Causal reasoning and confidence-driven gas over-limit risk judgment method and device

Abstract

The invention provides a causal reasoning and confidence-driven gas overrun risk judgment method and device, and the method comprises the following steps: firstly constructing a scenario semantic context based on underground monitoring data and an equipment state, and inputting the scenario semantic context into a fine-tuned large language model to generate a risk evolution causal chain containing hidden intermediate nodes; then, respectively calculating a historical case support degree, a real-time data goodness of fit and a knowledge fragment matching degree, and obtaining a comprehensive confidence coefficient based on arithmetic average logic; and the system automatically executes a hierarchical response strategy of full-automatic confirmation, man-machine collaborative research and judgment or low-confidence suppression according to a comparison result of the comprehensive confidence and a preset threshold value, and performs model closed-loop optimization by using feedback data. According to the method, by introducing a recessive logic completion and three-dimensional confidence verification mechanism, interpretability analysis and reliability quantification of the gas over-limit risk are achieved, and the problems that a traditional method is high in false alarm rate and lacks physical basis are effectively solved.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of coal mine safety monitoring and disaster early warning, and in particular to a cause-effect reasoning and confidence-driven gas overrun risk determination method and device. BACKGROUND

[0002] Coal mine gas accidents are one of the main disasters threatening coal safety production, and building an efficient and accurate gas overrun risk determination and early warning system is crucial for ensuring the safety of underground personnel and production continuity.

[0003] Existing gas monitoring and early warning technologies mainly rely on traditional sensor networks, triggering alarms by setting fixed gas concentration thresholds or rate thresholds. Although this method is simple and direct, it is essentially a passive defense, often issuing an alarm when the gas concentration has reached a dangerous level, leaving a short disposal window for on-site personnel. And due to the complex electromagnetic environment in the mine, sensor drift or signal interference often leads to frequent false alarms, and long-term repeated invalid alarms can easily cause dispatchers to develop a paralyzing mentality.

[0004] With the development of artificial intelligence technology, time series prediction algorithms based on recurrent neural networks or long short-term memory networks are widely used in gas concentration trend prediction. However, such data-driven deep learning models usually exhibit "black box" characteristics, only outputting numerical prediction results, and cannot explain the physical causes of gas changes. In the face of complex underground ventilation networks, existing prediction models are difficult to deduce hidden fault states such as wind flow short circuit, abnormal opening of air door, or failure of local ventilation machine that are not directly monitored by sensors, lacking deep analysis capabilities for risk evolution logic.

[0005] In addition, although large language models and other generative artificial intelligence technologies have shown strong potential for logical reasoning, their application in the industrial safety field faces the serious challenge of the illusion problem, i.e., the reasoning results generated by the model are logically smooth but violate physical facts. Existing technologies lack an effective confidence quantification evaluation mechanism, and cannot cross-verify the reasoning results of the model from multiple dimensions such as historical case matching degree, real-time data consistency, and domain knowledge logic. The lack of such reliability evaluation makes it difficult for the system to dynamically adjust the response strategy based on the credibility of risk determination, and only a single alarm action can be performed, which cannot balance the efficiency of automated disposal and the seriousness of safety decisions. SUMMARY

[0006] The present disclosure aims to at least partially address one of the technical problems of existing gas overrun early warning technologies, which rely solely on numerical threshold triggering, resulting in high false alarm rates, traditional data-driven models lack physical causal explanation, and cannot reliably quantify the evaluation of intelligent reasoning results.

[0007] Therefore, this disclosure proposes a method and apparatus for determining the risk of gas over-limit based on causal reasoning and confidence level.

[0008] The first aspect of this invention provides a method for determining the risk of gas exceeding limits based on causal reasoning and confidence level.

[0009] This method first constructs a contextualized semantic context. It acquires monitoring data sets and equipment status sets from underground coal mines at preset time intervals. Unlike traditional single-point monitoring, this method aligns the monitoring data sets and equipment status sets with time windows and performs feature vectorization to generate high-dimensional query vectors. These high-dimensional query vectors are then used to retrieve domain knowledge fragments, safety regulations, and historical similar cases with a semantic distance less than a preset value from a pre-defined multimodal knowledge storage unit. Subsequently, according to a predefined semantic template, the monitoring data sets, equipment status sets, and retrieved external knowledge are assembled into standardized natural language description inputs. This process achieves deep integration of multi-source heterogeneous data and domain expert knowledge, providing a physically meaningful input foundation for subsequent reasoning.

[0010] Building upon this foundation, this method utilizes a pre-tuned, domain-specific language model to generate a causal chain for risk evolution. The language model identifies anomalous features in the monitoring dataset as logical trigger points and generates explicit evolutionary node sequences. The key innovation of this invention lies in the introduction of an implicit logic completion mechanism: when the language model detects logical breakpoints in the evolutionary node sequence that cannot be directly supported by the monitoring dataset, it generates implicit intermediate nodes representing unobserved physical processes based on domain physical laws and inserts these implicit intermediate nodes at the logical breakpoints, thereby constructing a closed-loop logical chain encompassing the initial trigger, intermediate evolutionary processes, and final result. This transforms risk assessment from a black-box output into an interpretable physical evolutionary path.

