Edge computing driven coal mine underground multi-parameter adaptive sensing and abnormal identification sensor network

Abstract

This invention relates to the field of coal mine safety monitoring and intelligent early warning technology, and discloses an edge computing-driven multi-parameter adaptive sensing and anomaly identification sensor network for underground coal mines. The network includes: multiple intrinsically safe intelligent sensor nodes for constructing a distributed monitoring network, comprising: a multi-parameter acquisition module for real-time acquisition of various monitoring parameters in the coal mine; an edge computing module deployed with a lightweight long-short-term relation-improved gated cyclic unit time-series prediction model, which adaptively corrects the background values ​​of monitoring parameters to dynamically generate threshold ranges and generate preliminary anomaly judgment results; a collaborative communication module that, after generating the preliminary anomaly judgment results, interacts with neighboring nodes to generate spatiotemporal collaborative verification results; and an anomaly decision module that, based on the preliminary anomaly judgment results and the spatiotemporal collaborative verification results, generates multi-parameter fusion identification results and outputs graded early warnings. This invention significantly improves the accuracy, real-time performance, and engineering feasibility of identifying precursors to dynamic disasters in underground coal mines.

Description

Technical Field

[0001] This invention relates to the field of coal mine safety monitoring and intelligent early warning technology, specifically to an edge computing-driven multi-parameter adaptive sensing and anomaly identification sensor network for underground coal mines. Background Technology

[0002] Dynamic disasters such as underground gas and rock bursts are major threats to coal mine safety in my country. Real-time and accurate monitoring and identification of disaster precursors are core aspects of coal mine safety control. Currently, the sensor monitoring networks widely used in underground coal mines generally adopt fixed threshold alarm mechanisms, with data acquisition and anomaly detection completed independently by a single node. This has revealed significant technical deficiencies in practical engineering applications.

[0003] First, the fixed threshold alarm mechanism has extremely poor environmental adaptability. Underground ventilation conditions, air pressure, mining progress, and other environmental factors are constantly and gradually changing. Fixed thresholds cannot adaptively adjust to these environmental changes, easily leading to systemic false alarms. Furthermore, manually configuring alarm thresholds relies on the experience of on-site personnel, lacking theoretical basis and making it difficult to adapt to the complex conditions of different mines and mining faces. Second, the independent identification by a single node has weak anti-interference capabilities. Frequent local disturbances such as blasting, equipment start-up and shutdown, and electromagnetic interference occur underground. Single nodes cannot distinguish between local disturbances and actual disaster precursors, lacking the spatiotemporal collaborative verification capability of multiple nodes. This results in high false alarm and missed alarm rates in the monitoring system, severely impacting the on-site personnel's trust in early warning information. Third, multi-parameter monitoring lacks fusion and identification at the physical mechanism level. Existing monitoring systems only achieve simple superposition of multi-parameter data, failing to establish correlation rules between parameters based on the physical mechanisms of coal mine disaster evolution. This results in insufficient ability to capture the precursor characteristics of complex dynamic disasters such as coal and gas outbursts and rock bursts, making accurate disaster prediction impossible. Fourth, artificial intelligence algorithms are difficult to deploy at the underground edge. Existing deep learning-based disaster identification algorithms mostly rely on cloud computing power to complete inference, resulting in high data transmission latency and high dependence on underground networks. Meanwhile, complex time-series prediction models have a large number of parameters and high computational complexity, making it impossible to achieve local real-time inference on low-power, low-computing-power edge hardware underground, thus failing to meet the millisecond-level real-time requirements of disaster monitoring.

[0004] Meanwhile, existing technologies include solutions that directly transplant time series prediction models from other fields such as finance and industry to coal mine monitoring scenarios. These solutions suffer from core problems such as imprecise domain mapping, lack of physical basis for constructing correlation structures, mismatch between model complexity and edge hardware computing power, and insufficient model generalization ability due to the scarcity of coal mine disaster samples. They cannot adapt to the special working environment and disaster evolution patterns in underground coal mines, making it difficult to implement in actual engineering projects. Summary of the Invention

[0005] The purpose of this invention is to overcome the core shortcomings of existing underground coal mine sensor networks, such as poor environmental adaptability, high false alarm and false negative rates, lack of collaborative identification capabilities, and difficulty in deploying artificial intelligence algorithms locally on low-power underground hardware. This invention provides an edge computing-driven multi-parameter adaptive sensing and anomaly identification sensor network for underground coal mines. This network embeds lightweight artificial intelligence algorithms based on the reconstruction of the physical mechanisms of coal mine disasters into intelligent sensor terminals. It achieves adaptive drift correction of background values ​​of monitoring parameters, spatiotemporal collaborative verification between nodes, and multi-parameter adaptive fusion identification, thereby accurately identifying and filtering various interference signals underground. This significantly improves the accuracy, real-time performance, and engineering feasibility of identifying precursors to dynamic disasters in underground coal mines, providing an efficient and reliable solution for disaster monitoring in complex underground coal mine environments.

[0006] To achieve the above objectives, the following technical solution is adopted:

[0007] This invention provides an edge computing-driven multi-parameter adaptive sensing and anomaly recognition sensor network for underground coal mines, comprising:

[0008] Multiple intrinsically safe smart sensor nodes for mining are connected to form a distributed monitoring network through a self-organizing network of a mining low-power wide area network.

[0009] Each of the intrinsically safe smart sensor nodes for mining applications includes:

[0010] Multi-parameter acquisition module, used to acquire various monitoring parameters in coal mines in real time;

[0011] The edge computing module is equipped with a lightweight long-short-term relationship-improved gated cyclic unit time-series prediction model based on the reconstruction of the physical mechanism of coal mine disasters. This model is used to process the monitoring parameters locally, realize adaptive drift correction of the background values ​​of the monitoring parameters to dynamically generate threshold ranges, and perform preliminary anomaly judgment on the monitoring parameters based on the threshold ranges to generate preliminary anomaly judgment results. The lightweight long-short-term relationship-improved gated cyclic unit time-series prediction model is used to integrate the long-term stable correlation between monitoring parameters determined by physical laws and disaster mechanisms, the short-term dynamic linkage caused by underground working condition disturbances, and the original time-series characteristics of the various monitoring parameters to predict the background value prediction vector of each monitoring parameter at the next moment, and realize adaptive drift correction of the background value based on the background value prediction vector.

[0012] The collaborative communication module is used to interact with neighboring nodes after the edge computing module generates an initial anomaly judgment result, and to perform spatiotemporal collaborative verification based on the interaction data to generate a spatiotemporal collaborative verification result.

[0013] The anomaly decision module is used to perform multi-parameter fusion identification based on the initial anomaly judgment result and the spatiotemporal collaborative verification result, generate multi-parameter fusion identification result, and output hierarchical warning based on the multi-parameter fusion identification result.

[0014] Furthermore, the lightweight long-short-term relationship-improved gated recurrent unit time series prediction model adopts a structure combining long-short-term feature physical fusion and a lightweight recurrent neural network, specifically including:

[0015] The long-term relationship feature extraction unit is used to perform feature fusion on the multi-parameter original feature vector at time t based on a pre-constructed long-term association weight matrix with physical basis, and generate a long-term relationship fusion feature vector at time t to extract the long-term stable association relationship.

