Mine disaster early warning method and system based on machine learning

Abstract

The invention relates to a mine disaster early warning method and system based on machine learning. The method comprises the following steps: detecting node response time based on a periodic heartbeat signal to obtain an invalid node list; the periodic heartbeat signal is actively sent by a control node and is used for indicating all nodes to feed back the node response time; the invalid node list is used for representing invalid nodes in the network topological graph; based on the invalid node list and a real-time network topological graph, calculating an optimal path through a Q-learning algorithm to obtain an optimized communication path table; based on the optimized communication path table, synchronizing early warning information through an improved Raft protocol to obtain a global consistent early warning information set; and planning an optimal escape route based on the global consistent early warning information set and a three-dimensional digital twinborn model of the mine to obtain an early warning instruction set, and sending the early warning instruction set to an escape terminal. By adopting the method, the effects of quickly achieving consistency of the early warning information in the topological network and dynamically adjusting the escape path can be achieved.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent early warning, and particularly relates to a mine disaster early warning method and system based on machine learning. Background Art

[0002] With the development of mine mining technology, in order to prevent losses caused by mine disasters, mine early warning technology has emerged. Currently, mine early warning mainly relies on a sensor network deployed fixedly and a centralized data processing architecture. A typical solution uses wired or basic wireless communication. Sensors periodically upload data to the ground control center, and alarms are triggered by threshold comparison. Most existing systems use static routing. After a node fails, manual intervention is required for maintenance. Information synchronization depends on a polling mechanism, and it often takes several minutes to achieve global data consistency. The escape route planning depends on a pre-stored static map and cannot dynamically avoid newly added risk points. There are problems such as slow speed in achieving global data consistency and rigid escape route planning. Summary of the Invention

[0003] Based on this, in view of the above technical problems, it is necessary to provide a mine disaster early warning method and system based on machine learning that can quickly unify global data and dynamically adjust the escape route.

[0004] In a first aspect, the present application provides a mine disaster early warning method based on machine learning, including:

[0005] Detecting the node response time based on a periodic heartbeat signal to obtain a list of failed nodes; the periodic heartbeat signal is actively sent by a control node and is used to instruct all nodes to feedback the node response time; the list of failed nodes is used to represent the failed nodes in the network topology graph;

[0006] Calculating an optimal path through the Q-learning algorithm based on the list of failed nodes and the real-time network topology graph to obtain an optimized communication path table;

[0007] Synchronizing early warning information through an improved Raft protocol based on the optimized communication path table to obtain a globally consistent early warning information set;

[0008] Planning an optimal escape route based on the globally consistent early warning information set and the three-dimensional digital twin model of the mine to obtain an early warning instruction set, and sending the early warning instruction set to an escape terminal.

[0009] Further, calculating an optimal path through the Q-learning algorithm based on the list of failed nodes and the real-time network topology graph to obtain an optimized communication path table includes:

[0010] Updating the original network topology graph based on the list of failed nodes to obtain an effective topological subgraph;

[0011] Based on the effective topological subgraph, the candidate path set is obtained through the following formula:

[0012] Q(s,a)←Q(s,a)+α[r+γmaxQ(s′,a′)-Q(s,a)]

[0013] where Q(s,a) is the expected cumulative reward for choosing action a in state s. The path with the maximum sum of expected cumulative rewards is determined as the candidate path set. α is the learning rate, r is the immediate reward, γ is the discount factor, and maxQ(s′,a′) is the maximum expected reward among all possible actions in the next state;

[0014] The stability of the candidate path set is evaluated by improving the XGBoost model to obtain the evaluation result;

[0015] Based on the evaluation result, the paths with the top 10% stability are determined as the hierarchical path set;

[0016] Based on the constraint conditions, the Pareto front of the hierarchical path set is calculated through the following formula:

[0017] w1×Q score +w2×(1 - stability_risk)

[0018] where w1 and w2 are dynamic weights, Q score is the Q - value score, and stability_risk is the path failure risk probability;

[0019] The paths that constitute the Pareto front are determined as the optimized path sequence;

[0020] Based on the optimized path sequence and the real - time network topology graph, the path scores are calculated to obtain the optimized communication path table.

[0021] Furthermore, the improved XGBoost model is trained by the following method:

[0022] Based on the real - time failure paths and historical path data, the proportion of failure paths in the most recent hour is statistically calculated to obtain the failure frequency;

[0023] Based on the failure frequency, the weights of the failure paths are dynamically adjusted to obtain the dynamic class weight table;

[0024] Based on the dynamic class weight table, the loss function is obtained through the following formula:

[0025] Loss=-α t (1 - p t ) γ log(p t )

[0026] Among them, Loss is the loss function, and α t is the dynamic class weight, p t is the model prediction probability, γ is the focusing parameter, and (1 - p t ) γ is the modulation factor, and log(p t ) is the cross-entropy based loss;

[0027] Extract the features of the candidate path set to obtain a multi-dimensional feature vector, and splice the multi-dimensional feature vector and the historical failure paths to obtain a training data set;

[0028] Based on the training data set, use the K-means clustering algorithm to divide the candidate paths into K groups to obtain grouped paths;

[0029] Based on the grouped paths, use each group of candidate paths as the training data set to train an independent XGBoost sub-model to obtain the trained XGBoost sub-model;

[0030] Integrate all the trained XGBoost sub-models to obtain the trained improved XGBoost model.

[0031] Furthermore, based on the optimized communication path table, synchronize the warning information through the improved Raft protocol to obtain a globally consistent warning information set, including:

[0032] Based on the optimized communication path table, determine each node with information forwarding ability as a relay node;

[0033] Calculate the credit scores of each relay node through the following formula:

[0034] F = 0.6×a + 0.3×b + 0.1×c

[0035] Among them, F is the credit score, a is the online rate, b is the processing ability, and c is the historical synchronization success rate;

[0036] Based on the credit score and the node version number, obtain the Leader node;

[0037] Receive the commit log list fed back by the Leader node;

[0038] Among them, the commit log list is generated by the Leader node spreading the warning information to each sub-node through the optimal communication path table and each sub-node feeding back the version number of each sub-node to the Leader node;

[0039] Based on the commit log list and the global version number, compare the version differences to obtain a list of different nodes;

[0040] Push the missing log entries to the differential nodes in the differential node list and generate a hash based on the log entries; the missing log entries are used to update the warning information and instruct the differential nodes to feedback the differential node hashes;

[0041] Receive the differential node hashes and compare whether the differential node hashes are the same as the quorum node hashes. If they are the same, the verification is passed and a globally consistent warning information set is generated.

