Intelligent coal mine management system based on multiple agents

Abstract

The invention relates to the technical field of intelligent management, in particular to a coal mine intelligent management system based on multiple agents, which comprises a graph construction module, a mapping module, a mapping module and a mapping module, a semantic training module is utilized to construct a coal mine mapping model with aligned time sequence and text on the basis of a fine-grained contrast learning strategy of a significance mask, a double-layer agent cooperation mechanism triggered by a monitoring query module is combined, and a standard query agent maps real-time sensing monitoring data into natural language fault semantics and retrieves associated security constraints. And the reasoning agent performs logic comparison based on the large language model to generate a control instruction and issues and executes the control instruction, so that automatic decision and response of coal mine safety management are realized. According to the invention, automatic decision and response of coal mine safety management are realized through multiple agents.

Description

Technical Field

[0001] This invention relates to the field of intelligent management technology, specifically to a coal mine intelligent management system based on multi-agent systems. Background Technology

[0002] Modern large-scale coal mines have generally established automated control platforms based on programmable logic controllers and data acquisition, monitoring and control systems, and aggregated data to ground dispatch centers for centralized management. Some have even built regional coal mine big data platforms with a capacity of hundreds of millions of tons, connecting thousands of devices and tens of thousands of collection points, realizing the construction of full-domain perception and data foundation.

[0003] However, existing coal mine management systems still have significant limitations in intelligent decision-making. First, the traditional centralized control framework relies on pre-programmed static logic rules. This rigid control mode is difficult to cope with dangerous and rapidly changing unstructured environments. Although management algorithms have been introduced, they are mostly limited to single-point fault alarms or simple threshold judgments, lacking the ability to make logical reasoning in complex scenarios.

[0004] Meanwhile, coal mine industry management systems need to possess relevant professional knowledge and implicit troubleshooting experience formed through long-term practice. Currently, most of these key decision-making bases exist in unstructured forms such as PDF documents and paper records. Existing digital systems cannot directly read, understand, and transform them into executable management strategies, resulting in a disconnect between data and knowledge. This still requires secondary analysis and decision-making by manual dispatchers, which seriously restricts production efficiency and response speed.

[0005] To address this, a coal mine intelligent management system based on multi-agent systems is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a coal mine intelligent management system based on multiple agents, which enables intelligent management of coal mines through multiple agents.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A multi-agent-based intelligent management system for coal mines includes:

[0009] Knowledge graph construction module: acquires unstructured multimodal historical data of coal mines, processes the historical data using a hybrid processing unit, and generates a structured knowledge graph of coal mines;

[0010] Semantic training module: includes a temporal encoder and a text encoder; the temporal encoder and text encoder are jointly trained using a fine-grained contrastive learning strategy based on saliency masks to generate saliency masks; based on the saliency masks, a fine-grained cross-correlation matrix between temporal local features and text local features is calculated to construct a coal mine mapping model;

[0011] Monitoring and query module: Real-time collection of sensor monitoring data from the coal mine production site, triggering a two-layer intelligent agent collaboration mechanism, including a standard query intelligent agent and an inference intelligent agent; the standard query intelligent agent uses the coal mine mapping model to map the sensor monitoring data into natural language fault semantics, and retrieves the associated coal mine safety constraints in the coal mine knowledge graph;

[0012] Inference control module: The inference agent receives the natural language fault semantics and the coal mine safety constraints, generates coal mine control instructions, and issues them for execution.

[0013] Preferably, the hybrid processing unit includes a cleaning subunit and an extraction subunit. The cleaning subunit performs noise filtering on the historical data, uses a wavelet threshold denoising algorithm to smooth the electromagnetic interference noise of the sensor monitoring data, and uses a coal mine-specific terminology library to perform spelling correction and standardization on the text data.

[0014] The extraction subunit uses BERT-BiLSTM-CRF to extract entities and their relationships, identifies the association between equipment entities and faults from the text, and maps the measurement point IDs of the sensor monitoring data as attributes to the corresponding equipment entities, generating a coal mine knowledge graph containing a topology structure of equipment-sensor-fault.

[0015] Preferably, the coal mine knowledge graph consists of entity nodes and relation edges;

[0016] The entity nodes include device entity nodes, sensor measurement point nodes, fault semantic nodes, and safety constraint nodes.

[0017] The relational edges include monitoring mapping edges connecting the device entity nodes and sensor measurement point nodes, fault association edges connecting the device entity nodes and fault semantic nodes, and procedure constraint edges connecting the fault semantic nodes and safety constraint nodes; the safety constraint nodes store numerical thresholds, alarm levels, and forced shutdown command attributes.

[0018] Preferably, both the temporal encoder and the text encoder include a feature extraction layer and a nonlinear projection head;

[0019] The temporal encoder is constructed based on a multi-layer dilated causal convolutional network. It extracts multi-scale temporal features from sensor data through convolutional kernels with different dilation rates, preserves local contextual information in the time dimension, and generates a temporal feature sequence.

[0020] The text encoder is built based on a pre-trained Transformer architecture, extracts word embedding vectors from the fault description text, and combines them with positional encoding to generate a text feature sequence containing semantic dependencies.

[0021] The nonlinear projection head maps the temporal feature sequence and the text feature sequence to the same dimensional metric space.

[0022] Preferably, the multilayer dilated causal convolutional network consists of N stacked temporal residual blocks; each temporal residual block contains two parallel paths, the main path containing a dilated causal convolutional layer, a weight normalization layer and an activation function layer connected in sequence, and the residual path is used for gradient propagation;

[0023] The expansion rate of the dilated causal convolutional layer increases exponentially with the increase of the number of network layers, thus constructing a multi-scale time receptive field that can cover the instantaneous cutting impact of coal mining equipment from the millisecond level to the mechanical wear trend of the hour level.

[0024] The main path filters downhole environmental background noise through a gating mechanism and prioritizes the extraction of impulse response features related to equipment failure.

[0025] Preferably, the specific process of jointly training the temporal encoder and the text encoder using a fine-grained contrastive learning strategy based on saliency masks is as follows: the contribution of each time step in the temporal feature sequence to fault classification is calculated using a gradient backpropagation mechanism to generate a saliency mask, and at the same time, the text saliency mask is extracted according to the self-attention weight distribution of the text encoder.

