Coal mine underground gas anomaly inspection method based on multi-dimensional data fusion

Abstract

This invention discloses a method for inspecting underground gas anomalies in coal mines based on multidimensional data fusion, comprising the following steps: acquiring multidimensional inspection data and generating a multidimensional observation input structure; constructing a joint set of state variables and building an incremental factor graph inference structure using the iSAM2 algorithm; constructing an anomaly-switchable factor structure within the incremental factor graph inference structure and completing adaptive reconstruction using the iSAM2 algorithm; extracting ventilation topology relationships and inspection spatial partitioning relationships, and constructing a hierarchical Bayesian tree structure; performing local relinearization and incremental smoothing calculations during incremental data updates to generate joint inference results; constructing an inspection information gain evaluation structure based on the joint inference results and generating an information gain distribution; generating inspection update instructions based on the information gain distribution and applying them to the next round of inspection behavior, forming a closed-loop inspection mechanism. This invention improves the accuracy of gas anomaly identification and inspection efficiency.

Description

Technical Field

[0001] This invention relates to the field of coal mine safety monitoring and gas control technology, and in particular to a method for underground gas anomaly inspection in coal mines based on multi-dimensional data fusion. Background Technology

[0002] Methane gas in coal mines is one of the main hazards affecting safe production. Abnormal changes in methane concentration are characterized by their suddenness, wide spread, and complex evolution. To prevent methane accidents, existing technologies typically combine fixed sensor monitoring with manual or semi-automatic inspections to continuously monitor and assess the underground methane situation. However, these technologies still suffer from incomplete sensing, delayed response, and insufficient decision-making basis in complex underground environments.

[0003] Existing gas inspection technologies are mostly based on a single data source or a limited number of data types, focusing on threshold judgment of gas concentration itself. They lack a comprehensive characterization of ventilation status, environmental parameters, and spatial location, making it difficult to accurately reflect the formation mechanism and propagation characteristics of gas anomalies. Furthermore, inspection data exhibits significant differences in temporal sampling frequency, spatial distribution, and acquisition methods. Existing methods have limited ability to handle the temporal consistency and spatial correlation of multi-source data, easily leading to unstable or distorted condition assessment results.

[0004] Furthermore, existing technologies utilize inspection results in a relatively static manner, primarily relying on post-event analysis or fixed rule triggers. They lack mechanisms for dynamic inference and feedback adjustments based on continuous observation results, making it difficult to adapt to actual working conditions where gas conditions change constantly over time and space. Under complex tunnel structures and ventilation conditions, existing inspection strategies often fail to effectively guide subsequent inspection activities, hindering further improvements in inspection efficiency and anomaly identification capabilities.

[0005] Therefore, how to provide a method for detecting gas anomalies in coal mines based on multi-dimensional data fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for detecting gas anomalies in coal mines based on multidimensional data fusion. This invention constructs a multidimensional observation input structure around multidimensional inspection data, introduces a joint set of state variables and an incremental factor graph inference structure, and combines the iSAM2 algorithm to achieve incremental inference updates of multidimensional states. Furthermore, it adaptively reconstructs the observation factor set through an anomaly-switchable factor structure. Simultaneously, a hierarchical Bayesian tree structure is constructed based on ventilation topology and inspection spatial partitioning to constrain the propagation range of incremental updates. Based on the joint inference results, an inspection information gain evaluation structure is formed to generate inspection update instructions, thus constructing a closed-loop inspection mechanism. This invention achieves the co-evolution of gas anomaly identification and inspection behavior, possessing advantages such as strong continuity of state inference, high targeting of anomaly perception, and outstanding adaptive capability of inspection decision-making.

[0007] The coal mine underground gas anomaly inspection method based on multi-dimensional data fusion according to an embodiment of the present invention includes the following steps: Acquire multidimensional inspection data generated during the inspection process, perform time alignment and spatial correlation processing, and generate a multidimensional observation input structure; Based on the multidimensional observation input structure, a joint state variable set is constructed. Using the joint state variable set as nodes, an incremental factor graph inference structure is constructed by combining the iSAM2 algorithm, which includes the observation factor set and the observation residual set. An anomaly-switching factor structure is constructed in the incremental factor graph inference structure. Based on the observation residual set, the iSAM2 algorithm is used to perform incremental inference updates and replacement updates are performed on the observation factor set to complete the adaptive reconstruction of the incremental factor graph inference structure. Based on the multidimensional observation input structure, the ventilation topology and inspection spatial partitioning relationship are extracted, and the joint state variable set is hierarchically organized to construct a hierarchical Bayesian tree structure for ventilation propagation perception. For each incremental update of the multidimensional observation input structure, the local relinearization and incremental smoothing calculations of the iSAM2 algorithm are performed by combining the anomaly switchable factor structure and the hierarchical Bayesian tree structure to generate joint inference results; Based on the joint inference results, an inspection information gain evaluation structure is constructed, the anomaly uncertainty gradient is calculated, and the information gain distribution is generated. Based on the information gain distribution, inspection update instructions are generated and applied to the next round of inspection behavior, forming a closed-loop inspection mechanism that coordinates the evolution of gas anomaly identification and inspection behavior.

[0008] Optionally, the generation of the multidimensional observation input structure includes: During the inspection, gas concentration data, ventilation status data, environmental parameter data, and inspection posture data are acquired to form multi-dimensional inspection data. All multi-dimensional inspection data have corresponding sampling time and spatial location identifiers, forming sampling time series and spatial location series. Based on the sampling time identifier, multidimensional inspection data falling within the same time window are uniformly time-calibrated. Multidimensional inspection data that do not fall within the current time window and have not yet participated in time calibration are incorporated into subsequent time windows for time alignment. Spatial correlation processing is performed based on inspection pose data. A unified roadway zoning coordinate system is used as the spatial correlation benchmark to map gas concentration data, ventilation status data, and environmental parameter data to the roadway zoning spatial location consistent with the inspection pose data. Determine the spatial zoning relationship of the inspection based on the lane zoning coordinate system; The multidimensional inspection data are arranged sequentially according to the sampling time identifier, and the corresponding spatial location sequence, gas observation sequence, ventilation observation sequence and environmental observation sequence are bound one-to-one with the sampling time sequence as the main index to construct a multidimensional observation input structure.

