Search

CN-122022070-A - Multi-agent collaborative deduction method and system for oilfield emergency scene based on visual anomaly perception

CN122022070ACN 122022070 ACN122022070 ACN 122022070ACN-122022070-A

Abstract

Provided are an oilfield emergency scene multi-agent collaborative deduction method and system based on visual anomaly perception, which relate to the field of high-risk industrial scene safety production and emergency treatment deduction. The method solves the problems of unquantized visual abnormality sensing, low judging precision and the like in the existing oilfield emergency treatment deduction field, provides an end-to-end solution for oilfield emergency deduction by performing double-source data processing, initializing multi-agent deduction environment, collaborative reinforcement learning deduction, quantifying abnormal diffusion and rescue progress, evaluating deduction results, constructing a closed loop flow of visual display, realizing deep fusion of visual abnormality sensing and multi-agent emergency deduction, constructing a six-layer closed loop architecture by taking full-flow visualization as an operation carrier, realizing data intercommunication and result feedback among all layers, and is also suitable for the fields of emergency event sensing and collaborative treatment deduction such as fire, leakage and explosion.

Inventors

  • LV JIALIANG
  • ZHANG JIAN
  • BAI LU

Assignees

  • 大庆安瑞达科技开发有限公司

Dates

Publication Date
20260512
Application Date
20260403

Claims (10)