[0011] To ensure the reliability of the inference results, this method introduces a multi-dimensional confidence assessment mechanism. This mechanism does not rely on single-dimensional verification but instead calculates historical case support, real-time data consistency, and knowledge fragment matching separately. Historical case support is determined by calculating the maximum cosine similarity between the current working condition feature vector and the feature vectors of historical real accident cases; real-time data consistency is determined by extracting predicted trends from the causal chain to construct a theoretical expected sequence and calculating its linear correlation with the actual sensor observation sequence; and knowledge fragment matching is determined by detecting the presence ratio of logical path triples in the causal chain in the coal mine knowledge graph. Finally, the method uses an arithmetic mean calculation model to fuse the above three dimensions of indicators to obtain a comprehensive confidence score that characterizes the credibility of the current judgment result.

[0012] Based on the overall confidence level, this method implements a tiered response strategy. The overall confidence level is compared with preset high-confidence and low-confidence thresholds: if it exceeds the high-confidence threshold, it enters a fully automatic confidence mode, directly generating and issuing early warnings and handling suggestions; if it falls between the two, it enters a human-machine collaborative judgment mode, requesting expert terminal review; if it falls below the low-confidence threshold, it enters a low-confidence suppression mode, blocking alarms and recording anomaly logs. Furthermore, the method includes a closed-loop feedback update mechanism. Based on actual operating conditions or expert feedback, it calculates the deviation loss value, encapsulates verified cases into a database, and dynamically adjusts the internal normalization parameters for confidence level calculation, achieving continuous iterative optimization of the system.

[0013] A second aspect of the present invention provides a device for determining the risk of gas exceeding limits based on causal reasoning and confidence level.

[0014] The device includes a data acquisition and preprocessing module, a contextualized semantic context construction module, a causal reasoning calculation module, a multi-dimensional confidence assessment module, a hierarchical decision-making and response control module, and a multimodal knowledge storage unit.

[0015] The data acquisition and preprocessing module is configured to acquire monitoring data sets and equipment status sets from underground coal mines at preset time intervals to ensure the temporal continuity of the data. The contextualized semantic context construction module is connected to the data acquisition and preprocessing module and the multimodal knowledge storage unit. It is configured to vectorize the data into features and construct a standardized contextualized semantic context by combining the retrieved domain knowledge.

[0016] The causal reasoning computation module is the core computing unit of the device. It is loaded with a large language model that has been fine-tuned with domain knowledge. It is configured to receive contextualized semantic context, capture element associations through an attention mechanism, and generate a risk evolution causal chain containing implicit intermediate nodes using built-in physical laws.

[0017] The multi-dimensional confidence assessment module is configured to perform historical matching, data trend verification, and graph verification in parallel, respectively outputting the support of historical cases, the consistency of real-time data, and the matching degree of knowledge fragments. It also calculates the comprehensive confidence score based on the arithmetic mean logic, thereby achieving quantitative auditing of the model output results.

[0018] The hierarchical decision-making and response control module is configured to automatically route to three different processing logics based on the comparison results of comprehensive confidence and threshold: fully automated confirmation, human-machine collaborative judgment, or low-confidence suppression. The multimodal knowledge storage unit, serving as the data support foundation, is configured to store domain knowledge fragments, safety procedure guidelines, historical similar cases, and a coal mine domain knowledge graph, and supports knowledge base updates during the closed-loop feedback process.

[0019] This invention provides a method and apparatus for determining the risk of gas exceeding limits based on causal reasoning and confidence level. It has the following beneficial effects: 1. This invention addresses the problem of traditional monitoring systems relying solely on sensor numerical thresholds or lacking interpretability in black-box models by constructing a risk evolution causal chain containing hidden intermediate nodes. Utilizing a large language model fine-tuned with domain knowledge, the system can deduce unobserved intermediate states such as airflow short-circuiting and insufficient effective airflow based on physical laws such as fluid mechanics and gas transport when logical breakpoints exist in explicit monitoring data. This reconstructs the complete physical evolution path of the accident, thereby improving the ability to identify hidden disaster-causing factors under complex operating conditions and enhancing logical interpretability.

[0020] 2. This invention suppresses the risks of sensor noise interference and non-factual illusions from generative models by establishing a three-dimensional orthogonal confidence quantification evaluation system based on historical case support, real-time data consistency, and knowledge fragment matching. This system combines empirical historical backtesting, positivist data trend verification, and rationalist graph logic validation to comprehensively cross-validate risk assessment results, ensuring that early warning signals have statistical significance and are supported by physical facts, thereby reducing the system's false alarm and false negative rates.

[0021] 3. This invention employs a hierarchical decision-making and response mechanism based on confidence thresholds, achieving a dynamic balance between the efficiency of automated processing and the seriousness of human decision-making. For high-confidence risks, a fully automated, confident response is executed to secure the crucial processing time; for medium-confidence risks, human-machine collaborative judgment is introduced to supplement expert experience; and low-confidence results are suppressed to avoid production disruptions. Combined with a closed-loop feedback optimization mechanism based on deviation loss, adaptive updates to the knowledge base and continuous iteration of the inference model are achieved, enabling the system to adapt to the dynamic spatiotemporal changes in mine geological conditions. Attached Figure Description

[0022] The above and / or additional aspects and advantages of this disclosure will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, in which: Figure 1 This is a schematic diagram of the overall process flow of the method provided in the embodiments of this disclosure; Figure 2 This is a schematic diagram of the multi-dimensional confidence quantification evaluation logic provided in the embodiments of this disclosure; Figure 3 This is a structural connection block diagram of the device provided in the embodiments of this disclosure.

