Search

CN-121981315-A - Bracket working surface area division and prediction method based on digital twin and multiple agents

CN121981315ACN 121981315 ACN121981315 ACN 121981315ACN-121981315-A

Abstract

The invention discloses a method for dividing and predicting a bracket working surface area facing to a complex geological environment of a deep mine, the method utilizes digital twinning and multiple agents to establish a bracket working surface area division and prediction system architecture, and comprises a perception layer, an interaction layer, a twinning layer and an application layer. The method comprises the steps of carrying out coarse initial division on a bracket working surface area by adopting a bandwidth self-adaptive MeanShift clustering algorithm, carrying out secondary fine division on the coarse division area by adopting a K-Means clustering algorithm, adding an online updating mechanism so as to carry out area dynamic division on new acquired data, and carrying out prediction on the divided area by adopting a TCN-differential attention mechanism-Informer model structure which is processed in a two-way parallel manner and is fused and output. The intelligent decision and dynamic feedback mechanism is integrated, and a method for dividing the area of the working surface of the bracket and predicting the load is established.

Inventors

  • ZHANG FAN
  • YAN SHUAISHUAI
  • CHENG HAIXING

Assignees

  • 中国矿业大学(北京)
  • 中煤能源研究院有限责任公司

Dates

Publication Date
20260505
Application Date
20251218

Claims (4)

