A mine production risk early warning method, system, device and medium

Abstract

The application discloses a mine production risk early warning method, system, device and medium, and relates to the technical field of coal mine safety; the method comprises the following steps: collecting structured time sequence monitoring data, unstructured text data and coal mine background description text in the field of coal mine safety; extracting entities and relations from the unstructured text data, and constructing a coal mine safety knowledge graph; training a time sequence network by using the structured time sequence monitoring data, and extracting abnormal sensing features based on a prediction residual; dynamically querying the knowledge graph according to the abnormal sensing features, and obtaining knowledge enhanced features; encoding the coal mine background description text to obtain text semantic features, and extracting time sequence evolution features of the structured time sequence monitoring data by using a prediction model; splicing the four types of features to input a multi-layer fully connected neural network, and outputting a multi-dimensional vector to represent the occurrence probability of multiple disasters; and dynamically warning according to the disaster occurrence probability. The application can solve the problems of difficult integration of coal mine safety knowledge, poor warning pertinence and low intelligence.

Description

Technical Field

[0001] This invention relates to the field of coal mine safety technology, and in particular to a method, system, equipment and medium for early warning of production risks in underground coal mines. Background Technology

[0002] In recent years, with the continuous increase in the depth and intensity of coal resource mining, the safety situation in underground coal mines has become increasingly severe. Coal mine disasters are characterized by multiple sources, coupling, and suddenness, mainly including five types: gas disasters, surrounding rock disasters, mine fires, mine water disasters, and dust disasters. Existing coal mine safety management mainly relies on manual experience and static procedures, which are insufficient to cope with the complex and ever-changing underground environment and the risks of multiple coupled disasters.

[0003] The existing coal mine safety early warning technology system still faces several bottlenecks that urgently need to be addressed. First, knowledge acquisition is difficult; professional knowledge in the field of coal mine safety is scattered across numerous documents, regulations, and case studies, lacking effective knowledge integration and management methods. Second, the early warning system lacks specificity; existing methods struggle to tailor risk assessments to the specific engineering context of each mine. Finally, the level of intelligence is low; existing early warning systems are unable to perform inference analysis of complex causal chains and dynamic risk assessment. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a method, system, equipment, and medium for early warning of production risks in underground coal mines, which can solve the problems of difficulty in knowledge integration, insufficient targeting of early warning, and low level of intelligence in the field of coal mine safety.

[0005] To achieve the above objectives, the present invention provides a method for early warning of production risks in underground coal mines, comprising: Acquire multi-source heterogeneous data in the field of coal mine safety; the multi-source heterogeneous data includes structured time-series monitoring data, unstructured text data, and coal mine background description text; Based on the LightRAG framework, entities and relationships are extracted from the unstructured text data to construct a coal mine safety knowledge graph; The LSTM network is trained using the structured time-series monitoring data to obtain a prediction model, and anomaly perception features are obtained by analyzing the prediction residuals of the prediction model. The knowledge graph is dynamically queried based on the anomaly perception features to retrieve associated disaster chain paths and entities, and the retrieval results are encoded as knowledge enhancement features. Semantic encoding is performed on the coal mine background description text to obtain text semantic features, and the temporal evolution features of the structured time-series monitoring data are extracted using the prediction model. The text semantic features, temporal evolution features, anomaly perception features, and knowledge enhancement features are concatenated and then input into a multi-layer fully connected neural network to obtain a multi-dimensional vector; each dimension of the multi-dimensional vector represents the probability of occurrence of a coal mine disaster. When the probability of any coal mine disaster exceeds the corresponding preset threshold, a dynamic early warning is triggered.

[0006] Optionally, entities and relationships are extracted from the unstructured text data based on the LightRAG framework to construct a coal mine safety knowledge graph, including: Based on the disaster chain theory, core entities and semantic relationships are defined to form an ontology model; the core entities include disaster-causing factors, disaster categories, disaster manifestations, monitoring indicators, and prevention and control measures. The unstructured text data is vectorized and segmented into text blocks to obtain multiple text blocks; Based on the ontology model, the LightRAG framework is used to call the large language model to perform semantic parsing on each text block in order to extract entities and relations and generate corresponding descriptive text, forming enhanced triples. The system performs deduplication and merging of entities pointing to the same real-world object in different enhanced triples, and performs semantic fusion on multiple descriptive texts corresponding to the same entity after deduplication and merging. After semantic fusion, the knowledge is stored in a graph database to form a coal mine safety knowledge graph.

[0007] Optionally, the LSTM network is trained using the structured time-series monitoring data to obtain a prediction model, including: The structured time-series monitoring data is subjected to sliding window processing, standardization, and data augmentation to obtain training samples; A stacked LSTM network is constructed, and the training samples are used to train the stacked LSTM network for multi-index synchronous prediction to obtain a prediction model.

[0008] Optionally, anomaly detection features can be obtained by analyzing the prediction residuals of the prediction model, including: Calculate the prediction residuals of the prediction model and standardize the prediction residuals. Based on the standardized prediction residuals, the mean residual, residual volatility, and residual trend are extracted as anomaly detection features.

[0009] Optionally, the knowledge graph is dynamically queried based on the anomaly perception features to retrieve associated disaster chain paths and entities, and the retrieval results are encoded as knowledge enhancement features, including: The monitoring indicators that exceed the preset abnormal threshold among the abnormality perception features are identified as key abnormality indicators. Based on the aforementioned key anomaly indicators, three types of query statements are dynamically generated, targeting background risk priors, anomaly indicator correlations, and spatial location risks. Execute the query statement to retrieve disaster chain paths and entities associated with the key anomaly indicators from the knowledge graph, and obtain a subgraph structure; The topological structure information of the subgraph structure is encoded into a fixed-dimensional knowledge-enhancing feature vector, which serves as the knowledge-enhancing feature.

