Mine emergency communication heterogeneous mobile edge computing self-organizing collaborative task offloading method based on sgc-fedmappo

Abstract

The application discloses a mine emergency communication heterogeneous mobile edge computing self-organizing cooperative task offloading method based on SGC-FedMAPPO, belongs to the cross technical field of mine emergency communication and heterogeneous mobile edge computing, and aims to solve the core problems that the existing underground edge computing task offloading scheme architecture is prone to paralysis, the security constraint is not embedded in the algorithm bottom layer, and safe and reliable cooperative offloading cannot be realized based on underground original physical parameters. The method directly takes underground gas, temperature and vibration original physical parameters as inputs, does not need to calculate preposition normalized safety risk values, and solves the optimal cooperative offloading strategy in a centerless scene through four core links of lightweight SGC topology reconstruction, frequency domain difference MoE link prediction, logarithmic barrier double security hard constraint and decentralized Gossip aggregation, and constructs a federated multi-agent reinforcement learning model with security hard constraint. The application can realize self-organizing safe cooperative offloading of heterogeneous nodes after a catastrophe.

Description

Technical Field

[0001] This invention belongs to the cross-technical field of mine emergency communication and heterogeneous mobile edge computing (MEC). Specifically, it relates to a method for mine-use intrinsically safe heterogeneous edge nodes to form a decentralized self-organizing network and distribute collaborative task offloading after the macro base station and fiber optic backbone network are damaged in disaster scenarios such as mine gas explosion and roof collapse. It can be applied to ensure computing power for low-latency and high-reliability critical services during the emergency rescue phase of mine disasters. Background Technology

[0002] With the deepening of smart mine construction and the popularization of mobile edge computing technology, computationally intensive tasks in emergency rescue scenarios after mine disasters need to be safely and efficiently offloaded locally to ensure the real-time performance and reliability of emergency responses. However, how to achieve reliable collaborative task offloading in the extreme environment of mine disasters—decentralized, highly non-stationary, and with stringent security constraints—has become a core challenge facing the field of mine emergency communications.

[0003] Traditional edge computing task offloading methods in mining are primarily designed for normal underground production scenarios. They mostly employ centralized architectures relying on macro base stations, edge aggregation nodes, or cloud platforms, failing to adapt to extreme post-disaster environments. Damage to core communication infrastructure after a disaster can directly lead to a complete system paralysis. Furthermore, existing decentralized solutions have not achieved effective self-organization and adaptation to address the dynamic topology changes, random node failures, and network fragmentation issues encountered in disaster scenarios, and are unable to achieve distributed collaborative scheduling of heterogeneous edge nodes. In terms of safety and compliance, existing solutions do not embed the hard safety boundaries of underground operations stipulated in the "Coal Mine Safety Regulations" into the algorithm's underlying layer. They only use post-processing rules to screen and avoid risks, and rely heavily on pre-processing normalized risk value mapping for safety assessment. This makes them susceptible to interference from complex post-disaster scenarios and lacks robustness. In terms of link prediction and computing power adaptation, existing solutions mostly use time-domain sequence models for link state prediction, which cannot effectively capture the early warning signs of channel frequency domain energy collapse. This can easily lead to incorrect offloading decisions. Furthermore, existing complex intelligent optimization solutions have excessive computing power overhead and cannot adapt to the real-time operational needs of mining equipment with low computing power remaining after a disaster. These deficiencies directly lead to core problems such as system architecture paralysis, emergency rescue service interruption, high safety risks in offloading decisions, and difficulty in implementing solutions in underground environments during disaster scenarios. They fail to meet the core requirement of "the safety red line cannot be crossed" in mine emergency rescue.

[0004] To address these shortcomings, this invention proposes a self-organizing and collaborative task offloading method for heterogeneous mobile edge computing in mine emergency communication based on SGC-FedMAPPO (Simplified Graph Convolution-Federated Multi-Agent Proximal Policy Optimization) (graph convolution + federated multi-agent reinforcement learning), in order to overcome the aforementioned deficiencies in existing technologies. Summary of the Invention

[0005] The purpose of this invention is to address the core pain points of macro base station failure, dynamic topology changes, strong channel non-stationarity, and stringent security constraints after a mine disaster, as well as the insufficient adaptability of existing technologies to extreme underground disaster scenarios. This invention provides a self-organizing collaborative task offloading method for heterogeneous mobile edge computing based on SGC-FedMAPPO for mine emergency communication. This method overcomes the limitations of existing solutions that rely on pre-normalized risk value mapping, directly using the original underground physical safety parameters as system input. It integrates four key technologies: lightweight SGC topology reconstruction, frequency-domain differential-driven MoE link survival probability prediction, logarithmic barrier function dual security hard constraints, and decentralized weighted Gossip aggregation. This enables secure and efficient heterogeneous node collaborative task offloading in the decentralized extreme scenario of macro base station and fiber optic backbone network damage after a mine disaster, providing stable and reliable distributed computing power for critical post-disaster emergency rescue services in mines.

[0006] The applicable scenarios for this invention are subject to the following preconditions: After a mine disaster, the macro base station and fiber optic backbone network are damaged and ineffective, there is no central control node, no global status awareness, and the heterogeneous edge nodes can only self-organize through direct D2D communication in the mine; all nodes are intrinsically safe mining equipment with gas, temperature, and vibration sensing capabilities and edge computing capabilities; and the invention strictly adheres to the hard safety boundaries of the "Coal Mine Safety Regulations": the lower explosive limit concentration of gas. Mining equipment's maximum withstand temperature Destructive vibration threshold .

[0007] To achieve the above-mentioned technical objectives, the specific technical solution of the present invention includes the following five steps connected in sequence:

[0008] Step 1: Multi-source physical parameter acquisition and node safety validity pre-verification: Collect the original physical parameters of the mine, perform binary judgment on the validity of intrinsically safe heterogeneous edge nodes in the mine based on safety hard constraints, isolate disaster-failed nodes, generate an initial set of valid nodes and complete neighbor discovery and initial communication adjacency matrix construction.

[0009] Step 2, Attenuation-aware topology reconstruction: Based on the initial communication adjacency matrix constructed in Step 1, an attenuation-aware channel matrix is ​​constructed to address the attenuation characteristics of electromagnetic wave propagation in underground roadways. A lightweight simplified graph convolution without nonlinear activation is used to perform multi-hop smoothing on the node communication features. Combined with the node validity determination results from Step 1, a binary federated aggregation mask is generated to delineate the logically trusted neighbor domain of each node.

[0010] Step 3: Hybrid expert model link confidence prediction: Based on the logically trusted neighborhood domain defined in Step 2, perform fast Fourier transform on the link signal-to-noise ratio time series, extract time-domain statistical features and frequency-domain energy difference features, use a hybrid expert model with Top-1 sparse activation to output the link survival probability in the next scheduling cycle, and combine the trusted neighborhood domain to generate row-normalized federated aggregation confidence weights.