[0042] Further, based on the globally consistent warning information set and the 3D digital twin model of the mine, plan the optimal escape route to obtain a warning instruction set, including:

[0043] Extract the features of the globally consistent warning information set to obtain a feature matrix;

[0044] Classify the feature matrix based on a preset level through a random forest classifier to obtain a classification result;

[0045] Construct a grid map based on the 3D digital twin model of the mine and the real-time network topology map to obtain a weighted map of the mine;

[0046] Based on the multi-objective A* algorithm, calculate the weighted map through the following formula to obtain an optimized path set:

[0047] f(n) = g(n) + 1.2h(n) + 0.5r(n)

[0048] where f(n) is the cost of each path, g(n) is the actual movement cost, h(n) is the heuristic estimate, and r(n) is the risk coefficient;

[0049] Generate a warning instruction set based on the classification result and the optimized path set.

[0050] Further, detect the node response time based on the periodic heartbeat signal to obtain a list of failed nodes, including:

[0051] Preprocess the node response time based on the sensor data to obtain a standardized data set; the standardized data set includes the node id, response time, and sensor status;

[0052] Predict the failure probability of the standardized data set through an LSTM model to obtain a prediction result;

[0053] Obtain the dynamic network delay threshold based on the current network load;

[0054] Determine the node failure situation based on the dynamic network delay threshold and the prediction result to obtain a preliminary list of failed nodes;

[0055] Generate a list of failed nodes based on the node dependency relationship and the preliminary failure list.

[0056] Further, after planning the optimal escape route based on the globally consistent early warning information set and the 3D digital twin model of the mine, obtaining the early warning instruction set, and sending the early warning instruction set to the escape terminal, it further includes:

[0057] Receiving the received confirmation list fed back by the terminal;

[0058] Based on the received confirmation list, quickly marking suspicious data through a Bloom filter to obtain a preliminary verification result;

[0059] Verifying the hash value of the preliminary verification result to obtain a deep verification result;

[0060] Based on the historical operation logs, judging the similarity of the preliminary verification result through a twin neural network to obtain a similarity comparison result;

[0061] Based on the preliminary verification result, the deep verification result and the similarity comparison result, generating a trusted terminal information set.

[0062] Further, based on the historical operation logs, judging the similarity of the preliminary verification result through a twin neural network to obtain a similarity comparison result, including:

[0063] Extracting the features of the preliminary verification result to obtain a preprocessed data set;

[0064] Inputting the preprocessed data set into the encoder to obtain an embedding vector;

[0065] Retrieving the normal embedding vectors of similar devices from the reference library;

[0066] Calculating the similarity between the embedding vector and the normal embedding vector through the following formula:

[0067]

[0068] where sim is the similarity, W query is the embedding vector, and E ref is the normal embedding vector;

[0069] Comparing the similarity with a preset threshold to obtain a similarity comparison result.

[0070] In a second aspect, the present application further provides a mine disaster early warning system based on machine learning, including:

[0071] A response module, configured to detect the node response time based on a periodic heartbeat signal to obtain a list of failed nodes; the periodic heartbeat signal is actively sent by the control node and is used to instruct all nodes to feedback the node response time; the list of failed nodes is used to represent the failed nodes in the network topology diagram;

[0072] A communication module, which is used to calculate the optimal path through the Q-learning algorithm based on the list of failed nodes and the real-time network topology map, and obtain an optimized communication path table;

[0073] A synchronization module, which is used to synchronize warning information through an improved Raft protocol based on the optimized communication path table, and obtain a globally consistent warning information set;

[0074] A warning module, which is used to plan the optimal escape route based on the globally consistent warning information set and the three-dimensional digital twin model of the mine, obtain a warning instruction set, and send the warning instruction set to the escape terminal.

[0075] Thirdly, the present application also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the method provided in the first aspect of the present application are implemented.

[0076] The above-mentioned mine disaster warning method and system based on machine learning detect the node response time based on the periodic heartbeat signal to obtain a list of failed nodes; the periodic heartbeat signal is actively sent by the control node to indicate that all nodes feedback the node response time; the list of failed nodes is used to represent the failed nodes in the network topology map; based on the list of failed nodes and the real-time network topology map, the optimal path is calculated through the Q-learning algorithm to obtain an optimized communication path table; based on the optimized communication path table, the warning information is synchronized through the improved Raft protocol to obtain a globally consistent warning information set; based on the globally consistent warning information set and the three-dimensional digital twin model of the mine, the optimal escape route is planned to obtain a warning instruction set, and the warning instruction set is sent to the escape terminal. Through the above technical means, the effects of fast and strictly unified warning information and dynamic planning of escape routes are achieved, the reliability of the warning information dissemination process is strengthened, and the safety is guaranteed. Description of the Drawings

[0077] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following will briefly introduce the drawings required for use in the description of the embodiments or related technologies. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0078] Figure 1 It is the flow chart of the mine disaster warning method based on machine learning of the present invention;

[0079] Figure 2 It is the system diagram of the mine disaster warning system based on machine learning of the present invention; Detailed Embodiments

[0080] To make the objectives, technical solutions and advantages of the present application more clear and understandable, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0081] In one embodiment, as Figure 1 shown, a mine disaster early warning method based on machine learning is provided. In this embodiment, an example is given where this method is applied to a server. It can be understood that this method can also be applied to a terminal, and can also be applied to a system including a server and a terminal, and is implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0082] Step 101, detecting the node response time based on the periodic heartbeat signal to obtain a list of failed nodes; the periodic heartbeat signal is actively sent by the control node to indicate that all nodes feedback the node response time; the list of failed nodes is used to characterize the failed nodes in the network topology diagram.

[0083] Specifically, the periodic heartbeat signal is a detection message periodically broadcast by the control node, which is sent once every 30 seconds and the period is adjustable. The control node is the authoritative coordinator in the network and is responsible for monitoring the status of the entire network. The node response time is the timestamp from sending the heartbeat to receiving the ACK. Nodes with a continuous three-time response time greater than 1500 ms are defined as failed nodes. The network topology diagram is used to characterize the node connection relationship. The control node broadcasts detection packets at 30-second intervals, carrying a unique increasing serial number for the entire network, indicating each node to immediately return an ACK, along with the local resource utilization rate.

[0084] Step 102, calculating the optimal path based on the list of failed nodes and the real-time network topology diagram through the Q-learning algorithm to obtain an optimized communication path table.

[0085] Specifically, the Q-learning algorithm is a reinforcement learning algorithm based on the Markov decision process. The real-time network topology diagram is a dynamically updated network connection status diagram, which includes real-time link quality data. The optimized communication path table is a structured routing table storing the optimal path and the backup path. The Dijkstra algorithm is used to generate the initial shortest path, and then the Q value is updated through the Q-learning algorithm to calculate the optimal path, so as to obtain the optimized communication path table.

