Well site hazard source visual identification and control method and system

Abstract

The invention relates to the technical field of industrial safety and risk management, in particular to a well site hazard source visual identification and control method and system. The method comprises the following steps: acquiring multi-source data of a well site operation environment, generating a standardized image through preprocessing and coordinate correction, dividing semantic function sub-regions according to a spatial position and a safety rule, extracting multi-dimensional state vectors in the sub-regions, and fusing the multi-dimensional state vectors into a global semantic state matrix; personnel and equipment are continuously tracked, tracks of the personnel and the equipment are divided into behavior segments to construct multi-dimensional time sequence characteristics, and dynamic modeling of potential dangerous behaviors is realized by establishing a space-time interaction network and a behavior chain; high-risk fragments are extracted based on the behavior chain, and a risk state evolution model is constructed to quantify risks and identify hidden risks; generating a self-adaptive early warning and intelligent cooperative control strategy according to the dynamically calculated risk threshold and priority, and feeding back a control effect to the system to form a closed-loop optimization mechanism; according to the invention, the initiative and reliability of well site safety management are improved.

Description

Technical Field

[0001] This invention relates to the field of industrial safety and risk management technology, specifically to a method and system for visual identification and control of well site hazards. Background Technology

[0002] With the continuous improvement of industrial safety production standards and the rapid development of intelligent technologies, the safety management of high-risk operating environments such as oil and gas well sites is constantly moving towards digitalization and intelligence; the advancement of technologies such as computer vision, multi-sensor fusion, behavior recognition and risk modeling has provided a wealth of means to achieve in-depth perception and understanding of operating scenarios.

[0003] Chinese invention patent application CN118298496A discloses a method for identifying hazards in construction areas based on video and semantics, including: S1, installing smart cameras in the construction area to ensure coverage of the entire construction area; S2, collecting video data through the smart cameras, including the behavior of construction workers and environmental information of the construction site; S3, performing image recognition processing on the collected video data based on the Faster R-CNN algorithm to detect and identify hazardous behaviors; S4, establishing an ontology semantic network based on unsafe behaviors, and performing semantic reasoning on the image recognition results according to construction risk rules to identify hazardous behaviors.

[0004] Well site operating environments are characterized by clearly defined areas, diverse equipment types, frequent human-machine interactions, and dynamic correlations of risk factors. This requires safety management systems to integrate multi-source heterogeneous information, understand the semantics of complex scenarios, analyze temporal behavior patterns, and achieve early prediction and closed-loop control of risks. Current technological advancements have laid a solid foundation for real-time monitoring and analysis of personnel behavior, equipment status, and environmental parameters. They have also provided clear research directions and technical requirements for further developing comprehensive active safety systems that integrate semantic understanding, behavioral chain analysis, risk evolution modeling, and intelligent control. Summary of the Invention

[0005] The purpose of this invention is to address the problems existing in the background technology by proposing a visual identification and control method and system for well site hazards.

[0006] The technical solution of this invention: a method for visual identification and control of well site hazards, comprising the following specific implementation steps: S1. Collect multi-view video and sensor data of the well site, perform preprocessing and coordinate correction to generate standardized images; divide the well site into several semantic functional sub-regions according to spatial location, operation density and safety rules; extract multi-dimensional state vectors of personnel, equipment and environment in each semantic functional sub-region and fuse them into a global semantic state matrix; S2. Continuously track personnel and equipment within the well site, divide their trajectories into behavioral segments, and extract multi-dimensional temporal features; construct a multi-object spatiotemporal interaction network and integrate the behavioral segments into object behavior chains; S3. Select high-risk behavior segments from the object's behavior chain and construct a multi-dimensional hazard feature vector; map the high-risk segments to the hazard potential space, combine historical behavior and environmental factors to predict the hazard development trend through a dynamic model, and form a hazard potential evolution matrix; calculate the joint risk degree based on the evolution matrix and the interaction network, identify hidden risks and classify hazard levels; S4. Dynamically calculate risk thresholds and trigger graded early warnings based on real-time hazard levels; calculate hazard priorities by combining hazard levels, hidden risks, and evolution trends; generate intelligent collaborative control strategies based on early warning levels and hazard priorities; and feed back the execution effects of control actions to the semantic state matrix and hazard evolution matrix to adjust risk thresholds and control strategies, thereby achieving closed-loop optimization.

[0007] Preferably, in step S1, the preprocessing and coordinate correction to generate a standardized image specifically involves: The original video frames captured by each camera are subjected to noise filtering, perspective correction, and illumination normalization to obtain intermediate images; By using the pre-calibrated intrinsic and extrinsic parameter matrices of each camera, the pixel coordinates of each intermediate image are mapped to the global coordinate system of the well site to obtain a standardized image.