[0011] Step 4, Construction of Markov Decision Process under Security Constraints: Based on the original physical parameters in Step 1 and the link survival probability in Step 3, define an observation state space that includes the physical security state of nodes, the task queue state, the security state of neighbors, and the link survival probability, as well as an action space that includes local computation and neighbor node offloading. Construct a latency and energy consumption calculation model for heterogeneous node task offloading, introduce a logarithmic barrier function to transform the security hard constraints of the original physical parameters into continuous penalty terms, and construct an instantaneous reward function with security constraints.

[0012] Step 5, Model Solving under Topological Consistency Regularization Constraints: Based on the Markov decision process constructed in Step 4, a near-end policy optimization pruning loss function with topological Laplace regularization term is constructed. Each node updates the policy network based on local experience data and exchanges model parameters with trusted neighbors only through one-hop mining D2D direct communication. Decentralized weighted Gossip aggregation is performed based on the confidence weight to output the optimal collaborative task offloading strategy.

[0013] Furthermore, the intrinsically safe heterogeneous edge node for mining includes a fixed emergency communication gateway for mining, an underground inspection robot, a portable terminal for rescue personnel, and intrinsically safe sensing nodes for mining. All types of nodes meet the explosion-proof standards for intrinsically safe equipment for mining.

[0014] Furthermore, step 1 specifically includes:

[0015] Step 1.1 Acquisition of raw physical parameters: Each edge node acquires the gas volume concentration, ambient temperature, and vibration amplitude of its current environment in real time through intrinsically safe sensors for mining, forming a physical safety state vector, which is directly used as the raw state input for multi-agent reinforcement learning;

[0016] Step 1.2 Hard determination of node validity: Based on the safety hard constraints stipulated in the "Coal Mine Safety Regulations" and supporting national standards, the validity of nodes is determined by binary method. A node is determined to be a valid node only when it simultaneously meets the following conditions: vibration amplitude is lower than the destructive vibration threshold stipulated in the national standard, gas concentration is lower than the lower explosive limit stipulated in the national standard, and ambient temperature is lower than the equipment limit tolerance temperature stipulated in the national standard. Otherwise, it is determined to be a failed node and is silently isolated.

[0017] Furthermore, step 2 specifically includes:

[0018] Step 2.1 Construction of Attenuation Sensing Channel Matrix in Underground Roadway: Combining the initial communication adjacency matrix and physical attenuation matrix of the underground roadway, an attenuation sensing channel matrix that conforms to the spatial propagation characteristics of the mine is constructed. The elements of the physical attenuation matrix are calculated from the spatial distance between nodes and the electromagnetic wave attenuation coefficient of the underground roadway.

[0019] Step 2.2 SGC Multi-hop Feature Smoothing: A lightweight SGC without nonlinear activation is used to smooth the node communication features in multiple hops, reducing the computational overhead of the algorithm and adapting to the computational power limitations of mining equipment;

[0020] Step 2.3 Binarized Federation Aggregation Mask Generation: Based on the smoothed features, calculate the cosine similarity between nodes, and combine it with the node validity determination results from Step 1 to generate a binarized federated aggregation mask, delineating the logically trustworthy neighborhood of each node.

[0021] Furthermore, step 3 specifically includes:

[0022] Step 3.1 Frequency domain energy difference feature extraction: Perform fast Fourier transform on the link signal-to-noise ratio time series to extract the spectral energy difference features between adjacent time points, capture the precursor features of fast channel fading after a disaster, and identify the risk of link interruption in advance compared with time domain statistical features;

[0023] Step 3.2 Construction of Sparse Gated Hybrid Expert Model: The MoE model is constructed using the Top-1 sparse activation mechanism to reduce computing power overhead, adapt to the computing power limitations of mining equipment, and output the survival probability of the link in the next scheduling cycle.

[0024] Step 3.3 Federated Aggregate Confidence Weight Generation: Combine the trusted neighbor domain mask generated in Step 2 to generate row-normalized federated aggregation confidence weights that satisfy the convergence conditions of decentralized Gossip aggregation;

[0025] Step 3.4 Training and Deployment of the MoE Model: The MoE model is deployed using an offline pre-training + online fine-tuning approach. During the pre-training phase, measured underground channel data from historical disaster scenarios are used as samples, and the link interruption within a future scheduling cycle is used as a binary classification label. A binary cross-entropy loss function is employed to jointly train the gated network and the expert network. During the online operation phase, each node performs incremental fine-tuning of the model based on locally collected real-time channel data. The fine-tuning process is executed only locally on the node and does not require global data interaction.

[0026] Furthermore, step 4 specifically includes:

[0027] Step 4.1 Basic definition of Markov decision process: Define the observation state space and unloading action space for matching the centerless scenario, delineate the feasible domain of safe actions based on the safety threshold of the original physical parameters, and prohibit unloading tasks to high-risk nodes from the action level;

[0028] Step 4.2 Calculation of Task Offloading Latency and Energy Consumption for Heterogeneous Nodes: To address the differences in computing power among heterogeneous nodes in the mine, a calculation model for task offloading latency and energy consumption that fits the edge computing scenario is constructed.

[0029] Step 4.3 Construction of instantaneous reward function with logarithmic barrier function: Introduce logarithmic barrier function to transform the security hard constraint into a continuous penalty term at the bottom layer of the algorithm, realizing dual hard constraints of physical security and network security.

[0030] Furthermore, in the feasible domain of safe actions in step 4.1, the safety interval determination threshold of the target unloading node is the safety warning threshold specified in the "Coal Mine Safety Regulations"; in the instantaneous reward function in step 4.3, the input of the logarithmic barrier function includes the safety margin of the target node and the probability of link interruption. When the target node approaches the safety threshold or the probability of link survival approaches 0, the logarithmic barrier function outputs a negative penalty value.

[0031] Furthermore, step 5 specifically includes:

[0032] Step 5.1 Local policy update with topological Laplace regularization: Each node updates the policy network using only local empirical data. A topological Laplace regularization term is added to the PPO pruning loss function to force the model parameters of trusted neighbors to converge, reduce the variance of decentralized aggregation, and accelerate policy convergence.

[0033] Step 5.2 Decentralized Weighted Gossip Aggregation: Each node exchanges model parameters with trusted neighbors through only one-hop mining D2D direct communication, performs weighted Gossip aggregation, and completes distributed policy collaborative update.

[0034] The beneficial effects of this invention are:

[0035] Addressing the unique communication environment and emergency security requirements of IoT in mining, this invention integrates lightweight graph convolution and decentralized federated reinforcement learning to achieve collaborative task offloading and optimal scheduling efficiency for heterogeneous edge nodes in disaster scenarios, possessing the following core advantages:

[0036] This invention directly embeds the hard thresholds of the "Coal Mine Safety Regulations" into the algorithm's underlying layer, eliminating the need for pre-normalized risk value calculations and avoiding the robustness issues of the mapping model being affected by scene interference. Through hard isolation of the feasible domain of safe actions and soft penalty of the logarithmic barrier function, it achieves dual hard constraints on physical safety and network security, completely intercepting task offloading behavior towards high-risk areas, thoroughly avoiding the risk of transmitting data to high-risk disaster areas, and fully meeting the core requirement of "safety red lines cannot be crossed" in emergency scenarios.