[0010] Optionally, the prediction model is used to extract the temporal evolution features of the structured time-series monitoring data, including: The structured time-series monitoring data within a preset time window before the current moment is input into the prediction model; Extract the hidden state of the last time step of the prediction model; The hidden state is mapped to a fixed-dimensional temporal evolution feature vector through a fully connected layer, which serves as the temporal evolution feature.

[0011] Optionally, the structured time-series monitoring data includes at least one of the following: temperature, wind speed, dust concentration, oxygen concentration, methane concentration, carbon dioxide concentration, micro-vibration energy, roof pressure, roof displacement, water pressure, water level, and water temperature.

[0012] This invention also provides a production risk early warning system for underground coal mines, comprising: The data acquisition module is used to acquire multi-source heterogeneous data in the field of coal mine safety; the multi-source heterogeneous data includes structured time-series monitoring data, unstructured text data, and coal mine background description text. The multi-source feature generation module is used for: Based on the LightRAG framework, entities and relationships are extracted from the unstructured text data to construct a coal mine safety knowledge graph; The LSTM network is trained using the structured time-series monitoring data to obtain a prediction model, and anomaly perception features are obtained by analyzing the prediction residuals of the prediction model. The knowledge graph is dynamically queried based on the anomaly perception features to retrieve associated disaster chain paths and entities, and the retrieval results are encoded as knowledge enhancement features. Semantic encoding is performed on the coal mine background description text to obtain text semantic features, and the temporal evolution features of the structured time-series monitoring data are extracted using the prediction model. The disaster probability prediction module is used to concatenate the text semantic features, temporal evolution features, anomaly perception features, and knowledge enhancement features, and then input them into a multi-layer fully connected neural network to obtain a multi-dimensional vector; each dimension of the multi-dimensional vector represents the probability of occurrence of a coal mine disaster. The early warning module is used to trigger a corresponding dynamic early warning when the probability of any coal mine disaster exceeds the corresponding preset threshold.

[0013] The present invention also provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for early warning of production risks in underground coal mines.

[0014] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described method for early warning of production risks in underground coal mines.

[0015] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: The underground coal mine production risk early warning method provided by this invention acquires multi-source heterogeneous data, including structured time-series monitoring data, unstructured text data, and coal mine background description text. Based on the LightRAG (Lightweight Retrieval-Augmented Generation) framework, it extracts entities and relationships from the unstructured text data and constructs a coal mine safety knowledge graph. This integrates professional knowledge in the field of coal mine safety scattered in literature, regulations, and cases, solving the problems of difficulty in knowledge acquisition and lack of effective knowledge integration and management methods in the prior art.

[0016] By semantically encoding the background description text of the coal mine to obtain textual semantic features, and simultaneously using structured time-series monitoring data to train an LSTM prediction model to extract time-series evolution features, risk assessment can integrate personalized engineering background information such as the mine's own geological conditions and mining technology. This improves the pertinence and personalization of early warning and makes up for the shortcomings of existing early warning methods that are difficult to combine with specific mine engineering background for risk assessment.

[0017] By obtaining anomaly perception features through prediction residual analysis of the LSTM prediction model, and using these features to drive dynamic querying of associated disaster chain paths and entities from the knowledge graph, knowledge-enhanced features are generated. Then, textual semantic features, temporal evolution features, anomaly perception features, and knowledge-enhanced features are concatenated and fused, and multi-dimensional vectors representing the probability of occurrence of various coal mine disasters are output through a multi-layer fully connected neural network. This realizes an intelligent process from anomaly perception of real-time monitoring data to causal chain reasoning analysis of complex disasters, and enables dynamic assessment and early warning of concurrent risks of multiple disasters, thereby improving the intelligent prediction and decision support capabilities of coal mine multi-disaster coupled risks. Attached Figure Description

[0018] The above and other objects, features and advantages of the present invention will become more apparent from the more detailed description of exemplary embodiments of the invention in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same parts.

[0019] Figure 1 This is a schematic diagram of the method flow for early warning of production risks in underground coal mines, as shown in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the knowledge graph construction process based on LightRAG, as shown in an embodiment of the present invention. Figure 3 This is a schematic diagram of a coal mine disaster classification system shown in an embodiment of the present invention; Figure 4 This is a schematic diagram of the coal mine safety knowledge graph network structure shown in an embodiment of the present invention; Figure 5 This is a schematic diagram of the module structure of an underground coal mine production risk early warning system according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1 , Figure 1 This is a flowchart illustrating a method for early warning of production risks in underground coal mines. This method can be executed by computer equipment, specifically by a terminal or server alone, or by both. In this embodiment, the method is illustrated using a server as an example.

[0022] Methods for early warning of production risks in underground coal mines include: S101: Acquire multi-source heterogeneous data in the field of coal mine safety.

[0023] Multi-source heterogeneous data includes structured time-series monitoring data, unstructured text data, and coal mine background description text.

[0024] Structured time-series monitoring data refers to the time series of environmental monitoring indicators collected in real time by various downhole sensors. This data is used for training and real-time prediction of time-series forecasting models. Specifically, it may include real-time historical data of at least one of the following monitoring indicators: temperature, wind speed, dust concentration, oxygen concentration, methane concentration, carbon dioxide concentration, microseismic energy, roof pressure, roof displacement, water pressure, water level, and water temperature. These indicators reflect the real-time state of the downhole environment from different dimensions, providing data sources for identifying abnormal operating conditions and assessing disaster risks.