[0016] The short-term relationship feature extraction unit is used to construct a short-term correlation degree diagonal matrix based on the parameters calculated within the sliding window and the Pearson correlation coefficient, and to perform feature fusion on the original multi-parameter feature vector at time t to generate a short-term relationship fusion feature vector at time t, so as to extract the short-term dynamic linkage relationship.

[0017] The lightweight gated loop unit is used to input the multi-parameter original feature vector, long-term relationship fusion feature vector and short-term relationship fusion feature vector after concatenation, and output the temporal fusion hidden state at time t through the calculation of update gate, reset gate and candidate hidden state;

[0018] A single-layer multilayer perceptron is used to perform a linear transformation on the temporal fusion hidden state and output the background value prediction vector of each monitoring parameter at time t+1.

[0019] Furthermore, the adaptive drift correction of the background value includes the generation of a dynamic baseline and an adaptive threshold:

[0020] The edge computing module calculates the smooth mean μ of each monitoring parameter and the standard deviation σ of the measured data within the sliding window based on the background value prediction vector output by the lightweight long-short-term relationship-improved gated loop unit time-series prediction model, combined with the historical measured data within the sliding window.

[0021] The edge computing module is also used to determine the optimal threshold coefficient offline based on the subject operating characteristic curve of historical abnormal data in the mine. Based on the smoothed mean μ and the standard deviation σ, an anomaly detection upper limit threshold is dynamically generated. and lower threshold The threshold range is obtained.

[0022] Furthermore, the adaptive drift correction of the background value also includes a baseline adaptive update and anomaly locking mechanism:

[0023] The edge computing module uses the measured values ​​of monitoring parameters and the upper limit threshold. and lower threshold To determine the relative relationships, perform the following operations:

[0024] When the measured value of the monitoring parameter remains within the threshold range, it is determined to be a gradual environmental change scenario. The smooth mean μ and standard deviation σ are updated according to a preset cycle to achieve adaptive drift of the alarm baseline.

[0025] When the measured value of a monitoring parameter exceeds the threshold range in a single instance or the measured values ​​of three consecutive sampling steps are close to the threshold, it is determined to be an abnormal change scenario. The current baseline state is immediately locked, the update of the smoothed mean μ and standard deviation σ is stopped, and the abnormal initial judgment mode is triggered.

[0026] Furthermore, the spatiotemporal collaborative verification includes time dimension verification and spatial dimension verification;

[0027] The time dimension verification includes: downsampling and feature extraction of the microseismic signal, and determining whether it is transient interference and filtering it, or continuous abnormality and triggering the spatial dimension verification, based on the duration of the abnormal signal, energy change rate, signal fallback speed, or matching degree with a preset interference feature library.

[0028] The spatial dimension verification includes: the node that triggers the initial anomaly judgment broadcasts a query request to neighboring nodes through the collaborative communication module, and based on the received feedback data from neighboring nodes, combined with the spatial evolution physical laws of coal mine disasters, collaboratively determines whether the current anomaly is a false alarm or a real disaster precursor, and generates the spatiotemporal collaborative verification result.

[0029] Furthermore, the spatial dimension verification also includes communication latency fault tolerance and degradation processing mechanisms:

[0030] Set a timeout period for neighbor node feedback. Nodes that fail to respond within the timeout period are marked as communication abnormal nodes and will not participate in collaborative identification.

[0031] If the number of valid neighboring nodes participating in the identification is lower than a preset threshold, it will be downgraded to local multi-parameter fusion identification. In this case, the spatiotemporal collaborative verification result will directly indicate downgrade processing.

[0032] Furthermore, the multi-parameter fusion identification includes:

[0033] A multi-parameter association rule base is established based on the physical evolution mechanism of different dynamic disaster types in coal mines. The rule base includes core association parameters and their initial weights.

[0034] Based on the degree of deviation between the measured values ​​of the monitoring parameters and the dynamic thresholds, the single-parameter anomaly probability of each core associated parameter is calculated.

[0035] The weights of the core correlation parameters are periodically updated based on the historical number of true positive samples and false positive samples.

[0036] The comprehensive anomaly probability is calculated based on the single-parameter anomaly probability and its adaptive weight.

[0037] Based on the magnitude of the comprehensive anomaly probability, a four-level graded early warning output is executed.

[0038] Furthermore, the multi-parameter acquisition module includes a gas sensor, a micro-vibration probe, a temperature sensor, and a pressure sensor, and all sensors meet the intrinsically safe certification requirements for underground coal mines.

[0039] Furthermore, the edge computing module uses an STM32H7 series chip as its microcontroller unit. The lightweight long-short-term relationship-improved gated loop unit model is deployed in this microcontroller unit after being quantized by INT8, and its single-sample inference time is less than 5ms.

[0040] The collaborative communication module adopts a mining low-power wide area network LoRa-PRO module, which operates in the 433MHz unlicensed frequency band and has an end-to-end communication latency of less than 80ms.

[0041] Furthermore, the anomaly decision-making module includes an intrinsically safe audible and visual alarm and a mining fourth-generation mobile communication technology module, used to trigger local alarms and upload early warning information to the ground safety monitoring center, respectively.

[0042] Compared with the prior art, the present invention achieves the following beneficial effects:

[0043] 1. The model design possesses rigorous physical rationality and scenario adaptability: This invention reconstructs the long-term and short-term relationship based on the evolution mechanism and physical laws of coal mine disasters - the improved gated cyclic unit (LSR-IGRU) model, redefines the connotation of the long-term and short-term relationship adapted to coal mine monitoring scenarios, strips away redundant features and structures from non-coal mining fields, and all parameter correlations and identification rules have clear coal mine engineering and physical basis, solving the logical defects of cross-domain model porting; at the same time, through extremely lightweight design, the number of model parameters is compressed to less than 10k, perfectly adapting to the computing power and storage requirements of underground low-power microcontroller units (MCUs).

[0044] 2. Edge deployment offers excellent feasibility and real-time performance: All algorithms in this invention are computed locally on the edge side of the intelligent sensor node. The lightweight LSR-IGRU model, after INT8 quantization, has a single-sample inference time of less than 5ms on the STM32H743 MCU and an end-to-end early warning latency of less than 100ms. It does not rely on cloud computing power or underground network transmission, fundamentally solving the problems of high transmission latency and high network dependence of traditional artificial intelligence algorithms, and fully meeting the real-time requirements of underground disaster monitoring in coal mines.

[0045] 3. High Accuracy and Strong Anti-interference Capability in Disaster Identification: This invention achieves smooth updates of the alarm baseline as the downhole environment gradually changes through an adaptive background value drift correction algorithm, fundamentally avoiding systemic false alarms caused by gradual environmental changes. Through a spatiotemporal collaborative verification mechanism, it achieves transient interference filtering in the time dimension and multi-node collaborative logic verification in the spatial dimension, effectively eliminating misjudgments caused by local disturbances. Through a multi-parameter adaptive fusion identification model, it establishes association rules based on the physical mechanism of disasters, avoiding false alarms caused by single-parameter anomalies. Validated on an independent test set, the multi-parameter anomaly identification accuracy of this invention reaches 95.2%, with a false alarm rate as low as 2.1%, significantly outperforming traditional time-series algorithms such as Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs) and fixed threshold methods.

[0046] 4. The technical solution possesses complete reproducibility and verifiability: This invention establishes the construction standards, statistical characteristics, sample size, and full-process preprocessing specifications for the model training dataset. It also discloses the hyperparameters, hardware environment, and performance verification results for model training. Third parties can construct similar coal mine monitoring datasets based on the technical solution disclosed in this invention to reproduce the model and experimental results. The model performance data is based on independent test sets of real underground coal mine monitoring data, possessing objectivity and verifiability.