[0037] This invention employs a frequency-domain differential-driven sparse-gated hybrid expert model. Compared with traditional time-domain LSTM, GRU and other sequence prediction schemes, it can identify early characteristics of channel spectrum collapse in advance, significantly extend the link interruption early warning window, and significantly improve the prediction accuracy of link survival probability. It effectively solves the problems of offloading decision errors and emergency service transmission interruptions caused by the strong non-stationary characteristics of the channel after a disaster, and improves the reliability of collaborative offloading decision.

[0038] This invention replaces the traditional GCN with SGC that has no nonlinear activation and adopts the MoE model with Top-1 sparse activation. The overall computing power cost of the algorithm is significantly reduced compared with the existing complex schemes. It can run stably and in real time on low-computing-power embedded nodes in mining, and has excellent adaptability to underground field deployment. It can provide reliable distributed computing power support for emergency rescue after mine disasters. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the system architecture and method of the present invention;

[0040] Figure 2 This is a schematic diagram of the deployment and collaborative unloading of heterogeneous edge nodes in a mine disaster scenario according to the present invention;

[0041] Figure 3 This is a diagram illustrating the macroscopic spatiotemporal evolution of the method of the present invention before and after a disaster; Figure 4 This is a flowchart of the present invention. Detailed Implementation

[0042] The embodiments of the present invention will be further described below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the contents described.

[0043] Example 1, Figures 1-4A self-organizing collaborative task offloading method for heterogeneous mobile edge computing in mine emergency communication, based on SGC-FedMAPPO, is applied to a decentralized self-organizing scenario after a mine disaster where macro base stations and fiber optic backbone networks are damaged. All nodes in the scenario are intrinsically safe heterogeneous edge nodes for mining, equipped with gas, temperature, and vibration sensing capabilities as well as edge computing capabilities. Nodes communicate with each other via direct device-to-device (D2D) connections. The method includes the following steps:

[0044] Step 1: Multi-source physical parameter acquisition and node safety validity pre-verification: The original physical parameters of gas concentration, ambient temperature and vibration amplitude collected by intrinsically safe sensors in mines are directly used as system inputs. There is no need for pre-normalized safety risk value calculation. Based on the safety hard constraints stipulated in the "Coal Mine Safety Regulations", the validity of intrinsically safe heterogeneous edge nodes in mines is binarized and determined, catastrophic failure nodes are isolated, a safety boundary is provided for the subsequent whole process, an initial set of valid nodes is generated and neighbor discovery and initial communication adjacency matrix construction are completed.

[0045] Furthermore, the intrinsically safe heterogeneous edge node for mining includes a fixed emergency communication gateway for mining, an underground inspection robot, a portable terminal for rescue personnel, and intrinsically safe sensing nodes for mining. All types of nodes meet the explosion-proof standards for intrinsically safe equipment for mining.

[0046] Furthermore, step 1 specifically includes:

[0047] Step 1.1 Acquisition of raw physical parameters: Each edge node acquires data in real time using intrinsically safe sensors for mining applications. The gas volume concentration, ambient temperature, and vibration amplitude of the environment at any given time form a physical safety state vector, which is directly used as the raw state input for multi-agent reinforcement learning.

[0048]

[0049] In the formula: For nodes exist The volumetric concentration of gas collected at all times, in units of ; For nodes exist The ambient temperature is collected at all times, in units of ; For nodes exist Vibration amplitude values ​​collected at specific times, in units of ; , This represents the total number of heterogeneous edge nodes in the wellbore.

[0050] Step 1.2 Hard determination of node validity: Based on the safety hard constraints stipulated in the national standard for coal mine safety, "Coal Mine Safety Regulations" and supporting national standards, the validity of nodes is determined by binarization. A node is determined to be a valid node only if it simultaneously meets the following conditions: vibration amplitude is lower than the destructive vibration threshold stipulated in the national standard, gas concentration is lower than the lower explosive limit stipulated in the national standard, and ambient temperature is lower than the equipment limit tolerance temperature stipulated in the national standard. Otherwise, it is determined to be a failed node and is silently isolated.

[0051] Specifically, node validity is determined using binary methods; failed nodes are directly isolated and do not participate in any subsequent network formation or offloading processes; this occurs in the initial moments after a disaster. All nodes complete initial self-checks and validity determinations, generate an initial set of valid nodes, and complete neighbor discovery and initial communication adjacency matrix construction.

[0052]

[0053] In the formula: The representative node is valid and can participate in collaborative unloading and federated aggregation; If a node fails, it will be silently isolated.

[0054] Step 2, Attenuation-Aware Topology Reconstruction: Based on the initial communication adjacency matrix constructed in Step 1, an attenuation-aware channel matrix is ​​constructed to address the attenuation characteristics of electromagnetic wave propagation in underground roadways. To address the dynamic changes in topology and the limited computing power of mining equipment in a centerless scenario, a lightweight simplified graph convolution (SGC) without nonlinear activation is used to perform multi-hop smoothing on node communication features. Combined with the node validity determination results from Step 1, a binary federated aggregation mask is generated to delineate the logically trusted neighbor domain of each node, providing a topological foundation for subsequent federated aggregation and offloading decisions.

[0055] Furthermore, step 2 specifically includes:

[0056] Step 2.1 Construction of Attenuation Sensing Channel Matrix in Underground Roadway: Combining the initial communication adjacency matrix and physical attenuation matrix of the underground roadway, an attenuation sensing channel matrix that conforms to the spatial propagation characteristics of the mine is constructed. The elements of the physical attenuation matrix are calculated from the spatial distance between nodes and the electromagnetic wave attenuation coefficient of the underground roadway.

[0057]

[0058] In the formula: for 3D original communication adjacency matrix, If and only if node and If within the range of one-hop D2D communication, otherwise 0; for The downhole physical decay matrix, elements , For nodes and The Euclidean distance of the alleyway space. The electromagnetic wave attenuation coefficient for underground roadways, with a value range of... ; This is the Hadamard product, which is the element-wise multiplication of matrices.

[0059] Step 2.2 SGC Multi-hop Feature Smoothing: A lightweight SGC without nonlinear activation is used to smooth the node communication features in multiple hops, reducing the computational overhead of the algorithm and adapting to the computational limitations of mining equipment.

[0060]

[0061] In the formula: for A 3D identity matrix is ​​used to introduce node self-loops to ensure that feature aggregation includes its own information; for The degree matrix, where the diagonal elements are the sum of the elements in the corresponding rows; For the renormalized smooth adjacency operator, This is a multi-hop aggregation order, adapted to underground roadway topology settings, with a value range of... Rank; for Original communication feature matrix For single-node communication feature dimensions; for The linear feature projection matrix has no nonlinear activation, ensuring the algorithm is lightweight.

[0062] Step 2.3 Binarized Federated Aggregation Mask Generation: Based on the smoothed features, calculate the cosine similarity between nodes. Combined with the node validity determination results from Step 1, generate a binary federated aggregation mask to delineate the logically trusted neighbor domain of each node.

[0063]

[0064]

[0065] In the formula The topological similarity threshold is set to a threshold that ensures the average node degree of the trusted neighbor domain meets the preset connectivity requirements. Representative node For nodes Trusted neighbors, defining trusted neighbor domains .