[0025] Unstructured text data refers to professional literature in the field of coal mine safety that exists in natural language form, used to construct a coal mine safety knowledge graph. This can include academic literature, accident investigation reports, and coal mine safety regulations, etc.

[0026] Coal mine background description text refers to static textual data describing the specific engineering characteristics of a mine, serving as the static context for model input. This type of text may include descriptive documents such as mine geological conditions, mining technology, and roadway layout, reflecting the inherent risk characteristics and engineering background of the target mine.

[0027] S102: Extract entities and relationships from unstructured text data based on the LightRAG framework to construct a coal mine safety knowledge graph.

[0028] In applications, the LightRAG framework can be used to extract and organize knowledge from unstructured text data, constructing a coal mine safety knowledge graph. This transforms domain knowledge scattered across documents, regulations, and case studies into structured, queryable knowledge representations. See also... Figure 2 The knowledge graph construction process based on LightRAG includes: first, preprocessing unstructured text data (cleaning, word segmentation), then identifying entities and relationships through a lightweight extraction module (joint or pipeline NER&RE), constructing triples, and storing them in a graph database after knowledge fusion and disambiguation to form a coal mine safety knowledge graph.

[0029] For example, the specific composition and quantity distribution of unstructured text data can be seen in Table 1. These data cover multiple levels from theoretical mechanisms to practical cases, providing a rich corpus foundation for constructing high-quality knowledge graphs.

[0030] Table 1

[0031] Specifically, S102 above includes: Based on the disaster chain theory, core entities and semantic relationships are defined to form an ontology model; the core entities include disaster-causing factors, disaster categories, disaster manifestations, monitoring indicators, and prevention and control measures. Unstructured text data is vectorized and segmented into multiple text blocks. Based on the ontology model, the LightRAG framework is used to call the large language model to perform semantic parsing on each text block in order to extract entities and relations and generate corresponding descriptive text, forming enhanced triples. The system performs deduplication and merging of entities pointing to the same real-world object in different enhanced triples, and performs semantic fusion on multiple descriptive texts corresponding to the same entity after deduplication and merging. After semantic fusion, the knowledge is stored in a graph database to form a coal mine safety knowledge graph.

[0032] Based on the disaster chain theory, an ontology model is defined, encompassing a five-layer structure: "causing factor - disaster category - disaster manifestation - monitoring indicators - prevention and control measures," and the entity types and semantic relationships are standardized. See also... Figure 3 The coal mine disaster classification system, based on the disaster chain theory, divides coal mine disasters into natural dynamic disasters (natural internal dynamics and natural external dynamics) and anthropogenic external dynamics. It encompasses five major coal mine disasters: mine flooding, surrounding rock disasters, gas disasters, dust disasters, and mine fires, clearly presenting the hierarchical relationship between disaster-causing factors, disaster categories, and disaster manifestations, and corresponding to their direct manifestations. This ontology model serves as the top-level architecture for knowledge organization during knowledge graph construction. On the one hand, it provides clear target types and semantic constraints for subsequent large language models to extract entities and relations from text blocks; on the other hand, it provides a unified data model for knowledge fusion and graph storage of the extracted results, enabling the semantic integration and association of knowledge extracted from different sources and different text blocks within a unified ontology framework.

[0033] In the process of knowledge graph construction, unstructured text data can be preprocessed. Text encoding models are used to encode documents such as literature, accident investigation reports, and safety regulations into dense semantic vectors. The text is then segmented according to preset word lengths and overlap ratios, dividing long texts into several text blocks suitable for model processing, thus constructing a semantic corpus. For example, the BAAI / bge-m3 model can be used to encode text data into 1024-dimensional vectors, and segmented into blocks with a word length of 512 and an overlap rate of 10%. The sources of unstructured text data suitable for knowledge graph construction are extensive, specifically covering multiple categories such as authoritative literature, accident case reports, and coal mine safety regulations, providing a rich corpus foundation for knowledge extraction.

[0034] After text segmentation, the LightRAG framework is used to perform semantic parsing on each text block. The LightRAG framework can invoke a large language model to automatically extract entities and relations from the text blocks under the constraints and guidance of an ontology model, generating corresponding descriptive text for each entity and relation, thus forming enhanced triples. The large language model is an AI model with semantic understanding and text generation capabilities, such as the RWKV (Receptance Weighted KeyValue) model. The descriptive text provides a natural language explanation of the semantic connotations of entities or relations, endowing knowledge graph nodes and edges with richer semantic information, which helps improve the accuracy and interpretability of subsequent knowledge retrieval and reasoning.

[0035] Since unstructured text data originates from multiple sources, the descriptions of the same entity or relationship may differ, be redundant, or overlap. Therefore, after extracting a large number of enhanced triples, knowledge fusion processing is required on these extraction results from multiple text sources. A MapReduce incremental fusion strategy can be employed, comparing newly extracted entities with existing entities in a graph database (such as Neo4j), merging duplicate entities pointing to the same real-world object, and semantically fusing and refining multiple descriptive texts associated with the merged entity. This eliminates data redundancy and semantic conflicts, improving the consistency and accuracy of knowledge. Storing the semantically fused knowledge in a graph database forms a coal mine safety knowledge graph.