[0023] The module includes: 101. Data acquisition and preprocessing module; 102. Contextualized semantic context construction module; 103. Causal reasoning calculation module; 104. Multi-dimensional confidence assessment module; 105. Hierarchical decision-making and response control module; 106. Closed-loop feedback optimization module; and 107. Multimodal knowledge storage unit. Detailed Implementation

[0024] Embodiments of this disclosure are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this disclosure, and should not be construed as limiting this disclosure.

[0025] It should be noted that all information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals involved in this disclosure are authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. The acquisition, transmission, storage, use, and processing of data in the technical solution of this disclosure all comply with the relevant provisions of national laws and regulations.

[0026] It should be noted that in the embodiments disclosed herein, certain software, components, models, and other existing solutions in the industry are mentioned. These should be considered as exemplary and are intended only to illustrate the feasibility of implementing the technical solutions disclosed herein. However, they do not mean that the applicant has used or necessarily used such solutions.

[0027] See attached document Figure 1 - Appendix Figure 3 The present invention provides a causal reasoning and confidence-driven gas over-limit risk determination device, which includes: a data acquisition module, a contextualized semantic context construction module, a causal reasoning module, a confidence assessment module, a decision response module, a feedback optimization module, and a storage unit.

[0028] The data acquisition module communicates with the sensor network and monitoring substations underground in the coal mine. The data acquisition module is configured to acquire monitoring data sets in real time at preset time intervals. With device status set Monitoring data set This includes data from gas concentration sensors, wind speed sensors, negative pressure sensors, temperature sensors, and carbon monoxide sensors. Equipment status set. This includes signals indicating the on / off status of local ventilation fans, the opening / closing status of dampers, and the on / off status of power-off devices. The data acquisition module cleans and denoises the acquired raw signals and transmits the processed, standardized data to the contextualized semantic context construction module.

[0029] The contextualized semantic context building module is connected to the data acquisition module and storage unit. The contextualized semantic context building module is configured based on the current monitoring data set. With device status set A feature vector is generated, and this feature vector is used to perform retrieval within the storage unit. The retrieved content includes: domain knowledge fragments. Safety regulations and similar historical cases Domain knowledge fragments Includes gas geology theory and fluid physics rules; safety regulations are based on. Includes preset gas concentration limits and operational procedures; similar historical cases. It includes records of gas exceedance events from historical databases where the distance to the current feature vector is less than a preset value. The contextualized semantic context building module concatenates and serializes the above structured data with unstructured text to generate standardized natural language description input. .

[0030] The causal reasoning module is connected to the contextualized semantic context building module. The causal reasoning module loads a large language model fine-tuned with knowledge from the coal mining domain. The causal reasoning module is configured to receive natural language description input. And use a large language model to generate a causal chain of risk evolution. Risk evolution causal chain It includes initial trigger nodes, intermediate evolution process nodes, and final result nodes. The causal reasoning module is also configured to generate implicit intermediate nodes not included in the monitoring data based on the knowledge base pattern in the storage unit. and hidden intermediate nodes Insert risk evolution causal chain In the logical sequence, a complete logical closed loop is formed.

[0031] The confidence assessment module is connected to the causal inference module and the storage unit. The confidence assessment module is configured to receive the risk evolution causal chain. And calculate the overall confidence level of the chain. The confidence assessment module uses the following calculation formula: ; In the above formula, This represents the support of historical cases and is configured to calculate the semantic similarity value between the current working condition feature vector and the feature vector of historical real accident cases in the storage unit. This represents the real-time data consistency, configured as the normalized value of the Pearson correlation coefficient between the trend sequence of key parameter changes predicted by the causal inference module and the real-time time series data acquired by the data acquisition module. This represents the matching degree of knowledge fragments and is configured to calculate the causal chain of risk evolution. The percentage of overlap between each logical node and the pre-stored domain knowledge graph path in the storage unit.

[0032] The decision response module is connected to the confidence assessment module. The decision response module has a preset high confidence threshold. With low confidence threshold The decision response module is configured to use the overall confidence level. Compare with the above threshold: If Greater than The decision response module generates an early warning signal that includes the risk level, causal chain, and handling recommendations; if Between and In between, the decision response module generates a review request signal containing the current data and the inference chain, and sends it to the expert terminal; if Less than The decision response module generates an exception log signal.

[0033] The feedback optimization module is connected to the decision response module and the storage unit. The feedback optimization module is configured to acquire actual operating condition results from external input. The feedback and correction results from the expert terminal are also considered. The feedback optimization module calculates the prediction results from the causal inference module and the actual operating results. The deviation value is calculated, and the historical case feature vector pattern in the storage unit is updated based on the deviation value. The feedback optimization module is also configured to adjust the calculation weight parameters of each sub-indicator in the confidence assessment module according to the expert correction results.

[0034] The storage unit is connected to the aforementioned modules and is used to store historical monitoring data, a vectorized knowledge base, a domain knowledge graph, a safety procedure text database, and system operation logs. The causal reasoning and confidence-driven gas over-limit risk assessment device in this embodiment can be deployed in an electronic device. This electronic device includes a processor and a memory. The memory stores computer program instructions, and when the processor executes these instructions, it implements the functions of the aforementioned modules. The electronic device transmits data to the downhole monitoring substation via an industrial Ethernet or fiber optic ring network.