[0066] Step 3, Hybrid Expert Model Link Confidence Prediction: Based on the logically trusted neighborhood domain defined in Step 2, perform a Fast Fourier Transform on the link signal-to-noise ratio time series to extract time-domain statistical features and frequency-domain energy difference features. Use a Top-1 sparse activation Hybrid Expert Model (MoE) to output the link's survival probability in the next scheduling cycle. Combine the trusted neighborhood domain to generate row-normalized federated aggregation confidence weights.

[0067] Furthermore, addressing the issues of strong non-stationarity and high link interruption probability in underground channels after a disaster, a hybrid expert model (MoE) driven by both time and frequency domain features is adopted to accurately capture link interruption precursors and output the link survival probability, providing a reliable basis for subsequent federated aggregation and offloading decisions; step 3 specifically includes:

[0068] Step 3.1 Frequency domain energy difference feature extraction: Perform Fast Fourier Transform (FFT) on the link signal-to-noise ratio time series to extract the spectral energy difference features between adjacent time points, capture the precursor features of fast channel fading after a disaster, and identify the risk of link interruption in advance compared with time domain statistical features;

[0069]

[0070] In the formula: for Time Node arrive The link signal-to-noise ratio time series was processed with a Hanning window and then subjected to FFT; This is a Fast Fourier Transform (FFT) operator. In this step, the signal-to-noise ratio (SNR) time series is taken from the most recent 64 SNR sampling points, corresponding to a 320ms historical window; the Fast Fourier Transform uses a 64-point FFT, outputting a single-sided spectrum; the spectral energy is defined as the sum of the squares of the amplitudes at all frequency points, i.e., the total signal energy.

[0071] Step 3.2 Construction of Sparse Gated Hybrid Expert Model: A Top-1 sparse activation mechanism is used to construct the MoE model, reducing computational overhead, adapting to the computational limitations of mining equipment, and outputting the survival probability of the link in a future scheduling cycle.

[0072]

[0073]

[0074]

[0075] In the formula: This is the time-domain statistical feature vector of the signal-to-noise ratio sequence, containing the mean, variance, peak value, and kurtosis. This is a vector concatenation operation; , Here are the weight matrix and bias terms of the gated network; The noise is Gaussian noise with a mean of 0 and a variance of 0.01 to ensure sparse activation of experts; The total number of expert networks, adapted to the computing power limit settings of mining equipment, and the range of values. indivual; For the k-th expert network, a 2-layer fully connected network is adopted; The Sigmoid activation function guarantees The physical meaning is the probability that the link will not be interrupted within a future scheduling cycle.

[0076] Step 3.3 Federated Aggregate Confidence Weight Generation: Combining the trusted neighbor domain mask generated in Step 2, row-normalized federated aggregation confidence weights are generated to satisfy the convergence condition of decentralized Gossip aggregation.

[0077]

[0078] In the formula: satisfy , for nodes To the neighbors Aggregate weights.

[0079] Step 3.4 Training and Deployment of the MoE Model: The MoE model is deployed using an offline pre-training + online fine-tuning approach. During the pre-training phase, measured underground channel data from historical disaster scenarios are used as samples, and the link interruption within a future scheduling cycle is used as a binary classification label. A binary cross-entropy loss function is employed to jointly train the gated network and the expert network. During the online operation phase, each node incrementally fine-tunes the model based on locally collected real-time channel data to adapt to the real-time changing channel environment after the disaster. The fine-tuning process is performed only locally on the node, without the need for global data interaction, thus adapting to decentralized scenarios.

[0080] Step 4, Construction of Markov Decision Process under Security Constraints: Based on the original physical parameters in Step 1 and the link survival probability in Step 3, define an observation state space that includes the physical security state of nodes, the task queue state, the security state of neighbors, and the link survival probability, as well as an action space that includes local computation and neighbor node offloading. Construct a latency and energy consumption calculation model for heterogeneous node task offloading, introduce a logarithmic barrier function to transform the security hard constraints of the original physical parameters into continuous penalty terms, and construct an instantaneous reward function with security constraints.

[0081] Furthermore, to address the absolute safety requirements of emergency scenarios, the original physical parameters from step 1 are directly embedded into security constraints to construct a Markov decision process with a logarithmic barrier function. This achieves dual hard constraints on physical and network security, absolutely avoiding high-risk areas and vulnerable links. Step 4 specifically includes:

[0082] Step 4.1 Basic Definition of Markov Decision Process: Define the observation state space and unloading action space for matching a centerless scenario, delineate the feasible domain of safe actions based on the safety threshold of the original physical parameters, and prohibit unloading tasks to high-risk nodes at the action level; specifically including:

[0083] State space: nodes The observation status includes its own physical security status. The local task queue length, the physical security status of one-hop neighbors, and the link survival probability are all considered, but there is no global state awareness, which is in line with the constraints of a decentralized scenario.

[0084] Action Space: Nodes Uninstallation action , This represents local computation of the task. The task is unloaded to the node. ;

[0085] Feasible domain for safe actions: a safety threshold directly bound to the original physical parameters, prohibiting unloading to high-risk nodes at the action level.

[0086]

[0087] In the formula: For nodes A secure and reliable set of target nodes for unloading; nodes For a system to be considered safe and effective, it must simultaneously meet the following conditions: , , ; The safety warning threshold is set based on the lower limit of gas explosion. , The safety warning threshold is set based on the equipment's extreme tolerance temperature. Unloading is only allowed when both the node itself and the target node are in a safe zone; otherwise, it is forced to degenerate into local computation.

[0088] Step 4.2 Calculation of Task Offloading Latency and Energy Consumption for Heterogeneous Nodes: Considering the computing power differences among heterogeneous nodes in the mine, a calculation model for task offloading latency and energy consumption tailored to edge computing scenarios is constructed.

[0089]

[0090]

[0091] In the formula: The amount of data for the task to be unloaded, in bits; The computational load of the task to be unloaded, in CPU cycles; For nodes CPU calculation frequency, in Hz; The minimum CPU frequency required for the task to execute; For nodes and The communication bandwidth, measured in bits per second; For nodes The CPU energy efficiency coefficient is related to the chip architecture; For nodes The radio frequency transmission power, measured in W.

[0092] Step 4.3 Construction of the instantaneous reward function with a logarithmic barrier function: By introducing a logarithmic barrier function, the hard security constraint is transformed into a continuous penalty term at the bottom layer of the algorithm, achieving dual hard constraints of physical security and network security.

[0093]

[0094]

[0095] In the formula: The comprehensive safety margin for a single scheduling operation, with a value range of... The closer it is to 1, the higher the risk. As a priority weight, emergency rescue tasks are assigned higher values, while environmental monitoring tasks are assigned lower values; , For normalized weight coefficients, satisfying This is used to balance the optimization priorities of latency and energy consumption; This is a safety penalty coefficient; To ensure numerical stability, a small amount of logarithmic function input should be avoided; when the target node's gas / temperature approaches the limiting threshold, or the link is about to be interrupted, When the number of terms approaches negative infinity, a severe penalty is incurred, forcing the agent to absolutely avoid high-risk areas and vulnerable links.