[0009] Optionally, the construction of the incremental factor graph inference structure includes: A joint state variable set is constructed based on a multidimensional observation input structure. The joint state variable set includes location state variables, gas state variables, ventilation state variables, and environmental state variables. Using the joint set of state variables as nodes, an incremental factor graph inference structure is constructed by combining the iSAM2 algorithm, where the iSAM2 algorithm serves as the incremental update calculation mechanism. In the incremental factor graph inference structure, a set of observed factors is constructed; Based on the set of observed factors, an observation residual set is constructed in the incremental factor graph inference structure.

[0010] Optionally, the adaptive reconstruction of the incremental factor graph inference structure includes: In the incremental factor graph inference structure, an abnormal switchable factor structure is constructed based on the observation factor set. The abnormal switchable factor structure consists of a normal factor set, a working condition factor set, and an abnormal factor set, and each is configured with a corresponding set of activation state variables. Residual evolution analysis is performed based on the observed residual set to obtain the residual change sequence of the observed residual set, and residual evolution characteristics are constructed based on the residual change sequence; Based on the residual change sequence, windowed statistics are performed within a preset judgment window to determine the time consistency of residual changes corresponding to multiple consecutive sampling times and generate time consistency features. Based on the residual evolution characteristics and time consistency characteristics, the iSAM2 algorithm is used to perform incremental inference updates on the set of active state variables to determine the activation state of the abnormal switchable factor structure at the current sampling time. Based on the update results of the set of active state variables, the set of observed factors is replaced and updated. When the residual statistic of the corresponding observed factor exceeds the preset threshold within the continuous sampling time and meets the time consistency judgment condition, the observed factor is switched to the set of abnormal factors. The updated set of observed factors is reconfigured into the incremental factor graph inference structure, and the constraint relationship of the anomaly switchable factor structure in the incremental factor graph inference structure is adjusted accordingly, thus completing the adaptive reconstruction of the incremental factor graph inference structure.

[0011] Optionally, the construction of the hierarchical Bayesian tree structure includes: Based on the ventilation observation sequence in the multidimensional observation input structure, ventilation connection information and ventilation flow direction information corresponding to the ventilation status data are extracted to construct ventilation topology; Based on the spatial position sequence in the multidimensional observation input structure, the spatial zoning relationship of the inspection is constructed according to the positional distribution of the inspection pose data in the roadway space. Based on the ventilation topology and inspection space partitioning, hierarchical organization is performed on the set of joint state variables, and joint state variables that are located in the same inspection space partitioning relationship and have direct connectivity in the ventilation topology are organized into a set of nodes at the same level. Based on the ventilation topology, the hierarchical node set is divided into upstream and downstream levels to construct a hierarchical Bayesian tree structure for ventilation propagation perception. Based on the hierarchical Bayesian tree structure, the variable elimination order and the scope of influence of incremental updates in the iSAM2 algorithm are limited, and the hierarchical Bayesian tree structure for ventilation propagation sensing is constructed and constrained.

[0012] Optionally, the generation of the joint inference result includes: After receiving new multidimensional inspection data in the multidimensional observation input structure, an incremental data update process is triggered to incorporate the new multidimensional inspection data into the multidimensional observation input structure. Based on the multidimensional observation input structure after data integration, according to the current configuration state of the anomaly switchable factor structure, the corresponding set of observation factors is loaded into the incremental factor graph inference structure, and the set of joint state variables participating in this update is identified within the hierarchical range defined by the hierarchical Bayesian tree structure. Within the scope of variable elimination order and incremental update influence determined by the hierarchical Bayesian tree structure, the iSAM2 algorithm is used to perform local relinearization calculation on the set of joint state variables participating in this update, and to perform local linear approximation update on the set of affected observation factors and corresponding constraint relationships. After completing the local relinearization calculation, the iSAM2 algorithm is used to perform incremental smoothing calculation on the incremental factor graph inference structure, update the state of the joint state variable set, and update the observation residual set simultaneously. Based on the updated joint set of state variables and the set of observation residuals, state reading processing and uncertainty extraction processing are performed to generate joint inference results.

[0013] Optionally, the generation of the information gain distribution includes: Based on the joint inference results, the spatial uncertainty distribution corresponding to the inspection spatial partitioning relationship and the abnormal activation state corresponding to the set of activation state variables are extracted as input elements. The inspection spatial partitioning relationship is used as the computation domain to construct the inspection information gain evaluation structure. In the inspection information gain evaluation structure, spatial difference processing is performed on the spatial uncertainty distribution along the partition boundary of the inspection spatial partition relationship to calculate the uncertainty change between adjacent inspection spatial partition relationships and generate an abnormal uncertainty gradient. Weighted processing is performed on the abnormal uncertainty gradient based on the abnormal activation state, and the activation intensity corresponding to the abnormal activation state is mapped to the abnormal uncertainty gradient to form a weighted abnormal uncertainty gradient. Based on the weighted anomaly uncertainty gradient, information gain calculation is performed on the spatial partitioning relationship of each inspection within the inspection information gain evaluation structure to generate an information gain distribution.

[0014] Optionally, the construction of the closed-loop inspection mechanism includes: Based on the information gain distribution, the information gain values ​​corresponding to each inspection spatial partition relationship are sorted, and the candidate inspection spatial partition relationship set is determined in descending order of information gain value. Based on the spatial connectivity between the inspection spatial partition relationships, the candidate inspection spatial partition relationship set is processed to generate a path node sequence. Based on information gain distribution and abnormal activation state, for each inspection spatial partition relationship in the path node sequence, the node dwell time sequence is determined according to the combination relationship between the intensity of abnormal activation state and the level of spatial uncertainty distribution. Based on the path node sequence and node dwell time sequence, an inspection update instruction is constructed to constrain the inspection path and inspection rhythm of the next round of inspection behavior, and is issued to the inspection execution unit to execute the corresponding inspection behavior. After completing the inspection behavior corresponding to the inspection update command, new multi-dimensional inspection data is collected, forming a closed-loop inspection mechanism for the co-evolution of gas anomaly identification and inspection behavior during continuous inspection.

[0015] The beneficial effects of this invention are: First, this invention constructs a multi-dimensional observation input structure and introduces a joint set of state variables and an incremental factor graph inference structure, enabling multi-source inspection information to participate in inference updates under unified spatiotemporal constraints. This allows for the maintenance of temporal and spatial consistency of state estimation even with continuous incorporation of multi-dimensional data, effectively avoiding the anomaly identification lag and state drift problems caused by fragmented processing of multi-source data in existing technologies, and significantly improving the stability and continuity of underground gas anomaly inspection results in coal mines.