  1. 1. An oilfield emergency scene multi-agent collaborative deduction method based on visual anomaly perception is characterized by comprising the following steps of: step 1, configuring data parameters in a visual interface, constructing synthetic data based on the selected data parameters to generate and process double-source data with real data, preprocessing the double-source data, and inputting the preprocessed double-source data into PADIM visual anomaly detection models; Step 2, PADIM, extracting an oilfield emergency image to be detected by a visual anomaly detection model, calculating a pixel anomaly score and a global anomaly score S, and generating an anomaly thermodynamic diagram; based on the abnormality detection threshold T, completing abnormality qualitative judgment, if the abnormality is judged, extracting an abnormality position P (x, y) and an abnormality type, and outputting an abnormality score, an abnormality thermodynamic diagram, the abnormality position and the abnormality type as deduction parameters; Step 3, the outputted abnormal positions P (x, y) and the abnormal types are imported into a multi-agent emergency deduction environment, the abnormal positions are marked as dangerous areas, personalized initialization of the deduction environment is completed, and environment initial observation information is generated; Step 4, generating a collaborative decision action by the intelligent agent according to the environment initial observation information generated in the step 3, updating the position of the intelligent agent, optimizing resource allocation according to the scheduling action, calculating the rescue progress increment delta Pr of the single deduction step, and updating the real-time rescue progress Pr; if the rescue progress Pr is less than 1, simulating abnormal spatial diffusion according to an abnormal diffusion rule, and updating the dangerous area distribution of the scene grid; calculating the total rewards R of the single deduction step according to the multi-agent collaborative rewards function, and using the total rewards R for strengthening the optimization feedback of the learning model; Repeating the steps until the rescue progress Pr is more than or equal to 1 or the number of dangerous areas exceeds a threshold value or the maximum deduction steps are reached, and stopping deduction.
  2. 2. The multi-agent collaborative deduction method for the oilfield emergency scene based on visual anomaly perception according to claim 1 is characterized in that the synthetic data in the step 1 is used for generating tagged image data of three anomalies of fire, leakage and explosion by simulating oilfield geographical scenes and has anomaly type, position and degree tagging information, and real data is used for supporting oilfield emergency image loading and analysis in a standardized format and realizing seamless switching of double data sources.
  3. 3. The multi-agent collaborative deduction method for the oilfield emergency scene based on visual anomaly perception according to claim 1, wherein the PADIM visual anomaly detection model extracts depth features of oilfield emergency images through a backbone network, achieves quantification of anomaly scores by adopting PCA dimension reduction and Markov distance calculation, generates an anomaly region thermodynamic diagram, completes anomaly qualitative judgment based on a preset threshold, and outputs anomaly scores, anomaly thermodynamic diagrams and anomaly positions.
  4. 4. The method for deducing the multi-agent cooperation of the emergency scene of the oil field based on visual anomaly perception according to claim 3, wherein the method for realizing the quantification of anomaly score by adopting PCA dimension reduction and Markov distance calculation is characterized by comprising the following steps: Wherein, the Is the inverse of the covariance matrix.
  5. 5. The visual anomaly perception-based oilfield emergency scene multi-agent collaborative deduction method according to claim 1, wherein the method for calculating the pixel anomaly score and the global anomaly score S in the step 2 is as follows: generating a pixel anomaly score as an anomaly thermodynamic diagram, carrying out normalization processing on the maximum value of the thermodynamic diagram to obtain a global anomaly score S, and representing the anomaly degree of the whole image: 。
  6. 6. the method for collaborative deduction of multiple agents in an oilfield emergency scene based on visual anomaly perception according to claim 1, wherein the method for calculating the total prize R of a single deduction step according to the collaborative prize function of multiple agents in step 4 is as follows: Wherein the weight coefficient , For the purposes of rewarding the patrol agent, Is awarded for the rescue agent, In order to schedule the agent rewards, A penalty is given for the danger zone, And (5) rewarding is completed for rescue.
  7. 7. The multi-agent collaborative deduction method for the oilfield emergency scene based on visual anomaly perception according to claim 1, wherein in the step 4, resource allocation is optimized according to scheduling actions, and the method for calculating the rescue progress increment deltapr of a single deduction step is as follows: Wherein, the For the distance from the rescue agent to the abnormal position, L is the maximum side length of the oilfield scene grid, and the rescue progress is real-time , 。
  8. 8. An oilfield emergency scene multi-agent collaborative deduction system based on visual anomaly perception, which is characterized by comprising: the dual-source data processing module is used for configuring data parameters in the visual interface, constructing synthetic data based on the selected data parameters to generate and process dual-source data with real data, preprocessing the dual-source data, inputting the preprocessed dual-source data into PADIM visual anomaly detection model, PADIM the visual anomaly quantitative perception module is used for extracting an oilfield emergency image to be detected by the PADIM visual anomaly detection model, calculating pixel anomaly score and global anomaly score S, and generating an anomaly thermodynamic diagram; based on the abnormality detection threshold T, completing abnormality qualitative judgment, if the abnormality is judged, extracting an abnormality position P (x, y) and an abnormality type, and outputting an abnormality score, an abnormality thermodynamic diagram, the abnormality position and the abnormality type as deduction parameters; The multi-agent emergency deduction environment module is used for leading the outputted abnormal positions P (x, y) and the abnormal types into the multi-agent emergency deduction environment, marking the abnormal positions as dangerous areas, completing personalized initialization of the deduction environment and generating environment initial observation information; The multi-agent cooperative reinforcement learning training module is used for generating cooperative decision actions according to the environment initial observation information agents generated by the multi-agent emergency deduction environment module, updating the positions of the agents, optimizing resource allocation according to scheduling actions, calculating the rescue progress increment delta Pr of a single deduction step and updating the real-time rescue progress Pr; if the rescue progress Pr is less than 1, simulating abnormal spatial diffusion according to an abnormal diffusion rule, and updating the dangerous area distribution of the scene grid; calculating the total rewards R of the single deduction step according to the multi-agent collaborative rewards function, and using the total rewards R for strengthening the optimization feedback of the learning model; Repeating the steps until the rescue progress Pr is more than or equal to 1 or the number of dangerous areas exceeds a threshold value or the maximum deduction steps are reached, and stopping deduction.
  9. 9. A computer storage medium having stored thereon a computer program, which when executed by a processor performs the method of any of claims 1-7.
  10. 10. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor executing the program to implement the method of any of claims 1-7.