[0096] Furthermore, in the feasible domain of safe actions in step 4.1, the safety interval determination threshold of the target unloading node is the safety warning threshold specified in the "Coal Mine Safety Regulations"; in the instantaneous reward function in step 4.3, the input of the logarithmic barrier function includes the safety margin of the target node and the probability of link interruption. When the target node approaches the safety threshold or the probability of link survival approaches 0, the logarithmic barrier function outputs a negative penalty value.

[0097] Step 5: Solving the model under topological consistency regularization constraints: Based on the Markov decision process constructed in Step 4, a proximal policy optimization (PPO) pruning loss function with topological Laplace regularization term is constructed. Each node updates the policy network based on local experience data and exchanges model parameters with trusted neighbors only through one-hop mining D2D direct communication. Based on the confidence weight, decentralized weighted Gossip aggregation is performed to output the optimal collaborative task offloading strategy.

[0098] Furthermore, for decentralized scenarios, a decentralized federated multi-agent proximal policy optimization framework is constructed. Convergence is accelerated through topological Laplacian regularization, and fully distributed policy co-evolution is achieved through weighted Gossip aggregation, eliminating the dependency on a central server. Step 5 specifically includes:

[0099] Step 5.1 Local Policy Update with Topological Laplace Regularization: Each node updates the policy network using only local empirical data. A topological Laplace regularization term is added to the PPO pruning loss function to force the model parameters of trusted neighbors to converge, reducing the variance of decentralized aggregation and accelerating policy convergence.

[0100]

[0101]

[0102] In the formula: For nodes Local policy network parameters; This represents the probability ratio between the old and new strategies. The advantage function estimate is calculated based on the generalized advantage estimation (GAE) and is used to evaluate the quality of the current action relative to the average level. Set a PPO pruning threshold to limit the policy update magnitude and ensure training stability; represents the topological regularization coefficients, and the regularization term is a graph Laplace quadratic form, which implements topological consistency constraints.

[0103] Step 5.2 Decentralized Weighted Gossip Aggregation: Each node exchanges model parameters with trusted neighbors through a single-hop mining D2D direct communication, performs weighted Gossip aggregation, and completes distributed policy collaborative updates. This process is completely independent of a central server, achieving distributed policy collaboration. The aggregation cycle is synchronized with the task scheduling cycle, adapting to millisecond-level emergency scheduling needs in underground mining environments.

[0104]

[0105] In the formula: The node generated in step 3 For nodes The federated aggregation confidence weights, with the weight matrix satisfying row normalization constraints, guarantee the convergence of the decentralized Gossip aggregation process.

[0106] Example 2: This example addresses a gas explosion disaster scenario in a fully mechanized coal mining face, with corresponding details provided. Figure 2 Scenario deployment diagram and appendix Figure 3 The spatiotemporal evolution process of the SGC-FedMAPPO algorithm, and the specific implementation steps are as follows:

[0107] S1 Multi-source physical parameter acquisition and node safety validity pre-verification: Collect original physical parameters of the mine, perform binary judgment on the validity of intrinsically safe heterogeneous edge nodes in the mine based on safety hard constraints, isolate disaster-failed nodes, generate an initial set of valid nodes and complete neighbor discovery and initial communication adjacency matrix construction;

[0108] This step corresponds to the attached document. Figure 1 ① The neighbor discovery process of the Hello frame, and the appendix Figure 3 The node hard isolation process following the T=50 catastrophe specifically includes:

[0109] S1.1 Acquisition of raw physical parameters;

[0110] Throughout the entire period, all nodes, with a scheduling cycle of 100ms, collect three core physical parameters in real time—gas volume concentration, ambient temperature, and vibration amplitude—using intrinsically safe sensors for mining, forming a physical safety state vector. (Corresponding to formula (1) mentioned above in this specification), directly used as the original state input of the algorithm. In this embodiment, a total of 18 intrinsically safe heterogeneous edge nodes for mining are deployed within the disaster impact range (corresponding to attached...). Figure 2 Appendix Figure 3The nodes (N1~N18) are divided into three categories: high-power nodes consist of 4 fixed emergency communication gateways (N3, N8, N10, N12), which is the preferred configuration in this embodiment. The CPU frequency is 1.2GHz and the computing power is 1200DMIPS. They are deployed at the connection roadway between the main transport roadway and the working face. Medium-power nodes consist of 6 underground inspection robots (N1, N4, N13, N16, N17, N18), which is the preferred configuration in this embodiment. The CPU frequency is 800MHz and the computing power is 500DMIPS. They are deployed on the inspection path of the fully mechanized mining face and the roadway. Weak-power nodes consist of 8 portable terminals for rescue personnel / intrinsically safe sensor nodes for mining (N2, N5, N6, N7, N9, N11, N14, N15), which is the preferred configuration in this embodiment. The CPU frequency is 400MHz and the computing power is 150DMIPS. They are deployed with the underground workers and fixed monitoring points.

[0111] S1.2 Hard determination of node validity;

[0112] Let the time of the catastrophe be T=50, which corresponds to the initial time of the algorithm's execution. All nodes complete hardware self-tests and parameter acquisition. Using the node validity hard determination formula (2) mentioned above in this manual, node N5 is determined to be a failed node and is directly and silently isolated, not participating in any subsequent networking and offloading process. At the same time, nodes N1, N2, N12, and N13 with increased risk are determined to be valid nodes, but are included in the high-risk monitoring range. All valid nodes complete one-hop neighbor discovery by broadcasting Hello frames and generate the initial communication adjacency matrix. This provides a foundation for subsequent topology reconstruction; corresponding appendices Figure 3 At time T=50, the failed node is quickly isolated, the original topology connection is interrupted, and the network enters the topology reconstruction preparation phase. In this embodiment, the safety boundary is strictly set in accordance with the "Coal Mine Safety Regulations": lower limit of gas explosion. Safety warning threshold Equipment's ultimate temperature tolerance Safety warning threshold Destructive vibration threshold .

[0113] S2 Attenuation-aware SGC federated aggregation topology reconstruction: Based on the initial communication adjacency matrix constructed in step 1, an attenuation-aware channel matrix is ​​constructed to address the attenuation characteristics of electromagnetic wave propagation in underground roadways. Lightweight simplified graph convolution without nonlinear activation is used to perform multi-hop smoothing on node communication features. Combined with the node validity determination results in step 1, a binary federated aggregation mask is generated to delineate the logically trusted neighbor domain of each node.

[0114] This step corresponds to the attached document. Figure 1 ③ SGC Trusted Topology Mask The generation process, and the appendix Figure 3 The topological self-organizing process during the disaster recovery phase from T=51 to T=60 specifically includes:

[0115] S2.1 Construction of attenuation sensing channel matrix in underground roadways;

[0116] A physical attenuation matrix is ​​constructed based on the actual spatial distance between nodes in the underground roadway. Combining the initial adjacency matrix, the attenuation-aware channel matrix is ​​generated using the aforementioned formula (3) in this specification. To match the electromagnetic wave attenuation characteristics of underground roadways and eliminate topological noise caused by non-line-of-sight propagation; in this embodiment, the electromagnetic wave attenuation coefficient of underground roadways is... Pick It is adapted to the spatial propagation characteristics of fully mechanized mining face roadways.