[0035] See attached document Figure 2In this embodiment, the construction process of contextualized semantic context specifically includes three stages: spatiotemporal alignment of multi-source heterogeneous data, feature vectorization retrieval, and structured prompt engineering assembly.

[0036] First, in the spatiotemporal alignment stage of multi-source heterogeneous data, the system performs time window synchronization processing on data from sensors with different sampling frequencies. Because the sampling periods of the gas concentration sensor, wind speed sensor, and equipment status sensor differ, the system sets a unified time reference point. And define a time sliding window (in (The preset historical backtracking duration is, for example, 5 minutes). The system extracts the set of all monitoring data within this time window. With device status set The missing sampling points were filled using a linear interpolation algorithm, and a moving average filtering algorithm was used to remove non-logical numerical jumps caused by transient electromagnetic interference. The processed data was encapsulated into a structured state tensor containing temporal characteristics, which explicitly records the state of each monitoring indicator at [the specified timeframe]. The rate of change and trend direction of values ​​over a period of time, such as the upward slope of gas concentration and the variance of wind speed fluctuations.

[0037] Secondly, in the feature vectorization retrieval stage, the system utilizes the Retrieval Enhancement Generation (RAG) technology architecture to map the aforementioned structured state tensor into a high-dimensional query vector. Specifically, the system has a built-in pre-trained encoder that converts numerical state tensors into indexed text describing natural language (e.g., gas concentration increasing exponentially and local ventilators being shut down). This indexed text is then input into the embedding layer to generate a dimension... (e.g., 768-dimensional) query vector The system will query vectors. It matches the data with a vector database in the storage unit. The vector database pre-stores fragments of domain knowledge. Safety regulations and similar historical cases The converted key-value pair vector. The system uses the cosine similarity algorithm to calculate... The system calculates the distance to each key vector in the database and sorts them by similarity score from highest to lowest. The system then selects items with similarity scores higher than a preset search threshold. The former Each entry serves as valid external knowledge support information. This process ensures that the knowledge input into the large model is not static, but rather a dynamic subset of knowledge strongly correlated with the current operating conditions.

[0038] Finally, in the structured prompting engineering assembly stage, the system integrates the above information into a standardized input stream that can be parsed by a large language model according to a predefined semantic template. This semantic template contains four strictly defined field areas: the role definition domain, the real-time observation domain, the external knowledge domain, and the task instruction domain. The role definition domain is injected with system identity constraint instructions, limiting the model to reasoning only based on physical facts; the real-time observation domain is filled with cleaned current... Time data and External knowledge domains are filled in with the retrieved and filtered data. , and The text content and task instruction domain explicitly require the model to output a causal chain in a specific format. The system performs token length validation on the assembled text. If the length exceeds the model context window limit, the content of the external knowledge domain is truncated based on the retrieval similarity score, prioritizing the retention of historical cases with the highest similarity and mandatory safety procedure clauses, ultimately generating a complete contextualized semantic context. And transmit it to the causal reasoning module.

[0039] In this embodiment, the construction of the causal reasoning mechanism is not merely a simple model invocation process, but a deep processing flow encompassing domain-adaptive fine-tuning, explicit causal extraction, and implicit logic completion. First, the large language model used in this embodiment has undergone domain-supervised fine-tuning. Before deployment, the system constructed a training dataset containing hundreds of thousands of instruction-response pairs in the coal mine safety domain. This dataset originates from annual coal mine gas accident investigation reports, expert experience summaries, and physical simulation exercise records. Through low-rank adaptation (LoRA) or full-parameter fine-tuning techniques, the model's weight parameters are adjusted to optimize its processing capabilities. During input, the model can suppress the general probability of language generation and instead prioritize the physical laws of fluid dynamics and gas transport. This enables the model to identify working conditions in coal mines, such as establishing a strong correlation between sudden drops in local ventilation fan current and duct disconnection, rather than simply treating them as numerical fluctuations.

[0040] Secondly, in the explicit causal extraction stage, the model receives contextualized semantic context. Utilizing multi-head attention mechanism to monitor input data With equipment status A weighted analysis is performed. The model identifies anomalous features in the data as logical trigger points (i.e., the starting nodes of the causal chain). For example, a step drop in the wind speed sensor reading at a certain moment. Subsequently, the model combines the retrieved procedures from the context. Case Studies The system generates a series of explicit evolutionary nodes according to the time sequence. The system forces the model to output in a structured thought chain format, clearly distinguishing between the cause, intermediate state and final consequence, forming a preliminary logical framework.

[0041] Finally, and this is the core technical feature of this embodiment, lies in the implicit logic completion mechanism. In actual working conditions, sensors cannot cover all physical spaces or states (e.g., they cannot directly monitor whether the airflow is short-circuited or whether the damper is leaking air). When the model detects a logical breakpoint in the preliminary logic framework, i.e., the previous node... (If the damper opening status signal is true) the next node cannot be directly derived. (For example, when the gas concentration in the return airway increases), and there is a lack of sensor data to support this, the model activates the hidden node generation logic. Based on pre-trained domain knowledge, the model infers physical processes that must exist but have not been directly observed (i.e., hidden intermediate nodes). For example, airflow short-circuiting or insufficient effective airflow. The model will... By inserting the breakpoints in the logical chain, a continuous, closed-loop, complete causal chain of risk evolution is constructed. This chain not only yielded conclusions but also reconstructed the complete physical path of the accident's evolution, providing a verifiable logical basis for subsequent confidence assessments.