[0036] For example, the constructed knowledge graph can contain approximately 2,200 entities and 5,500 relationships. The graph covers the complete disaster chain path through multi-hop path traversal. The retrieval fidelity is evaluated using the RAGAS framework. The evaluation experiment selected 8 typical disaster cases, with 5 representative accident cases for each disaster type, for a total of 40 evaluation samples. The evaluation results are shown in Table 2, with an average fidelity of 0.883.

[0037] Table 2

[0038] See Figure 4 In the constructed coal mine safety knowledge graph network structure, blue nodes represent spatial locations, orange nodes represent disaster manifestations, red nodes represent key indicators, and arrows represent secondary disaster chains, intuitively presenting the network of connections between disaster entities, relationships, and indicators.

[0039] S103: The LSTM network is trained using structured time-series monitoring data to obtain a prediction model, and the anomaly perception features are obtained by analyzing the prediction residuals of the prediction model.

[0040] To address the issues of low intelligence and difficulty in automatically identifying abnormal states from massive monitoring data in existing technologies, this invention trains an LSTM network using acquired structured time-series monitoring data to construct a prediction model capable of capturing the normal evolution of the downhole environment, and extracts abnormal perception features based on the prediction residuals of this model.

[0041] The prediction model is obtained by training an LSTM network using structured time-series monitoring data, specifically including: The structured time-series monitoring data is processed by sliding window, standardized, and augmented to obtain training samples; A stacked LSTM network is constructed, and the stacked LSTM network is trained with training samples to perform multi-index synchronous prediction training to obtain a prediction model.

[0042] Before training the model, the structured time-series monitoring data can be preprocessed. For example, sliding sampling can be performed on each monitoring indicator with a 10-minute window and a 5-minute step size, Z-score standardization can be used, and Gaussian noise with a standard deviation of 0.05 can be injected to enhance the robustness of the model. To address the problem of sparse outlier samples, an oversampling strategy can be used to balance the dataset.

[0043] By sampling various monitoring indicators through a sliding window with a preset time window and sliding step size, continuous time-series data is transformed into multiple sets of input-output sample pairs, enabling the model to learn the mapping relationship from historical observations to future states. Based on this, the sampled data is standardized, for example using Z-score standardization, to eliminate the influence of differences in units and numerical ranges between different monitoring indicators. Simultaneously, to enhance the model's generalization ability and robustness, data augmentation can be performed on the training data, such as injecting Gaussian noise of a certain intensity into the standardized data. For the data imbalance problem where abnormal samples are far fewer than normal samples in actual working conditions, an oversampling strategy can be used to balance the distribution of normal and abnormal samples in the dataset, preventing the model from biased towards learning normal working conditions and weakening its sensitivity to abnormal patterns.

[0044] In terms of model construction, a multi-indicator synchronous prediction model is built based on an LSTM network. For example, a two-layer LSTM network with 128 hidden units per layer can be constructed to perform multi-step synchronous prediction of 12 monitoring indicators. Model training aims to minimize the error between the predicted and true values, and the network parameters are iteratively updated using the backpropagation algorithm until the model converges. After training, the model's average R-value on the test set is [value missing]. 2 The value reached 0.9693, indicating that the model has excellent fitting and predictive capabilities for monitoring data under normal operating conditions.

[0045] After obtaining the trained prediction model, it is used to predict real-time monitoring data, and the residuals between the predicted and actual observed values ​​are calculated. When abnormal fluctuations occur in the downhole environment, the actual observed values ​​deviate from the normal trajectory predicted by the model, manifested as an abnormal increase or intensified fluctuations in the prediction residuals. Therefore, the prediction residuals contain key information that can significantly distinguish between normal and abnormal operating conditions. By performing statistical analysis on the prediction residuals, the original sensor monitoring signals are transformed into more discernible anomaly perception features, thus providing a quantitative basis for subsequent risk warnings. The Mann-Whitney U test verifies that there is a significant statistical difference (p<0.001) between the residual features of normal and abnormal periods, confirming the effectiveness of the anomaly perception features extracted based on the prediction residuals in reflecting abnormal system states.

[0046] Specifically, anomaly detection features are obtained by analyzing the prediction residuals of the prediction model, including: Calculate the prediction residuals of the prediction model and standardize the prediction residuals. Based on the standardized prediction residuals, the mean residual, residual volatility, and residual trend are extracted as anomaly detection features.

[0047] Among them, the mean residual is the average value of the standardized residuals within the prediction time window, reflecting the overall magnitude of the systematic deviation of the monitoring indicator from the normal level during this period; the residual volatility describes the degree of fluctuation of the residual within the prediction time window, reflecting the instability characteristics of the abnormal state; and the residual trend is the trend term obtained by fitting the residual sequence with linear regression, reflecting the direction and speed of the abnormal evolution, that is, whether the abnormal state is continuously aggravated, tending to ease, or remaining stable.

[0048] S104: Dynamically query the knowledge graph based on anomaly perception features to retrieve associated disaster chain paths and entities, and encode the retrieval results as knowledge enhancement features.

[0049] In the application, the coal mine safety knowledge graph is dynamically queried based on the anomaly perception characteristics. The disaster chain paths and entities associated with the current abnormal state are retrieved from the knowledge graph, and the retrieval results are encoded as knowledge-enhanced features, thereby transforming static domain knowledge into feature representations that are dynamically matched with real-time working conditions.

