thresholding, ordinary Bayesian network methods, and the method of this invention were used for processing, and the comparative data are shown in Table 1.

[0116] Table 1 Comparison of Coal Mine Terminal Risk Scoring Results

[0117] Sample categories Sample size On-site verification of residual risk mean Traditional threshold method score mean Mean score of ordinary Bayesian network Average score of this invention The average deviation between the present invention and on-site verification Risk increased and barriers failed to activate 420 88.6 90.2 86.9 88.1 0.5 Risk has increased and barriers have been activated. 465 54.3 82.7 71.4 55.8 1.5 Moderate risk and barrier failure 315 76.8 58.6 67.3 75.1 1.7 All samples combined 1200 72.4 78.9 74.7 72.8 0.4

[0118] As shown in Table 1, the traditional threshold method significantly overestimates residual risk in samples where the risk level has increased and the barrier has been activated, and significantly underestimates residual risk in samples where the risk level is moderate but the barrier has failed. Although the ordinary Bayesian network method is an improvement over the traditional threshold method, the scoring results still have a large deviation from the on-site verification results because it does not combine the barrier state description item to perform residual decomposition and state enhancement marginalization during the propagation process. The average scores of the method of the present invention in the three types of samples reached 88.1, 55.8 and 75.1 respectively, which are closer to the average residual risk of 88.6, 54.3 and 76.8 of the on-site verification. The overall average deviation of all samples is only 0.4. At the same time, in actual statistics, the risk level determination accuracy of the method of the present invention reached 94.7%, which is significantly higher than the 72.6% of the traditional threshold method and the 83.8% of the ordinary Bayesian network method. This shows that the present invention can effectively solve the problems of insufficient barrier effect and inaccurate residual risk characterization in existing methods, and can more realistically reflect the actual risk level of the coal mine terminal.

[0119] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for scoring terminal risks in coal mines based on Bayesian networks, characterized in that, Includes the following steps: Step 1: Collect coal mine terminal risk data, align execution time and associate terminal identifiers to form a terminal risk dataset; Step 2: Extract risk variable nodes, barrier variable nodes, and terminal scoring nodes from the terminal risk dataset, construct a Bayesian network, and statistically analyze the node state relationships to form a node conditional probability table; Step 3: Perform deontization, triangulation, and clique partitioning on the Bayesian network to form a connection tree, and form a partition set by the intersection variables of adjacent clique nodes; Step 4: Calculate the probability estimate of the cluster nodes based on the node conditional probability table, and generate barrier state description items based on the state of the barrier variable nodes in the partition set; Step 5: The improved Shafer-Shenoy algorithm is used to construct propagation messages between adjacent clique nodes. At the split set, the barrier state description term is used to perform residual decomposition on the propagation messages to form penetrating residual components. Step 6: Perform state enhancement marginalization on the penetrating residual component within the delimiter variable domain, form a delimiter propagation message, and pass the delimiter propagation message to the adjacent clique node; Step 7: Aggregate the propagation messages of the segmented set and the probability estimates of the cluster nodes where the terminal scoring node is located, perform combined calculations, and form the terminal residual posterior risk distribution; Step 8: Perform score mapping and level classification on the distribution of terminal residual posterior risk to form the coal mine terminal risk score result.

2. The method for scoring terminal risks in coal mines based on Bayesian networks according to claim 1, characterized in that, Step one specifically involves: The risk records generated by the coal mine terminal within a preset collection period are collected, and the terminal identifier, collection timestamp, status marker and numerical field are extracted from the risk records to form the original risk record set. Time mapping is performed on the collection timestamps in the original risk record set according to a unified time benchmark, and the original risk record set is arranged in order according to the time mapping results to form a time-aligned record set; The time-aligned record set is associated and merged according to the terminal identifier, and the merged records are then filled with missing information, corrected for anomalies, and standardized in terms of fields to form a standardized risk record set. The risk records are arranged in order of collection timestamps to form a terminal risk dataset.

3. The method for scoring terminal risks in coal mines based on Bayesian networks according to claim 1, characterized in that, Step two specifically involves: State markers and numerical fields are extracted from the terminal risk dataset. State markers and numerical fields that reflect changes in risk status are categorized to form risk status items. State markers and numerical fields that reflect changes in barrier status are categorized to form barrier status items. State markers and numerical fields that reflect changes in terminal risk scoring status are categorized to form scoring status items, thus obtaining the node extraction record set. Risk variable nodes are established based on risk status items extracted from the record set of nodes; barrier variable nodes are established based on barrier status items extracted from the record set of nodes; and terminal scoring nodes are established based on scoring status items extracted from the record set of nodes, thus forming a node set. The node states in the node set are paired and statistically analyzed according to the collection timestamp order to determine the state association between risk variable nodes and barrier variable nodes, the state association between risk variable nodes and terminal scoring nodes, and the state association between barrier variable nodes and terminal scoring nodes, forming a node relationship record set. A Bayesian network is constructed by connecting risk variable nodes, barrier variable nodes, and terminal scoring nodes based on the node relationship record set. The probability distribution of risk variable nodes, barrier variable nodes, and terminal scoring nodes under conditional states is statistically analyzed based on the node relationship record set to form a node conditional probability table.

4. The method for scoring terminal risks in coal mines based on Bayesian networks according to claim 1, characterized in that, Step three specifically involves: Extract risk variable nodes, barrier variable nodes, terminal scoring nodes, and node connection relationships from the Bayesian network. Convert the directed connections in the node connection relationships into undirected connections, and supplement undirected connections for nodes that all point to the same node to form a deontological network. In the deontological network, nodes are selected for elimination in order of increasing number of node connections. For each node to be eliminated, connecting edges are added to the adjacent nodes to form a complete connection. After the connecting edges are added, the node to be eliminated is deleted, thus forming a triangulated network. In the triangulated network, retrieve the node groups where there is a connecting edge between any two nodes, merge the node groups to form clique nodes, and form clique partitioning results according to the node group affiliation relationship; Based on the clique partitioning results, calculate the intersection variables and the number of intersection variables between clique nodes. Select clique node connection relationships in descending order of the number of intersection variables. Delete clique node connection relationships that form closed loops to form a connection tree. Determine the intersection variables between adjacent clique nodes as the separator set.