[0008] Preferably, in step S1, extracting the multidimensional state vectors of personnel, equipment, and environment within each semantic function sub-region specifically includes: The system counts the number of personnel identified through visual detection and tracking within a sub-area; it acquires sensor data or visual features of equipment within the sub-area and determines equipment operating status indicators through a state mapping function. Obtain environmental parameters such as temperature, humidity, and smoke concentration within the sub-region; Calculate the interaction between people and equipment, and between people and environmental factors; Interaction metrics include the duration and frequency of people's proximity to high-risk equipment.

[0009] Preferably, in step S2, the specific basis for dividing the trajectory into behavioral segments includes: When a change in the action code of an object is detected to exceed a preset action change threshold, it is classified. When an object's speed change is detected to exceed a preset speed fluctuation range, it is divided into segments. The classification is made when a significant change in the interaction distance between the object and high-risk equipment or other personnel is detected.

[0010] Preferably, in step S2, constructing the multi-object spatiotemporal interaction network specifically involves: Nodes are defined by behavior segments, and edges are defined by the interaction relationships between behavior segments. The weight of each edge is determined by the interaction type weight, the environmental risk correction factor, and the average spatial distance between the objects corresponding to the behavioral fragments.

[0011] Preferably, in step S3, the conditions for screening high-risk behavioral segments include: The object performed a high-risk action within the behavioral segment; the object's average spatial location was close to a high-risk facility; and environmental indicators exceeded the safe range during the time period in which the behavioral segment took place. The risk level of a high-risk behavior segment is calculated by weighted summation of action risk factors, spatial proximity risk factors, and environmental risk factors. Among them, the action hazard factor is determined based on the frequency and weight of actions within the segment, the spatial proximity hazard factor is determined based on the average distance to the hazardous facility, and the environmental hazard factor is determined based on the quantitative value and corresponding weight of each environmental indicator.

[0012] Preferably, in step S4, the dynamic calculation of the risk threshold specifically involves: For each object or area, the risk is assessed based on the basic safety threshold, the recent average risk level, the magnitude of risk fluctuations, environmental factors, and the historical cumulative risk value. The current risk threshold is obtained by weighted summation and used to determine whether the real-time hazard level triggers an early warning.

[0013] Preferably, in step S4, calculating the hazard priority specifically involves: For each object or area, its real-time hazard level, the slope of the hazard situation evolution trend, the object's contribution to the overall hidden risk, and the risk level of its associated object set are calculated. The danger priority score is calculated by weighted summation, and the priority is divided into high, medium and low levels based on the score.

[0014] Preferably, in step S4, generating the intelligent collaborative control strategy specifically involves: Based on the warning level and hazard priority, the intelligent control strategy generation function is invoked to generate a set of control actions for a single object or area; For several high-risk objects that have interactive relationships, a unified set of group collaborative control strategies is generated based on the weight of each object in the group through a collaborative control strategy generation function. Control actions include audible and visual alarms, mechanical shutdown, and personnel evacuation.

[0015] The technical solution of the present invention: a visual recognition and control system for well site hazards, used to execute the visual recognition and control method for well site hazards as described in claim 1, comprising: The scene semantic partitioning and state mapping module is used to collect multi-source data from the well site, perform preprocessing and coordinate correction, divide semantic functional sub-regions and extract multi-dimensional state vectors of each sub-region, and generate and update the global semantic state matrix. The danger situation evolution analysis module is used to continuously track objects and divide trajectories based on the global semantic state matrix, and to construct multi-dimensional temporal features, spatiotemporal interaction networks and behavior chains of behavioral fragments; The risk assessment and priority management module is used to screen high-risk segments from the behavioral chain, build a dangerous situation evolution model to quantify risks and predict trends, and identify potential high-risk object groups or dangerous areas. The intelligent control and closed-loop optimization module is used to dynamically calculate risk thresholds and object priorities, generate and execute adaptive early warning and intelligent collaborative control strategies, and adjust system parameters based on feedback information of execution results to achieve closed-loop optimization.

[0016] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects: This invention designs a visual recognition and control method and system for well site hazards. First, by fusing multi-view visual data and sensor information, and dynamically semantically partitioning and mapping the operational scene, the system achieves a high-precision, structured understanding of the complex well site environment. This overcomes the limitations of traditional monitoring methods, which suffer from a single perspective and isolated information, providing a reliable and semantically rich input foundation for subsequent analysis. Second, based on a temporal behavior chain and multi-object interaction network constructed through continuous behavior tracking, the system can deeply capture subtle and dynamic interaction patterns between personnel and equipment, enabling early identification and correlation analysis of individual abnormal behaviors and potential group hazards, significantly improving hazard identification and control. The system ensures the timeliness and accuracy of risk detection; furthermore, by establishing a risk situation evolution model and integrating a hidden risk identification mechanism, the system can not only quantitatively assess real-time risks but also predict the dynamic development trend of risks, achieving a leap from passive alarm to proactive prediction; finally, relying on adaptive threshold setting, multi-objective priority assessment, and intelligent collaborative control strategy generation, the system can perform differentiated and precise interventions for different levels of risks, and continuously optimize early warning thresholds and control strategies in conjunction with a closed-loop feedback mechanism, thereby comprehensively improving the initiative, intelligence level, and overall reliability of well site safety management, forming a proactive safety protection system that can evolve and continuously improve itself. Attached Figure Description