[0086] Step 103, synchronizing the early warning information based on the optimized communication path table through the improved Raft protocol to obtain a globally consistent early warning information set.

[0087] Specifically, the improved Raft protocol is a distributed consistency protocol that introduces credit score weighting and batch submission optimization based on the classical Raft consensus algorithm. The global consistent warning information set is the latest warning data set that all network nodes reach a consensus on. The warning information is the disaster information generated by the warning system, which can include data such as disaster type and scale. The Leader node is determined by the credit score and the version number of the warning information stored in each node. The Leader node divides the warning information into blocks, attaches a check code, and transmits the warning information to other nodes in parallel through the main path and the backup path. When the number of nodes that feedback and confirm the received information is greater than the threshold, the log is marked as committed, and the global consistent warning information set is obtained.

[0088] Step 104, plan the optimal escape route based on the global consistent warning information set and the three-dimensional digital twin model of the mine, obtain the warning instruction set, and send the warning instruction set to the escape terminal.

[0089] Specifically, the three-dimensional digital twin model is a virtualized mapping model of the mine's physical environment, integrating geographic information, equipment status, and real-time monitoring data. The optimal escape route is a dynamically planned route that comprehensively considers path length, safety risk, and traffic capacity. The warning instruction set is a structured data set containing escape routes, safety tips, and operation instructions. The improved A* algorithm is called every 10 seconds to recalculate the optimal path, and 3 candidate paths are calculated in parallel. The one with the highest comprehensive score is selected as the optimal escape route.

[0090] Exemplarily, the escape terminal can include miner personal positioning devices, roadway emergency navigation terminals, intelligent helmet systems, mobile rescue terminals, and emergency broadcast devices, etc.

[0091] The mine disaster warning method based on machine learning provided in this embodiment obtains the list of failed nodes by detecting the node response time based on the periodic heartbeat signal; the periodic heartbeat signal is actively sent by the control node to indicate that all nodes feedback the node response time; the list of failed nodes is used to represent the failed nodes in the network topology graph; based on the list of failed nodes and the real-time network topology graph, the optimal path is calculated through the Q-learning algorithm to obtain the optimized communication path table; based on the optimized communication path table, the warning information is synchronized through the improved Raft protocol to obtain the global consistent warning information set; based on the global consistent warning information set and the three-dimensional digital twin model of the mine, the optimal escape route is planned to obtain the warning instruction set, and the warning instruction set is sent to the escape terminal. By these technical means, it realizes the beneficial effects of quickly and strictly unifying the warning information in the network topology, dynamically planning the escape path, strengthening the reliability of the warning information dissemination process, and ensuring safety.

[0092] In one embodiment, based on the list of failed nodes and the real-time network topology map, the optimal path is calculated through the Q-learning algorithm to obtain an optimized communication path table, including:

[0093] Step 201, update the original network topology map based on the list of failed nodes to obtain an effective topology subgraph.

[0094] Specifically, the list of failed nodes records the currently unavailable nodes. The original network topology map is a weighted graph model describing the connection relationships of all nodes. The effective topology subgraph is the available network subgraph after removing the failed nodes and their associated links. Traverse the adjacency list of the original topology map, delete all nodes in the list of failed nodes and their associated edges, perform connectivity detection on the remaining nodes to ensure that the subgraph is strongly connected, generate a new adjacency list and mark the key nodes to obtain the effective topology subgraph.

[0095] Step 202, based on the effective topology subgraph, obtain a set of candidate paths through the following formula:

[0096] Q(s,a)←Q(s,a)+α[r+γmaxQ(s′,a′)-Q(s,a)]

[0097] Where Q(s,a) is the expected cumulative reward for selecting action a in state s. The path with the maximum sum of expected cumulative rewards is determined as the set of candidate paths. α is the learning rate, r is the immediate reward, γ is the discount factor, and maxQ(s′,a′) is the maximum expected reward among all possible actions in the next state.

[0098] Specifically, the set of candidate paths is a set of potential optimized paths obtained through Q-learning exploration. By initializing the Q-table, an initial Q value is assigned to the link of each node pair. With a 90% probability, the path with the highest Q value is selected, and with a 10% probability, a new path is randomly explored. The Q value is updated according to the formula and iterated until convergence. Finally, the path with the highest Q value is determined as the candidate path.

[0099] Step 203, evaluate the stability of the set of candidate paths through an improved XGBoost model to obtain an evaluation result.

[0100] Specifically, the improved XGBoost model is an enhanced gradient boosting decision tree model used to predict the path failure risk. The stability evaluation result is a quantitative result of the robustness of the path in the dynamic network environment. Extract the feature vectors for the candidate paths, input the features into the improved XGBoost model, and output the stability score of each path to quantitatively evaluate the path robustness and identify the vulnerable paths that are susceptible to network fluctuations.

[0101] Step 204, based on the evaluation result, determine the paths with the top 10% stability as the hierarchical path set.

[0102] Specifically, the hierarchical path set is a subset of highly reliable paths screened by the stability threshold. The stability scores of all candidate paths are sorted in descending order, and the 90th percentile is calculated as the threshold. Paths with scores greater than the threshold are retained to generate the hierarchical set.

[0103] Step 205, based on the constraint conditions, calculate the Pareto front of the hierarchical path set through the following formula:

[0104] w1×Q score +w2×(1 - stability_risk)

[0105] where w1 and w2 are dynamic weights, Q score is the Q - value score, and stability_risk is the probability of path failure risk.

[0106] Specifically, the Pareto front is the set of optimal solutions in a multi - objective optimization problem, where the improvement of any one objective requires the degradation of other objectives. The dynamic weights are the objective weights adaptively adjusted according to the network state. Calculate the comprehensive scores for the hierarchical paths, use fast non - dominated sorting to identify the Pareto optimal solutions, remove the dominated solutions, and retain the frontier solution set.

[0107] Step 206, determine the paths constituting the Pareto front as the optimized path sequence.

[0108] Specifically, the optimized path sequence is the final routing scheme obtained by sorting the Pareto front solutions according to the priority. Sort the Pareto front solutions in descending order of the comprehensive score, remove the redundant paths that are completely included in other paths, and store the result as the optimized path sequence.

[0109] Step 207, based on the optimized path sequence and the real - time network topology map, calculate the path scores to obtain the optimized communication path table.

[0110] Specifically, the optimized communication path table is the communication decision basis for storing the optimal paths and metadata. Calculate the final scores for each path in the optimized path sequence, fill the path table in descending order of the scores, and mark the path types to obtain the optimized communication path table.

[0111] This embodiment significantly improves the reliability and adaptability of the mine communication network while ensuring real - time performance through multi - stage optimization.