Description

Multi-agent collaborative deduction method and system for oilfield emergency scene based on visual anomaly perception Technical Field The invention relates to the technical field of high-risk industrial scene safety production and emergency treatment deduction, in particular to an oilfield emergency scene multi-agent collaborative deduction method and system based on visual anomaly perception. Background The emergency treatment deduction of the oil field is a key link of the safety production of the oil field, and the core of the emergency treatment deduction is to verify the feasibility of an emergency plan and optimize the cooperative strategy of a plurality of treatment subjects by simulating the development and treatment process of an emergency. The existing oilfield emergency deduction technology is limited by factors such as perceptibility, algorithm design, data support and the like, has a plurality of core defects, is difficult to match with the actual requirements of complex oilfield emergency scenes, and is specifically expressed as follows: the visual anomaly perception is unquantized, and the judgment precision is low, the conventional visual recognition method is adopted for emergency anomaly detection of the existing oil field, so that qualitative judgment of whether anomaly exists can be only realized, unquantized anomaly score and anomaly area thermodynamic diagram cannot be used for accurately representing anomaly degree and position, and basic information provided for emergency deduction is insufficient; The multi-agent cooperative non-targeted rewarding mechanism has poor deduction effect, namely the existing multi-agent emergency deduction does not design a differential cooperative rewarding function according to roles of different agents such as inspection, rescue, scheduling and the like, the agent decision does not have definite optimization and guidance, the rescue efficiency is low, the cooperativity is poor, and the deduction result is disjointed with the actual emergency disposal requirement; The oilfield emergency data is scarce, the model generalization capability is weak, the oilfield real emergency event sample collection is limited by factors such as safety, scenes and the like, the sample quantity is small, the labeling difficulty is high, the abnormality detection and reinforcement learning model generalization capability which simply depends on real data training is poor, and the adaptation to emergency scenes of different oilfield is difficult; Abnormal perception and multi-agent deduction are disjoint, and a closed-loop mechanism is not provided, wherein the prior art does not realize linkage triggering of a visual abnormal perception result and multi-agent emergency deduction, perception information such as abnormal positions, types and the like cannot be directly used as initialization parameters of a deduction environment, and the perception and deduction form an 'information island', so that the deduction is not enough in reality and pertinence; the emergency scene simulation distortion and no dynamic abnormal diffusion mechanism exist, namely the existing emergency deduction is mainly static in setting for the development simulation of abnormal events, the abnormal space diffusion processes such as fire and leakage are not dynamically simulated according to the rescue progress, the time sensitivity of emergency treatment cannot be reflected, and the deduction result has low reference value. In summary, the existing oilfield emergency treatment deduction field has the problems of unquantized visual anomaly perception, low judgment precision, multi-agent cooperation non-targeted rewarding mechanism, poor deduction effect, disjoint anomaly perception and multi-agent deduction, no closed-loop mechanism, simulated distortion of emergency scenes, no dynamic anomaly diffusion mechanism and the like. Disclosure of Invention The method aims to solve the problems that visual anomaly perception is unquantized, judging precision is low, a multi-agent cooperative non-targeted rewarding mechanism is poor in deduction effect, anomaly perception and multi-agent deduction are disjointed, a closed-loop mechanism is not available, emergency scene simulation distortion is avoided, a dynamic anomaly diffusion mechanism is not available in the field of oilfield emergency treatment deduction in the prior art. The invention provides a multi-agent collaborative deduction method and a system for an oilfield emergency scene based on visual anomaly perception, which are realized by the following technical scheme for solving the technical problems: The invention provides an oilfield emergency scene multi-agent collaborative deduction method based on visual anomaly perception, which comprises the following steps: step 1, configuring data parameters in a visual interface, constructing synthetic data based on the selected data parameters to generate and process double-source data with real data, p