[0047] 5. Convenient and maintainable engineering implementation: This invention establishes the selection, interface definition, and measured performance indicators of each hardware module of the intelligent sensor node. All hardware meets the intrinsically safe certification requirements for underground coal mines and can be directly adapted to the installation specifications of existing coal mine safety monitoring systems. It designs engineering mechanisms such as communication delay fault tolerance and degradation due to insufficient effective neighbor nodes, perfectly adapting to the complex wireless communication environment and operation scenarios underground. At the same time, the model parameters and weights support offline iterative updates without the need to redeploy hardware equipment, greatly reducing the operation and maintenance costs and operational difficulty of the mine, and demonstrating excellent engineering implementation.

[0048] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0049] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0050] Figure 1 This is a schematic diagram of the system architecture of an edge computing-driven multi-parameter adaptive sensing and anomaly recognition sensor network in underground coal mines according to an embodiment of the present invention.

[0051] Figure 2 This is a schematic diagram of the lightweight long-short-term relationship-improved gated recurrent unit (LSR-IGRU) model in an embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of the background value adaptive drift correction process in an embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram of the spatiotemporal collaborative verification and multi-parameter fusion identification process in an embodiment of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0055] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0056] Figure 1 This is a schematic diagram of the system architecture of an edge computing-driven multi-parameter adaptive sensing and anomaly recognition sensor network in underground coal mines, according to an embodiment of the present invention. Figure 1As shown, the edge computing-driven multi-parameter adaptive sensing and anomaly identification sensor network for underground coal mines provided by this invention consists of several intrinsically safe intelligent sensor nodes 100 with edge computing capabilities, forming a distributed monitoring network through a self-organizing LoRa network. Each intelligent sensor node has clearly defined hardware selection, interfaces, and performance parameters. At the software level, a lightweight long-short-term relationship-improved gated cyclic unit (LSR-IGRU) time-series prediction model based on coal mine physical mechanisms is deployed. The core technical solution includes five main parts: hardware module definition and selection, background value adaptive drift correction based on the lightweight LSR-IGRU model, spatiotemporal collaborative verification, multi-parameter fusion identification, and engineering deployment and operation and maintenance processes. The specific implementation methods for each part are as follows:

[0057] (I) Hardware Module Definition and Selection

[0058] The intelligent sensor node 100 of this invention consists of a multi-parameter acquisition module 110, an edge computing module 120, a collaborative communication module 130, and an anomaly decision-making module 140. The core functions, hardware selection, interface definitions, and measured performance indicators of each module are clearly defined as follows. All hardware meets the intrinsically safe certification requirements for underground coal mines, as shown in Table 1.

[0059] Table 1

[0060]

[0061] Among them: the multi-parameter acquisition module 110 is used to acquire various monitoring parameters in coal mines in real time, as shown in Table 1; specifically, the multi-parameter acquisition module 110 includes a gas sensor, a micro-vibration probe, a temperature sensor and a pressure sensor, and all sensors meet the intrinsic safety certification requirements for underground coal mines.

[0062] The edge computing module 120 is equipped with a lightweight long-short-term relationship-improved gated cyclic unit time-series prediction model based on the reconstruction of the physical mechanism of coal mine disasters. This model is used to process monitoring parameters locally, realize adaptive drift correction of the background values ​​of monitoring parameters to dynamically generate threshold ranges, and perform preliminary anomaly judgment on the monitoring parameters based on the threshold ranges to generate preliminary anomaly judgment results. The lightweight long-short-term relationship-improved gated cyclic unit time-series prediction model is used to integrate the long-term stable correlation between monitoring parameters determined by physical laws and disaster mechanisms, the short-term dynamic linkage relationship caused by underground working condition disturbances, and the original time-series characteristics of various monitoring parameters to predict the background value prediction vector of each monitoring parameter at the next moment, and realize adaptive drift correction of the background value based on the background value prediction vector.

[0063] The collaborative communication module 130 is used to interact with neighboring nodes after the edge computing module generates an initial anomaly judgment result, and to perform spatiotemporal collaborative verification based on the interaction data to generate spatiotemporal collaborative verification results.

[0064] The anomaly decision module 140 is used to perform multi-parameter fusion identification based on the initial anomaly judgment result and the spatiotemporal collaborative verification result, generate multi-parameter fusion identification result, and output hierarchical early warning based on the multi-parameter fusion identification result.

[0065] (II) Lightweight Long-Short-Term Relationship Based on Physical Mechanism Reconstruction - Improved Gated Recurrent Unit (LSR-IGRU) Model and Adaptive Drift Correction for Background Value

[0066] Regarding the construction of edge computing module 120: The long-short-term relationship-improved gated cyclic unit (LSR-IGRU) model proposed in this invention has the core principle of fusing the long-short-term correlation and time-series characteristics of monitoring parameters to complete time-series prediction. By mining the long-term stable correlation between parameters determined by physical laws and disaster mechanisms, and the short-term dynamic linkage caused by underground working condition disturbances, the two types of correlation features are deeply fused with the original time-series characteristics of the parameters and then input into the improved gated cyclic unit (GRU) to complete time-series prediction. This invention reconstructs the long-short-term relationship connotation of the model based on the physical mechanism of coal mine disaster evolution, strips away redundant features and structures in non-coal mining fields, adapts to the underground low-power microcontroller unit (MCU) through an extremely lightweight design, and realizes accurate prediction and adaptive drift correction of the background value of underground monitoring parameters based on this model. The specific implementation steps are steps S1, S2, and S3. Figure 2 The diagram shown is a structural schematic of the lightweight long-short-term relationship-improved gated recurrent unit (LSR-IGRU) model in an embodiment of the present invention.

[0067] Step S1: Lightweighting of long-term and short-term relationships - Reconstruction and adaptation of the improved gated recurrent unit (LSR-IGRU) model.

[0068] Furthermore, the lightweight long-short relationship-improved gated recurrent unit time series prediction model adopts a structure that combines long-short feature physical fusion with a lightweight recurrent neural network. Specifically, it includes: a long-term relationship feature extraction unit, used to extract long-term relationship weights based on a pre-constructed, physically-based long-term association weight matrix. For the multi-parameter original feature vector at time t Perform feature fusion to generate a long-term relation fusion feature vector at time t. The system extracts long-term stable associations; the short-term relationship feature extraction unit is used to construct a diagonal matrix of short-term association degrees based on the parameters calculated within the sliding window and the Pearson correlation coefficient. For the multi-parameter original feature vector at time t Perform feature fusion to generate a short-term relation fusion feature vector at time t. To extract short-term dynamic linkages; a lightweight gated loop unit is used to process multi-parameter original feature vectors. Long-term relationship fusion feature vector Feature vector fusion with short-term relationships After concatenating the input, and through the calculation of update gates, reset gates, and candidate hidden states, the temporal fusion hidden state at time t is output. Single-layer multilayer perceptron, used for temporal fusion of hidden states. Perform a linear transformation to output the background value prediction vector of each monitoring parameter at time t+1. The specific implementation process is as follows:

[0069] Based on the coal mine safety engineering and disaster evolution mechanism, the long-term / short-term relationship of monitoring parameters is defined, and a core structure of long-term and short-term feature physical fusion + single-layer lightweight gated recurrent unit (GRU) + single-layer multilayer perceptron (MLP) is designed. All calculations are linear / nonlinear tensor operations, which are adapted to embedded artificial intelligence (AI) inference frameworks. The specific implementation is steps S11, S12 and S13.

