Petroleum scene operator violation action detection method based on posture recognition

Abstract

The invention provides a petroleum scene operator violation action detection method based on posture recognition, and relates to the technical field of action recognition, and the method comprises the steps: obtaining a joint point position sequence, hazard source dynamic data and an operation process state identifier of an operator; performing time sequence evolution analysis based on the joint point position sequence to obtain a subsequent attitude evolution trend; establishing an operator prediction motion envelope and a hazard source prediction motion envelope according to a subsequent attitude evolution trend and hazard source dynamic data, and calculating a minimum distance evolution curve of the operator prediction motion envelope and the hazard source prediction motion envelope in a future time window to obtain a bidirectional interaction risk value representing a relative motion collision probability; and according to the bidirectional interaction risk value and the operation process state identifier, generating a violation judgment result containing a violation emergency degree and an early warning opportunity through a context-dependent risk weight rule. According to the invention, the prospective risk assessment of the dynamic hazard source can be realized, and the accuracy of violation detection and the timeliness of early warning are improved.

Description

Technical Field

[0001] This invention relates to action recognition technology, and more particularly to a method for detecting unauthorized actions of workers in oilfield scenarios based on posture recognition. Background Technology

[0002] Oilfield work sites are typically characterized by high risk, complex working environments, and numerous hazards. Workers face various safety risks during operations, including mechanical collisions, falls from heights, and electric shocks. However, most existing posture recognition methods are based on static posture determination, only able to identify the worker's current movement state and unable to predict or analyze subsequent movement trends. This results in significant warning delays, failing to provide sufficient response time before a hazard occurs. This passive detection mode often cannot prevent accidents in time when faced with rapidly moving hazards or sudden dangerous scenarios. Oilfield work sites also contain numerous dynamic hazards, such as moving machinery, hoisted objects, and transport vehicles. These hazards have complex relative motion relationships with workers, which current technologies cannot effectively identify and avoid. Summary of the Invention

[0003] This invention provides a method for detecting unauthorized actions of workers in oilfield scenarios based on posture recognition, which can solve the problems in the prior art.

[0004] A first aspect of this invention provides a method for detecting unauthorized actions of workers in oilfield scenarios based on posture recognition, comprising:

[0005] Acquire the sequence of key locations for workers, dynamic data of hazard sources, and status indicators of work processes;

[0006] Temporal evolution analysis is performed based on the joint position sequence to obtain the subsequent attitude evolution trend;

[0007] Based on the subsequent attitude evolution trend and the dynamic data of the hazard source, a predicted motion envelope for the worker and a predicted motion envelope for the hazard source are established. By calculating the minimum distance evolution curves of the predicted motion envelopes for the worker and the hazard source within a future time window, a two-way interactive risk value is obtained, which characterizes the probability of relative motion collision between the worker and the dynamic hazard source.

[0008] Based on the two-way interactive risk value and the work process status identifier, a violation judgment result is generated through context-dependent risk weight rules. The risk weight rules assign different risk levels to the same two-way interactive risk value according to the work process status identifier. The violation judgment result includes the urgency of the violation and the timing of the warning.

[0009] The steps for performing temporal evolution analysis based on the joint position sequence to obtain the subsequent attitude evolution trend include:

[0010] The actual posture is extracted from the joint position sequence, and an inertial posture is constructed assuming that active control has stopped. By comparing the actual posture with the inertial posture, the posture evolution is decomposed into intention-driven components and inertial constraint components.

[0011] The intent-driven components are mapped to a predefined job action semantic space to identify the currently executed action type and its completion degree, and the subsequent changes of the intent-driven components are predicted based on the evolution pattern of the action type.

[0012] Physical motion constraints and environmental contact constraints are applied to the inertial constraint components to generate predictions of subsequent changes in the inertial constraint components;

[0013] The fusion weights of the intention-driven component and the inertial constraint component are dynamically adjusted according to the prediction time length to obtain the joint position prediction sequence. The uncertainty range of each predicted position is marked based on the action semantic recognition confidence. The joint position prediction sequence and its uncertainty range are used as the subsequent attitude evolution trend.