[0117] S2.2 SGC multi-hop feature smoothing;

[0118] The node communication characteristics are smoothed using the second-order SGC in formula (4) above in this specification to eliminate communication noise and output the smoothed node feature matrix. This algorithm has no nonlinear activation layer, and its computational cost is only a fraction of that of traditional GCN. It can run in real time on portable terminals with weak computing power; in this embodiment, the SGC multi-hop aggregation order It is adapted to the topology of underground linear tunnels.

[0119] S2.3 Binarized federated aggregation mask generation;

[0120] The feature cosine similarity between nodes is calculated using formula (5) mentioned above in this manual. Combined with the node validity determination result in step 1, a binary federated aggregation mask is generated using formula (6) mentioned above in this manual. The trusted neighbor domains of each node are defined; among them, nodes with increased risk N1, N2, N12, and N13 are only included in the temporary trusted domain by low-risk nodes, while the failed node N5 is excluded from the trusted neighbor domain by all nodes, thus achieving isolation of high-risk nodes at the topology level; corresponding to the attached Figure 3 The evolutionary process: During the disaster recovery phase, the algorithm uses SGC to reconstruct the topology, re-establishing trusted connections between valid nodes, and gradually restoring the network from a fragmented state after the disaster to a connected self-organizing network.

[0121] S3 Frequency Domain Differential Driven MoE Link Confidence Prediction: Based on the logically trusted neighbor domain defined in step 2, a fast Fourier transform is performed on the link signal-to-noise ratio time series to extract time domain statistical features and frequency domain energy difference features. A hybrid expert model with Top-1 sparse activation is used to output the survival probability of the link in the next scheduling cycle. The trusted neighbor domain is combined to generate row-normalized federated aggregation confidence weights.

[0122] This step corresponds to the attached document. Figure 1 The second part of the FFT+MoE link survival probability prediction stage, and the third part of the confidence weighting. The generation process specifically includes:

[0123] S3.1 Extraction of frequency domain energy difference features;

[0124] Each node collects the signal-to-noise ratio time series of its one-hop neighbor links, takes the most recent 64 sampling points (corresponding to a 320ms historical window), performs a 64-point FFT after adding a Hanning window, and extracts the spectral energy difference features between adjacent time points using the formula (7) mentioned above in this manual. This feature can capture early signs of channel spectrum collapse caused by dust and shock waves after a disaster, and can identify link interruption risks two scheduling cycles in advance compared to time-domain statistical features.

[0125] S3.2 Construction of sparse gated hybrid expert model;

[0126] The time-domain statistical features and frequency-domain differential features are concatenated and input into the Top-1 sparse activation MoE model of the aforementioned formulas (8)-(10) in this specification, and the link survival probability is output. Its physical meaning is the probability that the link will not be interrupted within the next scheduling cycle; among which, the attached Figure 2 In this example, the dashed weak links have a significantly lower survival probability than the solid usable links; links with a high risk of interruption are assigned a low survival probability. The number of expert networks in this embodiment... It is adapted to the computing power limitations of mining equipment.

[0127] S3.3 Generation of confidence weights for federated aggregation;

[0128] Based on the trusted topology mask generated by S2, row-normalized federated aggregation confidence weights are generated using the formula (11) mentioned above in this specification. The higher the link survival probability and the higher the topology similarity, the greater the corresponding aggregation weight, which provides a reliable basis for subsequent federated aggregation and offloading decisions.

[0129] S3.4 Training and Deployment of Hybrid Expert Models;

[0130] The MoE model is deployed using an offline pre-training + online fine-tuning approach. During the pre-training phase, publicly available channel measurement datasets from typical domestic coal mine disaster scenarios are used as samples. The binary classification label is whether the link will be interrupted within the next 100ms, and the gating network and expert network are jointly trained using a binary cross-entropy loss function. During the online execution phase, each node incrementally fine-tunes the model based on locally collected real-time channel data to adapt to the real-time changing channel environment after a disaster. The fine-tuning process is performed locally without global data interaction, making it fully adaptable to decentralized scenarios.

[0131] S4 Construction of a local Markov decision process with a logarithmic barrier function: Based on the original physical parameters in step 1 and the link survival probability in step 3, an observation state space including the physical security state of nodes, the task queue state, the security state of neighbors and the link survival probability are defined, as well as an action space including local computation and neighbor node offloading. A latency and energy consumption calculation model for heterogeneous node task offloading is constructed. A logarithmic barrier function is introduced to transform the security hard constraints of the original physical parameters into continuous penalty terms, and an instantaneous reward function with security constraints is constructed.

[0132] This step corresponds to the attached document. Figure 1 Mid-task unloading direction With local computing The decision-making process specifically includes:

[0133] S4.1 Basic definition of Markov decision process;

[0134] Each node, based on its own and trusted neighbors' original physical parameters, delineates the feasible domain for safe actions using the aforementioned formula (12) in this specification: only when the gas concentration and temperature of the target node are both below the safety warning threshold are they included in the legal unloading target; risk-increased nodes N1, N2, N12, and N13 are excluded from the legal unloading target set because their parameters exceed the safety warning threshold, and all nodes can only unload tasks to safe nodes or force local calculations; in this embodiment, the core emergency services are the location calculation of trapped personnel, real-time data analysis of environmental monitoring, and emergency voice coding. The tasks arrive according to a Poisson distribution. As a preferred setting in this embodiment, the arrival rate is 5 tasks / second, and the maximum tolerable latency for a single task is 500ms. The priority of personnel location and emergency voice services is higher than that of environmental monitoring services, corresponding to the task priority weights. Differentiated settings.

[0135] S4.2 Calculation of task unloading latency and energy consumption on heterogeneous nodes;

[0136] Based on the differences in computing power among heterogeneous nodes, a task unloading latency and energy consumption calculation model is constructed using the formulas (13)-(14) mentioned above in this manual. Different latency-energy consumption weights are set for emergency services of different priorities, prioritizing the low latency requirements of high-priority services such as personnel positioning and emergency voice.

[0137] S4.3 Construction of instantaneous reward function with logarithmic barrier function;

[0138] By defining the instantaneous reward function with a logarithmic barrier function using formulas (15)-(16) mentioned above in this specification, a hard safety constraint is embedded into the algorithm's underlying layer: when the algorithm attempts to offload tasks to nodes with increased risk or weak links during the exploration process, the overall safety margin is considered. Approaching 1, the logarithmic barrier term generates a penalty value approaching negative infinity, forcing the algorithm to absolutely avoid high-risk nodes and vulnerable links, achieving a dual hard constraint on both physical and network security; corresponding to the appendix Figure 2 Scenario: Nodes N1 and N4 within the fully mechanized mining face are forced to perform local calculations due to their location in an area of ​​increased risk. Security nodes N3 and N10 will offload the task to neighboring security nodes N2 and N17. All unloading actions avoid interrupted links and high-risk nodes. In this embodiment, the security penalty coefficient... Delay weight Energy consumption weight Priority will be given to ensuring the low latency requirements of emergency operations.