[0070] Step S11: Reconstruction of the coal mine scene based on the core idea of ​​the model

[0071] Through the core logic of "integrating long-short-term relationship for time-series prediction" of the long-short-term relationship-improved gated loop unit (LSR-IGRU), based on the coal mine disaster evolution mechanism and physical laws, the long / short-term relationship of multiple parameters is defined. All the relationships have clear physical meanings. Parameters without physical relationships are not connected to avoid redundant calculations: (1) Long-term relationship: Stable relationship between parameters determined by physical laws and coal mine disaster evolution mechanism, such as gas concentration and gas pressure being negatively correlated, and temperature and gas desorption being positively correlated. This type of relationship does not change with short-term changes in underground working conditions; (2) Short-term relationship: Short-term linkage between parameters affected by underground mining working conditions and local disturbances, such as blasting vibration causing short-term changes in micro-vibration signal and gas concentration, and equipment start-up and shutdown causing short-term fluctuations in local temperature and gas pressure. This type of relationship only exists when working condition disturbances occur, and has timeliness and locality.

[0072] Step S12: Physical definition and characteristic calculation of long / short-term relationship

[0073] Based on expert knowledge in the coal mining field and the evolution law of disasters, feature extraction methods for long-term and short-term relationships are constructed respectively. Feature fusion is performed only on parameter pairs with physical correlation, which greatly reduces the amount of computation, specifically (1) and (2).

[0074] (1) Long-term relationship feature extraction

[0075] Based on the "Coal Mine Safety Regulations", the "Detailed Rules for the Prevention and Control of Coal and Gas Outbursts", and the knowledge of experts in the field of coal mine dynamic disasters, a long-term correlation weight matrix with physical basis was pre-constructed. ,in m represents the number of monitoring parameters, and the data is taken during project implementation. (Including gas concentration, microseismic energy, temperature, and air pressure); matrix elements This represents the long-term correlation between any two monitoring parameters e and f. Values ​​are assigned only to parameters with a clear physical correlation; otherwise, a value of 0 is assigned. Here, e represents the row in the matrix, considered the source parameter; and f represents the column in the matrix, considered the target parameter. Specifically, this represents the long-term impact of the f-th monitoring parameter on the e-th monitoring parameter. A typical weight matrix for engineering implementation (m=4, gas concentration, microseismic energy, temperature, and air pressure) is shown in Table 2 below:

[0076] Table 2

[0077]

[0078] The formula for calculating the characteristics of long-term relationships is: ;

[0079] in Let m be the original feature vector with multiple parameters at time t; The long-term relation fusion feature vector at time t has a dimension of m; A long-term correlation weight matrix with physical basis; The original characteristic value of the first monitoring parameter at time t; The original characteristic value of the second monitoring parameter at time t; : The original characteristic value of the m-th monitoring parameter at time t; m: the number of monitoring parameters, taken for engineering implementation. .

[0080] (2) Short-term relationship feature extraction

[0081] A Pearson correlation coefficient calculation method with a fixed time window and low update frequency is adopted. Short-term correlation is calculated only for parameter pairs with short-term linkages, avoiding the computational cost of time-step calculations. The specific implementation rules and calculation methods are as follows:

[0082] ① Sliding window length: 30 s (based on a sampling frequency of 1 Hz, with a total of 30 data points, adapted to the time scale of downhole working condition disturbances);

[0083] ② Calculation objects: Only for parameter pairs with short-term linkage, such as micro-seismic energy-gas concentration, equipment temperature-local ambient temperature;

[0084] ③ Correlation calculation: Calculate the Pearson correlation coefficient of the target parameter pairs within the sliding window, as the short-term correlation. ;

[0085] ④ Update frequency: Updated every 30 seconds, not calculated step by step;

[0086] ⑤ The formula for calculating short-term characteristics is: ,in, A diagonal matrix of short-term correlations (dimensions) ), Let m be the short-term relation fusion feature vector at time t (dimension m).

[0087] Step S13: Definition of core formulas and parameters for model inference

[0088] The original time-series features, long-term relationship features, and short-term relationship features are concatenated and input into a lightweight gated recurrent unit (GRU). After passing through a single-layer multilayer perceptron (MLP), the predicted normal background values ​​of each monitored parameter at the next time step are output. All formulas are tensor operations that can be directly implemented by the embedded microcontroller unit (MCU). The core formulas are as follows:

[0089] ;

[0090] in, Update gate, dimension This is consistent with the number of neurons in the hidden layer of the lightweight GRU. Reset Gate, Dimension This is consistent with the number of neurons in the hidden layer of the lightweight GRU. Candidate hidden state, dimension ; The final output of the temporal fusion hidden state at time t, with dimensions... ; : Temporal fusion of hidden states and dimensions at different moments ; : Input the weight matrix corresponding to the features, dimension m is the number of monitoring parameters ( Initialization uses a uniform distribution of Xavier parameters; The weight matrix corresponding to the hidden layer, dimensions Initialization uses Xavier uniform distribution; Bias vector, dimension The initial value is 0; Multilayer Perceptron (MLP) weight matrix, dimensions Multilayer Perceptron (MLP) bias vector, dimension ; sigmoid: sigmoid activation function; tanh: hyperbolic tangent activation function; Hadama product (element-level product); Feature splicing operation; : The time-based background value prediction vector (dimension m) is the final inference output of the model.

[0091] In step S1, the number of neurons in the lightweight GRU hidden layer is set to 8, the number of neurons in the single-layer MLP hidden layer is set to 4, and the activation function is the linear rectified function 6 (ReLU6) to adapt to the quantization requirements of embedded models. After quantization, the number of parameters in the model is <10k, which meets the requirements for edge hardware deployment.

[0092] Step S2: Lightweighting of long-short-term relationships - Training and validation of the improved gated recurrent unit (LSR-IGRU) model.

[0093] The training, validation and testing datasets for the model are constructed based on actual monitoring scenarios in coal mines, covering three types of data: normal operating conditions, disaster precursors and interference. The dataset construction standards, statistical characteristics, training environment and training steps are clearly defined. The model performance is validated based on an independent test set, specifically implemented in steps S21, S22 and S23.

[0094] Step S21: Construction and preprocessing of training / test datasets

[0095] (1) Basic information of the dataset The dataset covers three types of data: normal operating conditions, disaster precursors, and interference. One time series sample consists of multi-parameter data (model input) for 15 consecutive time steps + the true value (model label) at the 16th time step. The dataset is divided into normal: disaster precursor: interference data in a 7:1:2 ratio, and into training set: validation set: test set in an 8:1:1 ratio. The test set is an independent dataset. Specific information is shown in Table 3 below:

[0096] Table 3

[0097]

[0098] The statistical feature engineering for the dataset selected four core monitoring parameters (m=4): gas concentration, microseismic energy, ambient temperature, and downhole gas pressure. The sampling range, mean, standard deviation, data units, and labeling rules for each parameter are clearly defined in Table 4 below, providing a unified standard for dataset construction.