[0017] Figure 1 This is a flowchart of a method for visual identification and control of well site hazards proposed in this invention; Figure 2 This is a system architecture diagram of a well site hazard source visual recognition and control system proposed in this invention. Detailed Implementation

[0018] Example 1, as Figure 1 As shown, the present invention proposes a visual identification and control method for well site hazards, the specific implementation steps of which are as follows: S1. By fusing multi-view video and sensor data, the complex working environment of the well site is divided into semantic functional sub-regions, and multi-dimensional working state vectors of each sub-region are extracted. Combined with a dynamic update mechanism, a continuous, high-dimensional global semantic state matrix is ​​formed, providing high-precision and quantifiable input for subsequent hazard identification and behavior analysis. The specific implementation process is as follows: S11. Deploy multi-view cameras and sensors in key areas of the well site to collect personnel, equipment, and environmental parameters. Perform noise reduction, perspective correction, and illumination normalization on video frames. Unify data from different perspectives to the global coordinate system of the well site to generate analyzable standardized images, providing reliable input for semantic partitioning. Specifically: Multi-view high-definition cameras are deployed at the well site, including but not limited to key areas such as the wellhead operation area, power equipment area, hazardous chemical storage area, and personnel access area. At the same time, sensors are deployed on key equipment and environmental parameters to collect information such as temperature, humidity, smoke concentration, and equipment vibration. The camera's field of view coverage ensures no blind spots, and sensor data and video stream are collected synchronously. The captured video Standardization is performed to generate analyzable data in a unified coordinate system: ; in, Represents the original video frame pixel matrix at time t; 'c' represents the pixel coordinates in two-dimensional space, corresponding to the row and column positions of the camera image; 'c' represents the color channel, such as RGB three-channel or grayscale; 'v' represents the camera number, distinguishing multi-view video sources. This represents a standardized image after preprocessing and coordinate correction. The video frame preprocessing function includes noise filtering, perspective / distortion correction, illumination normalization, and frame synchronization; H represents the camera intrinsic and extrinsic parameter calibration matrix, which obtains the camera's focal length, rotation matrix, translation matrix, etc., and is used to map the data of each camera to the well site global coordinate system. S12. Based on the standardized images and well site spatial coordinates from step S11, the well site is divided into several functional sub-zones, such as the wellhead operation area, power equipment area, hazardous chemical storage area, and personnel access area. These sub-zones are formed based on spatial location, operation frequency, and safety rules, creating dynamic semantic partitions to provide context for subsequent state quantification. Specifically: The well site is divided into N functional sub-regions. Each sub-area corresponds to a different work type, such as wellhead operation area S1, power equipment area S2, hazardous chemical storage area S3, personnel passage area S4, etc. The division is based on, but is not limited to: spatial location L (two-dimensional plane coordinates); work density D. i (Average activity of personnel and equipment per unit time); Work safety regulations R i (Safe distance, work sequence, and hazard proximity threshold); Define a semantic partitioning function: ; ; in, This represents the pixel set of the i-th functional sub-region, which divides the well site according to operation type, equipment function, and safety rules. Each sub-region can be mapped to a video screen or global coordinate space. represents the semantic partitioning mapping function; L represents the well site global coordinate matrix, which represents the position of each spatial point in the well site. It is obtained from camera calibration and field measurement and is used to map pixel coordinates to actual spatial positions. This indicates the activity density of a sub-region, representing the amount of activity within that sub-region per unit time. Sub-region The number of people at time t; Sub-region The number of devices at time t; Indicates the length of the statistical time window; This represents the set of safety and operational rules corresponding to the sub-area, including but not limited to safety distances, work sequences, and hazard proximity thresholds. S13. Extract multi-dimensional operational state vectors within each sub-region, including personnel numbers, equipment operating status, environmental indicators, and personnel-equipment-environment interaction metrics, to achieve cross-modal state representation and provide quantifiable high-level semantic features for hazard potential analysis. That is, for each sub-region... Extracting multidimensional state vectors : ; ; ; in, This represents the operational state vector of the i-th sub-region at time t; This indicates an estimate of the number of people within a sub-region; This represents the probability of a person being present at time t for pixel (x, y). Indicates the operating status of equipment within the sub-area; This represents device sensor data or visual features; This represents a status mapping function that maps raw data to device operating status indicators (such as normal, abnormal, and warning). This represents a vector of environmental indicators for a sub-region, including but not limited to environmental parameters such as temperature, humidity, and smoke concentration. It represents the measurement of personnel-equipment-environment interactions, such as the duration of personnel approaching high-risk equipment, the frequency of actions, and the frequency of environmental anomaly triggers; S14. The state vectors of each sub-region are combined to form a global semantic state matrix. A smooth update mechanism is used to combine historical states with new observations to ensure the temporal continuity and dynamic stability of the states. This provides reliable and continuous input for subsequent behavioral fragment extraction and hazard evolution analysis, achieving dynamic-continuous-high-dimensional semantic mapping. Specifically: Define the global semantic state matrix : ; Considering personnel movement, equipment status fluctuations, and environmental changes, a smooth update mechanism is introduced: ; in, The global semantic state matrix is ​​composed of the state vectors of each sub-region. This represents the new observation state vector of the sub-region at time t+1; This represents the state smoothing coefficient, with a value range of [0,1]. This represents the updated value of the sub-region state vector, which is a weighted fusion of historical states and new observation data to form a continuous time-series state.