[0042] In this embodiment, the confidence assessment module does not simply perform a binary judgment on the model output, but rather constructs a ternary orthogonal quantitative evaluation system based on statistics and knowledge graph technology. The system receives the risk evolution causal chain generated by the preceding steps. The system calculates the support of historical cases, the consistency of real-time data, and the matching degree of knowledge fragments, and finally obtains the overall confidence level by arithmetic average.

[0043] First, regarding The system calculates (historical case support) by performing feature space mapping and similarity measurement. The system will then use the current... and generated Jointly encoded into a high-dimensional feature vector This vector contains not only numerical features such as current gas concentration and wind speed, but also semantic features of events such as damper opening and power outages. The system traverses the historical case vector library in the storage unit. ,calculate With each The cosine similarity is calculated. The system selects the maximum value from the calculated results as the cosine similarity. The value is 1. If the value is close to 1, it indicates that the current risk evolution pattern has been repeated and recorded in history, and has a very high a priori credibility; if the value is close to 0, it indicates that the current working condition belongs to a rare or completely new unknown pattern, and the system will reduce the weight of its reliance on historical experience.

[0044] Secondly, targeting The system calculates the (real-time data consistency) by performing time-series trend verification. (System analysis) The evolution nodes in the data are used to extract predictive descriptions containing changes in physical quantities. For example, it is predicted that gas concentration will rise non-linearly after wind speed decreases. The system defines a verification time window. Obtain the actual sensor observation sequence within this window. And construct the theoretical expected sequence based on the prediction model generated by the inference chain. The system uses the Pearson correlation coefficient algorithm to calculate the linear correlation between two sequences. Considering that negative correlation usually represents prediction error in this scenario, the system... Perform normalization processing, that is This calculation process strictly constrains the numerical range to the [0,1] interval, thereby quantifying the degree to which actual feedback from the physical world supports logical reasoning.

[0045] Secondly, regarding The calculation of (knowledge fragment matching degree) involves the system performing graph-based logical path verification. A knowledge graph of the coal mining field is pre-built in the storage unit. In this system, nodes represent physical entities or states (such as gas, explosion limits), and edges represent physical laws or procedural constraints (such as causing, exceeding). The system will generate causal chains. Parse into a set of triple paths The system checks one by one. Does each path in exist in In this process, if the inferred causal relationship (e.g., nitrogen causing a gas explosion) has no corresponding edge in the knowledge graph or violates the mutual exclusion constraints in the graph, the path is marked as invalid. The value of ExpertKnowledge is determined by the ratio of the number of valid paths to the total number of causal chain paths. This mechanism is equivalent to introducing a digital expert to logically refute the illusions of physics that arise from the large model.

[0046] Finally, based on the normalized values ​​of the three dimensions mentioned above, the system performs a final comprehensive confidence score calculation. Using the calculation formula, the system outputs a scalar. This value dynamically balances the judgments from three perspectives: empiricism (historical cases), positivism (real-time data), and rationalism (expert knowledge), ensuring that the final risk assessment result is robust.

[0047] In this embodiment, the decision response module is not merely a simple threshold trigger, but a decision center with dynamic routing and distribution capabilities, aiming to balance automation efficiency with the seriousness of security decisions. The system presets two strict confidence thresholds: a high-confidence threshold... (For example, set to 0.85) and low confidence threshold (For example, set to 0.45), the overall confidence level The decision is mapped to three mutually exclusive decision intervals, and differentiated response strategies are executed for each interval.

[0048] First, when the overall confidence level At this point, the system enters fully automatic confidence mode. In this mode, the system determines the causal chain of risk evolution. Highly reliable, it can serve as a decision-making basis without human intervention. The system immediately generates a red-level risk warning signal and distributes it to the downhole substation and the surface dispatch screen via industrial control protocols (such as Modbus or OPCUA). More importantly, the output is not just an alarm signal, but a structured intelligent analysis report. This report includes: Clear risk level assessment; complete The system includes text explaining the physical path of the accident; specific evidence chain data, such as listing the IDs of similar historical cases and fitting graphs of current gas concentration versus predicted curves; directly cited clause numbers from the "Coal Mine Safety Regulations"; and targeted handling recommendations generated by the large model (such as immediately cutting off power to a specific area and evacuating personnel). Furthermore, the system can be configured to directly trigger power outage interlocking control logic, cutting off non-intrinsically safe power to the threatened area within milliseconds, thus proactively blocking the risk.

[0049] Secondly, when the overall confidence level is at During the specified interval, the system activates a human-machine collaborative assessment mode. In this mode, if the system determines that a risk exists but the chain of evidence is insufficient to support fully automated decision-making, a manual review mechanism is triggered. Instead of directly issuing a mine-wide alarm, the system generates a verification assessment package, which is then pushed to the monitoring terminal of a safety expert. This assessment package is presented in a split-screen format on the user interface: the left side displays the causal reasoning chain generated by the system. The system displays the potential consequences derived from the data, along with corresponding real-time sensor curves, video surveillance footage, and summaries of similar historical cases on the right. Experts input confirmation, correction, or rejection commands via the terminal interface. The system treats the expert's input as the final decision command and executes it accordingly; if the expert chooses to correct, the system records the expert's revised reasoning logic as a strongly supervised sample for subsequent model optimization.