[0016] Secondly, this invention introduces an anomaly switchable factor structure into the incremental factor graph inference structure, and combines the residual evolution characteristics and time consistency characteristics of the observation residual set on the sampling time series to adaptively replace and update the observation factor set, thereby realizing dynamic perception and structural reconstruction of the abnormal state. This allows the abnormal factors to continuously participate in the subsequent incremental inference process, thus avoiding the misjudgment and missed judgment problems caused by existing methods that rely only on fixed thresholds or single-moment judgments, and significantly enhancing the pertinence and reliability of gas anomaly identification.

[0017] Furthermore, this invention constructs an inspection information gain evaluation structure based on joint inference results. It generates an information gain distribution through the collaborative calculation of spatial uncertainty distribution, abnormal activation state and inspection spatial partitioning relationship, and generates inspection update instructions accordingly. This forms a closed-loop inspection mechanism for the co-evolution of gas anomaly identification and inspection behavior, enabling the inspection path and dwell strategy to be dynamically adjusted according to the abnormal situation. This effectively improves the efficiency of inspection resource utilization and anomaly coverage capability, and overcomes the limitations of fixed inspection paths and delayed response in the prior art. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a coal mine underground gas anomaly inspection method based on multi-dimensional data fusion proposed in this invention. Figure 2 This is a schematic diagram of the abnormal switchable factor structure and the incremental factor graph inference structure in this invention; Figure 3 This is a schematic diagram of the inspection information gain evaluation structure and closed-loop inspection mechanism in this invention. Detailed Implementation

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

[0020] refer to Figures 1-3A method for detecting gas anomalies in coal mines based on multi-dimensional data fusion includes the following steps: The system acquires multidimensional inspection data generated during the inspection process. This multidimensional inspection data includes gas concentration data, ventilation status data, environmental parameter data, and inspection pose data. The environmental parameter data includes temperature data, humidity data, and dust concentration data. The system performs time alignment processing and spatial correlation processing on the multidimensional inspection data to form a multidimensional observation input structure. This multidimensional observation input structure includes sampling time series, spatial location series, gas observation series, ventilation observation series, and environmental observation series. A joint state variable set is constructed based on a multidimensional observation input structure. The joint state variable set includes location state variables, gas state variables, ventilation state variables, and environmental state variables. Using the joint state variable set as nodes, an incremental factor graph inference structure is constructed in conjunction with the iSAM2 algorithm. The incremental factor graph inference structure uses the iSAM2 algorithm to perform incremental update calculations. The incremental factor graph inference structure includes an observation factor set and an observation residual set. An anomaly-switchable factor structure is constructed in the incremental factor graph inference structure. The anomaly-switchable factor structure includes a normal factor set, a working condition factor set, and an anomaly factor set, and also includes an active state variable set corresponding to the normal factor set, working condition factor set, and anomaly factor set, respectively. Based on the residual evolution characteristics and time consistency characteristics of the observation residual set in the sampling time series, the iSAM2 algorithm is used to perform incremental inference update on the active state variable set. According to the update result of the active state variable set, the observation factor set is replaced and updated to complete the adaptive reconstruction of the incremental factor graph inference structure. Ventilation topology is extracted based on ventilation observation sequences in the multidimensional observation input structure, and inspection spatial partitioning is extracted based on spatial location sequences. The ventilation topology and inspection spatial partitioning are used to perform hierarchical organization on the joint state variable set to construct a hierarchical Bayesian tree structure for ventilation propagation perception. The hierarchical Bayesian tree structure is used to limit the variable elimination order and incremental update influence range of the iSAM2 algorithm, so that uncertainty propagates directionally along the ventilation topology within the hierarchical Bayesian tree structure. For each incremental update of data generated when new multidimensional inspection data is incorporated into the multidimensional observation input structure, the local relinearization and incremental smoothing calculations of the iSAM2 algorithm are performed by combining the anomaly switchable factor structure and the hierarchical Bayesian tree structure. The joint state variable set and the observation residual set are updated to generate joint inference results. The joint inference results include gas state estimation results, anomalous activation states of the activation state variable set, and spatial uncertainty distribution corresponding to the inspection spatial partitioning relationship. Based on the joint inference results, an inspection information gain evaluation structure is constructed. The inspection information gain evaluation structure takes the spatial uncertainty distribution and the abnormal activation state as input, the inspection spatial partitioning relationship as the computation domain, calculates the abnormal uncertainty gradient based on the spatial difference of the spatial uncertainty distribution on the inspection spatial partitioning relationship, and generates the information gain distribution based on the abnormal uncertainty gradient and the abnormal activation state. Based on the information gain distribution, inspection update instructions are generated. These instructions include a path node sequence and a node dwell time sequence. The inspection update instructions are applied to the next round of inspection to collect new multidimensional inspection data. The new multidimensional inspection data is incorporated into the multidimensional observation input structure. The incremental inference update of the activation state variable of the anomaly switchable factor structure and the incremental smoothing update under the constraint of the hierarchical Bayesian tree structure are performed again on the incremental factor graph inference structure. Based on the updated joint inference results, the inspection information gain evaluation structure is calculated again, forming a closed-loop inspection mechanism for the co-evolution of gas anomaly identification and inspection behavior.