[0099] Table 4

[0100]

[0101] (3) Dataset acquisition, construction and preprocessing standards

[0102] ① Data Acquisition and Construction: Normal operating condition data is extracted from the target mine safety monitoring system during production periods without disasters or significant interference, covering different mining faces, seasons, and ventilation conditions; disaster precursor data includes retrospective labeled data of historical disaster events in the mine and simulated data collected by the coal mine dynamic disaster simulation test bench (built according to GB / T 25217-2010 standard); interference data consists of measured labeled data of scenarios such as underground blasting, equipment start-up and shutdown, and electromagnetic interference.

[0103] ② Data preprocessing: All data undergoes a unified preprocessing procedure, including 1 Hz time synchronization, Criteria for outlier removal, linear interpolation missing value completion, min-max normalization, time series sample construction, and missing value rate. The time sequence segments are directly removed.

[0104] Step S22: Model Training Environment and Training Steps

[0105] Model training is completed on a general-purpose hardware platform with no special computing power requirements. The training environment and the entire training process are clearly defined. All parameters are fixed values ​​to ensure the reproducibility of the training process, specifically (1) and (2).

[0106] (1) Training environment

[0107] The hardware platforms and software frameworks for model training, quantization, and edge deployment are all general-purpose, as shown in Table 5 below:

[0108] Table 5

[0109]

[0110] The edge computing module 120 uses an STM32H7 series chip as its microcontroller unit. The lightweight long-short relationship-improved gated loop unit model is deployed in this microcontroller unit after being quantized by INT8, and its single-sample inference time is less than 5ms.

[0111] (2) Training steps

[0112] Step S221: Complete the construction of the training set, validation set, and test set according to the dataset preprocessing standards in step S21. The test set is completely independent of the training set.

[0113] Step S222: Model initialization, weight matrix ( Initialization is performed using a uniform distribution of Xavier bias vectors. Initialize to 0;

[0114] Step S223: Set the training hyperparameters, the optimizer uses the Adaptive Moment Estimation Optimizer (Adam), and the initial learning rate... batch size Training epochs The loss function used is mean squared error (MSE).

[0115] Step S224: Model training and early stopping. Mini-batch gradient descent is used for training. After each epoch, the model accuracy is verified on the validation set. If the mean squared error (MSE) loss on the validation set does not decrease for three consecutive epochs, early stopping is triggered to avoid model overfitting.

[0116] Step S225: Model lightweighting. The trained model is converted into the Open Neural Network Exchange Format (ONNX), and INT8 quantization is performed using the STM32Cube.AI tool to remove redundant parameters and adapt to the hardware computing power of the STM32H743 MCU.

[0117] Step S226: Model deployment. The quantized model is burned into the flash memory of the edge computing module using the ST-Link simulator to complete the embedded deployment of the model.

[0118] Step S23: Model Performance Validation

[0119] The model performance was validated on an independent test set (17,143 samples), and compared with three baseline algorithms: Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), and Transformer. Validation metrics included mean absolute error (MAE) of gas concentration prediction, accuracy of multi-parameter anomaly detection, false alarm rate of multi-parameter anomaly detection, and inference time per sample. All inference times were measured values ​​using an STM32H743 microcontroller unit (MCU). Specific validation results are shown in Table 6 below.

[0120] Table 6

[0121]

[0122] In a real-world test environment using an STM32H743 microcontroller unit (MCU), the lightweight long short-term relation-improved gated recurrent unit (LSR-IGRU) model of this invention was compared with three baseline algorithms—Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), and Transformer—on an independent test set. The results showed that the LSR-IGRU model had the best overall performance. Regarding the mean absolute error (MAE) of gas concentration prediction, LSR-IGRU was 0.02%CH4, only slightly higher than GRU, and significantly better than the other two algorithms. In terms of multi-parameter anomaly recognition accuracy, LSR-IGRU achieved 95.20%, ranking first, with a corresponding false alarm rate of only 2.10%, the lowest among all compared algorithms. In terms of single-sample model inference time, LSR-IGRU's measured time was 4.2ms, similar to GRU (both less than 5ms), and much faster than the other two algorithms. Overall, the LSR-IGRU model achieves higher prediction accuracy and anomaly identification capabilities while maintaining fast inference performance adapted to edge hardware, making it more suitable for the application needs of disaster monitoring on the edge side of coal mines.

[0123] Step S3: Execution of adaptive drift correction for background values

[0124] Furthermore, the adaptive drift correction of the background value includes: (1) Generation of dynamic baseline and adaptive threshold: The edge computing module 120 calculates the smooth mean μ of each monitoring parameter and the standard deviation σ of the measured data within the sliding window based on the background value prediction vector output by the lightweight long-short-term relationship-improved gated recurrent unit time-series prediction model, combined with the historical measured data within the sliding window; The edge computing module 120 is also used to determine the optimal threshold coefficient offline based on the receiver operating characteristic curve of the mine's historical abnormal data. And based on the smoothed mean μ and standard deviation σ, an upper threshold for anomaly detection is dynamically generated. and lower threshold The threshold range is obtained. (2) Baseline adaptive update and anomaly locking mechanism: The edge computing module obtains the threshold range based on the measured value of the monitoring parameters and the upper limit threshold. and lower threshold Based on the relative relationship, the following operations are performed: When the measured value of the monitored parameter remains within the threshold range, it is determined to be a gradual environmental change scenario. The smoothed mean μ and standard deviation σ are updated according to a preset period to achieve adaptive drift of the alarm baseline. When the measured value of the monitored parameter exceeds the threshold range for a single time or the measured values ​​for three consecutive sampling steps are close to the threshold, it is determined to be an abnormal change scenario. The current baseline state is immediately locked, the update of the smoothed mean μ and standard deviation σ is stopped, and the initial abnormal judgment mode is triggered. The specific implementation process is as follows:

[0125] Based on the background value prediction results of the Lightweight Long Short-Term Relationship-Improved Gated Cyclic Unit (LSR-IGRU) model, and combined with the statistical characteristics of downhole measured data, a dynamic baseline and adaptive threshold are calculated to achieve smooth baseline updates and lock-in of abnormal states, thereby fundamentally avoiding systematic false alarms caused by gradual environmental changes. The specific implementation steps are S31, S32, and S33. Figure 3 The diagram shown is a schematic flowchart of the adaptive drift correction of background values ​​in an embodiment of the present invention.

[0126] Step S31: Dynamic baseline calculation

[0127] Background value based on model prediction By combining the measured data within the sliding window, the smoothed mean of each monitoring parameter is calculated. and standard deviation Sliding window length Configurable parameters for the project (optional) The calculation formula is as follows: This adapts to the environmental chamfering scale under different working conditions.

[0128] ;

[0129] in, : The smoothed mean of the i-th monitoring parameter at time t; 0.7: Smoothing coefficient, determined by offline training and optimization using historical mine data; The i-th monitoring parameter is in The model prediction value at time step L; L: the number of measured data points within the sliding window; : The measured value of the i-th monitoring parameter at time k.

[0130] Step S32: Adaptive generation of dynamic threshold

[0131] Instead of manually configuring sensitivity coefficients, the optimal threshold coefficients are determined offline using receiver operating characteristic (ROC) curves based on historical anomaly data from mines. This achieves an optimal balance between precision and recall in the model. The optimization results for different mine types are as follows: high-gas outburst mines Low-gas mines ; The threshold is updated quarterly based on newly collected anomaly data from the mine, and the formula for calculating the threshold is as follows:

[0132] ; ;

[0133] in, The upper limit threshold for anomaly detection of monitored parameters; The lower limit threshold for detecting anomalies in monitored parameters; Standard deviation of the measured data of the monitoring parameters within the sliding window.