[0050] Specifically, S104 above includes: Monitoring indicators that exceed the preset anomaly threshold in the anomaly perception features are identified as key anomaly indicators. Based on key anomaly indicators, three types of query statements are dynamically generated, targeting background risk priors, anomaly indicator correlations, and spatial location risks. Execute the query statement to retrieve disaster chain paths and entities associated with key anomaly indicators from the knowledge graph, and obtain the subgraph structure; The topological information of the subgraph structure is encoded into a fixed-dimensional knowledge-enhancing feature vector, which serves as the knowledge-enhancing feature.

[0051] When implementing dynamic knowledge query, the first step is to screen and determine abnormal indicators based on the abnormal perception features obtained in S103. The mean residuals of each monitoring indicator are iterated through. When the mean residual of a certain monitoring indicator exceeds a preset abnormal threshold, that indicator is determined as a key abnormal indicator at the current moment and included in the abnormal indicator set. The preset abnormal threshold can be determined based on statistical criteria and a validation set grid search; for example, the threshold can be set to 2.0 based on the 2σ criterion.

[0052] After identifying key anomaly indicators, Cypher queries for the graph database are dynamically generated based on specific combinations of indicators in the anomaly indicator set. The query statements are not constructed using a fixed template but are dynamically generated based on the types, quantities, and combinations of key anomaly indicators at the current moment, ensuring that the query results are closely related to the current actual working conditions. Specifically, three types of queries can be constructed: those for background risk priors, anomaly indicator associations, and spatial location risks. Background risk prior queries aim to retrieve knowledge related to the mine's inherent risk factors from the knowledge graph, such as querying potential disaster-causing factors and disaster susceptibility types based on the mine's existing geological structure characteristics and mining conditions. Anomaly indicator association queries aim to retrieve the disaster manifestations and disaster categories associated with the current key anomaly indicators in the knowledge graph, tracing the causal transmission relationship from the anomaly indicator to the corresponding disaster type along the disaster chain path. Spatial location risk queries aim to combine the underground deployment locations of sensors corresponding to the anomaly indicators to retrieve spatial information such as roadway layout, adjacent working faces, and geological structures in the area, assessing the spatial impact range of the anomaly and potential disaster risks. By combining the three types of queries mentioned above, knowledge information that is related to the current abnormal state in three dimensions—background prior, indicator association, and spatial location—can be retrieved from the knowledge graph.

[0053] Among these, queries targeting prior knowledge of background risks require comparing real-time monitoring indicators with risk level standards in expert knowledge. To this end, the system uses "environmental indicator grading standards" as a keyword and extracts threshold provisions from the "Coal Mine Safety Regulations" and related technical specifications from the knowledge base using the LightRAG framework. This is then combined with background descriptions of the current mine's geological conditions and mining processes to define risk level ranges for key indicators such as temperature, wind speed, dust, methane, and microseismic energy. See Table 3, which provides an example of a four-level grading standard for each environmental indicator, ranging from "low risk / normal" to "extremely high risk." This grading result, as prior knowledge, is dynamically embedded into the query statement to retrieve the hazard-causing factors, disaster manifestations, and prevention measures corresponding to the current risk level, thereby achieving personalized risk assessment for different mine engineering backgrounds.

[0054] Table 3

[0055] After obtaining the retrieved knowledge, the constructed query statement is executed to retrieve entities, relationships, and their descriptive texts associated with key anomaly indicators from the graph database, forming a subgraph structure highly relevant to the current anomaly situation. This subgraph structure contains various entities, such as disaster-causing factors, disaster categories, and disaster manifestations, that can be reached along the disaster chain path starting from the key anomaly indicators, as well as semantic relationships such as causality and association between entities, and is accompanied by descriptive text for each entity and relationship. This reveals the types of disasters that the current anomaly state may trigger and their causal transmission chains, providing a structured knowledge view for understanding the disaster-causing mechanisms behind the anomaly.

[0056] When transforming the retrieved knowledge into a feature representation that can be processed by a neural network, the topological information of the aforementioned subgraph structure is encoded and mapped into a fixed-dimensional knowledge-enhanced feature vector; for example, encoded as a fixed-length 24-dimensional knowledge feature vector. The encoding of topological information may include structural features such as the distribution of the number of different types of disaster entities in the subgraph and the sum of the strengths of associated paths. By organizing these graph-level statistics into fixed-length numerical vectors, the symbolic domain knowledge that originally existed in the form of graph data is transformed into a numerical feature representation suitable for neural network input.

[0057] S105: Semantically encode the background description text of the coal mine to obtain text semantic features, and use a prediction model to extract the temporal evolution features of the structured time-series monitoring data.

[0058] For extracting semantic features from the text, the coal mine background description text is input into a text encoder for semantic encoding to obtain its vector representation. The coal mine background description text includes descriptive documents such as mine geological conditions, mining technology, and roadway layout. These texts, in natural language form, record the inherent risk characteristics and engineering background of the mine itself, serving as an important basis for personalized risk assessment. The text encoder can employ a pre-trained language model to convert variable-length text descriptions into dense semantic vectors of fixed dimensions. For example, the BAAI / bge-m3 model can be used as the text encoder. After encoding the coal mine background description text, a dedicated projection layer maps it to a text semantic feature vector of a preset dimension (e.g., 256 dimensions). This feature vector condenses the core semantic information of the coal mine engineering background in a dense numerical form.

[0059] Specifically, predictive models are used to extract the temporal evolution features of structured time-series monitoring data, including: Input the structured time-series monitoring data within the preset time window before the current moment into the prediction model; Extract the hidden state of the last time step of the prediction model; The hidden states are mapped to fixed-dimensional temporal evolution feature vectors through fully connected layers, which serve as temporal evolution features.