[0026] The cosine similarity between each pair of elements in the temporal feature sequence and the text feature sequence is calculated to construct an original similarity matrix. The temporal saliency mask and the text saliency mask are then multiplied by an outer product to generate a filter matrix. The filter matrix and the original similarity matrix are then multiplied by a Hadamard product to generate a fine-grained cross-correlation matrix. The fine-grained cross-correlation matrix is ​​then pooled to obtain a global similarity score representing the matching degree of sample pairs. The contrastive loss function is used to maximize the global similarity score of positive sample pairs. The encoder parameters are then updated through backpropagation.

[0027] Preferably, the two-layer intelligent agent collaboration mechanism includes a standard query intelligent agent and an inference intelligent agent. The standard query intelligent agent has a built-in vector retrieval engine and a graph inference engine. It uses the time encoder in the coal mine mapping model to extract the high-dimensional feature vector of the sensing monitoring data, and uses an approximate nearest neighbor search algorithm to calculate the cosine similarity between the high-dimensional feature vector and the fault semantic node vector in the coal mine knowledge graph. Based on the similarity score, the top-K candidate fault semantics are selected from high to low.

[0028] For each candidate fault semantic, a breadth-first search strategy is used to traverse the path along the procedure constraint edges in the coal mine knowledge graph to obtain the associated coal mine safety constraints, and the path confidence score is calculated. The fault semantic with the highest confidence score is combined with the safety constraint and output to the reasoning agent.

[0029] Preferably, the reasoning agent concatenates the received natural language fault semantics, the coal mine safety constraints, and the device ID of the target equipment according to a preset format to form a decision prompt word; the decision prompt word is input into a large language model, and the logical reasoning capability of the large language model is used to compare the state value in the fault semantics with the numerical threshold in the coal mine safety constraints; when the state value exceeds the numerical threshold, the large language model directly extracts the pre-stored forced shutdown instruction attribute in the coal mine safety constraints, and fills the device ID into the reserved parameter bits of the instruction attribute, generating and issuing a coal mine control instruction for the specific equipment.

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0031] 1. This invention automatically cleans and extracts unstructured multimodal data from historical coal mine documents through a graph construction module and a hybrid processing unit. It uses BERT-BiLSTM-CRF to identify entity relationships and map measurement point IDs to construct a coal mine knowledge graph containing equipment-sensor-fault topology. This mechanism breaks down the barriers between sensor data and troubleshooting experience existing in PDF or paper form in traditional coal mine management. It transforms implicit expert knowledge into a machine-readable and searchable structured graph, providing a precise knowledge base for subsequent intelligent reasoning and significantly improving knowledge utilization.

[0032] 2. This invention employs a semantic training module that includes a temporal encoder and a text encoder. It utilizes a dilated causal convolutional network to extract multi-scale temporal features and employs a fine-grained contrastive learning strategy based on saliency masks to calculate a fine-grained cross-correlation matrix between temporal and textual features. Through this cross-modal joint training and alignment, the system can not only effectively filter environmental background noise such as downhole electromagnetic interference, but also accurately capture the deep semantic association between sensor data fluctuations and natural language fault descriptions. This enables the system to no longer be limited to simple threshold judgments in fault identification under complex working conditions, but to have the ability to understand fault semantics.

[0033] 3. This invention introduces a two-layer intelligent agent collaboration mechanism. The standard query agent uses vector retrieval and graph reasoning to accurately locate fault semantics and safety constraints, while the reasoning agent performs risk comparison based on the logical capabilities of the large language model and directly extracts the attributes of pre-stored forced shutdown instructions for parameter filling. This design combines the strong reasoning capabilities of the large model with the rigorous constraints of the knowledge graph, enabling the automatic completion of the entire process from state perception, procedure matching, logical verification to instruction issuance without human intervention. This significantly improves the response speed and decision-making safety of coal mine production when facing sudden or complex situations. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating a multi-agent-based intelligent management system for coal mines.

[0035] Figure 2 A flowchart illustrating the two-layer intelligent agent collaboration mechanism;

[0036] Figure 3 This is a schematic diagram of the structure of a coal mine intelligent management system based on multi-agent systems. Detailed Implementation

[0037] 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.

[0038] Example 1:

[0039] Please see Figure 1 This invention provides a coal mine intelligent management system based on multi-agent systems, the technical solution of which is as follows:

[0040] Knowledge graph construction module: acquires unstructured multimodal historical data of coal mines, processes the historical data using a hybrid processing unit, and generates a structured knowledge graph of coal mines;

[0041] Semantic training module: includes a temporal encoder and a text encoder; the temporal encoder and text encoder are jointly trained using a fine-grained contrastive learning strategy based on saliency masks to generate saliency masks; based on the saliency masks, a fine-grained cross-correlation matrix between temporal local features and text local features is calculated to construct a coal mine mapping model;

[0042] Monitoring and query module: Real-time collection of sensor monitoring data from the coal mine production site, triggering a two-layer intelligent agent collaboration mechanism, including a standard query intelligent agent and an inference intelligent agent; the standard query intelligent agent uses the coal mine mapping model to map the sensor monitoring data into natural language fault semantics, and retrieves the associated coal mine safety constraints in the coal mine knowledge graph;

[0043] Inference control module: The inference agent receives the natural language fault semantics and the coal mine safety constraints, generates coal mine control instructions, and issues them for execution.

[0044] The unstructured multimodal historical data includes text logs recording the historical operating status and fault repair process of underground coal mine equipment, PDF technical documents or paper scans containing equipment operation specifications and safety limits, expert troubleshooting experience database, and historical sensor monitoring time series data reflecting the physical characteristics of the equipment, such as current, voltage, temperature, vibration acceleration and hydraulic pressure, which are collected synchronously.

[0045] The hybrid processing unit includes a cleaning subunit and an extraction subunit. The cleaning subunit performs noise filtering on the historical data, uses a wavelet threshold noise reduction algorithm to smooth the electromagnetic interference noise of the sensor monitoring data, and uses a coal mine-specific terminology library to correct spelling errors and standardize the text data.

[0046] The extraction subunit uses BERT-BiLSTM-CRF to extract entities and their relationships, identifies the association between equipment entities and faults from the text, and maps the measurement point IDs of the sensor monitoring data as attributes to the corresponding equipment entities, generating a coal mine knowledge graph containing a topology structure of equipment-sensor-fault.