[0050] Finally, when the overall confidence level At this point, the system executes a low-confidence suppression mode. The system determines that the current model output originates from sensor noise interference or non-factual generation (i.e., hallucination) of a large language model, lacking physical factual support. To prevent false alarms from causing production interruptions or the "boy who cried wolf" effect reducing staff vigilance, the system blocks the alarm signal output interface, preventing the transmission of audible and visual alarms to the dispatch center. However, the system does not directly discard the inference result but marks it as an abnormal inference event or a low-confidence sample, recording it completely in the background audit log database. This log contains the input context at that time. The original inference chain generated by the model is provided for periodic review by technical personnel to analyze the reasons for the model producing low-confidence results (such as knowledge gaps or sensor malfunctions), thereby providing data support for the subsequent maintenance and iteration of the system.

[0051] See attached document Figure 1 and attached Figure 2 In this embodiment, the feedback optimization module is not a one-time offline training process, but a dynamically evolving system that runs parallel to online inference. This system constructs a prediction-real-time verification closed loop and uses incremental learning technology to continuously correct the completeness of the knowledge base and the deviations of the inference model, ensuring that the judgment method can adapt to the dynamic changes in mine geological conditions and ventilation networks over time.

[0052] First, the system executes real-time verification logic based on a time-delay window. The system performs a causal verification for each generated risk evolution chain. Set a valid verification time window (For example, inferring and predicting gas changes within the next 10 minutes, then...) The system will generate the predicted sequence from the causal inference module. Temporarily stored in the cache queue. When the system time advances to... At any given time, the feedback optimization module automatically retrieves the actual observation data for that time period from the data acquisition module. The system calculates the deviation loss value between the predicted sequence and the actual observed sequence. (For example, using mean squared error (MSE)). Simultaneously, the system acquires the expert review results recorded in the decision response module. If expert correction instructions exist, the system marks the expert's correction chain as a high-confidence true value label, and marks the original model output that has not been corrected by the expert as a sample to be validated.

[0053] Secondly, based on the verification results, the system performs adaptive incremental updates to the knowledge base. Regarding the deviation loss value... For ground truth labels that are below the preset tolerance or have been confirmed by experts as error-free, the system structures them into new standard cases. The system then extracts the contextual semantic context in which the case occurred. The feature vector and the finally confirmed risk evolution path are encapsulated into new key-value pairs. This data is then injected into the vector database of the storage unit. This process directly expands the support for historical cases. The computational benchmark library enables the system to achieve higher performance when similar operating conditions occur again in the future. Numerical analysis enables rapid risk identification and facilitates the digital accumulation of implicit experience.

[0054] Secondly, the system performs dynamic calibration of the confidence assessment parameters. Although the overall confidence calculation formula has a fixed structure, the system dynamically adjusts the normalization parameters within each sub-indicator based on historical feedback data. Specifically, the system statistically analyzes each sub-indicator over a past period (e.g., one week). The correlation between the accuracy of judgment and the accuracy of judgment. If it is found The real-time data fit was generally high in false alarm cases (indicating that simple data fitting can be deceptive), and the system will automatically adjust. The slope parameter of the Pearson correlation coefficient mapping function in the calculation logic is adjusted to reduce its sensitivity to local data fluctuations. In this way, the system fine-tunes the sensitivity of each factor without changing the core formula, thereby improving the overall confidence level. The numerical distribution is closer to the actual risk probability distribution.

[0055] Finally, the system performs low-frequency iterative fine-tuning of the large language model. The system maintains a hard sample set, specifically collecting those instances where the model's inferences are incorrect (i.e.,...). Cases that are significantly revised (e.g., extremely large) or have been substantially corrected by experts. During system downtime (e.g., maintenance shifts or nighttime), the feedback optimization module uses this sample set to perform instruction-based fine-tuning of the parameters of the large language model. By using the expert-corrected logic chain as the training target, the weight matrix of the attention layer in the model is adjusted using the backpropagation algorithm. This step aims to eliminate the model's type illusion at its root and correct its erroneous physical cognitive logic (e.g., correcting the model's one-sided logic that an increase in airflow necessarily leads to a decrease in gas concentration), thereby generating a more logically rigorous and physically accurate causal chain in the next inference cycle.

[0056] See attached document Figure 3 This embodiment provides a causal reasoning and confidence-driven gas over-limit risk assessment device. Physically, the device is a hardware and software combined system deployed on a coal mine safety monitoring center server or edge computing node. The device includes: a data acquisition and preprocessing module 101, a contextualized semantic context construction module 102, a causal reasoning calculation module 103, a multi-dimensional confidence assessment module 104, a hierarchical decision-making and response control module 105, a closed-loop feedback optimization module 106, and a multimodal knowledge storage unit 107.

[0057] The data acquisition and preprocessing module 101 establishes a physical connection with the underground environmental monitoring substation via an industrial Ethernet interface or fieldbus interface. The data acquisition and preprocessing module 101 is equipped with a signal demodulation unit and a digital filtering unit. The signal demodulation unit is used to analyze analog or frequency signals from the gas concentration sensor, wind speed sensor, negative pressure sensor, and carbon monoxide sensor, converting them into floating-point digital signals to form a monitoring data set. Simultaneously, analyze the switch signals for the on / off status of local ventilation fans, the on / off status of dampers, and the status of power-off devices to form a set of equipment statuses. The digital filtering unit is configured to use a moving average filtering algorithm to remove high-frequency noise from the signal and to perform linear interpolation to complete missing time series data, ensuring the temporal continuity of the input data. The data output terminal of the data acquisition and preprocessing module 101 is connected to the contextualized semantic context construction module 102.