[0019] S2. Based on the semantic state matrix of step S1, continuous behavior analysis is performed on personnel, equipment, and the environment within the well site. The object trajectory is divided into behavior segments, multi-dimensional temporal features are extracted, and a multi-object interaction network is constructed to form a dynamic behavior chain. This enables continuous capture and early identification of potential dangerous behaviors, providing core temporal input for hazard situation analysis. The specific implementation process is as follows: S21. Using the semantic sub-region information and visual data from step S1, perform multi-object detection to identify the position, movement, and speed of personnel and equipment. Then, continuously track these objects using Kalman or particle filtering to generate stable trajectories. Simultaneously, introduce semantic constraints to address occlusion and overlap issues, ensuring trajectory continuity. Specifically: Utilizing the semantic sub-region in step S1 With video information The personnel and key equipment within each sub-area are monitored, and the monitoring function formula is as follows: ; The detected objects are used to generate stable trajectories using a multi-object tracking algorithm. ; For objects that are occluded or temporarily disappear, a trajectory prediction model is introduced. This ensures the continuity of the trajectory; in, This represents the detection result of the j-th action object at time t, including position, speed, and action type; Represents the set of continuous trajectories of the j-th object; Indicates the current speed of the object; This indicates a trajectory prediction point, used for trajectory completion in cases of occlusion or temporary disappearance. S22. The continuous trajectory is divided into short-term behavior segments based on changes in action, speed fluctuations, and changes in interaction distance. A multi-dimensional temporal feature vector is constructed for each segment, including average position, speed, action type, interaction distance, and environmental state. This achieves high-precision characterization of potential dangerous behavioral trends and provides a foundation for constructing behavior chains. Specifically: Continuous trajectory Short-term behavioral segments are categorized based on action, speed, or spatial interaction. : ; The criteria for segmentation include, but are not limited to: Action change threshold: When the action code change exceeds the threshold, the segment is cut. Segmentation occurs when speed fluctuations exceed a preset range; Interaction distance: New fragments are generated when an object approaches high-risk equipment or other personnel. For each behavioral segment Constructing high-dimensional feature vectors: ; in, This represents the m-th behavior segment of the j-th object, a subsequence of a continuous trajectory; This represents the total number of behavior fragments in object j; Describing behavioral fragments eigenvectors; Indicates average spatial location; Indicates average speed; This indicates the average code for the action type; Indicates the average interaction distance; Indicates environmental conditions (temperature, humidity, smoke concentration, etc.); Indicates the frequency of actions; Indicates the duration of a segment; S23. Construct a multi-object spatiotemporal interaction network based on behavioral fragments, using fragments as nodes and weighting the distance, actions, and environment between objects to form edge weights. Analyze complex behavioral patterns, identify potentially dangerous behaviors such as multiple people gathering and people approaching high-risk equipment, and achieve dynamic modeling of multi-object interactions in complex scenarios, providing a logical basis for behavior chain generation. Specifically: Using behavioral fragments as nodes, a spatiotemporal interaction network is constructed. : ; ; ; ; High-risk complex behavioral patterns can be identified through network analysis, such as: multiple people gathering near hazardous chemical areas; people approaching high-speed operating equipment while the equipment vibrates abnormally; and abnormal frequency of continuous actions in unsafe environmental conditions. in, This represents a spatiotemporal interaction network, where node V represents a behavior segment and edge E represents the interaction relationship between segments. Represents a node and Interaction weights; Indicates the interaction type weight (e.g., people are more likely to be closer to high-speed equipment). This indicates an environmental risk correction factor (such as increased weighting for high temperature and high smog conditions). This represents the average spatial distance between objects; This represents the distance standardization coefficient; S24. Integrate the behavioral fragments and their interaction networks to generate an object behavior chain, and update the behavior chain in real time through a smooth dynamic update mechanism, taking into account both historical and new actions, to achieve continuous and predictable representation of potentially dangerous behaviors. Specifically: Integrate object fragments and their interaction networks to form a continuous chain of behaviors: ; Each chain of actions It includes the object's own time sequence segments and key interaction information with other objects, comprehensively describing the dynamic behavior of an individual in the working environment; Behavioral chains are updated over time, combining historical behavior with new observational data: ; in, This represents the complete behavior chain of object j, including all behavior segments and their interaction information; This represents the set of interactions between object j and other object k, with each edge representing a single interaction. Characterizes the strength of interactions or potential dangerous associations between objects; This represents the complete behavior chain of object j after the update at time t+1; This represents the historical behavior weighting smoothing coefficient, with a value range of [0,1]. It places greater emphasis on historical behavior and reduces the impact of short-term noise. It places greater emphasis on the latest behavioral snippets and responds quickly to changes in behavior; This represents the newly collected behavior chain segment and interaction information of object j at time t+1.