[0021] In this embodiment, the generation of the multidimensional observation input structure includes: During the inspection, gas concentration data, ventilation status data, environmental parameter data, and inspection posture data are acquired to form multi-dimensional inspection data. The environmental parameter data consists of temperature data, humidity data, and dust concentration data. All multi-dimensional inspection data have corresponding sampling time and spatial location identifiers, forming sampling time series and spatial location series. Based on the sampling time identifier, time alignment processing is performed on the multidimensional inspection data. The time alignment processing uses a fixed time window length to segment the multidimensional inspection data. The time window moves sequentially on the time axis according to a preset sliding step size. Multidimensional inspection data that fall within the same time window are uniformly time-calibrated. Multidimensional inspection data that do not fall within the current time window and have not yet participated in time calibration processing are incorporated into subsequent time windows to participate in time alignment processing. The fixed time window length and preset sliding step size are determined by the time resolution of the sampling time series. Specifically, the time interval set of adjacent sampling time identifiers is calculated for the sampling time series, and the median of the time interval set is taken as the basic time resolution. The fixed time window length is set as an integer multiple of the basic time resolution, and the preset sliding step size is set as an integer multiple of the basic time resolution and less than the fixed time window length. When the multidimensional inspection data has duplicate sampling time identifiers in adjacent time windows, the sampling time identifier is used as the key to perform deduplication and merging. The merging rule is to retain the records with complete data fields after time alignment under the same sampling time identifier. After completing the time alignment process, spatial association processing is performed on the multidimensional inspection data based on the inspection pose data. The spatial association processing uses a unified roadway partition coordinate system as the spatial association reference to map the gas concentration data, ventilation status data and environmental parameter data to the roadway partition spatial position consistent with the inspection pose data. The spatial association processing performs pose interpolation and observation projection in the roadway partition coordinate system: For each gas concentration data, ventilation status data, and environmental parameter data, its sampling time identifier is read, the pose record corresponding to the adjacent previous sampling time identifier and the next sampling time identifier is located in the inspection pose data, the inspection pose corresponding to the sampling time identifier is obtained through time linear interpolation, the observation projection point is calculated in the roadway partition coordinate system based on the inspection pose, the spatial position identifier corresponding to the observation projection point is written into the spatial position sequence, and the observation data is bound to the spatial position identifier; The inspection space zoning relationship is determined based on the roadway zoning coordinate system. The zoning granularity of the roadway zoning coordinate system is determined by the roadway topology and the inspection zoning table, so that each gas concentration data, ventilation status data and environmental parameter data corresponds to a unique roadway zoning identifier. Based on the multidimensional inspection data after time alignment and spatial correlation processing, the multidimensional inspection data are arranged sequentially according to the sampling time identifier. Using the sampling time series as the main index, the spatial location sequence, gas observation sequence, ventilation observation sequence and environmental observation sequence corresponding to the same sampling time are bound one-to-one to construct a multidimensional observation input structure. The multidimensional observation input structure serves as the input data basis for the subsequent construction of the joint state variable set and the incremental factor graph inference structure.

[0022] In this embodiment, the construction of the incremental factor graph inference structure includes: A joint state variable set is constructed based on a multidimensional observation input structure. The joint state variable set includes location state variables, gas state variables, ventilation state variables, and environmental state variables. The location state variables are updated only based on the spatial location sequence, the gas state variables are updated only based on the gas observation sequence, the ventilation state variables are updated only based on the ventilation observation sequence, and the environmental state variables are updated only based on the environmental observation sequence. Using the joint set of state variables as nodes, an incremental factor graph inference structure is constructed by combining the iSAM2 algorithm. The incremental factor graph inference structure is used to describe the constraint relationship between the joint set of state variables, and the iSAM2 algorithm is used as the incremental update calculation mechanism of the incremental factor graph inference structure. In the incremental factor graph inference structure, an observation factor set is constructed, which is used to constrain the spatial location sequence, gas observation sequence, ventilation observation sequence and environmental observation sequence to the corresponding location state variables, gas state variables, ventilation state variables and environmental state variables, respectively. Specifically, constructing the set of observation factors involves matching the observation data corresponding to the same sampling time with the corresponding state variables in the set of joint state variables. In the incremental factor graph inference structure, a location observation factor is constructed for each spatial location observation data, a gas observation factor is constructed for each gas observation data, a ventilation observation factor is constructed for each ventilation observation data, and an environmental observation factor is constructed for each environmental observation data. Each observation factor is then connected to the corresponding location state variable, gas state variable, ventilation state variable, and environmental state variable to form a set of observation factors used to describe the constraint relationship between the observation data and the state variables. Based on the set of observed factors, an observation residual set is constructed in the incremental factor graph inference structure. The observation residual set is used to characterize the constraint deviation between the observed factors and the corresponding state variables. Specifically, constructing the observation residual set includes: based on the observation data corresponding to each observation factor in the observation factor set and its constrained state variables, calculating the deviation between the observation data and the current estimated value of the state variables in the incremental factor graph inference structure; recording the location deviation corresponding to the location observation factor, the gas deviation corresponding to the gas observation factor, the ventilation deviation corresponding to the ventilation observation factor, and the environmental deviation corresponding to the environmental observation factor; and aggregating the deviations corresponding to each observation factor to form the observation residual set. The observation residual set is used to characterize the degree of deviation between the observation constraints and the state variables in the subsequent abnormal switchable factor structure and incremental update calculation process. The configuration of the observation factor set and the observation residual set in the incremental factor graph inference structure is completed, forming the structural basis for incremental update calculation based on the iSAM2 algorithm, which can be called later when multidimensional inspection data is incrementally incorporated.

[0023] In this embodiment, the adaptive reconstruction of the incremental factor graph inference structure includes: In the incremental factor graph inference structure, an abnormal switchable factor structure is constructed based on the observed factor set. The abnormal switchable factor structure consists of a normal factor set, a working condition factor set, and an abnormal factor set. A corresponding set of activation state variables is configured for the normal factor set, the working condition factor set, and the abnormal factor set. The set of activation state variables is used to characterize the activation state of each factor set in the current inference process. Based on the temporal order of the observed residual set in the sampling time series, residual evolution analysis is performed on the observed residual set within a preset time span. By calculating the residual change relationship corresponding to adjacent sampling times, the residual change sequence of the observed residual set within the preset time span is obtained, and residual evolution features are constructed based on the residual change sequence. The residual evolution features are used to characterize the change behavior of the observed residuals as the sampling time evolves. The residual evolution characteristics consist of monotonicity, volatility, and drift indices of the residual change sequence. The monotonicity index is calculated based on the sign consistency of the residual change sequence, the volatility index is calculated based on the absolute mean of the residual change sequence, and the drift index is calculated based on the cumulative sum of the residual change sequence. The preset time span is determined by the continuous time interval of the residual index table on the sampling time sequence. The continuous time interval covers the interval of continuous sampling time identifiers corresponding to the same spatial location identifier, thereby limiting the residual evolution characteristics to the observation consistency range of the same spatial location identifier. Based on the residual change sequence, windowed statistics are performed within a preset decision window. Time consistency is determined for the residual changes corresponding to multiple consecutive sampling times. The consecutive overthreshold count and consecutive overthreshold length within the decision window are calculated. Time consistency features are generated based on the consecutive overthreshold count and consecutive overthreshold length. The time consistency features are used to characterize the residual stability and persistence of the observed residual set within the decision window. The consecutive threshold count refers to the number of sampling time points in the residual change sequence where the residual change value exceeds the preset residual threshold within a preset judgment window. The consecutive threshold length refers to the longest consecutive sampling time span in the residual change sequence where the residual change value continuously exceeds the preset residual threshold within the preset judgment window. Based on the residual evolution characteristics and time consistency characteristics, the iSAM2 algorithm is used to perform incremental inference updates on the set of active state variables. The update of the set of active state variables is driven by the residual statistics of the corresponding observed factors in the set of observed residuals, which is used to determine the activation state of the abnormal switchable factor structure at the current sampling time. Based on the update results of the set of active state variables, the set of observed factors is replaced and updated. When the residual statistic of the corresponding observed factor exceeds the preset threshold within the continuous sampling time and meets the time consistency judgment condition, the observed factor is switched from the set of normal factors or the set of working condition factors to the set of abnormal factors, and the factors in the set of abnormal factors are included in the subsequent incremental inference process. The preset residual threshold is determined by the historical deviation distribution of the observed residual set: for the deviation sequence of the same observed factor under the same spatial location identifier, the quantile threshold of the deviation sequence is calculated and written into the threshold table. When the corresponding observed factor meets the time consistency judgment condition within the continuous sampling time, the quantile threshold corresponding to the observed factor is read from the threshold table as the preset residual threshold to participate in the replacement and update judgment, thereby limiting the replacement and update of the observed factor to a deterministic judgment process based on historical residual statistics. After replacing and updating the set of observed factors, the updated set of observed factors is reconfigured into the incremental factor graph inference structure based on the anomaly switchable factor structure. Correspondingly, the constraint relationship of the anomaly switchable factor structure in the incremental factor graph inference structure is adjusted to complete the adaptive reconstruction of the incremental factor graph inference structure.