[0112] In one of the embodiments, the improved XGBoost model is trained by the following method:

[0113] Step 301, based on the real - time failure paths and historical path data, count the proportion of failure paths in the most recent hour to obtain the failure frequency.

[0114] Specifically, the real-time failure path is the set of communication paths marked as unavailable within the current time window, and the historical path data is the historical record storing the status of all paths in the past time period. The failure frequency is a statistical indicator reflecting the failure probability of a path within the current period, determined by the ratio of the number of failure paths to the total number of paths.

[0115] Step 302: Dynamically adjust the weights of the failure paths based on the failure frequency to obtain a dynamic category weight table.

[0116] Specifically, the dynamic category weight table is a weight mapping table dynamically assigned according to the failure frequencies of different paths. If the path failure frequency is greater than the network average failure frequency, the weight increases.

[0117] Step 303: Based on the dynamic category weight table, obtain the loss function through the following formula:

[0118] Loss=-α t (1 - p t ) γ log(p t )

[0119] where Loss is the loss function, α t is the dynamic category weight, p t is the model prediction probability, γ is the focusing parameter, (1 - p t ) γ is the modulation factor, and log(p t ) is the cross-entropy base loss.

[0120] Specifically, the focal loss function is an improved cross-entropy loss function that reduces the weights of easily classified samples through the modulation factor and is used to train the improved XGBoost model.

[0121] Step 304: Extract the features of the candidate path set to obtain a multi-dimensional feature vector, and splice the multi-dimensional feature vector and the historical failure paths to obtain a training data set.

[0122] Specifically, the multi-dimensional feature vector is a set of features describing the path characteristics, including static attributes and dynamic indicators. The training data set is a labeled data set for model training. Extract features from the candidate path set to generate a multi-dimensional vector, and splice the labels of the historical failure paths with the feature vector to obtain the training data set.

[0123] Step 305: Based on the training data set, divide the candidate paths into K groups through the K-means clustering algorithm to obtain grouped paths.

[0124] Specifically, the K-means clustering algorithm is an unsupervised learning algorithm used to divide similar paths into the same group. The grouped paths are subsets of paths with similar characteristics. Paths can be grouped according to feature similarity, which facilitates targeted training of sub-models and improves the model's ability to distinguish heterogeneous paths.

[0125] Step 306: Based on the grouped paths, use each group of candidate paths as the training data set to train an independent XGBoost sub-model, and obtain the trained XGBoost sub-model.

[0126] Specifically, the XGBoost sub-model is an independent prediction model for a specific path group. According to the clustering grouping results, the training data is divided into K subsets. An independent XGBoost model is trained for each subset through focal loss, and the hyperparameters are adjusted through cross-validation to obtain the trained model.

[0127] Step 307: Integrate all the trained XGBoost sub-models to obtain the trained improved XGBoost model.

[0128] Specifically, the improved XGBoost model is a composite prediction model integrated by multiple sub-models. Weights are assigned to each sub-model, and the weights are proportional to the number of samples in their clustering groups. For test samples, the prediction results of each sub-model are combined to obtain the integrated composite prediction model.

[0129] This embodiment significantly improves the accuracy and real-time performance of mine communication network failure prediction through dynamic weight adjustment, feature clustering, and model integration.

[0130] In one of the embodiments, based on the optimized communication path table, the early warning information is synchronized through the improved Raft protocol to obtain a globally consistent early warning information set, including:

[0131] Step 401: Based on the optimized communication path table, determine each node with information forwarding ability as a relay node.

[0132] Specifically, a relay node is an intermediate node that undertakes the task of information forwarding and is usually located at a key hub position in the network topology. Analyze the optimized communication path table, extract the intermediate nodes in all paths, filter the node list, only retain the nodes with forwarding ability, and determine the remaining nodes as relay nodes.

[0133] Step 402: Calculate the credit score of each relay node through the following formula:

[0134] F = 0.6×a + 0.3×b + 0.1×c

[0135] where F is the credit score, a is the online rate, b is the processing capacity, and c is the historical synchronization success rate.

[0136] Specifically, the online rate is the proportion of the available duration of a node within the statistical period, reflecting the stability of the node. The processing capacity is a comprehensive performance indicator of the node, covering computing, storage, and communication capabilities. The historical synchronization success rate is the proportion of nodes that have successfully completed data synchronization in historical tasks.

[0137] Step 403: Obtain the Leader node based on the credit score and the node version number.

[0138] Specifically, the Leader node is a temporary master node responsible for coordinating data synchronization, with the highest decision-making power among the relay nodes. The node version number represents the version of the warning information in the node. The relay node broadcasts a campaign request to the control node, attaching the credit score and version number, and votes according to the election score. The node with the highest score is elected as the Leader. If there is a tie, the node with a higher version number is preferred.

[0139] Step 404: Receive the list of committed logs fed back by the Leader node.

[0140] Among them, the list of committed logs is generated by the Leader node spreading the warning information to each sub-node through the optimal communication path table, and each sub-node feeding back its own version number to the Leader node.

[0141] Specifically, the list of committed logs is a collection of data block sets to be synchronized distributed by the Leader node, containing operation instructions and version information. The Leader distributes the logs to each node according to the optimized communication path table, and each node returns an ACK, attaching the current local version number, to generate the list of committed logs.

[0142] Step 405: Based on the list of committed logs and the global version number, compare the version differences to obtain a list of differential nodes.

[0143] Specifically, the global version number is the identifier of the latest data version maintained by the core layer and is the benchmark for network-wide consistency. The list of differential nodes is a list of nodes whose versions lag behind the global version. Compare the global version number with the local version numbers of each node in the list of committed logs, and determine the nodes with differences as differential nodes.

[0144] Step 406: Push the missing log entries to the differential nodes in the list of differential nodes, and generate a hash based on the log entries; the missing log entries are used to update the warning information and instruct the differential nodes to feedback the differential node hash.

[0145] Specifically, the missing log entries are a set of data update operations that the differential nodes have not received. The differential node hash is the local data fingerprint generated by the node after synchronization is completed and is used for consistency verification. Push the missing logs to the differential nodes, and use RS erasure code to ensure the integrity of transmission. After the node applies the logs, calculate the local data hash and return it.

[0146] Exemplarily, when the version difference is greater than three versions, the set of complete data update operations is forcibly pushed to the differential nodes, overwriting the data status of the differential nodes for global synchronization.

[0147] Step 407: Receive the differential node hash and compare whether the differential node hash is the same as the quorum node hash. If they are the same, the verification is passed, and a globally consistent warning information set is generated.