[0047] The hybrid processing unit first activates the cleaning subunit to classify and preprocess the acquired unstructured multimodal historical data. For historical sensor monitoring time-series data, due to the complex downhole environment, the data often contains electromagnetic interference. The cleaning subunit uses a wavelet threshold denoising algorithm for processing. This algorithm selects the Dobesie wavelet system, which has compact support and orthogonality characteristics, as the wavelet basis function, and specifically sets the decomposition level to 4 levels. The tower algorithm is then used to perform multi-scale decomposition on the original sensor monitoring data. For each level of decomposition, the approximation coefficients of the previous level are separated by high-pass and low-pass filters to obtain detail coefficients representing the high-frequency components of the signal and approximation coefficients representing the low-frequency trend of the signal, resulting in 4 sets of detail coefficients and 1 set of approximation coefficients.

[0048] The cleaning subunit performs thresholding on the four sets of detail coefficients. It calculates the standard deviation of the noise using an estimation method based on the median of the absolute values ​​of the first-layer detail coefficients, and calculates a general adaptive threshold (0.6745 in this embodiment) in conjunction with the signal length. A soft thresholding function is used to shrink the detail coefficients. When the absolute value of a detail coefficient is less than the adaptive threshold, it is set to zero to filter out pure noise components. When the absolute value of a detail coefficient is greater than or equal to the adaptive threshold, its value is shrunk towards zero, i.e., the threshold value is subtracted, to retain the main fault characteristics and eliminate noise-induced glitches, while ensuring the continuity of the reconstructed signal. The cleaning subunit uses the processed four sets of detail coefficients and the lowest-level approximation coefficients to reconstruct the signal using an inverse wavelet transform algorithm, thereby obtaining clean sensor data that smoothly removes electromagnetic interference noise while completely preserving the impact characteristics of equipment faults.

[0049] The cleaning subunit, targeting text data such as text logs, technical documents, and expert experience databases, first converts non-text formats into editable text using optical character recognition (OCR) technology, and then performs spelling correction and standardization using a coal mine-specific terminology database. This database contains standard equipment names, component models, fault terminology abbreviations, and common names used in the coal mining industry. The cleaning subunit compares the identified text terms with the terminology database, automatically corrects typos generated by OCR, and replaces non-standard equipment common names with standard technical names.

[0050] The extraction subunit utilizes a composite neural network model consisting of BERT-BiLSTM-CRF (a pre-trained language model, a bidirectional long short-term memory network, and a conditional random field) to perform the task of extracting entities and their relationships. First, the cleaned text sequence is input into the pre-trained language model, converting characters or words into dense vector representations containing contextual semantic information. This vector sequence is then input into the bidirectional long short-term memory network to extract global features and long-distance dependencies from both forward and backward directions. The extracted features are then input into the conditional random field layer, and a decoding algorithm calculates the label sequence with the highest probability, thereby accurately identifying device entities, fault entities, and the logical relationships between them in the text. The BERT-BiLSTM-CRF composite model effectively solves the challenge of recognizing entities with large spans in complex text by using BERT to capture deep semantics, combining BiLSTM to extract long-distance dependency features, and optimizing label logic through CRF.

[0051] After completing the identification of entities and relationships, the extraction subunit performs attribute mapping and topology construction, reads the measurement point identifiers of the sensor monitoring data, and forcibly maps them to the equipment entities to which they belong as attribute fields, establishing a one-to-one correspondence between physical sensors and virtual equipment entities; on this basis, based on the identified entities and their interrelationships, it creates equipment entity nodes, sensor measurement point nodes, fault semantic nodes, and safety constraint nodes, and generates monitoring mapping edges connecting equipment and sensors, fault association edges connecting equipment and faults, and procedural constraint edges connecting faults and safety rules, generating a structured coal mine knowledge graph containing the topology of equipment, sensors, and faults;

[0052] The hybrid processing unit achieves deep cleaning and fusion of multi-source data. On the one hand, it adopts a wavelet denoising algorithm based on soft thresholds to filter out electromagnetic interference in the mine while fully preserving the transient impact characteristics of equipment failure, thus ensuring signal continuity and high signal-to-noise ratio. On the other hand, it integrates the coal mine terminology database and the BERT-BiLSTM-CRF model to accurately extract entity relationships and force mapping of sensor point IDs, establishing a structured association between the underlying discrete monitoring data and the upper-level abstract expert knowledge.

[0053] The coal mine knowledge graph consists of entity nodes and relation edges;

[0054] The entity nodes include device entity nodes, sensor measurement point nodes, fault semantic nodes, and safety constraint nodes.

[0055] The relational edges include monitoring mapping edges connecting the device entity nodes and sensor measurement point nodes, fault association edges connecting the device entity nodes and fault semantic nodes, and procedure constraint edges connecting the fault semantic nodes and safety constraint nodes; the safety constraint nodes store numerical thresholds, alarm levels, and forced shutdown command attributes.

[0056] Based on the structured data output by the extraction subunit, four types of entity nodes are instantiated. The equipment entity nodes, serving as the core hub of the graph, correspond to physical assets underground in the coal mine, specifically including coal mining machines, hydraulic supports, scraper conveyors, and crushers. Each node stores a unique equipment code and model parameters. The sensor measurement point nodes correspond to sensing elements installed in key parts of the equipment, including current transformers, temperature probes, vibration accelerometers, and hydraulic pressure gauges. Each node records the measurement point location number, data acquisition frequency, and measurement range. The fault semantic nodes correspond to standardized fault descriptions extracted from historical maintenance logs, such as cutting motor overload, hydraulic system leakage, bearing temperature exceeding limits, or communication module interruption. The safety constraint nodes correspond to mandatory clauses in coal mine safety regulations and operation manuals.

[0057] Based on the logical references and physical connections between the entity nodes, three types of oriented relational edges are constructed to form a topology network; a monitoring mapping edge connecting the device entity nodes and sensor measurement point nodes is established, describing the subordinate relationship between physical devices and sensing data sources, enabling the system to index all associated real-time data streams through the device entity; a fault association edge connecting the device entity nodes and fault semantic nodes is established, based on statistical analysis of massive historical maintenance logs, representing the types of faults that specific devices have experienced or are potentially prone to during historical operation, establishing the prior association between device objects and risk patterns;

[0058] Simultaneously, a procedural constraint edge is established to connect the fault semantic node and the safety constraint node, constructing a logical bridge from the description of the fault phenomenon to the specific handling rules; when a fault semantic node is activated during the reasoning process, the system can uniquely index to the corresponding safety constraint node along the procedural constraint edge.