[0024] In this embodiment, the construction of the hierarchical Bayesian tree structure includes: Based on the ventilation observation sequence in the multidimensional observation input structure, ventilation connection information and ventilation flow direction information corresponding to the ventilation state data are extracted. The ventilation connection information is used to represent the ventilation connectivity relationship between different spatial locations, and the ventilation flow direction information is used to represent the ventilation propagation direction. A ventilation topology relationship is constructed, and the topology nodes in the ventilation topology relationship correspond one-to-one with the location state variables in the joint state variable set. The ventilation connection information is extracted from the ventilation status data field in the ventilation observation sequence. The ventilation status data field includes wind direction identifier, wind speed value, and ventilation facility status identifier. The ventilation flow direction information is established based on the wind direction identifier and the adjacent spatial location identifier pair of the spatial location sequence. Specifically, the spatial location identifier is used as a topology node, and the wind direction consistency of the adjacent spatial location identifier pair is calculated. When the wind direction identifier is consistent on the adjacent spatial location identifier pair and the wind speed value satisfies the continuity constraint, the adjacent spatial location identifier pair is written into the connection set of the ventilation topology relationship, and the wind direction identifier is written into the connection direction attribute. Based on the spatial position sequence in the multidimensional observation input structure, according to the position distribution of the inspection pose data in the roadway space, the spatial position sequence is mapped to the preset inspection spatial partitioning rules to construct the inspection spatial partitioning relationship. The inspection spatial partitioning relationship is used to characterize the spatial partitioning belonging relationship of the joint state variable set. Based on the ventilation topology and inspection space partitioning, hierarchical organization is performed on the set of joint state variables, and joint state variables that are located in the same inspection space partitioning relationship and have direct connectivity in the ventilation topology are organized into a set of nodes at the same level. Based on the ventilation propagation direction in the ventilation topology, the hierarchical node set is divided into upstream and downstream levels to construct a hierarchical Bayesian tree structure for ventilation propagation perception. Different levels in the hierarchical Bayesian tree structure correspond to different ventilation propagation levels. Based on the hierarchical Bayesian tree structure, the variable elimination order of the iSAM2 algorithm is restricted. The set of joint state variables corresponding to the upstream level of ventilation propagation is given priority to participate in the variable elimination calculation, and the influence range of incremental update is restricted to the level range that has a ventilation topology association with the set of nodes of the level where the update occurs. This completes the construction and constraint configuration of the hierarchical Bayesian tree structure for ventilation propagation perception.

[0025] In this embodiment, the generation of the joint inference result includes: After receiving new multidimensional inspection data in the multidimensional observation input structure, integrity and time validity checks are performed on the new multidimensional inspection data. When the new multidimensional inspection data meets the preset minimum time span condition in the sampling time series and forms a continuous observation record in the spatial location series, an incremental data update process is triggered to incorporate the new multidimensional inspection data into the multidimensional observation input structure. Based on the multidimensional observation input structure after data integration, according to the current configuration status of the normal factor set, operating condition factor set and abnormal factor set in the anomaly switchable factor structure, the corresponding observation factor set is loaded into the incremental factor graph inference structure, and the joint state variable set participating in this update is identified within the hierarchical range defined by the hierarchical Bayesian tree structure. Within the scope of variable elimination order and incremental update influence determined by the hierarchical Bayesian tree structure, the iSAM2 algorithm is used to perform local relinearization calculation on the set of joint state variables participating in this update, and to perform local linear approximation update on the set of affected observation factors and corresponding constraint relationships. After completing the local relinearization calculation, the iSAM2 algorithm is used to perform incremental smoothing calculation on the incremental factor graph inference structure, update the state of the joint state variable set, and simultaneously update the observation residual set to reflect the change in constraint deviation between the current multidimensional observation input structure and the joint state variable set. Based on the updated joint state variable set and observation residual set, a joint inference result is generated by performing state reading processing and uncertainty extraction processing on the updated joint state variable set. The joint inference result includes the gas state estimation result obtained based on the gas state variable, the abnormal activation state obtained based on the activation state variable set, and the spatial uncertainty distribution obtained by aggregating the uncertainty distribution of the joint state variable set based on the inspection spatial partitioning relationship. The observation residual set is retained and used in the subsequent incremental inference update process of the activation state variable of the abnormal switchable factor structure. Specifically, the generation of the joint inference result includes: reading the current estimated values ​​of gas state variables from the joint state variable set and forming a gas state estimation result; reading the values ​​of the activated state variables corresponding to the normal factor set, the working condition factor set, and the abnormal factor set from the activated state variable set and forming an abnormal activation state; extracting the uncertainty distribution corresponding to each state variable from the joint state variable set; partitioning and merging the uncertainty distribution according to the inspection space partitioning relationship; performing partition aggregation calculation on the uncertainty distribution within the same inspection space partitioning relationship; outputting the spatial uncertainty distribution corresponding to the inspection space partitioning relationship; and combining the gas state estimation result, the abnormal activation state, and the spatial uncertainty distribution to form the joint inference result.