[0148] Specifically, the quorum nodes are a set of nodes that exceed a certain proportion in the network, representing the consensus majority. The globally consistent warning information set is the set of final states after the data of all network nodes are fully synchronized. The hash values of the quorum nodes are statistically calculated to generate a reference hash, and the differential node hash is compared with the reference hash. If they are consistent, it is marked as synchronization completed, and a globally consistent warning information set is generated.

[0149] In this embodiment, by improving the Raft protocol and blockchain verification, it is ensured that the reported data versions and contents across the network are completely consistent, meeting the requirement of zero error in warning information.

[0150] In one of the embodiments, based on the globally consistent warning information set and the three-dimensional digital twin model of the mine, the optimal escape route is planned to obtain a warning instruction set, including:

[0151] Step 501: Extract the features of the globally consistent warning information set to obtain a feature matrix.

[0152] Specifically, parse the text and numerical fields in the globally consistent warning information set, extract key parameters, perform one-hot encoding on non-numerical features, and normalize to eliminate the dimension difference, so as to convert complex warning information into a numerical matrix that can be processed by machines.

[0153] Step 502: Classify the feature matrix based on a preset level through a random forest classifier to obtain a classification result.

[0154] Specifically, the classification result is the division of the severity of the current disaster impact, which is used to guide different response strategies. Input the feature matrix into a pre-trained random forest model, and determine the final disaster level through the voting of multiple decision trees to obtain the classification result.

[0155] Step 503: Construct a grid map based on the three-dimensional digital twin model of the mine and the real-time network topology map to obtain a weighted map of the mine.

[0156] Specifically, the three-dimensional digital twin model is a virtual mapping of the mine physical environment with millimeter-level accuracy, integrating geological, equipment, and real-time sensor data. A weighted map is used to discretize the mine space into grids, where each grid is assigned a passage cost and a risk weight. The three-dimensional model is divided into uniform grids, the coordinates and attributes of each grid are marked, real-time topological data is fused, the grid weights are dynamically updated, an adjacency matrix is generated, and the movement cost between grids is recorded to obtain a weighted map.

[0157] Step 504, based on the multi-objective A* algorithm, calculate the weighted map through the following formula to obtain an optimized path set:

[0158] f(n) = g(n) + 1.2h(n) + 0.5r(n)

[0159] where f(n) is the cost of each path, g(n) is the actual movement cost, h(n) is the heuristic estimate, and r(n) is the risk coefficient.

[0160] Specifically, the multi-objective A* algorithm is an improved heuristic search algorithm that can optimize the path length, risk, and time simultaneously. The optimized path set is a set of paths with the lowest cost. By initializing the open list and the closed list, adding the starting point to the open list, iteratively selecting the node with the lowest cost for expansion, updating the cost and path of the adjacent grids, and backtracking the path when reaching the end point, the top three optimal paths are generated.

[0161] Step 505, based on the classification result and the optimized path set, generate an early warning instruction set.

[0162] Specifically, the early warning instruction set is a structured data packet containing escape routes, action instructions, and safety tips. Select a push strategy according to the disaster level, combine the optimized path with the emergency plan template to generate an early warning instruction set.

[0163] In this embodiment, through feature engineering, intelligent classification, and multi-objective path optimization, accurate classification is achieved, escape routes are dynamically planned, the timeliness of early warning after a mine disaster occurs is improved, and the safety of mine operations is increased.

[0164] In one of the embodiments, based on the detection of the node response time of the periodic heartbeat signal, a list of failed nodes is obtained, including:

[0165] Step 601, preprocess the node response time based on the sensor data to obtain a standardized data set; the standardized data set includes node id, response time, and sensor status.

[0166] Specifically, the sensor data are physical and environmental indicators collected during the operation of the node, which may include CPU temperature, memory occupancy, network packet loss rate, power supply voltage, etc.

[0167] Exemplarily, the preprocessing may include: denoising, data cleaning, and standardization, etc.

[0168] Step 602: Predict the failure probability of the standardized data set through the LSTM model to obtain a prediction result.

[0169] Specifically, the prediction result is the probability estimate of the node failing within the future time window. The standardized data set is segmented according to the time window to construct training samples, and the failure probability of each node is calculated in real-time through the pre-trained LSTM model to generate a prediction result table.

[0170] Step 603: Obtain a dynamic network delay threshold based on the current network load.

[0171] Specifically, the network load is a comprehensive indicator of the current network resource usage, which may include bandwidth utilization rate, queue depth, and number of concurrent connections, etc. The dynamic network delay threshold is the upper limit of the node response time tolerance dynamically adjusted according to the real-time load. Real-time collect network load indicators, calculate the comprehensive load rate, and dynamically adjust the delay threshold according to the piecewise function.

[0172] Step 604: Based on the dynamic network delay threshold and the prediction result, determine the node failure situation to obtain a preliminary list of failed nodes.

[0173] Specifically, the preliminary list of failed nodes is a set of temporarily failed nodes determined based on the dynamic threshold and the prediction result. Compare the actual response time of each node with the dynamic threshold, combine the LSTM prediction result, and apply the determination condition to generate a preliminary list, which records the failure time and the triggering condition.

[0174] Step 605: Generate a list of failed nodes based on the node dependency relationship and the preliminary failure list.

[0175] Specifically, the node dependency relationship is the logical or physical dependency between nodes in the network, such as routing paths and data synchronization relationships. The list of failed nodes is the set of finally confirmed unavailable nodes, including directly failed and dependently failed nodes. Starting from the preliminary list, traverse the dependency graph to mark indirectly failed nodes, verify whether the nodes meet the dependent failure rules, and update the final list.

[0176] This embodiment realizes the effects of forward-looking detection, accurate judgment, and perception of dependent nodes by combining time series prediction and dynamic thresholds, improves the forward-looking and accuracy of early warning, and avoids the spread of local failures.

[0177] In one of the embodiments, after planning the optimal escape route based on the globally consistent early warning information set and the three-dimensional digital twin model of the mine to obtain an early warning instruction set and sending the early warning instruction set to the escape terminal, it further includes:

[0178] Step 701, receive the reception confirmation list fed back by the terminal.

[0179] Specifically, the reception confirmation list is a set of structured confirmation information returned by the terminal device after receiving the warning instruction.

[0180] Step 702, based on the reception confirmation list, quickly mark the suspicious data through a Bloom filter to obtain a preliminary verification result.

[0181] Specifically, a Bloom filter is a probabilistic data structure used to efficiently detect whether an element belongs to a certain set. The preliminary verification result is a subset of the terminal confirmation data marked as suspicious or trustworthy. Preload the legal terminal IDs into the Bloom filter, traverse the reception confirmation list, perform a membership check on each terminal ID, and output the labeled confirmation list.