[0059] For the aforementioned safety constraint nodes, the system extracts key risk control indicators from technical documents and national standards in a structured manner and stores them as node attributes. The specific implementation process includes writing numerical threshold attributes to clearly define the critical values ​​between normal operation and fault states of the equipment, such as the upper limit of the alarm value for motor stator temperature or the lower limit of the initial support force of the hydraulic support; writing alarm level attributes to classify risks into four levels: general, relatively severe, serious, and extremely serious, to guide the priority ranking of alarm signals; and writing forced shutdown command attributes, which store a pre-compiled, machine-readable control code or natural language command template, for direct invocation by the reasoning agent when the monitored value exceeds the threshold and the risk level reaches the set standard.

[0060] By constructing a four-dimensional topology network that includes devices, sensors, faults, and safety constraints, a full-link logical index is realized, from physical device monitoring to fault risk warning and then to safety procedure handling. In particular, by embedding numerical thresholds and forced shutdown instructions in safety constraint nodes, static text procedures are transformed into executable knowledge that machines can directly call. This ensures that the system can match the handling plan and trigger emergency braking in milliseconds when a fault occurs, effectively constructing an automated knowledge closed loop of perception-diagnosis-control.

[0061] The semantic training module includes a temporal encoder and a text encoder, both of which include a feature extraction layer and a non-linear projection head.

[0062] The temporal encoder is constructed based on a multi-layer dilated causal convolutional network. It extracts multi-scale temporal features from sensor data through convolutional kernels with different dilation rates, preserves local contextual information in the time dimension, and generates a temporal feature sequence.

[0063] The text encoder is built based on a pre-trained Transformer architecture, extracts word embedding vectors from the fault description text, and combines them with positional encoding to generate a text feature sequence containing semantic dependencies.

[0064] The nonlinear projection head maps the temporal feature sequence and the text feature sequence to the same dimensional metric space.

[0065] The time-series encoder is constructed based on a multi-layer dilated causal convolutional network. Its core lies in stacking multiple dilated convolutional layers to capture long-term dependencies in sensor data. The system first inputs preprocessed historical sensor monitoring time-series data into the encoder. To expand the receptive field without reducing temporal resolution, the network employs a dilated convolution mechanism, inserting holes between elements of the convolutional kernel. The expansion rate of the dilated causal convolutional layers increases exponentially with the depth of the network layers; the expansion rate is 1 for the first layer, 2 for the second, 4 for the third, and so on, enabling higher-level networks to cover trend changes ranging from instantaneous impacts of milliseconds to trends of several hours. Simultaneously, to ensure temporal causality and prevent future information leakage, the system zero-padding the left side of the sequence before convolution operations, ensuring that the output at the current moment depends only on the input data at the current and past moments. After multi-layer convolution processing, the output is a temporal feature sequence that retains temporal dimension information.

[0066] The text encoder is built on a pre-trained Transformer architecture to extract semantic features from unstructured fault text. The system first uses a word segmenter to segment the input fault description text into a sequence of characters or words, and then converts it into high-dimensional word embedding vectors by looking up a table. Considering that the Transformer architecture itself does not have the ability to process sequence order, the system introduces positional encoding technology to generate sine and cosine position vectors with the same dimension as the word embedding vectors, and adds them to the word embedding vectors to give the model the ability to capture word order information. The system uses a multi-head self-attention mechanism to calculate the association weights between different words in the text sequence, capture long-distance semantic dependencies, and generate a text feature sequence containing rich contextual semantics.

[0067] The nonlinear projection head is connected after the outputs of the temporal encoder and the text encoder, and is used to map heterogeneous features to a unified metric space. The nonlinear projection head adopts a multilayer perceptron structure consisting of two fully connected layers and a nonlinear activation function. The first fully connected layer compresses or expands the high-dimensional features output by the encoder to an intermediate dimension, and increases the nonlinear expressive power of the model by modifying the linear unit activation function. The second fully connected layer maps the features to the final target dimension. Through this process, the temporal feature sequences and text feature sequences that were originally in different feature spaces are transformed into vector sequences of the same dimension. The nonlinear projection head maps heterogeneous temporal and text features to a unified metric space, and the multilayer perceptron structure and activation function enhance the nonlinear expressive power of the model.

[0068] By utilizing a multi-layer dilated causal convolutional network, a multi-scale receptive field covering instantaneous impacts at the millisecond level to trend evolution at the hour level is constructed. This effectively solves the problem that traditional models cannot take into account both local mutations and long-term dependencies without reducing temporal resolution. In conjunction with a nonlinear projection head, physical time-series signals and textual semantic features are mapped to a unified metric space, achieving deep mathematical alignment of heterogeneous data.

[0069] The multilayer dilated causal convolutional network consists of N stacked temporal residual blocks; each temporal residual block contains two parallel paths, the main path containing a dilated causal convolutional layer, a weight normalization layer and an activation function layer connected in sequence, and the residual path is used for gradient propagation.

[0070] The expansion rate of the dilated causal convolutional layer increases exponentially with the increase of the number of network layers, thus constructing a multi-scale time receptive field that can cover the instantaneous cutting impact of coal mining equipment from the millisecond level to the mechanical wear trend of the hour level.

[0071] The main path filters downhole environmental background noise through a gating mechanism and prioritizes the extraction of impulse response features related to equipment failure.

[0072] The time-series encoder specifically adopts a network structure consisting of five time-series residual blocks stacked vertically. Preprocessed sensor monitoring data is input into the first time-series residual block. In order to construct a multi-scale time receptive field that can cover the range from millisecond-level instantaneous truncation impact to hour-level mechanical wear trends, the dilated causal convolutional layers inside these five time-series residual blocks adopt exponentially increasing dilation rate parameters, specifically set to 1, 2, 4, 8, and 16 respectively. Combined with a convolutional kernel size of 3, the effective receptive field of the top layer of the network can completely cover or even exceed the length of the input time window, ensuring that the model can capture both local abrupt changes and understand long-term evolution patterns.