[0058] The contextualized semantic context construction module 102 is connected to the data acquisition and preprocessing module 101 and the multimodal knowledge storage unit 107. This module integrates a vector encoder and a cueing engineering assembler. The vector encoder is configured to process the received... and The vector is mapped to a high-dimensional feature vector, and this vector is used to perform an approximate nearest neighbor search (ANN) in the vector database of the multimodal knowledge storage unit 107 to retrieve domain knowledge fragments with a semantic distance less than a preset threshold. Safety regulations and similar historical cases The engineer assembler is configured to concatenate the above search results with real-time data according to a predefined structured template, generating a large language model input stream that includes role constraints, observed facts, and external knowledge. .

[0059] The causal reasoning calculation module 103, connected to the contextualized semantic context construction module 102, is the core computing unit of the device. This module loads a large language model parameter file fine-tuned with knowledge from the coal mine safety domain. The causal reasoning calculation module 103 is configured to perform deep neural network inference operations. Using this as input, an attention mechanism is used to capture the non-linear relationships between various elements, generating a text sequence describing the risk evolution path, i.e., the risk evolution causal chain. This module also includes a logic completion subunit, which, when logical breakpoints exist in the monitoring data, invokes implicit knowledge stored within the model to generate implicit intermediate nodes representing unobserved physical processes. and embed it In this process, a closed-loop logical chain is formed.

[0060] The multi-dimensional confidence assessment module 104 is connected to the causal reasoning calculation module 103 and the multimodal knowledge storage unit 107. This module is configured to perform multi-dimensional confidence assessment on the generated... This module performs reliability quantification calculations. It contains three parallel computational subunits: a historical matching degree calculation subunit, a data trend verification subunit, and a knowledge graph verification subunit. The historical matching degree calculation subunit is configured to calculate the semantic similarity between the current working condition and historical cases, and output... The data trend verification subunit is configured to calculate the Pearson correlation coefficient between the predicted curve and the real-time data, and output... The knowledge graph validation subunit is configured to validate the effectiveness of the reasoning path within the knowledge graph, and outputs... The multi-dimensional confidence assessment module 104 ultimately performs a weighted average calculation to output the overall confidence score. Its calculation logic strictly corresponds to the formula: ; The hierarchical decision-making and response control module 105 is connected to the multi-dimensional confidence assessment module 104. This module is configured with a dual-threshold comparator and a signal routing distributor. The comparator will input... With the preset high threshold and low threshold Perform numerical comparison. The signal routing distributor executes hierarchical actions based on the comparison results: when... At that time, a control command containing a complete cause-and-effect chain and handling suggestions is sent to the dispatch screen through the alarm output interface, and a power outage interlock can be triggered in the area via the relay interface; when At that time, a review request data packet containing the data to be verified is sent to the expert terminal through the network communication interface, and feedback instructions from the expert are received; when When this happens, the alarm signal output is blocked, and the current inference event is marked as an abnormal log and written to the storage unit.

[0061] The closed-loop feedback optimization module 106 is connected to the hierarchical decision and response control module 105 and the multimodal knowledge storage unit 107. This module is configured to receive actual operating condition results. The module includes a loss calculation unit and a parameter update unit. The loss calculation unit calculates the deviation between the predicted result and the actual result; the parameter update unit writes the new case feature vector into the multimodal knowledge storage unit 107 to update the retrieval library based on the magnitude of the deviation and the content of expert corrections, and dynamically adjusts the weight coefficients of each calculation subunit in the multidimensional confidence assessment module 104.

[0062] The multimodal knowledge storage unit 107 is connected to the above modules to provide data read and write support. This unit physically includes a relational database (storing structured time-series data), a vector database (storing embedded text knowledge and case vectors), and a graph database (storing the node-edge relationships of the domain knowledge graph), providing knowledge support for the entire device.

Claims

1. A method for determining the risk of gas exceeding limits driven by causal reasoning and confidence level, characterized in that, Includes the following steps: S1. Acquire a set of monitoring data and a set of equipment status data from underground coal mines at preset time intervals, and construct a contextualized semantic context based on the set of monitoring data and the set of equipment status data. S2. Input the contextualized semantic context into a large language model that has been fine-tuned with domain knowledge in advance to generate a risk evolution causal chain that includes an initial trigger node, intermediate evolution process nodes and a final result node. S3. Based on the aforementioned risk evolution causal chain, calculate the historical case support, real-time data consistency, and knowledge fragment matching degree respectively, and calculate the comprehensive confidence level based on the historical case support, real-time data consistency, and knowledge fragment matching degree. S4. Compare the overall confidence level with the preset high confidence threshold and low confidence threshold, and execute a graded response strategy based on the comparison result; S5. Obtain actual working condition results or expert feedback results, and update the construction basis of the contextualized semantic context and the calculation parameters of the comprehensive confidence level based on the actual working condition results or expert feedback results.