5. The method for scoring terminal risks in coal mines based on Bayesian networks according to claim 1, characterized in that, Step four specifically involves: Based on the clique nodes in the connection tree, extract the risk variable nodes, barrier variable nodes and terminal score nodes contained in the clique nodes, and extract the conditional probability records corresponding to the risk variable nodes, barrier variable nodes and terminal score nodes from the node conditional probability table to form a clique node probability record set; The conditional probability records in the clique node probability record set are jointly permuted according to the combination of node states contained in the clique node. The probability values ​​corresponding to the same combination of node states in the joint permutation result are merged to form the clique node probability estimate. Based on the segmentation set, extract the barrier variable nodes contained in the segmentation set, and extract the status flags and numerical fields corresponding to the barrier variable nodes from the terminal risk dataset to form a barrier status record set; The state markers and numerical fields in the barrier state record set are matched and categorized according to the barrier variable nodes contained in the delimiter set to form barrier state description items.

6. The method for scoring terminal risks in coal mines based on Bayesian networks according to claim 1, characterized in that, Step five specifically involves: Based on the connection relationships in the connection tree, adjacent clique node pairs are extracted, and for each pair of adjacent clique node pairs, the sending clique node, the receiving clique node, and the partition set located between the sending clique node and the receiving clique node are determined to form a propagation construction record set; An improved Shafer-Shenoy algorithm is used to construct propagation messages for the propagation construction record set. The improved Shafer-Shenoy algorithm is based on the probability estimate of the clique node corresponding to the sending clique node and uses the partition set between the sending clique node and the receiving clique node as the propagation constraint position. First, it performs aggregation calculation on the combination of node states in the sending clique node, and then forms the combined propagation volume of the sending clique node pointing to the partition set. For the combined propagation quantity, variable projection and state alignment are performed according to the partition set variables in the partition set, so that the node state combination in the combined propagation quantity is merged into the state combination corresponding to the partition set variable, forming the partition set constraint propagation quantity. Perform state matching between the partition set constraint propagation quantity and the barrier state description item corresponding to the partition set, and perform residual decomposition on each state combination in the partition set constraint propagation quantity according to the state matching result. Separate the propagation part that is reduced by the barrier state, and determine the propagation part that penetrates the barrier state as the penetration residual component.

7. The method for scoring terminal risks in coal mines based on Bayesian networks according to claim 1, characterized in that, Step six specifically involves: Based on the propagation construction record set, the penetration residual component, sending clique node, receiving clique node and delimiter variable are extracted, and then aggregated according to the correspondence between sending clique node, receiving clique node and delimiter variable to form the edge processing record set; For the penetrating residual components in the marginalized record set, variable retention and variable reduction are performed according to the separator set variable, so that the node state combination in the penetrating residual component is merged into the state combination corresponding to the separator set variable, forming the marginalized candidate record set. State enhancement marginalization is performed on the node state combination in the marginalization candidate record set. Merging calculation is performed on the propagation volume with the same value of the separator set variable state, and the merge calculation result is organized into the message volume corresponding to the value of the separator set variable state to form the separator set propagation message. The propagation relationship of the partition set propagation message is determined based on the receiving group node and partition set variables in the propagation construction record set, and the partition set propagation message is propagated to the receiving group node according to the propagation relationship.

8. The method for scoring terminal risks in coal mines based on Bayesian networks according to claim 1, characterized in that, Step seven specifically involves: Locate the clique node where the terminal scoring node is located, and collect the partitioned propagation messages transmitted to the clique node where the terminal scoring node is located and the probability estimate of the clique node corresponding to the clique node where the terminal scoring node is located, to form a combined processing record set; For the segmented set propagation messages and the probability estimates of the clique nodes in the combined processing record set, state matching and state alignment are performed according to the node state combination in the clique node where the terminal scoring node is located, so that the message volume in the segmented set propagation message and the probability value in the clique node probability estimate correspond to the same node state combination, forming a combined candidate record set. Perform combination calculations on the message volume and probability values ​​in the candidate record set according to the node state combination, determine the terminal posterior probability value corresponding to the node state combination, and form the terminal posterior record set. Merging and probability normalization are performed on the terminal scoring node state values ​​corresponding to the terminal posterior probability values ​​in the terminal posterior record set to form a posterior probability set corresponding to each state value of the terminal scoring node. The set is then arranged according to the state values ​​of the terminal scoring nodes to form the terminal residual posterior risk distribution.

9. The method for scoring terminal risks in coal mines based on Bayesian networks according to claim 1, characterized in that, Step eight specifically involves: Extract the posterior probability set corresponding to each state value of the terminal scoring node from the terminal residual posterior risk distribution, and establish a correspondence between the state values ​​of the terminal scoring node and the posterior probability set to form a score mapping record set; According to the risk level order of the terminal scoring node status values, a status score is assigned to each status value of the terminal scoring node, and the status score is matched with the corresponding posterior probability set to form a scoring calculation record set; A weighted convergence calculation is performed based on the state scores and posterior probability set in the scoring calculation record set to form the coal mine terminal risk score, and the risk level is determined based on the score range corresponding to the coal mine terminal risk score. The coal mine terminal risk score is correlated and organized with the risk level to form the coal mine terminal risk score result.

Images