[0073] The core of the main path for the temporal residual block lies in filtering downhole environmental background noise through a gating mechanism. The main path first performs feature mapping on the input data using a dilated causal convolutional layer. To strictly ensure temporal causality—that is, the output at the current moment depends only on the input at past moments—the system zero-padding is applied to the left side of the sequence's temporal dimension before the convolution operation. The padding length is equal to the product of the convolution kernel size minus 1 and the current layer's dilation rate. The output after the convolution operation is copied twice and enters two parallel branches. The first branch is processed by a hyperbolic tangent activation function, serving as a filtering channel for extracting content features. The second branch is processed by a sigmoid activation function, with output values ​​between 0 and 1, serving as a gating channel for controlling information flow. A sigmoid activation function is used to construct the gating channel, adaptively adjusting the information flow. Combined with the Tanh channel, this accurately extracts content features, effectively suppressing complex downhole background noise like an adaptive filter, and prioritizing the retention of high-confidence fault pulses.

[0074] The gating filtering operation is performed by multiplying the outputs of the two branches element by element. The values ​​calculated by the gating channel are used to weight the features extracted by the filtering channel. When the monitoring data contains random electromagnetic noise unrelated to the fault, the gating channel will output a value close to 0 to suppress the transmission of this part of the signal. For impulse response features that are highly correlated with equipment faults, the gating channel will output a value close to 1 to retain them. The features after gating are further normalized by the weight normalization layer to accelerate network convergence and stabilize the training process.

[0075] The residual path of the temporal residual block is used for direct gradient propagation. When the number of channels in the input feature map is the same as the number of channels in the output feature map of the main path, the residual path directly adds the input of the residual block to the output of the main path element by element. When the number of channels is different, the residual path linearly projects the input feature map through a convolutional layer with a kernel size of 1 to align its channel dimensions with the output of the main path before adding them. In this embodiment, the number of channels of the convolutional filters in each layer of the network is uniformly set to 64, and a random deactivation operation with a dropout rate of 0.2 is introduced before the residuals are added to prevent the model from overfitting. The final result after addition is processed by a modified linear unit activation function and used as the final output of the current residual block.

[0076] By stacking five layers of temporal residual blocks with exponential growth rate, a full-time domain receptive field covering millisecond-level instantaneous impacts to hourly-level wear trends is constructed, effectively solving the technical problem that a single scale cannot simultaneously address sudden faults and chronic degradation. Combined with a unique dual-branch gating mechanism, nonlinear filtering of the feature stream is achieved, which can accurately suppress complex downhole electromagnetic background noise like an adaptive filter, prioritizing the passage of high-confidence fault pulse features.

[0077] The specific process of jointly training the temporal encoder and text encoder using a fine-grained contrastive learning strategy based on saliency masks involves calculating the contribution of each time step in the temporal feature sequence to fault classification using a gradient backpropagation mechanism to generate a saliency mask, and simultaneously extracting a text saliency mask based on the self-attention weight distribution of the text encoder.

[0078] The cosine similarity between each pair of elements in the temporal feature sequence and the text feature sequence is calculated to construct an original similarity matrix. The temporal saliency mask and the text saliency mask are then multiplied by an outer product to generate a filter matrix. The filter matrix and the original similarity matrix are then multiplied by a Hadamard product to generate a fine-grained cross-correlation matrix. The fine-grained cross-correlation matrix is ​​then pooled to obtain a global similarity score representing the matching degree of sample pairs. The contrastive loss function is used to maximize the global similarity score of positive sample pairs. The encoder parameters are then updated through backpropagation.

[0079] For the temporal feature sequence generated by the temporal encoder, the system temporarily connects an auxiliary classifier during the training phase. This classifier consists of fully connected layers. The system inputs the temporal feature sequence into the auxiliary classifier for forward propagation to predict the fault category and calculates the prediction error using the cross-entropy loss function. Gradient backpropagation is performed, but it stops after only propagating the gradient to the temporal feature sequence layer. The L2 norm of the gradient vector for each time step in the feature sequence is calculated, representing the contribution of the corresponding time step to fault identification. These contribution values ​​are then normalized to map the values ​​to the interval between zero and one, thereby generating a temporal saliency mask composed of floating-point numbers, where the closer the value is to one, the more critical the signal feature at that time step is.

[0080] Specifically, the generation steps of the temporal saliency mask are as follows: First, a fully connected classification network is constructed, and the temporal feature sequence is input into it to predict the fault category, and the loss value between the prediction result and the true label is calculated. Then, the system performs backpropagation to calculate the gradient vector of this loss value relative to each time step in the temporal feature sequence. Next, the system calculates the vector magnitude of the gradient vector at each time step to quantify the contribution of that time step to fault identification. Finally, the contribution values ​​of all time steps are standardized and mapped to a value range of zero to one, thus obtaining the temporal saliency mask. The text saliency mask is obtained by extracting the weight matrix of all attention heads in the last layer of the pre-trained model, averaging them, summing them row-wise, and normalizing them to obtain the importance weight sequence corresponding to each text word.

[0081] For the text feature sequence generated by the text encoder, the self-attention map of the last layer of the pre-trained transformer architecture is extracted. Since a multi-head attention mechanism is adopted (set to 8 attention heads in this embodiment), the weight matrices of these 8 attention heads are first averaged and pooled into a unified attention weight matrix, which objectively reflects the degree of attention the model pays to specific words when understanding the whole fault description. The weight values ​​are normalized to generate a text saliency mask corresponding to the length of the text sequence.

[0082] After obtaining the bimodal mask, a fine-grained cross-correlation matrix is ​​constructed. The cosine similarity between the feature vector of each time step in the temporal feature sequence and the word embedding vector of each word in the text feature sequence is calculated to generate an original similarity matrix with a dimension of time step length multiplied by text length. The temporal saliency mask (column vector) and the text saliency mask (row vector) are subjected to a matrix outer product operation to generate a filter matrix. Each element value in the filter matrix represents the joint importance of the corresponding time step-word pairing. The filter matrix and the original similarity matrix are subjected to a Hadamard product operation (i.e., element-wise multiplication) to obtain the fine-grained cross-correlation matrix. Through this operation, the similarity value is only retained when a key time step meets a key semantic word, while the similarity between noise signals and irrelevant words is suppressed by the low weight of the filter matrix.

[0083] Pooling is performed on the fine-grained cross-correlation matrix to obtain the global similarity score. In this embodiment, a logarithmic summation exponential pooling strategy is adopted, that is, the exponential values ​​of all elements in the matrix are summed first, and then the logarithm is taken, or the average value of the top 5 elements with the largest values ​​in the matrix is ​​calculated, thereby aggregating the two-dimensional matrix into a scalar score that can measure the degree of matching between the entire sensing data segment and the fault text description, which is the global similarity score of the positive sample pair. In the training batch (the batch size is set to 128 in this embodiment), the current sensing data is combined with other mismatched fault texts to construct negative sample pairs, and their similarity scores are calculated in the same way.