[0134] Step S33: Baseline Adaptive Update and Anomaly Locking

[0135] Based on the relative relationship between the measured values ​​of monitoring parameters and dynamic thresholds, baseline updates or anomaly locking operations are performed, with the specific rules as follows:

[0136] (1) Gradual change in environment: When the measured data is continuously in a state of change. Within the threshold range, the sliding window is updated every 5 minutes, with the mean... with standard deviation It drifts slowly and smoothly with the downhole environment, enabling adaptive adjustment of the alarm baseline;

[0137] (2) Abnormal mutation scenario: When the measured data exceeds the threshold range once or the three consecutive sampling steps are close to the threshold, the current baseline state is immediately locked and the mean is stopped. with standard deviation Upon update, the system enters the initial anomaly detection mode, triggering the subsequent spatiotemporal collaborative verification process.

[0138] (III) Spatiotemporal Co-verification Mechanism

[0139] Spatiotemporal collaborative verification includes time dimension verification and spatial dimension verification. Time dimension verification includes downsampling and feature extraction of microseismic signals, and determining whether an abnormal signal is transient interference and filtering it, or a continuous anomaly and triggering spatial dimension verification, based on the duration of the abnormal signal, energy change rate, signal fallback speed, or matching degree with a preset interference feature library. Spatial dimension verification includes: the node triggering the initial anomaly judgment broadcasts a query request to neighboring nodes through a collaborative communication module, and based on the received feedback data from neighboring nodes, combined with the spatial evolution physical laws of coal mine disasters, collaboratively determines whether the current anomaly is a false alarm or a precursor to a real disaster, generating the spatiotemporal collaborative verification result. A communication delay tolerance and degradation processing mechanism is implemented: a neighboring node feedback timeout is set; nodes that do not respond within the timeout period are marked as communication anomaly nodes and do not participate in collaborative identification; if the number of effective neighboring nodes participating in the identification is lower than a preset threshold, it is downgraded to local multi-parameter fusion identification, and the spatiotemporal collaborative verification result directly indicates degradation processing.

[0140] Regarding the construction of collaborative communication module 130: Addressing the issue of high-frequency sampling and wireless communication delay fluctuations in downhole microseismic signals, this module utilizes downsampling to extract core signal features and a communication delay tolerance mechanism to achieve time-space dual-dimensional anomaly verification, filtering transient interference and local misjudgments. Only anomaly signals conforming to the disaster evolution pattern are included in the subsequent identification process. Specific implementation steps are steps S4 and S5. (For example...) Figure 4 The diagram shown is a schematic flowchart of spatiotemporal collaborative verification and multi-parameter fusion identification in an embodiment of the present invention.

[0141] Step S4: Time dimension verification

[0142] The microseismic signal is downsampled and its features are extracted. Then, the abnormal signals of all monitoring parameters are identified for interference, and transient interference and continuous anomalies are distinguished. This preserves the core features of the disaster precursors while reducing the computational load of edge nodes. The specific implementation is as follows: steps S41 and S42.

[0143] Step S41: Microseismic signal downsampling and feature extraction

[0144] The original signal sampled by the microseismic probe at 10kHz was downsampled to 100Hz, and the core features such as peak value, energy, duration, and frequency distribution within each 10ms time window were extracted. Other environmental parameters (gas, temperature, and air pressure) were kept at a sampling frequency of 1Hz.

[0145] Step S42: Identification and Judgment of Interference and Anomalies

[0146] Extract indicators such as the duration of abnormal signals, rate of energy change, signal fallback speed, and feature library matching degree, and determine transient interference and persistent anomalies according to the following rules:

[0147] (1) Transient interference judgment: If the abnormal signal meets any of the following conditions: duration < 100ms, energy change and rapid return to the threshold within 3 sampling steps, or signal frequency matching degree > 80% with the underground blasting / equipment start-up and shutdown interference feature library, it is judged as transient interference, directly filtered, and the subsequent process is not triggered.

[0148] (2) Continuous anomaly determination: If the abnormal signal simultaneously meets the following conditions: duration > 1s, more than 5 consecutive sampling steps exceeding the dynamic threshold, and the signal characteristics matching the coal mine disaster precursor feature library > 60%, it is determined to be a potential disaster precursor and triggers spatial dimension collaborative verification.

[0149] Step S5: Spatial Dimension Verification

[0150] Based on the measured communication performance of the mining low-power wide area network (LoRa)-PRO module, a communication delay fault tolerance and degradation processing mechanism is designed. Combined with the spatial evolution physical laws of coal mine disasters, the collaborative logic verification of multiple intelligent sensor nodes is realized. The specific implementation is as follows: steps S51, S52, and S53.

[0151] Step S51: Neighbor node query broadcast

[0152] The intelligent sensor node A that triggers a potential disaster precursor broadcasts a query request to all neighboring nodes within a preset geographical radius R (50-500m, configurable for the project) through the collaborative communication module. The request includes the spatial coordinates of node A, the type of abnormal parameters, the measured data, and the preliminary judgment result of the abnormality.

[0153] Step S52: Communication Delay Fault Tolerance and Degradation Processing

[0154] Set the neighbor node feedback timeout to 100ms (based on the measured latency of 80ms for the LoRa module, with a 20ms fluctuation margin), and perform fault tolerance and degradation processing according to the following rules:

[0155] (1) If a neighboring node does not send back data within the timeout period, it is marked as a "communication abnormal node" and will not participate in collaborative identification;

[0156] (2) If the number of valid neighboring nodes participating in the identification is ≥3, spatial collaborative verification is performed normally;

[0157] (3) If the number of valid neighboring nodes participating in the identification is less than 3, it is downgraded to local multi-parameter fusion identification to avoid missed identification due to communication problems.

[0158] Step S53: Spatial Coordination Logic Verification

[0159] Node A collects feedback data from its effective neighboring nodes (real-time monitoring values ​​+ local analysis results), and combines this data with the spatial evolution physical laws of coal mine disasters (such as the decrease in gas concentration with distance and the decrease in microseismic energy with propagation distance) to perform identification. The identification results are divided into false alarms and real anomalies.

[0160] (1) False alarm judgment: If ≥80% of the effective neighboring nodes report no abnormal trend, and the abnormal signal characteristics of node A match the built-in "equipment fault model / local strong interference model" by more than 80%, it is judged as a false alarm, the alarm is automatically suppressed, the event log is recorded and the baseline adaptive update is restarted.

[0161] (2) Real anomaly determination: If ≥50% of the effective neighboring nodes report synchronous anomaly trends, or if the abnormal data of the neighboring nodes show a spatial gradient distribution that conforms to physical laws, it is determined to be a real disaster precursor and enters the multi-parameter fusion identification process.

[0162] (iv) Multi-parameter fusion identification mechanism

[0163] Regarding the construction of the anomaly decision module 140: Based on the physical evolution mechanism of different dynamic disaster types in coal mines, a multi-parameter association rule library with engineering basis is established, an adaptive update mechanism for parameter weights is designed, and the comprehensive anomaly probability is calculated through joint identification of multiple parameters to realize hierarchical early warning of disasters, avoid false alarms caused by single-parameter anomalies, and improve the accuracy of disaster precursor identification. The specific implementation steps are steps S6 and S7.