[0020] S3. Based on the behavioral chain and multi-object interaction network generated in step S2, high-risk behavioral segments are extracted, a dangerous situation evolution model is constructed, potential risks are quantified and hidden dangers are identified, dynamic visualization and closed-loop safety feedback are achieved, providing a basis for proactive early warning and safety control for well site operations. The specific implementation process is as follows: S31. Screen potential high-risk segments from the behavioral chain, construct a multi-dimensional hazard feature vector by combining action type, spatial proximity, and environmental factors, and quantify the hazard level of the segments to provide input for subsequent evolutionary modeling. Specifically: From the behavioral chain Screening for potentially high-risk fragments The conditions are combined with the type of action of the object, its spatial location, and its proximity to high-risk facilities: ; ; in, This represents the fragment risk function, i.e., the risk of object j in fragment. The overall risk level; Representing fragments The set of surrounding objects or critical facilities that have potentially dangerous interactions with object j; This represents the average spatial distance between object j and object or facility k during the segment; To represent a small constant, preventing division by zero; , and The weighting coefficients for each risk dimension can be adjusted according to the actual situation at the well site, and the sum of their weighting coefficients is 1. and Each represents a segment Time range; This indicates the action code or frequency of high-risk actions of object j at time t. This represents the weighting coefficient for the action type; Representing fragments The set of environmental factors involved (such as temperature, smoke, humidity, vibration); Indicates that environmental factor f in the fragment Quantization value; The weighting coefficient of environmental factor f; Construct a multi-dimensional feature vector for each high-risk segment. : ; S32. High-risk segments are mapped to a hazard potential space. Combining historical behavior, temporal trends, and environmental factors, a dynamic model (such as an RNN or autoregressive model) is used to predict the hazard development trend, forming a multi-object spatiotemporal hazard evolution matrix, specifically: Mapping high-risk segments to the danger potential space: ; Construct a multi-object danger evolution matrix for all object fragments: ; in, This represents the danger potential vector of a segment, including information such as danger level, risk growth trend, and duration; Represents the cumulative state vector of an object's historical behavior chain; The hazard evolution modeling function can be represented by RNN, LSTM or AR model (autoregressive model); This represents a multi-object hazard evolution matrix, where rows represent objects and columns represent segments. S33. Based on the evolutionary matrix and interaction network, calculate the joint risk degree among objects, and identify potentially high-risk object groups or dangerous areas through clustering or community detection to achieve latent risk identification and hazard level classification, specifically: Calculate the joint risk degree among objects using behavioral chain interaction information and evolution matrix: ; Joint risk matrix Clustering is performed, and the risk levels are divided into low, medium, and high, with dynamic updates: ; in, This represents the joint risk level of object j and object k, reflecting the potential group danger under multi-object interaction; The weights of the interaction network for behavioral segments, i.e., nodes. and Interaction weights; This indicates an environmental correction factor, which strengthens the joint risk level under high-risk environments; and These represent the number of high-risk behavior segments for objects j and k, respectively. This represents the normalization coefficient, which standardizes the joint risk level, making the risks of different objects comparable. This indicates latent risk outcomes, including potentially high-risk groups or dangerous areas; This represents a clustering function used to identify potentially high-risk groups and regions. S34. Map the evolution matrix and latent risks to the well site space to generate a heat map, visually displaying high-risk areas. Simultaneously, feed the risk information back to the semantic state matrix and behavior chain monitoring module, establishing the foundation for closed-loop safety management and proactive control strategies. Specifically: Evolutionary matrix and hidden risks Mapped to the well site space, a heat map of high-risk areas is generated: ; The implicit risk information is fed back to the semantic state matrix in step S1 and the behavior chain monitoring mechanism in step S2: ; in, The position mapping weight function represents the segment. The degree of influence on the well site grid point (x,y); This represents the cumulative hazard value at grid point (x,y), which can be used to generate two-dimensional or three-dimensional heat maps; J represents the total number of monitored work objects in the well site; This represents the updated semantic state matrix; This represents the feedback weight coefficient, which adjusts the degree of influence of risk feedback on the semantic state.