  1. 1. The method is characterized in that the method utilizes digital twin and multiple agents to establish a working face support region division and prediction system framework, adopts a clustering algorithm to divide a working face support region, realizes the dynamic division of a region of new acquired data through an online updating mechanism, predicts the divided region by adopting a TCN-differential attention mechanism-Informer model structure, and realizes the accurate prediction of the mine working face support load under complex working conditions by establishing a working face support region division and load prediction flow method; The system architecture of the working face support region dividing and predicting method based on the digital twin and multi-agent comprises a sensing layer, an interaction layer, a twin layer and an application layer; the sensing layer consists of multiple intelligent agents, cameras, a multi-source sensor, a signal acquisition system and a roof separation monitoring system, wherein the cameras, the multi-source sensor, the signal acquisition system and the roof separation monitoring system are arranged on the hydraulic support, the multiple intelligent agents monitor surrounding rock stress by sensing mine working face environment information, support load data acquisition, processing and fusion are realized through real-time interaction and cooperation between a digital twin body and a physical entity, and data transmission and synchronous feedback of an interaction layer are realized through establishing an online updated learning mechanism; The interaction layer consists of multiple agents, zigBee, loRa, wiFi, RS, an Ethernet, an OPC UA communication interface, a cloud edge cooperative gateway, an intelligent gateway and a man-machine interaction interface, wherein the communication connection agents are responsible for data standardization among communication protocols, the communication protocols used by different devices or systems are converted into uniform intermediate protocols to ensure data intercommunication among heterogeneous devices, the data fusion agents are responsible for coordinating cloud edge task allocation to realize the uploading cloud of instruction data and the edge processing of real-time tasks, the interaction mapping agents and the twin layer coordinate operation to establish a bidirectional mapping relation between a physical entity and a digital twin body by implementing a data conversion and space mapping algorithm, the multimode interaction agents guarantee the accurate correspondence of data and behaviors to ensure the consistency of a bracket physical system and a virtual model, the interaction management agents are responsible for overall cooperation of the agents, optimizing interaction flow and simultaneously completing interaction feedback with the twin layer; The system comprises a twin layer, a sensing layer, a surrounding rock coupling layer, a simulation layer, a surrounding rock coupling model, a dynamic optimization intelligent body, a digital-analog linkage intelligent body, a sensing layer, a surrounding rock coupling layer and a surrounding rock coupling model, wherein the twin layer consists of a plurality of intelligent bodies, a twin model and a twin scene, the plurality of intelligent bodies comprise a data service intelligent body, a simulation modeling intelligent body, a dynamic optimization intelligent body and an interactive management intelligent body, the data service intelligent body is responsible for modeling, optimizing and analyzing supporting data, simulation data, real-time acquisition data and twin data in a database, the simulation modeling intelligent body is mainly responsible for constructing a physical model, a behavior model, the simulation model and the surrounding rock coupling model, the three-dimensional physical model of the bracket is constructed through geometrical parameters and a physical structure, the behavior model and the dynamic behavior model are constructed, the physical model and the behavior model are driven to construct the simulation model on a visual platform, and finally the surrounding rock coupling model is fused with the twin model; The application layer mainly realizes intelligent prediction and autonomous decision on the hydraulic support supporting load through multiple intelligent agents, wherein a state monitoring intelligent agent outputs a supporting load prediction value in real time through an embedded load prediction model, a model algorithm intelligent agent realizes iterative update of the prediction model through dynamic adjustment of model parameters through a self-adaptive learning algorithm, a prediction diagnosis intelligent agent carries out multidimensional attribution analysis on an overrun prediction result, identifies model deviation or abnormal working conditions, a decision control intelligent agent judges consistency of a true value and the prediction value based on the prediction result and a set threshold range, if the prediction result is in a threshold interval, a prediction result is directly output and a system database is updated, if the prediction result exceeds a threshold value, an optimization learning mechanism is triggered, the optimization learning intelligent agent synchronizes the optimization result to a physical entity and the prediction model, and closed-loop control of a support 'data perception-load prediction-decision judgment-optimization feedback' is realized through division and cooperation among intelligent agents; The working face support region dividing and predicting method based on the digital twin and multiple agents comprises the following steps: Firstly, carrying out rough division on a bracket working surface area by adopting a bandwidth self-adaptive MeanShift clustering algorithm, carrying out fine division on the rough division area by adopting a K-Means clustering algorithm, and carrying out area dynamic division on newly acquired data by adopting an online updating mechanism; step 2, processing the divided areas by adopting a two-way parallel fusion algorithm, and predicting an output TCN-differential attention-Informer model structure by utilizing new acquired data; the working face support region division and prediction method based on the digital twin and multi-agent further comprises the following sub-steps: step 1-1, monitoring pressure data of front/rear stand columns of a plurality of supports on a mine working surface, continuously numbering the supports according to space sequences of the working surface, calculating the weighted average pressure of circulation time of each support in a certain time, and preprocessing the pressure data; In the formula, Weighting the average pressure for the cycle time; Is the first Pressure values at each instant; Is the first The time interval between the previous pressure and the next pressure at each moment; is the total number of time periods; Step 1-2, adopting a MeanShift clustering algorithm for self-adaptive bandwidth kernel density estimation, inputting a bracket number and corresponding cycle time weighted average pressure, initially dividing n clusters of a bracket supporting area by a cross verification method, introducing a space continuity merging rule, and finally generating m space continuous rough dividing areas; Step 1-3, selecting a median bracket number and front and rear column pressure data of the median bracket number for each rough divided area, constructing a feature set, determining an optimal cluster number K based on a contour coefficient method, realizing fine division of a bracket supporting area through a K-Means clustering algorithm, and restraining continuous bracket numbers in the same cluster; step 1-4, determining the area jump phenomenon of the classification result of the partial support area according to the probability of the area where the support load is and the space adjacent degree between the supports, and finally determining the support area; step 1-5, adding an online updating mechanism to the newly acquired data, solving the distance between the newly acquired data and the center of each cluster, and obtaining the area of the newly acquired data by increasing the probability that the data belongs to a certain cluster as the distance between the newly acquired data and the center of each cluster is smaller; The working face support region dividing and predicting method based on the digital twin and multi-agent is used for preprocessing pressure data, and comprises the following steps: step 2-1, invalid value processing, namely deleting data exceeding the support load range in the data table; step 2-2, interpolation is carried out on the data which are deleted individually by adopting an average value method; and 2-3, carrying out Kalman filtering noise reduction treatment on the effective and completed data: In the formula, Is that A predicted value of stent pressure at a moment; in the form of a state transition matrix, Is that A pressure optimal estimated value at the moment; for controlling the input matrix; is a system input; , Respectively is , A corresponding covariance; , covariance of process white noise and measurement white noise, respectively, subject to gaussian distribution; Is Kalman gain; For the moment of time Is a vector of observation of (a); is an observation matrix; According to the working face support region dividing and predicting method based on the digital twin and multi-agent, a predicting model adopts a two-way parallel processing and fusion output model structure formed by a TCN-differential attention mechanism and a Informer model, local time sequence characteristics and global dependency relations are respectively extracted, and finally a predicting result is generated through a fusion layer; the working face support region dividing and predicting method based on the digital twin and multiple agents adopts a TCN-differential attention mechanism plus Informer two-way parallel processing prediction model, and the algorithm flow is as follows: Step 3-1, a TCN-differential attention mechanism receives long-sequence input, a time constraint is constructed by utilizing causal convolution, a long-period and short-period dependency relationship is captured, gradient attenuation is relieved through residual connection, finally multichannel characteristics containing local time sequence characteristics are output, the multichannel characteristics are grouped, attention fraction differences of different groups are calculated, dynamic weights are generated by comparing the differences, and characteristic enhancement of a load sensitive area is realized; Step 3-2, the Informier model encoder receives long sequence input, the input long sequence is subjected to probability sparse self-attention and self-attention distillation operation, the formed attention characteristic diagram is sent to the multi-head attention of the decoder, the decoder receives the long sequence, a multi-head attention interaction mechanism is constructed, the history coding characteristic and the future position coding are fused, and a complete prediction sequence is generated through single forward propagation by adopting non-autoregressive prediction; And 3-3, the fusion layer performs dimension alignment on the local sensitive features of the TCN branches and the global features extracted by the Informer branches, calculates the feature contribution degree weight through the gating unit, realizes dynamic weighted fusion of the two types of features, maps the fused features to a target space through the full connection layer, and outputs a final prediction result.
  2. 2. The method for dividing and predicting the area of a working surface bracket based on digital twinning and multi-agent according to claim 1, further comprising the steps of: step 2-1, collecting real-time load data of the global hydraulic support by a sensing layer, cleaning noise data, extracting space-time characteristics and transmitting the data to a database; Step 2-2, dividing the bracket load into areas by using a clustering algorithm according to geological conditions and preprocessing data, generating codes for each area, and giving out dividing limits of each area; Step 2-3, building a unified prediction model, training the model by using global data, and optimizing a loss function to balance errors of different areas so as to prevent the model from deviating to a high-load area; Step 2-4, inputting data and encoding a bracket, synchronously outputting all area load predicted values by a model, and explaining the area contribution degree by a visualization tool; and 2-5, comparing the predicted result with a set threshold range, if the predicted result is within the threshold range, displaying the predicted result in a visual interface, and displaying the global load thermodynamic diagram in the digital twin platform, otherwise, respectively performing voltage regulation and regional weight parameter adjustment on the physical entity and the prediction algorithm model through dynamic feedback and optimization.
  3. 3. The multi-agent according to claim 1, wherein the multi-agent is composed of a plurality of single agents, and the single agents have the functions of sensing environment, learning online, making independent decisions and executing actions, can perform information interaction and coordination with other agents according to the change of the target and environment, and make adaptive feedback, and the multi-agents perform cooperative work together and realize the target through physical interaction and information sharing.
  4. 4. The method for dividing the area of the working face support according to claim 1, wherein the method is used for aiming at the problem that the working face load difference is obvious due to the long-term circulation disturbance effect and partition breaking of a rock stratum of a top plate of a downhole working face, the roadway sides at two ends of the working face start to weaken the support of overlying rocks in the middle of the working face along with the increase of the length of the working face, the sinking amount of the overlying rocks in the middle is gradually increased, and the peak area in the middle is gradually moved towards two ends.