[0060] For extracting temporal evolution features, structured temporal monitoring data from the previous preset time window (e.g., 30 minutes) is input into the LSTM prediction model in S103, which has been trained and has its parameters frozen. The hidden state of the last time step of this LSTM prediction model is taken. This hidden state is the comprehensive memory information retained by the LSTM network after processing the input data from all times within the time window, containing the dynamic changes and evolution trends of the monitoring data within that preset time window. Subsequently, this hidden state is feature-mapped through a fully connected layer, outputting a fixed-dimensional (e.g., 128-dimensional) temporal evolution feature vector, which serves as the temporal evolution feature.

[0061] S106: After concatenating the text semantic features, temporal evolution features, anomaly perception features, and knowledge enhancement features, the data is input into a multi-layer fully connected neural network to obtain a multi-dimensional vector.

[0062] In this multidimensional vector, each dimension represents the probability of a coal mine disaster occurring.

[0063] After obtaining four types of feature vectors that characterize the underground safety status from different modalities and dimensions—textual semantic features, temporal evolution features, anomaly perception features, and knowledge enhancement features—this invention constructs a multi-source heterogeneous information fusion model to deeply fuse these four types of features and directly outputs the probability of occurrence of various coal mine disasters through a deep neural network.

[0064] First, the four types of feature vectors are concatenated along their feature dimensions to form a unified fused feature vector. This concatenation operation directly integrates heterogeneous features from different processing channels at the dimensional level, preserving the complete information of each modality's features and providing a foundation for the subsequent network to learn higher-order interaction relationships between different features. For example, the dimensions of the feature vectors involved in the concatenation can be: 256-dimensional text semantic feature vector, 128-dimensional temporal evolution feature vector, 36-dimensional anomaly perception feature vector, and 24-dimensional knowledge enhancement feature vector.

[0065] Subsequently, the concatenated fused feature vector is input into a multi-layer fully connected neural network, where information compression and feature cross-validation are performed through multi-layer nonlinear transformations. This network can automatically learn the complex interaction patterns between features of different modalities, enabling deep fusion of the static engineering background represented by textual semantic features, the dynamic trends implied by temporal evolution features, the real-time abnormal signals reflected by anomaly perception features, and the domain causal knowledge provided by knowledge enhancement features in the high-level representation of the network.

[0066] For example, the multi-layer fully connected neural network may include a first fully connected layer and a ReLU activation function, a second fully connected layer and a ReLU activation function, which sequentially perform nonlinear transformation and information compression on the fused feature vector to gradually extract higher-level abstract representations.

[0067] In the output layer of the multi-layer fully connected neural network, the deep features are mapped to a specific disaster category space using the sigmoid activation function, ultimately generating a multi-dimensional vector. Each dimension of this multi-dimensional vector corresponds to the probability of occurrence of a specific coal mine disaster, enabling concurrent probability prediction for multiple types of disasters. For example, this multi-dimensional vector can be an 8-dimensional probability vector, corresponding to the probability of occurrence of rockburst, roof collapse, gas explosion, coal and gas outburst, mine fire, mine water hazard, dust hazard, and thermal hazard. Through this method, multi-source heterogeneous features originally scattered across different processing channels are fused into an end-to-end disaster probability output via a unified multi-layer fully connected neural network.

[0068] S107: When the probability of any coal mine disaster exceeds the corresponding preset threshold, a corresponding dynamic early warning is triggered.

[0069] After a multi-dimensional vector is output from a multi-layer fully connected neural network, each dimension of this vector quantifies the probability of occurrence of different types of coal mine disasters under the current conditions. To transform the probability prediction results into actionable early warning signals, this invention achieves dynamic early warning for multiple disasters by comparing the probability of occurrence of each disaster with a corresponding preset threshold.

[0070] In practical applications, different preset thresholds can be set for different types of coal mine disasters, either using the same preset threshold or different preset thresholds for each. The preset thresholds for different disasters can be configured independently based on their severity, frequency of occurrence, and historical statistical characteristics. For example, the preset threshold for rockburst can be set to 0.4, and the preset threshold for gas explosion can be set to 0.45. The preset thresholds for different disasters can also be initially set based on statistical analysis of historical accident data or the experience of domain experts, and can be dynamically corrected and optimized based on accumulated field monitoring data and early warning feedback results. For example, the initial preset threshold for all disasters can be temporarily set to 0.3, and the optimal preset threshold for each disaster can be gradually adjusted by comparing historical early warning records with actual disaster occurrences. The multidimensional vector output is traversed dimension by dimension. When the probability of any disaster occurring exceeds the preset threshold corresponding to that disaster, it is determined that the risk level of that type of disaster has significantly increased in the current state, triggering a dynamic early warning for that specific disaster.

[0071] The presentation of early warning information can go beyond simple probability values; it can also integrate key reasoning evidence generated in the preceding steps. For example, early warning information can include the type of disaster, the probability of its occurrence, key correlation anomalies from the knowledge-enhanced feature retrieval process, and relevant risk knowledge retrieved from the knowledge graph. In this way, the early warning results not only provide a probabilistic judgment at the numerical level but also offer the causal reasoning chain and domain knowledge supporting that judgment. This enhances the interpretability of the early warning results and helps on-site safety management personnel quickly understand the source of the risk and take targeted preventative measures after receiving the early warning signal.