[0139] S5 Solving SGC-FedMAPPO under topological consistency regularization constraints; Based on the Markov decision process constructed in step 4, constructing a near-end policy optimization pruning loss function with topological Laplace regularization term, each node updates the policy network based on local experience data, exchanges model parameters with trusted neighbors only through one-hop mining D2D direct communication, performs decentralized weighted Gossip aggregation based on the confidence weight, and outputs the optimal collaborative task offloading strategy.

[0140] This step corresponds to the attached document. Figure 1 The fourth D2D weighted Gossip aggregation stage, and the attached Figure 3 middle The policy convergence process during the steady-state reconstruction phase specifically includes:

[0141] S5.1 Local policy update with topological Laplace regularization;

[0142] Each node uses only locally collected empirical data to update its local policy network parameters using the total loss function with topological Laplacian regularization term in formula (17) of this specification. Simultaneously, the PPO pruning loss function in formula (18) of this specification limits the policy update magnitude, ensuring training stability. The regularization term forces the model parameters of trusted neighbors to converge, reducing the variance of decentralized aggregation and accelerating policy convergence. In this embodiment, a single scheduling cycle... PPO clipping threshold Topological regularity coefficients .

[0143] S5.2 Decentralized weighted Gossip aggregation;

[0144] Each node exchanges model parameters with trusted neighbors through only one-hop D2D communication, and performs weighted Gossip aggregation based on the confidence weight in step 3 using the formula (19) mentioned above in this manual. This completely eliminates the need for a central server, achieving distributed strategy collaborative evolution. The aggregation cycle is synchronized with the 100ms task scheduling cycle, adapting to the real-time requirements of underground emergency operations. (Corresponding appendix...) Figure 3 The evolutionary process: After the disaster, the algorithm underwent iterative optimization. When the policy update magnitude was less than [a certain value] within 20 consecutive scheduling cycles, [the algorithm was optimized]. When the time is reached, the strategy is considered to have converged; in this embodiment, the algorithm converges before time T=100. The system enters a steady-state reconstruction phase, the network topology stabilizes, and the task offloading strategy converges to the optimal level, achieving safe, low-latency collaborative offloading in disaster scenarios.

[0145] Verification of the implementation effect of this invention:

[0146] This embodiment verifies the effectiveness of the method of the present invention in the extreme scenario of a mine disaster without a central authority. It also compares the effectiveness with the core technical deficiencies of existing solutions cited in the background section. The verification results are as follows:

[0147] In terms of security performance, the method of this invention, through a dual mechanism of hard determination of node validity and logarithmic barrier security constraint, can complete the hard isolation of failed nodes within one scheduling cycle after a disaster occurs, and achieve 100% interception of task offloading behavior to high-risk areas. No invalid offloading behavior to failed or high-risk nodes occurred during the entire operation, fully meeting the core requirement of "safety red line cannot be crossed" in underground emergency scenarios. In contrast, existing centralized solutions do not embed the underground physical safety boundary into the algorithm layer and only avoid risks through post-screening. In disaster scenarios, offloading attempts to high-risk areas are prone to occur, posing security risks of business interruption and data loss.

[0148] In terms of convergence performance and disaster adaptability, the method of this invention adopts lightweight SGC topology reconstruction and topology consistency regularization constraints. In scenarios where the network topology changes drastically after a disaster, policy convergence can be completed within dozens of scheduling cycles, quickly adapting to the dynamic topology after the disaster. In contrast, existing centralized solutions rely on global state collection and central node scheduling. After a disaster, the destruction of core infrastructure will directly lead to the paralysis of the architecture, making it impossible to complete distributed topology reconstruction and policy convergence, and unable to adapt to the dynamic changes in disaster scenarios.

[0149] In terms of latency performance, the method of this invention, through one-hop D2D offloading and distributed collaborative scheduling, can stably control the end-to-end latency of core emergency rescue services within the maximum tolerable latency range of 500ms, fully meeting the low latency requirements of key services such as underground personnel positioning and emergency voice communication; while existing solutions require multi-hop relays to transmit data to the remaining macro base stations after a disaster, resulting in a significant increase in end-to-end latency, which cannot meet the real-time requirements of emergency services.

[0150] Regarding link reliability, the method of this invention, through frequency domain differential-driven MoE link prediction, can identify early signs of channel fast fading after a disaster, proactively avoid high-risk links, significantly reduce the probability of link interruption during task offloading, and effectively solve the problem of service transmission interruption caused by strong channel non-stationarity after a disaster. In contrast, existing solutions mostly use time-domain sequence models for link prediction, which cannot effectively capture the early characteristics of channel spectrum collapse, and are prone to making incorrect offloading decisions when the link is about to be interrupted, resulting in a significant increase in service transmission failure rate.

[0151] In terms of computing power adaptability, the method of this invention adopts the SGC with no nonlinear activation and the Top-1 sparse activation MoE model. The overall computing power overhead of the algorithm is low, and the algorithm running time in a single scheduling cycle is much shorter than the length of the scheduling cycle. It can run stably in real time on intrinsically safe mining equipment with low computing power in underground mines. In contrast, existing schemes based on traditional GCN and full activation expert models have high nonlinear computational complexity and large computing power overhead, and cannot achieve real-time scheduling on low computing power embedded nodes in underground mines.

[0152] The results show that the method of the present invention can achieve self-organizing networking and secure collaborative task offloading of heterogeneous edge nodes in extreme scenarios without a center after a mine disaster, and fully achieves the expected purpose of the invention.

[0153] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for offloading heterogeneous mobile edge computing self-organizing collaborative tasks in mine emergency communication based on SGC-FedMAPPO, characterized in that, Including the following steps: Step 1: Multi-source physical parameter acquisition and node safety validity pre-verification: Collect the original physical parameters of the mine, perform binary judgment on the validity of intrinsically safe heterogeneous edge nodes in the mine based on safety hard constraints, isolate disaster-failed nodes, generate an initial set of valid nodes and complete neighbor discovery and initial communication adjacency matrix construction. Step 2, Attenuation-aware topology reconstruction: Based on the initial communication adjacency matrix constructed in Step 1, an attenuation-aware channel matrix is ​​constructed to address the attenuation characteristics of electromagnetic wave propagation in underground roadways. A lightweight simplified graph convolution without nonlinear activation is used to perform multi-hop smoothing on the node communication features. Combined with the node validity determination results from Step 1, a binary federated aggregation mask is generated to delineate the logically trusted neighbor domain of each node. Step 3: Hybrid expert model link confidence prediction: Based on the logically trusted neighborhood domain defined in Step 2, perform fast Fourier transform on the link signal-to-noise ratio time series, extract time-domain statistical features and frequency-domain energy difference features, use a hybrid expert model with Top-1 sparse activation to output the link survival probability in the next scheduling cycle, and combine the trusted neighborhood domain to generate row-normalized federated aggregation confidence weights. Step 4, Construction of Markov Decision Process under Security Constraints: Based on the original physical parameters in Step 1 and the link survival probability in Step 3, define an observation state space that includes the physical security state of nodes, the task queue state, the security state of neighbors, and the link survival probability, as well as an action space that includes local computation and neighbor node offloading. Construct a latency and energy consumption calculation model for heterogeneous node task offloading, introduce a logarithmic barrier function to transform the security hard constraints of the original physical parameters into continuous penalty terms, and construct an instantaneous reward function with security constraints. Step 5, Model Solving under Topological Consistency Regularization Constraints: Based on the Markov decision process constructed in Step 4, a near-end policy optimization pruning loss function with topological Laplace regularization term is constructed. Each node updates the policy network based on local experience data and exchanges model parameters with trusted neighbors only through one-hop mining D2D direct communication. Decentralized weighted Gossip aggregation is performed based on the confidence weight to output the optimal collaborative task offloading strategy.