[0026] In this embodiment, the generation of the information gain distribution includes: Based on the joint inference results, the spatial uncertainty distribution corresponding to the inspection spatial partition relationship and the abnormal activation state corresponding to the set of activation state variables are extracted. Using the spatial uncertainty distribution and abnormal activation state as input elements and the inspection spatial partition relationship as the computation domain, an inspection information gain evaluation structure is constructed. The inspection information gain evaluation structure is used to uniformly evaluate the abnormal perception value of each inspection spatial partition relationship. In the inspection information gain evaluation structure, spatial difference processing is performed on the spatial uncertainty distribution along the partition boundary of the inspection spatial partition relationship to calculate the uncertainty change between adjacent inspection spatial partition relationships, and an abnormal uncertainty gradient is generated based on the uncertainty change. The abnormal uncertainty gradient is used to characterize the direction and magnitude of change of spatial uncertainty in the inspection spatial partition relationship. In the inspection information gain evaluation structure, the abnormal uncertainty gradient is weighted based on the abnormal activation state, and the activation intensity corresponding to the abnormal activation state is mapped to the abnormal uncertainty gradient. The abnormal uncertainty gradient corresponding to the inspection space partition relationship with higher abnormal activation state is given higher weight, forming an abnormal-driven weighted abnormal uncertainty gradient. Based on the weighted anomaly uncertainty gradient, information gain calculation is performed on each inspection spatial partition relationship within the inspection information gain evaluation structure. The potential for reducing uncertainty brought about by each inspection spatial partition relationship in subsequent inspections is evaluated, and an information gain distribution corresponding one-to-one with the inspection spatial partition relationship is generated.

[0027] In this embodiment, the construction of the closed-loop inspection mechanism includes: Based on the information gain distribution output by the inspection information gain evaluation structure, within the scope of the inspection spatial partition relationship, the information gain values ​​corresponding to each inspection spatial partition relationship are sorted. Candidate inspection spatial partition relationship sets are determined according to the information gain values ​​from high to low. Based on the spatial connectivity between the inspection spatial partition relationships, path organization processing is performed on the candidate inspection spatial partition relationship sets. Spatially continuous inspection spatial partition relationships are combined according to the connectivity order, and the corresponding spatial position sequences are sequentially concatenated to generate a path node sequence. The path node sequence is composed of multiple spatial position sequences corresponding to inspection spatial partition relationships. The spatial connectivity is determined by the partition adjacency table of the inspection spatial partition relationship. The partition adjacency table is constructed based on the partition boundary sharing relationship under the lane partition coordinate system. The partition boundary sharing relationship is determined by the coincidence of the boundary point set. When performing path organization processing, a candidate adjacency subgraph is constructed for the candidate inspection spatial partition relationship set. Starting from the inspection spatial partition relationship with the highest information gain value in the candidate adjacency subgraph, a connected link is generated according to the adjacency expansion rule. The adjacency expansion rule limits each step of expansion to select the next inspection spatial partition relationship that is adjacent to the inspection spatial partition relationship at the end of the current link and has the highest information gain value, thereby ensuring that the spatial continuity of the path node sequence and the information gain priority are satisfied simultaneously. Based on information gain distribution and abnormal activation state, for each inspection spatial partition relationship in the path node sequence, the corresponding spatial uncertainty distribution and abnormal activation state level are extracted. According to the combination relationship between the intensity of abnormal activation state and the level of spatial uncertainty distribution, the node dwell time sequence corresponding to the path node sequence is determined. The node dwell time sequence is used to characterize the dwell time allocation of inspection behavior at each path node. The determination of the node dwell time sequence adopts a deterministic allocation rule: a dwell weight value is calculated for each inspection space partition relationship in the path node sequence. The dwell weight value is calculated by the information gain value corresponding to the inspection space partition relationship and the activation intensity of the abnormal activation state. The dwell weight value is normalized to a dwell ratio within the path node sequence range. The preset inspection time budget is allocated according to the dwell ratio to obtain the node dwell time sequence. The preset inspection time budget is determined by the execution cycle of the inspection update instruction and is written into the node dwell time sequence when each round of inspection update instruction is generated. Based on the path node sequence and node dwell time sequence, an inspection update instruction is constructed. The inspection update instruction is used to constrain the inspection path and inspection rhythm of the next round of inspection behavior, and the inspection update instruction is sent to the inspection execution unit to execute the corresponding inspection behavior. After completing the inspection behavior corresponding to the inspection update command, new multidimensional inspection data is collected and incorporated into the multidimensional observation input structure. This triggers the incremental inference update of the activation state variable of the incremental factor graph inference structure under the constraint of the abnormal switchable factor structure, and performs incremental smoothing calculation within the update range defined by the hierarchical Bayesian tree structure. Based on the joint inference results after completing this round of incremental smoothing calculation, the inspection information gain evaluation structure is reconstructed and a new information gain distribution is generated. The new information gain distribution is fed back to the inspection update instruction generation process, thereby forming a closed-loop inspection mechanism for the co-evolution of gas anomaly identification and inspection behavior during continuous inspection.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to a coal mine underground gas anomaly inspection scenario. In this scenario, the roadway structure exhibits obvious spatial connectivity characteristics, ventilation conditions continuously fluctuate with changes in the roadway structure, and gas concentration shows significant coupled variation characteristics in both time and space. Traditional inspection methods mainly rely on fixed routes and manual experience, resulting in delayed adjustments to inspection paths. Responses to local anomalies depend on single-point thresholds, making it difficult to continuously and stably identify and locate gas anomalies under complex ventilation conditions. Furthermore, inspection resource allocation lacks specificity, leading to problems such as redundant inspections and insufficient coverage of key areas.