[0182] Step 703, verify the hash value of the preliminary verification result to obtain a deep verification result.

[0183] Specifically, the deep verification result is the final judgment result obtained by cryptographically verifying the integrity of the confirmation data. Extract the information hash field from the suspicious confirmation data; pull the original warning instruction from the core layer and recalculate the SHA-256 hash; compare the hash returned by the terminal with the hash of the core layer. If they are the same, mark it as verified passed. If they are different, mark it as verified failed and trigger an alarm to obtain the deep verification result.

[0184] Step 704, based on the historical operation log, judge the similarity of the preliminary verification result through a siamese neural network to obtain a similarity comparison result.

[0185] Specifically, a siamese neural network is a two-branch neural network used to measure the similarity of two input sequences. The similarity comparison result is a quantitative index of the matching degree between the current terminal behavior and the historical normal mode. Extract the behavior characteristics of this terminal in the past 30 days from the historical log to construct a benchmark mode, convert the current reception confirmation data into a feature vector of the same dimension, and calculate the similarity score through the pre-trained siamese network.

[0186] Step 705, generate a set of trusted terminal information based on the preliminary verification result, the deep verification result, and the similarity comparison result.

[0187] Specifically, the set of trusted terminal information is an authoritative list of terminal devices that have passed multi-level verification.

[0188] This embodiment realizes the effects of efficient filtering and intelligent risk control through multi-level joint verification, ensuring that only legal devices can participate in key operations.

[0189] In one of the embodiments, based on the historical operation log, judging the similarity of the preliminary verification result through a siamese neural network to obtain a similarity comparison result includes:

[0190] Step 801: Extract features from the preliminary verification results to obtain a preprocessed data set.

[0191] Exemplarily, the preprocessing process may include: (1) Normalization: Min - Max scaling to the interval [0, 1]; (2) Missing value filling: KNN interpolation to fill in missing items; (3) Dimensionality reduction: Dimensionality reduction using the PCA algorithm.

[0192] Step 802: Input the preprocessed data set into the encoder to obtain an embedding vector.

[0193] Specifically, the encoder is a deep - learning model that maps high - dimensional features to a low - dimensional semantic space, and the embedding vector is a low - dimensional dense vector representing the operating state of the device. Load the pre - trained encoder model, input the preprocessed data set into the encoder, and perform forward propagation to obtain the embedding vector, making the devices significantly separable in the embedding space.

[0194] Step 803: Retrieve the normal embedding vectors of similar devices from the reference library.

[0195] Specifically, the reference library is a dedicated database storing the historical normal - state embedding vectors of similar devices, and the normal embedding vectors are a set of benchmark embedding vectors generated by similar devices in a fault - free state. According to the current device model label, pull the set of normal embedding vectors of the corresponding category from the reference library. If there is multi - working - condition data, screen the closest subset according to the operating environment, and perform secondary clustering on the screened vector set to extract the representative central vector.

[0196] Step 804: Calculate the similarity between the embedding vector and the normal embedding vector through the following formula:

[0197]

[0198] where sim is the similarity, E query is the embedding vector, and E ref is the normal embedding vector.

[0199] Specifically, the cosine similarity is an index to measure the consistency of the directions of two vectors. For each vector to be measured, calculate its cosine similarity with all reference vectors, and take the highest similarity as the final value.

[0200] Step 805: Compare the similarity with a pre - set threshold to obtain a similarity comparison result.

[0201] Specifically, the preset threshold is a critical similarity value for determining whether the device status is normal, and the similarity comparison result is used for the binary determination of the device status. The dynamic threshold of the current device model is pulled from the configuration center, and the size of the similarity is compared and calculated with the threshold. If the similarity is greater than the threshold, it is marked as normal; if the similarity is less than the threshold, it is marked as abnormal, an alarm is triggered, the determination result is recorded in the audit log, and pushed to the operation and maintenance platform.

[0202] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are sequentially shown according to the indications of the arrows, these steps do not necessarily need to be executed in the order indicated by the arrows. Unless there is a clear description in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages do not necessarily need to be executed at the same moment, but can be executed at different moments. The execution order of these steps or stages does not necessarily need to be sequential, but can be executed alternately or alternately with at least a part of the steps or stages in other steps or other steps.

[0203] Based on the same inventive concept, the embodiments of the present application also provide a machine learning-based mine disaster warning system for implementing the above-mentioned machine learning-based mine disaster warning method. The implementation solutions provided by this system to solve problems are similar to the implementation solutions described in the above method. Therefore, the specific limitations in one or more embodiments of the machine learning-based mine disaster warning system provided below can refer to the limitations on the machine learning-based mine disaster warning method in the above text, and will not be repeated here.

[0204] In an exemplary embodiment, as Figure 2 shown, a machine learning-based mine disaster warning system 900 is provided, including:

[0205] A response module 901, configured to detect the node response time based on the periodic heartbeat signal to obtain a list of failed nodes; the periodic heartbeat signal is actively sent by the control node to instruct all nodes to feedback the node response time; the list of failed nodes is used to characterize the failed nodes in the network topology diagram;

[0206] A communication module 902, configured to calculate the optimal path through the Q-learning algorithm based on the list of failed nodes and the real-time network topology diagram to obtain an optimized communication path table;

[0207] A synchronization module 903, configured to synchronize the warning information through an improved Raft protocol based on the optimized communication path table to obtain a globally consistent warning information set;

[0208] The early warning module 904 is used to plan the optimal escape route based on the globally consistent early warning information set and the three-dimensional digital twin model of the mine, obtain the early warning instruction set, and send the early warning instruction set to the escape terminal.

[0209] Furthermore, the communication module 902 is also used for:

[0210] Updating the original network topology map based on the list of failed nodes to obtain an effective topological subgraph;

[0211] Based on the effective topological subgraph, use the following formula to obtain the candidate path set:

[0212] Q(s,a)←Q(s,a)+α[r+γmaxQ(s′,a′)-Q(s,a)]

[0213] where Q(s,a) is the expected cumulative reward for choosing action a in state s, determine the path with the maximum sum of expected cumulative rewards as the candidate path set, α is the learning rate, r is the immediate reward, γ is the discount factor, and maxQ(s′,a′) is the maximum expected reward among all possible actions in the next state;

[0214] Evaluate the stability of the candidate path set through an improved XGBoost model to obtain the evaluation result;

[0215] Based on the evaluation result, determine the paths with the top 10% stability as the graded path set;

[0216] Based on the constraint conditions, calculate the Pareto front of the graded path set through the following formula:

[0217] w1×Q score +w2×(1-stability_risk)

[0218] where w1 and w2 are dynamic weights, Q score is the Q-value score, and stability_risk is the path failure risk probability;

[0219] Determine the paths that constitute the Pareto front as the optimized path sequence;

[0220] Based on the optimized path sequence and the real-time network topology map, calculate the path score to obtain the optimized communication path table.