2. The method for offloading heterogeneous mobile edge computing self-organizing collaborative tasks in mine emergency communication based on SGC-FedMAPPO according to claim 1, characterized in that, The intrinsically safe heterogeneous edge node for mining includes a fixed emergency communication gateway for mining, an underground inspection robot, a portable terminal for rescue personnel, and intrinsically safe sensing nodes for mining. All types of nodes meet the explosion-proof standards for intrinsically safe equipment for mining.

3. The method for offloading heterogeneous mobile edge computing self-organizing collaborative tasks in mine emergency communication based on SGC-FedMAPPO according to claim 1, characterized in that, Step 1 specifically includes: Step 1.1 Acquisition of raw physical parameters: Each edge node acquires the gas volume concentration, ambient temperature, and vibration amplitude of its current environment in real time through intrinsically safe sensors for mining, forming a physical safety state vector, which is directly used as the raw state input for multi-agent reinforcement learning; Step 1.2 Hard determination of node validity: Based on the safety hard constraints stipulated in the "Coal Mine Safety Regulations" and supporting national standards, the validity of nodes is determined by binary method. A node is determined to be a valid node only when it simultaneously meets the following conditions: vibration amplitude is lower than the destructive vibration threshold stipulated in the national standard, gas concentration is lower than the lower explosive limit stipulated in the national standard, and ambient temperature is lower than the equipment limit tolerance temperature stipulated in the national standard. Otherwise, it is determined to be a failed node and is silently isolated.

4. The method for offloading heterogeneous mobile edge computing self-organizing collaborative tasks in mine emergency communication based on SGC-FedMAPPO according to claim 1, characterized in that, Step 2 specifically includes: Step 2.1 Construction of Attenuation Sensing Channel Matrix in Underground Roadway: Combining the initial communication adjacency matrix and physical attenuation matrix of the underground roadway, an attenuation sensing channel matrix that conforms to the spatial propagation characteristics of the mine is constructed. The elements of the physical attenuation matrix are calculated from the spatial distance between nodes and the electromagnetic wave attenuation coefficient of the underground roadway. Step 2.2 SGC Multi-hop Feature Smoothing: A lightweight SGC without nonlinear activation is used to smooth the node communication features in multiple hops, reducing the computational overhead of the algorithm and adapting to the computational power limitations of mining equipment; Step 2.3 Binarized Federation Aggregation Mask Generation: Based on the smoothed features, calculate the cosine similarity between nodes, and combine it with the node validity determination results from Step 1 to generate a binarized federated aggregation mask, delineating the logically trustworthy neighborhood of each node.

5. The method for offloading heterogeneous mobile edge computing self-organizing collaborative tasks in mine emergency communication based on SGC-FedMAPPO according to claim 1, characterized in that, Step 3 specifically includes: Step 3.1 Frequency domain energy difference feature extraction: Perform fast Fourier transform on the link signal-to-noise ratio time series to extract the spectral energy difference features between adjacent time points, capture the precursor features of fast channel fading after a disaster, and identify the risk of link interruption in advance compared with time domain statistical features; Step 3.2 Construction of Sparse Gated Hybrid Expert Model: The MoE model is constructed using the Top-1 sparse activation mechanism to reduce computing power overhead, adapt to the computing power limitations of mining equipment, and output the survival probability of the link in the next scheduling cycle. Step 3.3 Federated Aggregate Confidence Weight Generation: Combine the trusted neighbor domain mask generated in Step 2 to generate row-normalized federated aggregation confidence weights that satisfy the convergence conditions of decentralized Gossip aggregation; Step 3.4 Training and Deployment of the MoE Model: The MoE model is deployed using an offline pre-training + online fine-tuning approach. During the pre-training phase, measured underground channel data from historical disaster scenarios are used as samples, and the link interruption within a future scheduling cycle is used as a binary classification label. A binary cross-entropy loss function is employed to jointly train the gated network and the expert network. During the online operation phase, each node performs incremental fine-tuning of the model based on locally collected real-time channel data. The fine-tuning process is executed only locally on the node and does not require global data interaction.

6. The method for offloading heterogeneous mobile edge computing self-organizing collaborative tasks in mine emergency communication based on SGC-FedMAPPO according to claim 1, characterized in that, Step 4 specifically includes: Step 4.1 Basic definition of Markov decision process: Define the observation state space and unloading action space for matching the centerless scenario, delineate the feasible domain of safe actions based on the safety threshold of the original physical parameters, and prohibit unloading tasks to high-risk nodes from the action level; Step 4.2 Calculation of Task Offloading Latency and Energy Consumption for Heterogeneous Nodes: To address the differences in computing power among heterogeneous nodes in the mine, a calculation model for task offloading latency and energy consumption that fits the edge computing scenario is constructed. Step 4.3 Construction of instantaneous reward function with logarithmic barrier function: Introduce logarithmic barrier function to transform the security hard constraint into a continuous penalty term at the bottom layer of the algorithm, realizing dual hard constraints of physical security and network security.

7. The method for offloading heterogeneous mobile edge computing self-organizing collaborative tasks in mine emergency communication based on SGC-FedMAPPO according to claim 6, characterized in that, In the feasible domain of safe actions in step 4.1, the safety interval determination threshold of the target unloading node is the safety warning threshold specified in the "Coal Mine Safety Regulations"; in the instantaneous reward function in step 4.3, the input of the logarithmic barrier function includes the safety margin of the target node and the probability of link interruption. When the target node approaches the safety threshold or the probability of link survival approaches 0, the logarithmic barrier function outputs a negative penalty value.

8. The method for offloading heterogeneous mobile edge computing self-organizing collaborative tasks in mine emergency communication based on SGC-FedMAPPO according to claim 1, characterized in that, Step 5 specifically includes: Step 5.1 Local policy update with topological Laplace regularization: Each node updates the policy network using only local empirical data. A topological Laplace regularization term is added to the PPO pruning loss function to force the model parameters of trusted neighbors to converge, reduce the variance of decentralized aggregation, and accelerate policy convergence. Step 5.2 Decentralized Weighted Gossip Aggregation: Each node exchanges model parameters with trusted neighbors through only one-hop mining D2D direct communication, performs weighted Gossip aggregation, and completes distributed policy collaborative update.

Images