[0164] Step S6: Establish a multi-parameter association rule base for coal mine disasters

[0165] Based on the "Coal Mine Safety Regulations" and the practical experience of coal mine dynamic disaster prevention and control engineering, a multi-parameter association rule base was established for two major underground dynamic disasters: coal and gas outburst and rock burst. The core association parameters, initial weights, and core disaster identification rules were clarified. The initial weights were determined jointly by coal mine safety engineers' scores and offline data training. The association rule base can be configured according to the disaster type of the mine, as shown in Table 7 below.

[0166] Table 7

[0167]

[0168] Step S7: Multi-parameter fusion identification and hierarchical early warning

[0169] Multi-parameter fusion identification includes: establishing a multi-parameter association rule base based on the physical evolution mechanism of different dynamic disaster types in coal mines, the rule base containing core association parameters and their initial weights; calculating the single-parameter anomaly probability of each core association parameter based on the deviation between the measured values ​​of monitoring parameters and dynamic thresholds; periodically updating the weights of the core association parameters according to the number of historical true positive samples and false positive samples; and calculating the comprehensive anomaly probability based on the single-parameter anomaly probability and its adaptive weights. Based on the magnitude of the overall anomaly probability, a four-level tiered early warning output is executed. The specific implementation process is as follows:

[0170] The comprehensive anomaly probability is calculated based on the single-parameter anomaly probability and adaptive weight. The four-level disaster early warning is realized according to the magnitude of the comprehensive anomaly probability. The weight parameters are updated every quarter to ensure the accuracy of the identification. The specific implementation is as follows: steps S71, S72, S73, and S74.

[0171] Step S71: Calculation of single-parameter anomaly probability

[0172] For each monitoring parameter, the probability of single-parameter anomaly is calculated based on the degree of deviation between its measured value and the dynamic threshold. Range of values The greater the deviation between the measured value and the threshold, The closer it is to 1; the calculation method for the single-parameter anomaly probability is determined by offline fitting of historical mine data, and the engineering implementation adopts the linear fitting method.

[0173] Step S72: Adaptive update of parameter weights

[0174] The weights of each parameter are iteratively updated quarterly based on historical true positive (TP) and false positive (FP) data from the mine. This avoids a decrease in identification accuracy caused by fixed weights. After the update, all weights are normalized to ensure that the sum of the weights is 1. The update formula is as follows:

[0175] ;

[0176] in, : No. The updated weights of the core parameters; : No. The weights of each core parameter before the update; Learning rate, configured to 0.1 in the project; : No. The number of true positive samples for each core parameter; : No. The number of false positive samples for each core parameter. It is an index of a subset of core parameters in a specific disaster identification rule base.

[0177] Step S73: Calculation of Comprehensive Anomaly Probability

[0178] Based on single-parameter anomaly probability and adaptive weights, a multi-parameter joint comprehensive anomaly probability is calculated to reflect the comprehensive anomaly degree of disaster precursors. The calculation formula is as follows:

[0179] ;

[0180] in, Overall anomaly probability, range of values M: Number of core correlation parameters for the target disaster type; : Indicates that for a certain type of disaster (such as a coal and gas outburst), the first Adaptive weights for each core parameter; : No. The probability of anomalies in each core parameter.

[0181] Step S74: Tiered early warning output

[0182] The anomaly decision module 140 includes an intrinsically safe audible and visual alarm and a mining fourth-generation mobile communication technology module, which are used to trigger local alarms and upload early warning information to the ground safety monitoring center, respectively.

[0183] Based on the comprehensive anomaly probability The size of the alarm determines the level of four-tiered early warning, and the warning results are simultaneously output to the local audible and visual alarm and the ground safety monitoring center. Only when the overall anomaly probability is considered... The highest level local alarm will be triggered when the alarm is triggered to avoid excessive warnings. The specific warning rules are as follows:

[0184] (1) No abnormalities were found; normal monitoring and baseline adaptive updates will continue.

[0185] (2) Low-level warnings only record anomaly logs on local smart sensor nodes and do not upload them to the ground.

[0186] (3) Medium-level warning: abnormal information is uploaded to the ground security monitoring center, but there is no local audible or visual alarm.

[0187] (4) High-level disaster warnings immediately trigger local intrinsically safe audible and visual alarms, simultaneously uploading alarm information and multi-parameter measured data to the ground monitoring center, and synchronizing the warning information to neighboring nodes.

[0188] (v) Project deployment and operation and maintenance process

[0189] The technical solution of the present invention has a complete engineering deployment and daily operation and maintenance process. All steps are based on the actual operation specifications and hardware operation standards in coal mines and can be directly implemented. The specific implementation steps are steps S81 to S85.

[0190] Step S81: Hardware Deployment

[0191] Install intrinsically safe intelligent sensor nodes for mining according to the layout of the mine face, with a node spacing of 50-100m to adapt to the spatial scale of underground disaster propagation; complete the intrinsically safe certification, communication pairing and hardware debugging of all nodes to ensure that the data acquisition, calculation, communication and alarm functions of each module are working properly.

[0192] Step S82: Data Initialization

[0193] Import historical monitoring data from the target mine, process it according to the dataset preprocessing standards in step S21, and then calculate the initial smoothed mean of the model offline. Standard deviation Optimal threshold coefficient With multi-parameter initial weights The parameters are then burned into the flash memory of the edge computing module.

[0194] Step S83: Model Deployment and Integration

[0195] The lightweight long-short-term relationship-improved gated cyclic unit (LSR-IGRU) model, quantized with INT8, was burned into the edge computing module. Typical scenarios such as blasting, gas exceeding limits, and equipment failure were simulated underground to verify the effectiveness of the model's inference, collaborative communication, anomaly identification, and early warning output. Engineering configuration parameters (such as communication radius) were then adjusted and optimized based on actual mine conditions. Sliding window length ).

[0196] Step S84: Startup

[0197] After completing the joint debugging and parameter optimization, the sensor network enters the formal operation state. All algorithms are executed on the local edge side of the smart sensor nodes without relying on cloud computing power. Only medium and high-level early warning information is uploaded to the ground safety monitoring center.

[0198] Step S85: Daily Operations and Model Iteration

[0199] (1) The acquisition modules of each smart sensor node are calibrated monthly to ensure the accuracy of acquisition of parameters such as gas, micro-vibration, and temperature;

[0200] (2) Collect false alarm / missed alarm data from the mine every quarter and iteratively update the optimal threshold coefficient of the model. With multi-parameter adaptive weights ;

[0201] (3) Based on the newly collected monitoring data from the mine each year, the model is retrained and quantized according to the training process in step S22 to improve the generalization ability of the model. The redeployed model directly covers the original model parameters in the edge computing module.