Description

Bracket working surface area division and prediction method based on digital twin and multiple agents Technical Field The invention belongs to the field of digital twinning and artificial intelligence mine pressure and prediction, and particularly relates to a bracket working surface area dividing and predicting method based on digital twinning and multiple intelligent agents. Background In the deep mining process of the coal mine, stope roof and surrounding rock present typical multilayer discontinuous medium mechanical characteristics, and are obviously influenced by multiple geological disasters such as fault structures, rock burst and the like. Along with the increase of the advancing length of the working face, the structural mechanical response of the mining area has an evolution rule that the supporting effect of the roadway sides at the two ends on the middle overlying strata decays exponentially, so that the stratum subsidence of the middle area is in a nonlinear growth state, and the peak stress area gradually migrates to the stope boundary. The study shows that when the length of the working surface reaches 300M, the supporting stress of the hydraulic support group begins to appear in a saddle-shaped tri-peak value M shape, and when the length of the working surface reaches 350M, the saddle shape is more obvious. In order to meet the support pressure requirement of the deep well working face, dynamic division and refined prediction of the hydraulic support working face area are required so as to implement a more targeted support strategy. The digital twin technology is used as an intelligent simulation system for multidisciplinary deep integration, and full life cycle bidirectional interaction mapping of physical entities and virtual space is realized through high-precision dynamic model construction, multisource heterogeneous data real-time fusion and autonomous decision algorithm optimization. In the field of mine intellectualization, a closed loop system of 'multi-mode perception-high-fidelity simulation-self-adaptive regulation and control' is established based on a digital twin hydraulic support dynamics model, and dynamic response of the support under complex geological conditions is accurately simulated through coupling of real-time data driving and a mechanism model, so that core support is provided for intelligent support decision. The hydraulic support column pressure prediction analysis is realized, and the hydraulic support column pressure prediction analysis has important functions of improving support adaptability and ensuring stability of surrounding strata. Aiming at the problem of load prediction under the condition that load differences in different areas of a working surface are obvious due to long-term circulation disturbance action and partition fracture of a roof stratum of a deep well long working surface, the invention provides a TCN-differential attention mechanism-Informer model based on a Informer model, a time domain convolution network (TCN) is adopted to construct a multi-scale time sequence feature extraction module, a long-period evolution rule of roof pressure is effectively captured, a differential attention mechanism is designed, the attention score differences of different groups are calculated, dynamic weights are generated by comparing the differences, feature enhancement of a load sensitive area is realized, and a long-sequence prediction advantage of a Informer model is combined, so that the full-length pressure field space-time model of the working surface is established, and the accurate prediction of the pressure of a hydraulic support is realized. Disclosure of Invention In order to solve the technical problems, the invention aims to provide a support working surface area dividing and predicting method based on digital twin and multiple intelligent bodies, which solves the problems that the load difference of different areas of a working surface is obvious and the support load prediction under the dynamic update of the working surface area is difficult to realize due to the fact that the regional fracture exists in a top plate rock layer of a deep well long working surface. The method is based on a digital twin technology and a multi-agent system, the MeanShift algorithm and the K-Means algorithm are utilized to divide the area of the working face of the support, the TCN-differential attention mechanism-Informer model structure is adopted to predict the divided area, differential and accurate prediction of the support load under multiple working conditions is achieved, and the state sensing and intelligent decision level of the working face support system is effectively improved. The invention provides a method for dividing and predicting a working face region of a bracket based on digital twin and multiple agents, which utilizes the digital twin and multiple agents to establish a working face bracket region dividing and predicting system framework, ado