[0029] In this scenario, gas concentration data, ventilation status data, environmental parameter data, and inspection pose data collected during inspections are continuously incorporated into a multi-dimensional observation input structure. Through time alignment and spatial correlation processing, a unified input suitable for joint modeling is formed. A joint set of state variables is constructed based on this multi-dimensional observation input structure, and multi-source observation constraints are uniformly expressed in the incremental factor graph inference structure, ensuring that gas state variables, ventilation state variables, and spatial location state variables are continuously updated within the same inference framework. As inspection data is continuously incorporated, the system dynamically adjusts the switchable anomaly factor structure driven by the observation residual set. This allows the normal factor set, operating condition factor set, and anomaly factor set to be replaced and updated under the control of the activated state variable set, thereby avoiding interference from long-term anomaly observations on the overall inference stability.

[0030] In actual operation, ventilation observation sequences and spatial location sequences are used together to construct a hierarchical Bayesian tree structure for ventilation propagation sensing, enabling the joint state variable set to be organized hierarchically under spatial partitioning and ventilation topological constraints. As new inspection data arrives, the system performs local relinearization and incremental smoothing calculations within a defined hierarchical range, continuously generating joint inference results. These joint inference results not only reflect changes in gas state estimates but also form the spatial uncertainty distribution corresponding to the inspection spatial partitioning relationship and output abnormal activation states, thus providing a quantitative basis for subsequent inspection decisions.

[0031] During the inspection decision-making phase, the system constructs an inspection information gain evaluation structure based on joint inference results. It obtains the anomaly uncertainty gradient by differentially calculating the spatial uncertainty distribution across the inspection spatial partitioning relationship, and generates an information gain distribution by combining this with anomaly activation states. This information gain distribution constrains the generation of inspection update instructions, ensuring that the inspection path spatially prioritizes areas with significant uncertainty changes and high anomaly activation states. The inspection update instructions are applied to the next round of inspections, and newly collected data is again incorporated into the multidimensional observation input structure, forming a closed-loop process of co-evolution between gas anomaly identification and inspection behavior.

[0032] Within a continuous operating cycle, the gas anomaly identification capability and inspection efficiency of the method of this invention were statistically analyzed by comparing it with traditional fixed-route inspection methods and single-point threshold alarm-based inspection methods under the same inspection resource constraints. The following experimental data comparison results were obtained.

[0033] Table 1. Performance Comparison Results of Different Inspection Methods in Coal Mine Underground Gas Anomaly Inspection Scenarios

[0034] As shown in Table 1, the traditional fixed-route inspection method suffers from low anomaly identification accuracy and effective coverage due to the lack of adjustment of the inspection path according to environmental changes, and a high proportion of duplicate inspections, indicating significant redundancy in inspection resource allocation. The inspection method based on single-point threshold alarms improves anomaly identification accuracy, but false positives and false negatives still occur in areas with frequent ventilation disturbances, and the improvement in the time to first detection of anomalies is limited.

[0035] In the method of this invention, the average gas anomaly identification accuracy is improved to 91.2%, mainly due to the responsiveness of the anomaly switchable factor structure to the continuous residual evolution, enabling a more reasonable modeling of the anomaly state during the inference process. The average time for the first detection of anomaly areas is shortened to 13.6 minutes, reflecting the guiding role of the inspection information gain evaluation structure in adjusting inspection paths. The effective coverage rate of a single inspection is increased to 84.6%, while the proportion of repeated inspections is reduced to 12.9%, indicating that the inspection update instructions generated based on the information gain distribution achieve better inspection resource allocation at the spatial partitioning level. The reduction in spatial anomaly location deviation is due to the introduction of the hierarchical Bayesian tree structure to constrain ventilation propagation, making the spatial propagation of uncertainty more consistent with actual ventilation conditions.

[0036] 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 coal mine underground gas anomaly inspection method based on multi-dimensional data fusion, characterized in that, Includes the following steps: Acquire multidimensional inspection data generated during the inspection process, perform time alignment and spatial correlation processing, and generate a multidimensional observation input structure; Based on the multidimensional observation input structure, a joint state variable set is constructed. Using the joint state variable set as nodes, an incremental factor graph inference structure is constructed by combining the iSAM2 algorithm, which includes the observation factor set and the observation residual set. An anomaly-switching factor structure is constructed in the incremental factor graph inference structure. Based on the observation residual set, the iSAM2 algorithm is used to perform incremental inference updates and replacement updates are performed on the observation factor set to complete the adaptive reconstruction of the incremental factor graph inference structure. Based on the multidimensional observation input structure, the ventilation topology and inspection spatial partitioning relationship are extracted, and the joint state variable set is hierarchically organized to construct a hierarchical Bayesian tree structure for ventilation propagation perception. For each incremental update of the multidimensional observation input structure, the local relinearization and incremental smoothing calculations of the iSAM2 algorithm are performed by combining the anomaly switchable factor structure and the hierarchical Bayesian tree structure to generate joint inference results; Based on the joint inference results, an inspection information gain evaluation structure is constructed, the anomaly uncertainty gradient is calculated, and the information gain distribution is generated. Based on the information gain distribution, inspection update instructions are generated and applied to the next round of inspection behavior, forming a closed-loop inspection mechanism that coordinates the evolution of gas anomaly identification and inspection behavior.

2. The coal mine underground gas anomaly inspection method based on multi-dimensional data fusion according to claim 1, characterized in that, The generation of the multidimensional observation input structure includes: During the inspection, gas concentration data, ventilation status data, environmental parameter data, and inspection posture data are acquired to form multi-dimensional inspection data. All multi-dimensional inspection data have corresponding sampling time and spatial location identifiers, forming sampling time series and spatial location series. Based on the sampling time identifier, multidimensional inspection data falling within the same time window are uniformly time-calibrated. Multidimensional inspection data that do not fall within the current time window and have not yet participated in time calibration are incorporated into subsequent time windows for time alignment. Spatial correlation processing is performed based on inspection pose data. A unified roadway zoning coordinate system is used as the spatial correlation benchmark to map gas concentration data, ventilation status data, and environmental parameter data to the roadway zoning spatial location consistent with the inspection pose data. Determine the spatial zoning relationship of the inspection based on the lane zoning coordinate system; The multidimensional inspection data are arranged sequentially according to the sampling time identifier, and the corresponding spatial location sequence, gas observation sequence, ventilation observation sequence and environmental observation sequence are bound one-to-one with the sampling time sequence as the main index to construct a multidimensional observation input structure.