[0202] It should also be noted that, in the embodiments of this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0203] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. An edge computing driven multi-parameter adaptive sensing and abnormality identification sensor network for underground coal mine, characterized in that, include: Multiple intrinsically safe smart sensor nodes for mining are connected to form a distributed monitoring network through a self-organizing network of a mining low-power wide area network. Each of the intrinsically safe smart sensor nodes for mining applications includes: Multi-parameter acquisition module, used to acquire various monitoring parameters in coal mines in real time; The edge computing module is equipped with a lightweight long-short-term relationship-improved gated cyclic unit time-series prediction model based on the reconstruction of the physical mechanism of coal mine disasters. This model is used to process the monitoring parameters locally, realize adaptive drift correction of the background values ​​of the monitoring parameters to dynamically generate threshold ranges, and perform preliminary anomaly judgment on the monitoring parameters based on the threshold ranges to generate preliminary anomaly judgment results. The lightweight long-short-term relationship-improved gated cyclic unit time-series prediction model is used to integrate the long-term stable correlation between monitoring parameters determined by physical laws and disaster mechanisms, the short-term dynamic linkage caused by underground working condition disturbances, and the original time-series characteristics of the various monitoring parameters to predict the background value prediction vector of each monitoring parameter at the next moment, and realize adaptive drift correction of the background value based on the background value prediction vector. The lightweight long-short-term relationship-improved gated recurrent unit time-series prediction model adopts a structure combining long-short-term feature physical fusion with a lightweight recurrent neural network. Specifically, it includes: a long-term relationship feature extraction unit, used to fuse the multi-parameter original feature vectors at time t based on a pre-constructed, physically-based long-term correlation weight matrix, generating a long-term relationship fused feature vector at time t to extract the long-term stable correlation; a short-term relationship feature extraction unit, used to fuse the multi-parameter original feature vectors at time t based on a short-term correlation degree diagonal matrix constructed from parameters calculated within a sliding window and the Pearson correlation coefficient, generating a short-term relationship fused feature vector at time t to extract the short-term dynamic linkage; a lightweight gated recurrent unit, used to concatenate the multi-parameter original feature vectors, long-term relationship fused feature vectors, and short-term relationship fused feature vectors, and output the time-series fused hidden state at time t through update gates, reset gates, and candidate hidden state calculations; and a single-layer multilayer perceptron, used to perform a linear transformation on the time-series fused hidden state, outputting the background value prediction vectors of each monitoring parameter at time t+1. The collaborative communication module is used to interact with neighboring nodes after the edge computing module generates an initial anomaly judgment result, and to perform spatiotemporal collaborative verification based on the interaction data to generate a spatiotemporal collaborative verification result. The anomaly decision module is used to perform multi-parameter fusion identification based on the initial anomaly judgment result and the spatiotemporal collaborative verification result, generate multi-parameter fusion identification result, and output hierarchical warning based on the multi-parameter fusion identification result.

2. The edge computing driven multi-parameter adaptive sensing and abnormality identification sensor network for underground coal mine according to claim 1, characterized in that, The adaptive drift correction of the background value includes the generation of a dynamic baseline and an adaptive threshold: The edge computing module calculates the smooth mean μ of each monitoring parameter and the standard deviation σ of the measured data within the sliding window based on the background value prediction vector output by the lightweight long-short-term relationship-improved gated loop unit time-series prediction model, combined with the historical measured data within the sliding window. The edge computing module is further configured to determine an optimal threshold coefficient based on an offline optimization of a receiver operating characteristic curve of historical abnormal data of the mine And dynamically generate an upper threshold value for abnormality discrimination according to the smoothed mean μ and the standard deviation σ And a lower threshold value To obtain a threshold range.

3. The edge computing-driven multi-parameter adaptive sensing and anomaly recognition sensor network for underground coal mines according to claim 2, characterized in that, The adaptive drift correction of the background value also includes baseline adaptive update and anomaly locking mechanisms: The edge computing module uses the measured values ​​of monitoring parameters and the upper limit threshold. and lower threshold To determine the relative relationships, perform the following operations: When the measured value of the monitoring parameter remains within the threshold range, it is determined to be a gradual environmental change scenario. The smooth mean μ and standard deviation σ are updated according to a preset cycle to achieve adaptive drift of the alarm baseline. When the measured value of a monitoring parameter exceeds the threshold range in a single instance or the measured values ​​of three consecutive sampling steps are close to the threshold, it is determined to be an abnormal change scenario. The current baseline state is immediately locked, the update of the smoothed mean μ and standard deviation σ is stopped, and the abnormal initial judgment mode is triggered.

4. The edge computing-driven multi-parameter adaptive sensing and anomaly recognition sensor network for underground coal mines according to claim 1, characterized in that, The spatiotemporal collaborative verification includes time dimension verification and spatial dimension verification; The time dimension verification includes: downsampling and feature extraction of the microseismic signal, and determining whether it is transient interference and filtering it, or continuous abnormality and triggering the spatial dimension verification, based on the duration of the abnormal signal, energy change rate, signal fallback speed, or matching degree with a preset interference feature library. The spatial dimension verification includes: the node that triggers the initial anomaly judgment broadcasts a query request to neighboring nodes through the collaborative communication module, and based on the received feedback data from neighboring nodes, combined with the spatial evolution physical laws of coal mine disasters, collaboratively determines whether the current anomaly is a false alarm or a real disaster precursor, and generates the spatiotemporal collaborative verification result.

5. The edge computing-driven multi-parameter adaptive sensing and anomaly recognition sensor network for underground coal mines according to claim 4, characterized in that, The spatial dimension verification also includes communication latency tolerance and degradation processing mechanisms: Set a timeout period for neighbor node feedback. Nodes that fail to respond within the timeout period are marked as communication abnormal nodes and will not participate in collaborative identification. If the number of valid neighboring nodes participating in the identification is lower than a preset threshold, it will be downgraded to local multi-parameter fusion identification. In this case, the spatiotemporal collaborative verification result will directly indicate downgrade processing.

6. The edge computing-driven multi-parameter adaptive sensing and anomaly recognition sensor network for underground coal mines according to claim 1, characterized in that, The multi-parameter fusion identification includes: A multi-parameter association rule base is established based on the physical evolution mechanism of different dynamic disaster types in coal mines. The rule base includes core association parameters and their initial weights. Based on the degree of deviation between the measured values ​​of the monitoring parameters and the dynamic thresholds, the single-parameter anomaly probability of each core associated parameter is calculated. The weights of the core correlation parameters are periodically updated based on the historical number of true positive samples and false positive samples. The comprehensive anomaly probability is calculated based on the single-parameter anomaly probability and its adaptive weight. Based on the magnitude of the comprehensive anomaly probability, a four-level graded early warning output is executed.

7. The edge computing-driven multi-parameter adaptive sensing and anomaly recognition sensor network for underground coal mines according to claim 1, characterized in that, The multi-parameter acquisition module includes a gas sensor, a micro-vibration probe, a temperature sensor, and a pressure sensor, and all sensors meet the intrinsically safe certification requirements for underground coal mines.

8. The edge computing-driven multi-parameter adaptive sensing and anomaly recognition sensor network for underground coal mines according to claim 1, characterized in that, The edge computing module uses an STM32H7 series chip as its microcontroller unit. The lightweight long-short-term relationship-improved gated loop unit model is deployed in this microcontroller unit after being quantized by INT8, and its single-sample inference time is less than 5ms. The collaborative communication module adopts a mining low-power wide area network LoRa-PRO module, which operates in the 433MHz unlicensed frequency band and has an end-to-end communication latency of less than 80ms.

9. The edge computing-driven multi-parameter adaptive sensing and anomaly recognition sensor network for underground coal mines according to claim 1, characterized in that, The anomaly decision-making module includes an intrinsically safe audible and visual alarm and a mining fourth-generation mobile communication technology module, which are used to trigger local alarms and upload early warning information to the ground safety monitoring center, respectively.

Images