[0084] A contrastive loss function is used to update the network parameters. When calculating the loss, the temperature parameter is set to 0.07 to control the smoothness of the probability distribution. The numerator of the loss function is the exponential value of the global similarity score of positive sample pairs, and the denominator is the sum of the exponential values ​​of the scores of positive sample pairs and the exponential values ​​of the scores of all negative sample pairs in the batch. By minimizing this loss function, the temporal encoder and the text encoder are forced to bring the matched sensor-text pairs closer in the feature space and push the mismatched pairs further apart. The optimization process uses an adaptive moment estimation optimizer with an initial learning rate set to 0.001. The learning rate decays until the model converges, thus completing the construction of the coal mine mapping model.

[0085] By utilizing gradient backpropagation and self-attention mechanisms to generate bimodal saliency masks, and by accurately suppressing downhole background noise and meaningless function words through Hadamard product operations, a leap from coarse overall matching of heterogeneous features to precise alignment of local key features is achieved. Combined with a contrastive loss function with temperature parameters, the model is forced to explore subtle feature differences and pay attention to difficult negative samples.

[0086] See Figure 2The two-layer intelligent agent collaboration mechanism includes a standard query agent and an inference agent. The standard query agent has a built-in vector retrieval engine and a graph inference engine. It uses the time encoder in the coal mine mapping model to extract the high-dimensional feature vector of the sensing and monitoring data, and uses an approximate nearest neighbor search algorithm to calculate the cosine similarity between the high-dimensional feature vector and the fault semantic node vector in the coal mine knowledge graph. Based on the similarity score, the top-K candidate fault semantics are selected from high to low.

[0087] For each candidate fault semantic, a breadth-first search strategy is used to traverse the path along the procedure constraint edges in the coal mine knowledge graph to obtain the associated coal mine safety constraints, and the path confidence score is calculated. The fault semantic with the highest confidence score is combined with the safety constraint and output to the reasoning agent.

[0088] The vector retrieval engine preloads the time-series encoder trained in the coal mine mapping model and constructs a hierarchical navigation small-world index structure from the text feature vectors of all fault semantic nodes in the coal mine knowledge graph. The parameters of the index structure are set as follows: the maximum number of bidirectional connections for each node is 16, and the size of the dynamic candidate list when constructing the index is 200. When the monitoring query module inputs real-time sensor monitoring data (in this embodiment, a tensor with a time window of 128 is selected), the time-series encoder maps it into a high-dimensional feature vector with a dimension of 512. The vector retrieval engine then uses the cosine similarity metric to calculate the distance between the real-time feature vector and tens of thousands of fault semantic vectors in the index within milliseconds, and selects the top 5 most likely candidate fault semantics from high to low similarity scores (i.e., K in Top-K is 5) to form an initial screening list.

[0089] For these five candidate fault semantics, the standard query agent initiates a graph reasoning engine to perform knowledge association retrieval. The graph reasoning engine, based on a graph database interface, initiates a traversal request for each candidate fault semantic node using a breadth-first search strategy. The traversal constraints are set to allow expansion only along the direction of the procedure constraint edges, and the search depth is limited to two layers to prevent semantic drift. Through this process, the system can accurately locate the coal mine safety constraint nodes directly linked to the current potential fault, obtain the alarm levels and handling rules stored therein, and thus construct multiple sets of fault semantic-safety constraint reasoning paths.

[0090] The standard query agent then calculates the path confidence score to quantify the credibility of each inference path. The calculation logic uses a weighted summation method, specifically assigning a weight of 0.7 to the semantic similarity score obtained in the vector retrieval stage and a weight of 0.3 to the association strength weight of the protocol constraint edges in the knowledge graph. The two are then weighted and added together to obtain the comprehensive confidence score. If the calculated comprehensive score is lower than the preset confidence threshold (set to 0.85 in this embodiment), the system will automatically determine that the path is invalid and discard it to reduce the false alarm rate.

[0091] The standard query agent sorts all the retained valid paths and selects the set of fault semantic-safety constraint combinations with the highest confidence score as the final output. The result is encapsulated into a standardized data packet, which includes the fault name, the probability value of the match, the associated safety procedure ID, and the procedure text content. The data packet is then transmitted unidirectionally to the inference agent.

[0092] By utilizing a hierarchical navigation small-world index structure, millisecond-level responses to massive fault features are achieved, effectively solving the response delay problem in real-time coal mine monitoring. Combined with a breadth-first search strategy with depth constraints, the knowledge reasoning boundary is strictly constrained, effectively preventing semantic drift that may be caused by relying solely on semantic similarity, and ensuring the logical rigor of safety constraint associations. Through a multi-dimensional confidence gating mechanism that integrates semantic scores and graph weights, low-quality search results are automatically filtered out.

[0093] The reasoning agent concatenates the received natural language fault semantics, coal mine safety constraints, and target equipment ID according to a preset format to form decision prompt words. These decision prompt words are then input into a large language model, which uses its logical reasoning capabilities to compare the state value in the fault semantics with the numerical threshold in the coal mine safety constraints. When the state value exceeds the numerical threshold, the large language model directly extracts the pre-stored forced shutdown instruction attribute from the coal mine safety constraints and fills the equipment ID into the reserved parameter bits of the instruction attribute, generating and issuing a coal mine control instruction for the specific equipment.

[0094] The reasoning agent receives a structured data packet output by the previous level standard query agent, which includes natural language fault semantics, associated coal mine safety constraints, and a unique identifier for the target equipment. The system concatenates the above information according to a preset expert prompt word template. The prompt word template consists of four parts: role setting, observation status, procedural knowledge, and reasoning task. The system fills in the role setting with the coal mine safety specialist, the observation status with the fault semantics, and the procedural knowledge with the threshold clauses in the safety constraints, and finally forms a complete decision prompt word text, which clearly requires the model to judge whether the current state is in violation and to provide operational suggestions.

[0095] This embodiment uses a pre-trained language model with 13 billion parameters and performs supervised fine-tuning using a coal mine historical accident case database. During inference, the temperature parameter generated by the model is set to 0.1 to reduce randomness and ensure the determinism of the output results. The large language model performs semantic parsing on the input decision prompts, accurately extracts the current real-time state value and the critical threshold value specified in the safety constraints contained in the fault semantics, and performs numerical comparison operations in the high-dimensional semantic space. When the model confirms that the current state value is mathematically greater than or equal to the critical threshold value, it determines that the working condition belongs to a violation and dangerous state.