[0021] S4. By transforming the hazardous situation evolution matrix and implicit risk information from step S3 into real-time early warning and intelligent control strategies, a dynamic closed loop for well site safety management is achieved. The innovation lies in adaptive thresholds, multi-object priority assessment, intelligent collaborative control, and closed-loop optimization mechanisms, which improve the accuracy and initiative of safety management. The specific implementation process is as follows: S41. Based on the average and fluctuating risk situation, environmental factors, and historical risks, dynamically calculate the risk threshold for each object or area to achieve multi-level adaptive early warning. Different levels of early warning are triggered by comparing the risk level with the threshold, ensuring immediate response to high-risk behaviors. Specifically: For each work object or area i, a dynamic threshold is defined based on the hazard potential evolution matrix and implicit risk information: ; Define real-time hazard level : ; ; ; Real-time risk level Make a judgment: ; Level 1 High Risk: Immediate alarm, restriction of personnel access, automatic shutdown or locking of critical equipment; Level 2 Medium Risk: Remind operators to pay attention and increase the frequency of inspections or monitoring; Level 3 Low Risk: Record abnormal behavior and incorporate it into the behavior chain update; in, This represents the average dangerous potential value of object i within the time window [tT,t]; T represents the length of the selected historical time period. Indicates the standard deviation of the dangerous situation; , , and The weight coefficients can be adjusted through learning from historical data, expert experience, or reinforcement learning models, and their sum is 1. The dynamic risk threshold of object or region i at time t is used to determine whether an early warning is triggered. It is calculated by combining the basic threshold, the average value of recent dangerous situations, the fluctuation range, environmental factors, and historical risk accumulation. This represents the basic safety threshold, the default minimum risk limit, which is set by well site safety regulations or management experience. The variance represents the dangerous situation and reflects the volatility or instability of behavior; This represents the cumulative historical risk value, including but not limited to records of past accidents and abnormal operations. , , and This represents the weight coefficient of each factor in the dynamic threshold calculation. It is obtained through historical data analysis, expert experience, or machine learning training, and the sum of its weight coefficients is 1. Indicates the real-time danger level of an object or area at time t; Indicates the warning status level (level 1-3); S42. Combining the object's hazard level, implicit risk matrix, and hazard evolution trend, calculate priority scores to dynamically classify and manage high-risk objects and areas, mapping hazard priorities to high, medium, and low-level strategies to provide a basis for generating subsequent control actions. Specifically: Based on risk level, hidden risk matrix, and evolution trend, high-risk objects or areas are prioritized: ; Priorities are divided into three categories: high, medium, and low, each corresponding to a different control response: High priority: Take immediate proactive intervention (shutdown, isolation, evacuation); Medium priority: Increase inspection frequency, send reminder messages, restrict specific operations; Low priority: Continuous monitoring, incorporating historical behavior chain data, and waiting for the risk to evolve; in, Represents the danger priority score of an object or area; It represents the slope of the evolution trend of the danger situation, reflecting the rate of change of the danger level, and is fitted to the time series slope by historical danger situation data; It represents the contribution of an object to the overall implicit risk, calculated by combining the object's behavioral chain and interaction network within the group; , , and The priority calculation of each factor's weight is set by historical data analysis or expert experience and can be dynamically adjusted. The sum of its weight coefficients is 1. This refers to the set of neighboring or related objects of object i, that is, other objects that are potentially associated with object i in terms of behavior or space; S43. Generate intelligent control actions based on the warning level and hazard priority, including audible and visual alarms, mechanical shutdown, and personnel evacuation. Simultaneously, generate collaborative control strategies for highly correlated objects to achieve unified, safe, and optimal control of high-risk group behaviors, enhancing system intelligence. Specifically: Based on warning level and priority Generate intelligent control actions for single objects and regions: ; For multiple highly correlated risk objects, generate a collaborative control strategy: ; in, Represents the set of intelligent control actions for an object or region at time t; This indicates the real-time behavior status of other objects within region i; A set of collaborative control strategies representing a high-risk group; It indicates the weight of an object in group collaborative control, reflecting the object's contribution or criticality to the overall risk; This represents the intelligent control strategy generation function, which can be implemented by a rule engine, reinforcement learning, or deep learning model. It dynamically generates the optimal control action based on the input, achieving adaptive and intelligent control. S44. The effects of control actions are fed back to the semantic state matrix and the hazard evolution matrix in real time. Feedback is used to adjust future thresholds and control strategies, achieving a prediction-control-feedback closed loop. This enables the system to adaptively optimize hazard prediction accuracy and control effectiveness, improving the initiative and reliability of well site safety management. Specifically: After the control action is executed, provide feedback on the actual effect: ; ; Adjust future risk thresholds and control strategies based on actual feedback: ; ; This embodiment can utilize reinforcement learning or online optimization algorithms to optimize the accuracy of risk prediction and the effectiveness of control strategy execution; forming a prediction-control-feedback closed loop to achieve proactive safety management. in, and These represent the updated semantic state matrix and the danger potential evolution matrix, respectively. and These represent the actual impact of the control action on the object's behavioral state and dangerous situation, respectively, calculated using sensor data and behavioral monitoring. and This represents the feedback weight coefficient, which controls the speed of closed-loop optimization; This represents the adaptive adjustment coefficient for the risk threshold; This represents the learning rate, a parameter of the control strategy. This represents the adjusted risk threshold of object or region i at the next time t+1, used for the next round of risk comparison and early warning triggering; This represents the actual observed risk value of object or area i, which is collected in real time through sensors, behavior monitoring, hazard situation matrices, etc., providing feedback information and reflecting the true risk situation; This represents the risk value predicted by the system for object or region i. This represents the updated value of the j-th parameter in the control strategy at the next time step t+1; This represents the current value of the j-th parameter in the control strategy, i.e., the weight at the current moment, such as when calculating the weight for priority. , , and ; Let represent the gradient of the loss function corresponding to the j-th control policy parameter. The difference between the control policy output and the expected control effect is calculated and obtained through reinforcement learning, gradient descent, or online optimization.