2. The method for determining the risk of gas over-limit based on causal reasoning and confidence level according to claim 1, characterized in that, In step S1, constructing the contextualized semantic context specifically includes: The monitoring data set and the device status set are aligned with time windows and processed by feature vectorization to generate a high-dimensional query vector; The high-dimensional query vector is used to search in a pre-set multimodal knowledge storage unit to obtain domain knowledge fragments, security procedure basis and historical similar cases with a semantic distance less than a preset value; According to the predefined semantic template, the monitoring data set, equipment status set, domain knowledge fragments, safety procedure basis and historical similar cases are assembled into a standardized natural language description input, which serves as the contextualized semantic context.

3. The method for determining the risk of gas over-limit based on causal reasoning and confidence level according to claim 1, characterized in that, In step S2, generating the risk evolution causal chain specifically includes: The large language model is used to identify abnormal features in the monitoring dataset as logical trigger points to generate explicit evolution node sequences; When the large language model detects logical breakpoints in the evolution node sequence that cannot be directly supported by the monitoring data set, it generates implicit intermediate nodes representing physical processes that are not directly observed, based on the domain physical laws. The hidden intermediate node is inserted at the logical breakpoint of the evolution node sequence to form the risk evolution causal chain with a logical closed loop.

4. The method for determining the risk of gas over-limit based on causal reasoning and confidence level as described in claim 1, characterized in that, In step S3, calculating the support of historical cases specifically includes: The contextualized semantic context and the risk evolution causal chain are jointly encoded into a current working condition feature vector; Calculate the cosine similarity between the current operating condition feature vector and the feature vectors of historical real accident cases stored in the multimodal knowledge storage unit; The maximum value among the cosine similarity scores is selected as the support score for the historical cases.

5. The method for determining the risk of gas over-limit based on causal reasoning and confidence level according to claim 1, characterized in that, Step S3, calculating the real-time data consistency specifically includes: Analyze the causal chain of risk evolution, extract the trend prediction description of key parameters, and construct a theoretical expected sequence; Obtain the actual sensor observation sequence within the verification time window; The linear correlation between the theoretically expected sequence and the actual observation sequence of the sensor is calculated using the Pearson correlation coefficient algorithm; The linear correlation is normalized, and the processed value is determined as the real-time data consistency.

6. The method for determining the risk of gas over-limit based on causal reasoning and confidence level according to claim 1, characterized in that, In step S3, calculating the knowledge fragment matching degree specifically includes: The risk evolution causal chain is analyzed into a set of logical path triples; Each path in the set of logical path triples is checked one by one to see if it exists in the pre-set knowledge graph of the coal mining field; The ratio of the number of valid paths in the coal mining knowledge graph to the total number of paths in the risk evolution causal chain is calculated, and the ratio is determined as the matching degree of the knowledge fragment.

7. The method for determining the risk of gas over-limit based on causal reasoning and confidence level according to claim 1, characterized in that, In step S3, the overall confidence level is calculated in the following way: An arithmetic mean calculation model is used to sum and average the support of the historical cases, the consistency of the real-time data, and the matching degree of the knowledge fragments. The result is then used as the overall confidence level.

8. The method for determining the risk of gas over-limit based on causal reasoning and confidence level according to claim 1, characterized in that, Step S4, executing the tiered response strategy specifically includes: If the overall confidence level is greater than the high confidence threshold, the system enters a fully automatic confidence mode, generates an early warning signal containing the risk evolution causal chain and handling suggestions, and sends it to the control terminal. If the overall confidence level is between the low confidence threshold and the high confidence threshold, the system enters the human-machine collaborative judgment mode, generates a review request signal containing the data to be verified, and sends it to the expert terminal. If the overall confidence level is less than the low confidence threshold, the system enters a low confidence suppression mode, blocks the alarm signal output, and marks the current inference event as an abnormal log.

9. The method for determining the risk of gas over-limit based on causal reasoning and confidence level according to claim 1, characterized in that, Step S5 specifically includes: Calculate the deviation loss value between the predicted result of the risk evolution causal chain and the actual working condition result; When the deviation loss value is less than the preset tolerance or an expert confirmation instruction is received, the current contextualized semantic context and the risk evolution causal chain are encapsulated into a new case vector and stored in the multimodal knowledge storage unit. The internal normalization parameters of historical case support, real-time data consistency, and knowledge fragment matching are dynamically adjusted based on the deviation loss value when calculating the overall confidence level.

10. A gas over-limit risk determination device driven by causal reasoning and confidence level, characterized in that, The causal reasoning and confidence-driven method for determining the risk of gas exceeding limits, applied to any one of claims 1-9, includes: The data acquisition and preprocessing module is used to acquire monitoring data sets and equipment status sets from underground coal mines at preset time intervals. A contextualized semantic context construction module is used to construct a contextualized semantic context based on the monitoring data set and the device status set; The causal reasoning calculation module is used to load a large language model that has been fine-tuned with domain knowledge, receive the contextualized semantic context, and generate a risk evolution causal chain that includes an initial trigger node, intermediate evolution process nodes, a final result node, and hidden intermediate nodes. The multi-dimensional confidence assessment module is used to calculate the support of historical cases, the consistency of real-time data, and the matching degree of knowledge fragments, and to obtain the comprehensive confidence score based on the arithmetic mean logical calculation. The hierarchical decision and response control module is used to compare the comprehensive confidence level with preset high confidence thresholds and low confidence thresholds, and execute fully automatic confidence mode, human-machine collaborative judgment mode or low confidence suppression mode according to the comparison results. A multimodal knowledge storage unit is used to store domain knowledge fragments, safety regulations, historical similar cases, and knowledge graphs in the coal mining field.

Images