3. The coal mine underground gas anomaly inspection method based on multi-dimensional data fusion according to claim 1, characterized in that, The construction of the incremental factor graph inference structure includes: A joint state variable set is constructed based on a multidimensional observation input structure. The joint state variable set includes location state variables, gas state variables, ventilation state variables, and environmental state variables. Using the joint set of state variables as nodes, an incremental factor graph inference structure is constructed by combining the iSAM2 algorithm, where the iSAM2 algorithm serves as the incremental update calculation mechanism. In the incremental factor graph inference structure, a set of observed factors is constructed; Based on the set of observed factors, an observation residual set is constructed in the incremental factor graph inference structure.

4. The coal mine underground gas anomaly inspection method based on multi-dimensional data fusion according to claim 1, characterized in that, The adaptive reconstruction of the incremental factor graph inference structure includes: In the incremental factor graph inference structure, an abnormal switchable factor structure is constructed based on the observation factor set. The abnormal switchable factor structure consists of a normal factor set, a working condition factor set, and an abnormal factor set, and each is configured with a corresponding set of activation state variables. Residual evolution analysis is performed based on the observed residual set to obtain the residual change sequence of the observed residual set, and residual evolution characteristics are constructed based on the residual change sequence; Based on the residual change sequence, windowed statistics are performed within a preset judgment window to determine the time consistency of residual changes corresponding to multiple consecutive sampling times and generate time consistency features. Based on the residual evolution characteristics and time consistency characteristics, the iSAM2 algorithm is used to perform incremental inference updates on the set of active state variables to determine the activation state of the abnormal switchable factor structure at the current sampling time. Based on the update results of the set of active state variables, the set of observed factors is replaced and updated. When the residual statistic of the corresponding observed factor exceeds the preset threshold within the continuous sampling time and meets the time consistency judgment condition, the observed factor is switched to the set of abnormal factors. The updated set of observed factors is reconfigured into the incremental factor graph inference structure, and the constraint relationship of the anomaly switchable factor structure in the incremental factor graph inference structure is adjusted accordingly, thus completing the adaptive reconstruction of the incremental factor graph inference structure.

5. The coal mine underground gas anomaly inspection method based on multi-dimensional data fusion according to claim 1, characterized in that, The construction of the hierarchical Bayesian tree structure includes: Based on the ventilation observation sequence in the multidimensional observation input structure, ventilation connection information and ventilation flow direction information corresponding to the ventilation status data are extracted to construct ventilation topology; Based on the spatial position sequence in the multidimensional observation input structure, the spatial zoning relationship of the inspection is constructed according to the positional distribution of the inspection pose data in the roadway space. Based on the ventilation topology and inspection space partitioning, hierarchical organization is performed on the set of joint state variables, and joint state variables that are located in the same inspection space partitioning relationship and have direct connectivity in the ventilation topology are organized into a set of nodes at the same level. Based on the ventilation topology, the hierarchical node set is divided into upstream and downstream levels to construct a hierarchical Bayesian tree structure for ventilation propagation perception. Based on the hierarchical Bayesian tree structure, the variable elimination order and the scope of influence of incremental updates in the iSAM2 algorithm are limited, and the hierarchical Bayesian tree structure for ventilation propagation sensing is constructed and constrained.

6. The coal mine underground gas anomaly inspection method based on multi-dimensional data fusion according to claim 1, characterized in that, The generation of the joint inference result includes: After receiving new multidimensional inspection data in the multidimensional observation input structure, an incremental data update process is triggered to incorporate the new multidimensional inspection data into the multidimensional observation input structure. Based on the multidimensional observation input structure after data integration, according to the current configuration state of the anomaly switchable factor structure, the corresponding set of observation factors is loaded into the incremental factor graph inference structure, and the set of joint state variables participating in this update is identified within the hierarchical range defined by the hierarchical Bayesian tree structure. Within the scope of variable elimination order and incremental update influence determined by the hierarchical Bayesian tree structure, the iSAM2 algorithm is used to perform local relinearization calculation on the set of joint state variables participating in this update, and to perform local linear approximation update on the set of affected observation factors and corresponding constraint relationships. After completing the local relinearization calculation, the iSAM2 algorithm is used to perform incremental smoothing calculation on the incremental factor graph inference structure, update the state of the joint state variable set, and update the observation residual set simultaneously. Based on the updated joint set of state variables and the set of observation residuals, state reading processing and uncertainty extraction processing are performed to generate joint inference results.

7. The coal mine underground gas anomaly inspection method based on multi-dimensional data fusion according to claim 1, characterized in that, The generation of the information gain distribution includes: Based on the joint inference results, the spatial uncertainty distribution corresponding to the inspection spatial partitioning relationship and the abnormal activation state corresponding to the set of activation state variables are extracted as input elements. The inspection spatial partitioning relationship is used as the computation domain to construct the inspection information gain evaluation structure. In the inspection information gain evaluation structure, spatial difference processing is performed on the spatial uncertainty distribution along the partition boundary of the inspection spatial partition relationship to calculate the uncertainty change between adjacent inspection spatial partition relationships and generate an abnormal uncertainty gradient. Weighted processing is performed on the abnormal uncertainty gradient based on the abnormal activation state, and the activation intensity corresponding to the abnormal activation state is mapped to the abnormal uncertainty gradient to form a weighted abnormal uncertainty gradient. Based on the weighted anomaly uncertainty gradient, information gain calculation is performed on the spatial partitioning relationship of each inspection within the inspection information gain evaluation structure to generate an information gain distribution.

8. The coal mine underground gas anomaly inspection method based on multi-dimensional data fusion according to claim 1, characterized in that, The construction of the closed-loop inspection mechanism includes: Based on the information gain distribution, the information gain values ​​corresponding to each inspection spatial partition relationship are sorted, and the candidate inspection spatial partition relationship set is determined in descending order of information gain value. Based on the spatial connectivity between the inspection spatial partition relationships, the candidate inspection spatial partition relationship set is processed to generate a path node sequence. Based on information gain distribution and abnormal activation state, for each inspection spatial partition relationship in the path node sequence, the node dwell time sequence is determined according to the combination relationship between the intensity of abnormal activation state and the level of spatial uncertainty distribution. Based on the path node sequence and node dwell time sequence, an inspection update instruction is constructed to constrain the inspection path and inspection rhythm of the next round of inspection behavior, and is issued to the inspection execution unit to execute the corresponding inspection behavior. After completing the inspection behavior corresponding to the inspection update command, new multi-dimensional inspection data is collected, forming a closed-loop inspection mechanism for the co-evolution of gas anomaly identification and inspection behavior during continuous inspection.

Images