[0022] Example 2, as Figure 2 As shown, the present invention proposes a well site hazard source visual recognition and control system, which is used to execute a well site hazard source visual recognition and control method proposed in Embodiment 1, including: a scene semantic partitioning and state mapping module, a hazard situation evolution analysis module, a risk assessment and priority management module, and an intelligent control and closed-loop optimization module.

[0023] The scene semantic partitioning and state mapping module is used for high-precision semantic understanding and state mapping of well site operation scenarios. It collects panoramic images and point cloud data of the well site through high-definition cameras, LiDAR, and 3D depth sensing equipment, and performs partitioning and identification of different operation areas, equipment, and personnel, while mapping the state of each object in real time. The module has a built-in multi-scale visual recognition algorithm that can dynamically generate semantic partitioning maps, mark the operation objects, dangerous areas, and key equipment locations, and update the object behavior status in real time. The output of this module provides basic spatial information and object state input for subsequent hazard situation analysis, realizing a refined understanding of the well site operation environment. The hazard situation evolution analysis module is used to construct a hazard situation evolution matrix and perform latent risk analysis. It receives semantic partitioning and object status information, dynamically analyzes the action sequences of each work object, equipment operating status, and personnel behavior chains, calculates potential hazards, and generates hazard evolution trends. The module can also perform group behavior analysis, identify potential risks arising from multi-object interactions, and perform short-term prediction and evolution modeling of hazard situations by fusing historical behavioral data with environmental factors. The output is a hazard situation matrix and evolution trend for an object or region, providing input data for risk assessment and intelligent control. The risk assessment and priority management module transforms hazard situation analysis results into actionable risk assessment and priority strategies. It generates dynamic risk thresholds and hazard priorities by comprehensively evaluating the real-time hazard level, latent risk contribution, environmental factors, and historical risk data of each object or area through multi-factor analysis. The module implements multi-level hierarchical early warning, classifying objects or areas into high, medium, and low-risk categories based on priority and outputting the warning level, providing precise control decision-making basis for the intelligent control module. This module ensures that the system can adopt differentiated management strategies for different hazardous objects, achieving refined safety control. The intelligent control and closed-loop optimization module is used to generate, execute, and optimize intelligent control strategies. Based on risk priority and early warning level, combined with the characteristics of the work object group and environmental information, it generates single-object and group collaborative control actions, including audible and visual alarms, mechanical shutdown, area isolation, and personnel evacuation instructions. The module supports real-time feedback, inputting control execution effects, actual hazard changes, and behavior adjustment information into the system to achieve closed-loop optimization. Through adaptive learning and strategy optimization algorithms, the system dynamically adjusts risk thresholds, control actions, and strategy parameters to improve prediction accuracy and control effectiveness, achieving proactive and intelligent well site hazard management.

[0024] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for visual identification and control of well site hazards, characterized in that, The specific implementation steps include the following: S1. Collect multi-view video and sensor data of the well site, perform preprocessing and coordinate correction to generate standardized images; divide the well site into several semantic functional sub-regions according to spatial location, operation density and safety rules; extract multi-dimensional state vectors of personnel, equipment and environment in each semantic functional sub-region and fuse them into a global semantic state matrix; S2. Continuously track personnel and equipment within the well site, divide their trajectories into behavioral segments, and extract multi-dimensional temporal features; construct a multi-object spatiotemporal interaction network and integrate the behavioral segments into object behavior chains; S3. Select high-risk behavior segments from the object's behavior chain and construct a multi-dimensional hazard feature vector; map the high-risk segments to the hazard potential space, combine historical behavior and environmental factors to predict the hazard development trend through a dynamic model, and form a hazard potential evolution matrix; calculate the joint risk degree based on the evolution matrix and the interaction network, identify hidden risks and classify hazard levels; S4. Dynamically calculate risk thresholds and trigger graded early warnings based on real-time hazard levels; calculate hazard priority by combining hazard level, hidden risks, and evolution trends; generate intelligent collaborative control strategies based on early warning levels and hazard priorities. The effects of control actions are fed back to the semantic state matrix and the hazard evolution matrix to adjust risk thresholds and control strategies, thereby achieving closed-loop optimization.