[0072] Through the complete technical process described above, from multi-source heterogeneous data acquisition, knowledge graph construction, anomaly perception and knowledge retrieval, to multi-source feature encoding fusion, joint disaster probability prediction, and finally dynamic early warning output, this invention realizes intelligent prediction and decision support for the coupled risks of multiple disasters in underground coal mines, and enhances the early warning capability for the coupled risks of multiple disasters in complex and ever-changing underground environments.

[0073] Corresponding to the aforementioned application function implementation method embodiments, the present invention also provides an underground coal mine production risk early warning system and corresponding embodiments.

[0074] Please see Figure 5 , Figure 5 This is a schematic diagram of the module structure of a production risk early warning system for underground coal mines.

[0075] The underground coal mine production risk early warning system includes: The data acquisition module 51 is used to acquire multi-source heterogeneous data in the field of coal mine safety; the multi-source heterogeneous data includes structured time-series monitoring data, unstructured text data and coal mine background description text; Multi-source feature generation module 52 is used for: Based on the LightRAG framework, entities and relationships are extracted from unstructured text data to construct a coal mine safety knowledge graph; The LSTM network is trained using structured time-series monitoring data to obtain a prediction model, and the anomaly perception features are obtained by analyzing the prediction residuals of the prediction model. Based on anomaly detection features, the knowledge graph is dynamically queried to retrieve associated disaster chain paths and entities, and the retrieval results are encoded as knowledge-enhancing features. Semantic encoding is performed on the background description text of the coal mine to obtain text semantic features, and a prediction model is used to extract the temporal evolution features of the structured time-series monitoring data; The disaster probability prediction module 53 is used to concatenate text semantic features, temporal evolution features, anomaly perception features and knowledge enhancement features, and then input them into a multi-layer fully connected neural network to obtain a multi-dimensional vector; each dimension in the multi-dimensional vector represents the probability of occurrence of a coal mine disaster. The early warning module 54 is used to trigger a corresponding dynamic early warning when the probability of any coal mine disaster exceeds the corresponding preset threshold.

[0076] In one embodiment, when extracting entities and relationships from unstructured text data based on the LightRAG framework to construct a coal mine safety knowledge graph, the multi-source feature generation module 52 is specifically used for: Based on the disaster chain theory, core entities and semantic relationships are defined to form an ontology model; the core entities include disaster-causing factors, disaster categories, disaster manifestations, monitoring indicators, and prevention and control measures. Unstructured text data is vectorized and segmented into multiple text blocks. Based on the ontology model, the LightRAG framework is used to call the large language model to perform semantic parsing on each text block in order to extract entities and relations and generate corresponding descriptive text, forming enhanced triples. The system performs deduplication and merging of entities pointing to the same real-world object in different enhanced triples, and performs semantic fusion on multiple descriptive texts corresponding to the same entity after deduplication and merging. After semantic fusion, the knowledge is stored in a graph database to form a coal mine safety knowledge graph.

[0077] In one embodiment, when training an LSTM network using structured time-series monitoring data to obtain a prediction model, the multi-source feature generation module 52 is specifically used for: The structured time-series monitoring data is processed by sliding window, standardized, and augmented to obtain training samples; A stacked LSTM network is constructed, and the stacked LSTM network is trained with training samples to perform multi-index synchronous prediction training to obtain a prediction model.

[0078] In one embodiment, when obtaining anomaly perception features by analyzing the prediction residuals of the prediction model, the multi-source feature generation module 52 is specifically used for: Calculate the prediction residuals of the prediction model and standardize the prediction residuals. Based on the standardized prediction residuals, the mean residual, residual volatility, and residual trend are extracted as anomaly detection features.

[0079] In one embodiment, when dynamically querying the knowledge graph based on anomaly perception features to retrieve associated disaster chain paths and entities, and encoding the retrieval results as knowledge-enhancing features, the multi-source feature generation module 52 is specifically used for: Monitoring indicators that exceed the preset anomaly threshold in the anomaly perception features are identified as key anomaly indicators. Based on key anomaly indicators, three types of query statements are dynamically generated, targeting background risk priors, anomaly indicator correlations, and spatial location risks. Execute the query statement to retrieve disaster chain paths and entities associated with key anomaly indicators from the knowledge graph, and obtain the subgraph structure; The topological information of the subgraph structure is encoded into a fixed-dimensional knowledge-enhancing feature vector, which serves as the knowledge-enhancing feature.

[0080] In one embodiment, when extracting the temporal evolution features of structured time-series monitoring data using a prediction model, the multi-source feature generation module 52 is specifically used for: Input the structured time-series monitoring data within the preset time window before the current moment into the prediction model; Extract the hidden state of the last time step of the prediction model; The hidden states are mapped to fixed-dimensional temporal evolution feature vectors through fully connected layers, which serve as temporal evolution features.

[0081] Regarding the system in the above embodiments, the specific manner in which each unit module performs operations has been described in detail in the embodiments related to the method, and will not be elaborated further here.

[0082] In one exemplary embodiment, the present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments. The computer device may be a server or a terminal, and its internal structure diagram may be as follows: Figure 6As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data to be processed. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network. When the computer program is executed by the processor, it implements a method for constructing an intelligent expert knowledge base and providing risk early warning in underground coal mines.