[0096] Once a dangerous state is determined, the reasoning agent does not rely on a model for divergent text generation, but directly triggers a rule-based instruction extraction mechanism; it reads the pre-stored forced shutdown instruction attributes in the coal mine safety constraint node attributes. Internally, it stores a standardized key-value pair format instruction template. In this embodiment, the action type is: emergency shutdown, the target object is: reserved parameter bit, and the execution level is: highest level. The post-processing script of the large language model or system locks the reserved parameter bit in the template and forcibly fills the received target device unique identifier into that position, thereby generating a complete, specific, and specific control instruction data packet for a particular physical device.

[0097] The system transmits the generated key-value pair format instruction data packet to the industrial control gateway. The protocol conversion program in the gateway parses the data packet, queries the underlying physical address mapping table based on the device identifier, calls the industrial communication protocol code template that matches the field controller, converts the logical emergency stop action into a specific hexadecimal control message, and sends it to the production control system of the corresponding device through the industrial Ethernet, triggering the hardware-level power failure protection or emergency braking program.

[0098] It adopts a hybrid decision-making mechanism that combines large-scale model logical reasoning with rule-based instruction filling. It uses a large language model constrained by low-temperature parameters for semantic parsing and numerical comparison, which effectively suppresses the risk of hallucination while leveraging the advantages of logical judgment in complex contexts. Combined with pre-stored key-value pair templates for deterministic parameter slot filling, it ensures the absolute standard of control instructions in terms of syntax format, effectively solving the safety problem of generative algorithms being difficult to implement in high-risk industrial scenarios.

[0099] This invention significantly improves the intelligence level of coal mine management by constructing a structured graph and a two-layer intelligent agent collaboration mechanism. On the one hand, by using a fine-grained contrastive learning strategy based on saliency masks, the cross-correlation between time series and local text features is accurately calculated, effectively overcoming the semantic barriers between the underlying sensor data and the upper-level expert knowledge, and realizing accurate semantic mapping of fault features. On the other hand, through the hierarchical collaboration between standard query agents and inference agents, real-time state perception, safety procedure retrieval, and logical decision control are deeply decoupled yet closely linked, ensuring that the generated control commands not only conform to the real-time working conditions on site but also strictly follow coal mine safety constraints, realizing an automated closed loop from data monitoring to intelligent control.

[0100] Example 2:

[0101] This embodiment applies a multi-agent-based intelligent coal mine management system to address an overheating fault in the stator winding of the left cutting motor of a high-power MG900 / 2210-GWD coal mining machine. It details the application process of this system in actual production at a fully mechanized mining face. (See attached document.) Figure 3 ;

[0102] The graph construction module first accesses multimodal historical data from the fully mechanized mining face underground. The cleaning subunit in the hybrid processing unit reads the coal mining machine operation logs from the past three months and uses a wavelet threshold denoising algorithm (decomposition level 4, adaptive threshold 0.6745) to remove high-frequency electromagnetic interference from the frequency converter in the current and temperature sensor data of the left cutting motor. At the same time, the extraction subunit uses the BERT-BiLSTM-CRF model to process the equipment maintenance manual, identifying the entity left cutting motor (equipment ID: SM-L-001), stator temperature sensor (measuring point ID: PT100-05), and the fault description of overheating of the winding caused by cooling water blockage. The system forcibly maps the measuring point ID as an attribute to the equipment entity, constructing a local knowledge graph containing monitoring mapping edges and fault association edges, and solidifies the numerical threshold and mandatory command of the safety procedure for immediate shutdown when the motor winding temperature exceeds 155℃ in the safety constraint node.

[0103] The semantic training module initializes the temporal encoder and text encoder. The temporal encoder uses a 5-layer dilated causal convolutional network (dilation rates 1, 2, 4, 8, 16) to extract multi-scale temporal features of the motor temperature suddenly rising from a normal fluctuation of 85℃ to 140℃. The text encoder extracts semantic features of the rapid temperature rise of the motor. The system applies a fine-grained contrastive learning strategy based on saliency masks, and the gradient backpropagation mechanism identifies the temperature rise stage with the steepest slope in the temperature curve as the key time step, generating a high-weight temporal saliency mask. At the same time, it identifies "rapid" and "rise" in the text as key semantic words. By maximizing the global similarity score of the fine-grained cross-correlation matrices of the two, the trained coal mine mapping model can understand the deep mathematical alignment relationship between the waveform of the rapid temperature rise and the semantics of motor overheating.

[0104] During production, the monitoring and query module collects sensor data from measurement point ID PT100-05 in real time. When the temperature data stream is detected to abnormally rise from 90℃ to 158℃ within 30 seconds, a two-layer intelligent agent collaboration mechanism is triggered. The standard query agent first uses a trained coal mine mapping model to map this abnormal temperature rise time series data into a high-dimensional feature vector, and then uses a vector retrieval engine to match the natural language fault semantic with the highest similarity in the graph—"severe overcutting of the motor stator winding"; the agent uses a breadth-first search strategy to traverse the graph along the procedure constraint edges, locks the associated coal mine safety constraint nodes, and obtains the alarm level (extremely severe) and the corresponding safety threshold (155℃).

[0105] The reasoning agent receives the semantic description of "severe overheating of the stator winding of the cutting motor," the real-time state value (158℃), the safety constraint threshold (155℃), and the device ID (SM-L-001). The agent assembles these into a large language model with 13 billion parameters for inputting decision prompts. The model performs logical reasoning under a temperature parameter of 0.1, determining that the current state value of 158℃ exceeds the safety threshold of 155℃, constituting a dangerous violation. The model then stops divergent text generation, directly extracts the pre-stored "forced shutdown command template" from the safety constraint node attributes, and fills the template parameter bits with the device ID (SM-L-001), generating a standard control command data package for the left cutting motor of the coal mining machine.

[0106] The generated control command data packet is transmitted to the centralized control center of the fully mechanized mining face. The gateway parses the command, converts it into a hexadecimal control message using the Modbus TCP protocol, and sends it to the onboard PLC controller of the coal mining machine via industrial Ethernet. Upon receiving the message, the PLC triggers the high-voltage side circuit breaker of the frequency converter to trip within milliseconds, achieving an emergency shutdown of the left cutting motor. The entire process requires no manual dispatcher intervention, effectively preventing underground fire accidents caused by motor overheating and burnout, and verifying the system's automated decision-making and response capabilities in a real-world scenario.