2. The method for visual identification and control of well site hazards according to claim 1, characterized in that, In step S1, preprocessing and coordinate correction are performed to generate a standardized image, specifically as follows: The original video frames captured by each camera are subjected to noise filtering, perspective correction, and illumination normalization to obtain intermediate images; By using the pre-calibrated intrinsic and extrinsic parameter matrices of each camera, the pixel coordinates of each intermediate image are mapped to the global coordinate system of the well site to obtain a standardized image.

3. The method for visual identification and control of well site hazards according to claim 2, characterized in that, Step S1, extracting the multidimensional state vectors of personnel, equipment, and environment within each semantic function sub-region, specifically includes: The system counts the number of personnel identified through visual detection and tracking within a sub-area; it acquires sensor data or visual features of equipment within the sub-area and determines equipment operating status indicators through a state mapping function. Obtain environmental parameters such as temperature, humidity, and smoke concentration within the sub-region; Calculate the interaction between people and equipment, and between people and environmental factors; Interaction metrics include the duration and frequency of people's proximity to high-risk equipment.

4. The method for visual identification and control of well site hazards according to claim 3, characterized in that, In step S2, the specific basis for dividing its trajectory into behavioral segments includes: When a change in the action code of an object is detected to exceed a preset action change threshold, it is classified. When an object's speed change is detected to exceed a preset speed fluctuation range, it is divided into segments. The classification is made when a significant change in the interaction distance between the object and high-risk equipment or other personnel is detected.

5. The method for visual identification and control of well site hazards according to claim 4, characterized in that, In step S2, the construction of the multi-object spatiotemporal interaction network specifically involves: Nodes are defined by behavior segments, and edges are defined by the interaction relationships between behavior segments. The weight of each edge is determined by the interaction type weight, the environmental risk correction factor, and the average spatial distance between the objects corresponding to the behavioral fragments.

6. The method for visual identification and control of well site hazards according to claim 5, characterized in that, In step S3, the criteria for selecting high-risk behavioral segments include: The object performed a high-risk action within the behavioral segment; the object's average spatial location was close to a high-risk facility; and environmental indicators exceeded the safe range during the time period in which the behavioral segment took place. The risk level of a high-risk behavior segment is calculated by weighted summation of action risk factors, spatial proximity risk factors, and environmental risk factors. Among them, the action hazard factor is determined based on the frequency and weight of actions within the segment, the spatial proximity hazard factor is determined based on the average distance to the hazardous facility, and the environmental hazard factor is determined based on the quantitative value and corresponding weight of each environmental indicator.

7. The method for visual identification and control of well site hazards according to claim 6, characterized in that, In step S4, the dynamic calculation of the risk threshold is specifically as follows: For each object or area, the risk is assessed based on the basic safety threshold, the recent average risk level, the magnitude of risk fluctuations, environmental factors, and the historical cumulative risk value. The current risk threshold is obtained by weighted summation and used to determine whether the real-time hazard level triggers an early warning.

8. The method for visual identification and control of well site hazards according to claim 7, characterized in that, In step S4, the hazard priority is calculated as follows: For each object or area, its real-time hazard level, the slope of the hazard situation evolution trend, the object's contribution to the overall hidden risk, and the risk level of its associated object set are calculated. The danger priority score is calculated by weighted summation, and the priority is divided into high, medium and low levels based on the score.

9. The method for visual identification and control of well site hazards according to claim 8, characterized in that, In step S4, the intelligent collaborative control strategy is generated as follows: Based on the warning level and hazard priority, the intelligent control strategy generation function is invoked to generate a set of control actions for a single object or area; For several high-risk objects that have interactive relationships, a unified set of group collaborative control strategies is generated based on the weight of each object in the group through a collaborative control strategy generation function. Control actions include audible and visual alarms, mechanical shutdown, and personnel evacuation.

10. A well site hazard source visual recognition and control system, used to execute the well site hazard source visual recognition and control method according to any one of claims 1 to 9, characterized in that, include: The scene semantic partitioning and state mapping module is used to collect multi-source data from the well site, perform preprocessing and coordinate correction, divide semantic functional sub-regions and extract multi-dimensional state vectors of each sub-region, and generate and update the global semantic state matrix. The danger situation evolution analysis module is used to continuously track objects and divide trajectories based on the global semantic state matrix, and to construct multi-dimensional temporal features, spatiotemporal interaction networks and behavior chains of behavioral fragments; The risk assessment and priority management module is used to screen high-risk segments from the behavioral chain, build a dangerous situation evolution model to quantify risks and predict trends, and identify potential high-risk object groups or dangerous areas. The intelligent control and closed-loop optimization module is used to dynamically calculate risk thresholds and object priorities, generate and execute adaptive early warning and intelligent collaborative control strategies, and adjust system parameters based on feedback information of execution results to achieve closed-loop optimization.

Images