[0221] Furthermore, the improved XGBoost model is trained through the following method:

[0222] Based on the real-time failed paths and historical path data, count the proportion of failed paths in the most recent hour to obtain the failure frequency;

[0223] Dynamically adjust the weights of failure paths based on the failure frequency to obtain a dynamic category weight table;

[0224] Based on the dynamic category weight table, obtain the loss function through the following formula:

[0225] Loss=-α t (1 - p t ) γ log(p t )

[0226] where Loss is the loss function, α t is the dynamic category weight, p t is the model prediction probability, γ is the focusing parameter, (1 - p t ) γ is the modulation factor, and log(p t ) is the cross - entropy base loss;

[0227] Extract the features of the candidate path set to obtain a multi - dimensional feature vector, and splice the multi - dimensional feature vector and the historical failure paths to obtain a training data set;

[0228] Based on the training data set, divide the candidate paths into K groups through the K - means clustering algorithm to obtain grouped paths;

[0229] Based on the grouped paths, use each group of candidate paths as a training data set to train an independent XGBoost sub - model to obtain a trained XGBoost sub - model;

[0230] Integrate all the trained XGBoost sub - models to obtain a trained improved XGBoost model.

[0231] Furthermore, the synchronization module 903 is also used for:

[0232] Based on the optimized communication path table, determine each node with information forwarding ability as a relay node;

[0233] Calculate the credit score of each relay node through the following formula:

[0234] F=0.6×a + 0.3×b + 0.1×c

[0235] where F is the credit score, a is the online rate, b is the processing ability, and c is the historical synchronization success rate;

[0236] Based on the credit score and the node version number, obtain the Leader node;

[0237] Receive the commit log list fed back by the Leader node;

[0238] Among them, the submission log list is generated by the Leader node spreading warning information to each sub-node through the optimal communication path table, and each sub-node feeding back the version number of each sub-node to the Leader node;

[0239] Based on the submission log list and the global version number, compare the version differences to obtain a list of different nodes;

[0240] Push the missing log entries to the different nodes in the list of different nodes, and generate a hash based on the log entries; the missing log entries are used to update the warning information and instruct the different nodes to feedback the different node hashes;

[0241] Receive the different node hashes, and compare whether the different node hashes are the same as the legal majority node hashes. If they are the same, pass the verification and generate a globally consistent warning information set.

[0242] Furthermore, the warning module 904 is also used for:

[0243] Extract the features of the globally consistent warning information set to obtain a feature matrix;

[0244] Classify the feature matrix based on a preset level through a random forest classifier to obtain a classification result;

[0245] Construct a grid map based on the three-dimensional digital twin model of the mine and the real-time network topology map to obtain a weighted map of the mine;

[0246] Based on the multi-objective A* algorithm, calculate the weighted map through the following formula to obtain an optimized path set:

[0247] f(n) = g(n) + 1.2h(n) + 0.5r(n)

[0248] Among them, f(n) is the cost of each path, g(n) is the actual movement cost, h(n) is the heuristic estimate, and r(n) is the risk coefficient;

[0249] Generate a warning instruction set based on the classification result and the optimized path set.

[0250] Furthermore, the response module 901 is also used for:

[0251] Based on the sensor data, preprocess the node response time to obtain a standardized data set; the standardized data set includes the node id, response time, and sensor status;

[0252] Predict the failure probability of the standardized data set through an LSTM model to obtain a prediction result;

[0253] Based on the current network load, obtain a dynamic network delay threshold;

[0254] Based on the dynamic network delay threshold and prediction results, determine the node failure situation to obtain a preliminary list of failed nodes;

[0255] Based on the node dependency relationship and the preliminary failure list, generate a list of failed nodes.

[0256] Furthermore, the system further includes a verification module for:

[0257] Receive the reception confirmation list fed back by the terminal;

[0258] Based on the reception confirmation list, quickly mark suspicious data through a Bloom filter to obtain a preliminary verification result;

[0259] Verify the hash value of the preliminary verification result to obtain a deep verification result;

[0260] Based on the historical operation logs, judge the similarity of the preliminary verification result through a twin neural network to obtain a similarity comparison result;

[0261] Based on the preliminary verification result, the deep verification result, and the similarity comparison result, generate a set of trusted terminal information.

[0262] Furthermore, the verification module is also used for:

[0263] Extract the features of the preliminary verification result to obtain a preprocessed data set;

[0264] Input the preprocessed data set into the encoder to obtain an embedding vector;

[0265] Retrieve the normal embedding vectors of similar devices from the reference library;

[0266] Calculate the similarity between the embedding vector and the normal embedding vector through the following formula:

[0267]

[0268] where sim is the similarity, E query is the embedding vector, and E ref is the normal embedding vector;

[0269] Compare the similarity with a pre-set threshold to obtain a similarity comparison result.

[0270] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the above-mentioned mine disaster warning method based on machine learning are implemented.

[0271] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0272] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial descriptions of the method embodiments. The device embodiments described above are only illustrative. The components described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of the present disclosure. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0273] The above embodiments only represent several implementation manners of the embodiments of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the patent scope of the embodiments of the application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the embodiments of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the embodiments of the present application.

Claims

1. A mine disaster early warning method based on machine learning, characterized in that, The method includes: Detecting the node response time based on the periodic heartbeat signal to obtain a list of failed nodes; the periodic heartbeat signal is actively sent by the control node and is used to instruct all nodes to feedback the node response time; the list of failed nodes is used to characterize the failed nodes in the network topology graph; Calculating the optimal path through the Q-learning algorithm based on the list of failed nodes and the real-time network topology graph to obtain an optimized communication path table; Synchronizing the warning information based on the optimized communication path table through the improved Raft protocol to obtain a globally consistent warning information set; Planning the optimal escape route based on the globally consistent warning information set and the three-dimensional digital twin model of the mine to obtain a warning instruction set, and sending the warning instruction set to the escape terminal.