[0083] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0084] In one exemplary embodiment, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0085] In one exemplary embodiment, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0086] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0087] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0088] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0089] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0090] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for early warning of production risks in a underground coal mine, characterized in that, include: Acquire multi-source heterogeneous data in the field of coal mine safety; the multi-source heterogeneous data includes structured time-series monitoring data, unstructured text data, and coal mine background description text; Based on the LightRAG framework, entities and relationships are extracted from the unstructured text data to construct a coal mine safety knowledge graph; The LSTM network is trained using the structured time-series monitoring data to obtain a prediction model, and anomaly perception features are obtained by analyzing the prediction residuals of the prediction model. The knowledge graph is dynamically queried based on the anomaly perception features to retrieve associated disaster chain paths and entities, and the retrieval results are encoded as knowledge enhancement features. Semantic encoding is performed on the coal mine background description text to obtain text semantic features, and the temporal evolution features of the structured time-series monitoring data are extracted using the prediction model. The text semantic features, temporal evolution features, anomaly perception features, and knowledge enhancement features are concatenated and then input into a multi-layer fully connected neural network to obtain a multi-dimensional vector. Each dimension in the multidimensional vector represents the probability of a coal mine disaster occurring; When the probability of any coal mine disaster exceeds the corresponding preset threshold, a dynamic early warning is triggered.

2. The method according to claim 1, c h a r a c t e r i z e d b y, Based on the LightRAG framework, entities and relationships are extracted from the unstructured text data to construct a coal mine safety knowledge graph, including: Based on the disaster chain theory, core entities and semantic relationships are defined to form an ontology model; the core entities include disaster-causing factors, disaster categories, disaster manifestations, monitoring indicators, and prevention and control measures. The unstructured text data is vectorized and segmented into text blocks to obtain multiple text blocks; Based on the ontology model, the LightRAG framework is used to call the large language model to perform semantic parsing on each text block in order to extract entities and relations and generate corresponding descriptive text, forming enhanced triples. The system performs deduplication and merging of entities pointing to the same real-world object in different enhanced triples, and performs semantic fusion on multiple descriptive texts corresponding to the same entity after deduplication and merging. After semantic fusion, the knowledge is stored in a graph database to form a coal mine safety knowledge graph.

3. The method according to claim 1, c h a r a c t e r i z e d b y, The prediction model is obtained by training an LSTM network using the structured time-series monitoring data, including: The structured time-series monitoring data is subjected to sliding window processing, standardization, and data augmentation to obtain training samples; A stacked LSTM network is constructed, and the training samples are used to train the stacked LSTM network for multi-index synchronous prediction to obtain a prediction model.

4. The method according to claim 1, c h a r a c t e r i z e d b y, Anomaly detection features are obtained by analyzing the prediction residuals of the prediction model, including: Calculate the prediction residuals of the prediction model and standardize the prediction residuals. Based on the standardized prediction residuals, the mean residual, residual volatility, and residual trend are extracted as anomaly detection features.

5. The method according to claim 1, c h a r a c t e r i z e d b y The knowledge graph is dynamically queried based on the anomaly detection features to retrieve associated disaster chain paths and entities, and the retrieval results are encoded as knowledge enhancement features, including: The monitoring indicators that exceed the preset abnormal threshold among the abnormality perception features are identified as key abnormality indicators. Based on the aforementioned key anomaly indicators, three types of query statements are dynamically generated, targeting background risk priors, anomaly indicator correlations, and spatial location risks. Execute the query statement to retrieve disaster chain paths and entities associated with the key anomaly indicators from the knowledge graph, and obtain a subgraph structure; The topological structure information of the subgraph structure is encoded into a fixed-dimensional knowledge-enhancing feature vector, which serves as the knowledge-enhancing feature.

6. The underground coal mine production risk early warning method according to claim 1, characterized in that, The prediction model is used to extract the temporal evolution features of the structured time-series monitoring data, including: The structured time-series monitoring data within a preset time window before the current moment is input into the prediction model; Extract the hidden state of the last time step of the prediction model; The hidden state is mapped to a fixed-dimensional temporal evolution feature vector through a fully connected layer, which serves as the temporal evolution feature.

7. The underground coal mine production risk early warning method according to claim 1, characterized in that, The structured time-series monitoring data includes at least one of the following: temperature, wind speed, dust concentration, oxygen concentration, methane concentration, carbon dioxide concentration, microseismic energy, roof pressure, roof displacement, water pressure, water level, and water temperature.

8. A risk early warning system for underground coal mine production, characterized in that, include: The data acquisition module is used to acquire multi-source heterogeneous data in the field of coal mine safety; the multi-source heterogeneous data includes structured time-series monitoring data, unstructured text data, and coal mine background description text. The multi-source feature generation module is used for: Based on the LightRAG framework, entities and relationships are extracted from the unstructured text data to construct a coal mine safety knowledge graph; The LSTM network is trained using the structured time-series monitoring data to obtain a prediction model, and anomaly perception features are obtained by analyzing the prediction residuals of the prediction model. The knowledge graph is dynamically queried based on the anomaly perception features to retrieve associated disaster chain paths and entities, and the retrieval results are encoded as knowledge enhancement features. Semantic encoding is performed on the coal mine background description text to obtain text semantic features, and the temporal evolution features of the structured time-series monitoring data are extracted using the prediction model. The disaster probability prediction module is used to concatenate the text semantic features, temporal evolution features, anomaly perception features and knowledge enhancement features, and then input them into a multi-layer fully connected neural network to obtain a multi-dimensional vector; Each dimension in the multidimensional vector represents the probability of a coal mine disaster occurring; The early warning module is used to trigger a corresponding dynamic early warning when the probability of any coal mine disaster exceeds the corresponding preset threshold.

9. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the underground coal mine production risk early warning method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the underground coal mine production risk early warning method as described in any one of claims 1-7.

Images