[0107] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A coal mine intelligent management system based on multi-agent systems, characterized in that, include: Knowledge graph construction module: acquires unstructured multimodal historical data of coal mines, processes the historical data using a hybrid processing unit, and generates a structured knowledge graph of coal mines; Semantic training module: includes a temporal encoder and a text encoder; the temporal encoder and text encoder are jointly trained using a fine-grained contrastive learning strategy based on saliency masks to generate saliency masks; based on the saliency masks, a fine-grained cross-correlation matrix between temporal local features and text local features is calculated to construct a coal mine mapping model; Monitoring and query module: Real-time collection of sensor monitoring data from the coal mine production site, triggering a two-layer intelligent agent collaboration mechanism, including a standard query intelligent agent and an inference intelligent agent; the standard query intelligent agent uses the coal mine mapping model to map the sensor monitoring data into natural language fault semantics, and retrieves the associated coal mine safety constraints in the coal mine knowledge graph; Inference control module: The inference agent receives the natural language fault semantics and the coal mine safety constraints, generates coal mine control instructions, and issues them for execution.

2. The coal mine intelligent management system based on multi-agent technology according to claim 1, characterized in that, The hybrid processing unit includes a cleaning subunit and an extraction subunit. The cleaning subunit performs noise filtering on the historical data, uses a wavelet threshold noise reduction algorithm to smooth the electromagnetic interference noise of the sensor monitoring data, and uses a coal mine-specific terminology library to correct spelling errors and standardize the text data. The extraction subunit uses BERT-BiLSTM-CRF to extract entities and their relationships, identifies the association between equipment entities and faults from the text, and maps the measurement point IDs of the sensor monitoring data as attributes to the corresponding equipment entities, generating a coal mine knowledge graph containing a topology structure of equipment-sensor-fault.

3. The coal mine intelligent management system based on multi-agent technology according to claim 1, characterized in that, The coal mine knowledge graph consists of entity nodes and relation edges; The entity nodes include device entity nodes, sensor measurement point nodes, fault semantic nodes, and safety constraint nodes. The relational edges include monitoring mapping edges connecting the device entity nodes and sensor measurement point nodes, fault association edges connecting the device entity nodes and fault semantic nodes, and procedure constraint edges connecting the fault semantic nodes and safety constraint nodes; the safety constraint nodes store numerical thresholds, alarm levels, and forced shutdown command attributes.

4. The coal mine intelligent management system based on multi-agent technology according to claim 1, characterized in that, Both the temporal encoder and the text encoder include a feature extraction layer and a non-linear projection head; The temporal encoder is constructed based on a multi-layer dilated causal convolutional network. It extracts multi-scale temporal features from sensor data through convolutional kernels with different dilation rates, preserves local contextual information in the time dimension, and generates a temporal feature sequence. The text encoder is built based on a pre-trained Transformer architecture, extracts word embedding vectors from the fault description text, and combines them with positional encoding to generate a text feature sequence containing semantic dependencies. The nonlinear projection head maps the temporal feature sequence and the text feature sequence to the same dimensional metric space.

5. The coal mine intelligent management system based on multi-agent technology according to claim 4, characterized in that, The multilayer dilated causal convolutional network consists of N stacked temporal residual blocks; each temporal residual block contains two parallel paths, the main path containing a dilated causal convolutional layer, a weight normalization layer and an activation function layer connected in sequence, and the residual path is used for gradient propagation. The expansion rate of the dilated causal convolutional layer increases exponentially with the increase of the number of network layers, thus constructing a multi-scale time receptive field that can cover the instantaneous cutting impact of coal mining equipment from the millisecond level to the mechanical wear trend of the hour level. The main path filters downhole environmental background noise through a gating mechanism and prioritizes the extraction of impulse response features related to equipment failure.

6. The coal mine intelligent management system based on multi-agent technology according to claim 4, characterized in that, The specific process of jointly training the temporal encoder and text encoder using a fine-grained contrastive learning strategy based on saliency masks involves calculating the contribution of each time step in the temporal feature sequence to fault classification using a gradient backpropagation mechanism to generate a saliency mask, and simultaneously extracting a text saliency mask based on the self-attention weight distribution of the text encoder. The cosine similarity between each pair of elements in the temporal feature sequence and the text feature sequence is calculated to construct an original similarity matrix. The temporal saliency mask and the text saliency mask are then multiplied by an outer product to generate a filter matrix. The filter matrix and the original similarity matrix are then multiplied by a Hadamard product to generate a fine-grained cross-correlation matrix. The fine-grained cross-correlation matrix is ​​then pooled to obtain a global similarity score representing the matching degree of sample pairs. The contrastive loss function is used to maximize the global similarity score of positive sample pairs. The encoder parameters are then updated through backpropagation.

7. The coal mine intelligent management system based on multi-agent technology according to claim 1, characterized in that, The two-layer intelligent agent collaboration mechanism includes a standard query agent and an inference agent. The standard query agent has a built-in vector retrieval engine and a graph inference engine. It uses the time encoder in the coal mine mapping model to extract the high-dimensional feature vector of the sensing and monitoring data, and uses an approximate nearest neighbor search algorithm to calculate the cosine similarity between the high-dimensional feature vector and the fault semantic node vector in the coal mine knowledge graph. Based on the similarity score, the top-K candidate fault semantics are selected from high to low. For each candidate fault semantic, a breadth-first search strategy is used to traverse the path along the procedure constraint edges in the coal mine knowledge graph to obtain the associated coal mine safety constraints, and the path confidence score is calculated. The fault semantic with the highest confidence score is combined with the safety constraint and output to the reasoning agent.

8. The coal mine intelligent management system based on multi-agent technology according to claim 1, characterized in that, The reasoning agent concatenates the received natural language fault semantics, coal mine safety constraints, and target equipment ID according to a preset format to form decision prompt words; The decision prompt words are input into the large language model. The logical reasoning ability of the large language model is used to compare the state value in the fault semantics with the numerical threshold in the coal mine safety constraints. When the state value exceeds the numerical threshold, the large language model directly extracts the pre-stored forced shutdown instruction attribute in the coal mine safety constraints and fills the equipment ID into the reserved parameter bits of the instruction attribute to generate and issue a coal mine control instruction for the specific equipment.

Images