2. The mine disaster warning method based on machine learning according to claim 1, wherein, The calculating the optimal path through the Q-learning algorithm based on the list of failed nodes and the real-time network topology graph to obtain an optimized communication path table includes: Updating the original network topology graph based on the list of failed nodes to obtain an effective topology subgraph; Based on the effective topology subgraph, obtaining a candidate path set through the following formula: Q(s,a)←Q(s,a)+α[r+γmaxQ(s′,a′)-Q(s,a)] where Q(s,a) is the expected cumulative reward for selecting action a in state s, the path with the maximum sum of expected cumulative rewards is determined as the candidate path set, α is the learning rate, r is the immediate reward, γ is the discount factor, and maxQ(s′,a′) is the maximum expected reward among all possible actions in the next state; Evaluating the stability of the candidate path set through the improved XGBoost model to obtain an evaluation result; Based on the evaluation result, determining the paths with the top 10% stability as the hierarchical path set; Calculating the Pareto front of the hierarchical path set based on the constraint conditions through the following formula: w1×Q score +w2×(1 - stability_risk) Among them, w1 and w2 are dynamic weights, Q score is the Q-value score, and stability_risk is the path failure risk probability; Determining the paths constituting the Pareto front as the optimized path sequence; Calculating the path score based on the optimized path sequence and the real-time network topology graph to obtain the optimized communication path table.

3. The mine disaster warning method based on machine learning according to claim 2, characterized in that The improved XGBoost model is trained through the following method: Based on the real-time failed paths and historical path data, statistically calculating the proportion of failed paths in the most recent hour to obtain the failure frequency; Dynamically adjusting the weights of the failed paths based on the failure frequency to obtain a dynamic category weight table; Based on the dynamic category weight table, obtaining the loss function through the following formula: Loss=-α t (1 - p t ) γ log(p t ) Among them, Loss is the loss function, and α t is the dynamic class weight, p t is the model prediction probability, γ is the focusing parameter, (1 - p t ) γ is the modulation factor, and log(p t ) is the cross-entropy based loss; Extracting the features of the candidate path set to obtain a multi-dimensional feature vector, and splicing the multi-dimensional feature vector and the historical failed paths to obtain a training data set; Dividing the candidate paths into K groups through the K-means clustering algorithm based on the training data set to obtain grouped paths; Based on the grouped paths, using each group of candidate paths as a training data set to train an independent XGBoost sub-model to obtain a trained XGBoost sub-model; Integrating all the trained XGBoost sub-models to obtain the trained improved XGBoost model.

4. The mine disaster early warning method based on machine learning according to claim 1, wherein Based on the optimized communication path table, synchronize warning information through an improved Raft protocol to obtain a globally consistent warning information set, including: Based on the optimized communication path table, determine each node with information forwarding ability as a relay node; Calculate the credit score of each relay node through the following formula: F = 0.6×a + 0.3×b + 0.1×c Where, F is the credit score, a is the online rate, b is the processing capacity, and c is the historical synchronization success rate; Based on the credit score and the node version number, obtain the Leader node; Receive the committed log list feedback by the Leader node; Among them, the committed log list is generated by the Leader node spreading the warning information to each sub-node through the optimal communication path table, and each sub-node feeds back the version number of each sub-node to the Leader node; Based on the committed log list and the global version number, compare the version differences to obtain a list of different nodes; Push the missing log entries to the different nodes in the list of different nodes, and generate a hash based on the log entries; the missing log entries are used to update the warning information and instruct the different nodes to feedback the different node hashes; Receive the different node hashes, and compare whether the different node hashes are the same as the legal majority node hashes. If they are the same, pass the verification and generate the globally consistent warning information set.

5. The mine disaster warning method based on machine learning according to claim 1, characterized in that, Based on the globally consistent warning information set and the three-dimensional digital twin model of the mine, plan the optimal escape route to obtain a warning instruction set, including: Extract the features of the globally consistent warning information set to obtain a feature matrix; Classify the feature matrix through a random forest classifier based on a preset level to obtain a classification result; Based on the three-dimensional digital twin model of the mine and the real-time network topology map, construct a grid map to obtain a weighted map of the mine; Based on the multi-objective A* algorithm, calculate the weighted map through the following formula to obtain an optimized path set: f(n) = g(n) + 1.2h(n) + 0.5r(n) Where, f(n) is the cost of each path, g(n) is the actual movement cost, h(n) is the heuristic estimate, and r(n) is the risk coefficient; Based on the classification result and the optimized path set, generate the warning instruction set.

6. The mine disaster early warning method based on machine learning according to claim 1, wherein Based on the periodic heartbeat signal to detect the node response time, obtain a list of failed nodes, including: Based on the sensor data, preprocess the node response time to obtain a standardized data set; the standardized data set includes the node id, response time, and sensor status; Predict the failure probability of the standardized data set through an LSTM model to obtain a prediction result; Based on the current network load, obtain a dynamic network delay threshold; Based on the dynamic network delay threshold and the prediction result, determine the node failure situation to obtain a preliminary list of failed nodes; Based on the node dependency relationship and the preliminary failure list, generate the list of failed nodes.

7. The mine disaster warning method based on machine learning according to any one of claims 1 to 6, characterized in that, After planning the optimal escape route based on the globally consistent warning information set and the three-dimensional digital twin model of the mine to obtain a warning instruction set, and sending the warning instruction set to the escape terminal, it further includes: A received confirmation list fed back by the receiving terminal; Based on the received confirmation list, quickly mark suspicious data through a Bloom filter to obtain a preliminary verification result; Verify the hash value of the preliminary verification result to obtain a deep verification result; Based on historical operation logs, judge the similarity of the preliminary verification result through a Siamese neural network to obtain a similarity comparison result; Generate a trusted terminal information set based on the preliminary verification result, the deep verification result, and the similarity comparison result.

8. The mine disaster warning method based on machine learning according to claim 7, characterized in that The step of judging the similarity of the preliminary verification result through a Siamese neural network based on historical operation logs to obtain a similarity comparison result includes: Extract features from the preliminary verification result to obtain a preprocessed data set; Input the preprocessed data set into an encoder to obtain an embedding vector; Retrieve the normal embedding vectors of similar devices from the reference library; Calculate the similarity between the embedding vector and the normal embedding vector through the following formula: Among them, sim is the similarity, and W query is the embedding vector, and E ref is the normal embedding vector; Compare the similarity with a preset threshold to obtain the similarity comparison result.

9. A mine disaster early warning system based on machine learning, characterized in that, The system includes: A response module for detecting the node response time based on a periodic heartbeat signal to obtain a list of failed nodes; the periodic heartbeat signal is actively sent by the control node for instructing all nodes to feedback the node response time; the list of failed nodes is used to represent the failed nodes in the network topology diagram; A communication module for calculating an optimal path through a Q-learning algorithm based on the list of failed nodes and the real-time network topology diagram to obtain an optimized communication path table; A synchronization module for synchronizing early warning information through an improved Raft protocol based on the optimized communication path table to obtain a globally consistent early warning information set; An early warning module for planning an optimal escape route based on the globally consistent early warning information set and the three-dimensional digital twin model of the mine to obtain an early warning instruction set and send the early warning instruction set to the escape terminal.

10. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images