[0014] The step of decomposing attitude evolution into intention-driven components and inertial constraint components by comparing the actual attitude with the inertial attitude includes:

[0015] Extract the joint position difference between adjacent moments from the joint position sequence, and calculate the actual motion velocity and actual motion acceleration of each joint;

[0016] Based on the current motion velocity of each joint, a velocity decay coefficient is applied to simulate the motion decay after the active control stops. The joint position at the next moment is calculated based on the decayed velocity, and the inertial motion velocity of each joint under the inertial attitude is obtained. The inertial acceleration is calculated based on the inertial motion velocity.

[0017] The difference between the actual motion velocity and the inertial motion velocity of each joint is calculated as the velocity domain intention component, and the difference between the actual motion acceleration and the inertial acceleration is calculated as the acceleration domain intention component.

[0018] The velocity domain intention component and the acceleration domain intention component are smoothed to obtain the intention drive component.

[0019] Subtracting the position change corresponding to the intention-driven component from the joint position sequence yields the inertial constraint component.

[0020] Based on the subsequent posture evolution trend and the dynamic data of the hazard source, the steps for establishing the predicted motion envelope of the worker and the predicted motion envelope of the hazard source include:

[0021] The motion direction vectors and motion determinism levels of the joints in the subsequent posture evolution trend are extracted as personnel motion features, and the motion paths and spatial occupancy ranges of the hazard sources in the dynamic data of the hazard sources are extracted as hazard source motion features;

[0022] The relative approach velocity between the movement characteristics of the personnel and the movement characteristics of the hazard source is calculated, the spatiotemporal coupling strength is constructed to quantify the relative motion trend, and the interaction moment when the spatial distance reaches a local minimum is identified;

[0023] Based on the motion determinism level, the initial envelope radius of each joint is determined to form an initial envelope. When the spatiotemporal coupling strength is greater than a set value, the envelope boundary is expanded towards the hazard source. The expansion range is proportional to the spatiotemporal coupling strength, thus obtaining the predicted motion envelope of the operator.

[0024] Based on the motion characteristics of the hazard source, a spatial envelope of the hazard source is constructed, and the envelope size is adjusted based on motion predictability to obtain the predicted motion envelope of the hazard source;

[0025] At the interaction time, when the spatial distance between the predicted motion envelope of the operator and the predicted motion envelope of the hazard source is less than a set distance threshold, the boundary of the predicted motion envelope of the operator in the direction of the hazard source is expanded and the predicted motion envelope of the hazard source is extended.

[0026] The steps for calculating the minimum distance evolution curves of the predicted motion envelope of the worker and the predicted motion envelope of the hazard source within a future time window to obtain the bidirectional interactive risk value include:

[0027] The spatial locations of personnel joints are extracted from the predicted motion envelope of the workers as key monitoring points, and hazardous boundary surfaces are extracted from the predicted motion envelope of the hazard sources. The minimum geometric distance between each key monitoring point and the hazardous boundary surface is calculated to obtain multiple location distance sequences.

[0028] Based on the movement characteristics of the personnel and the movement characteristics of the hazard source, a relative movement direction vector is calculated. When the relative movement direction vector points close together, a directional attenuation coefficient is applied to the distance sequence of the body parts; when the relative movement direction vector points far apart, a directional gain coefficient is applied to the distance sequence of the body parts.

[0029] Differential weights are assigned to the location distance sequences of different key monitoring points based on the spatiotemporal coupling strength, with higher weights for the corresponding location distance sequences corresponding to greater spatiotemporal coupling strength.

[0030] The distance sequences of each part after directional adjustment and weight allocation are weighted and fused to obtain the comprehensive minimum distance evolution curve; the bidirectional interaction risk value is calculated by weighting the reciprocal of the global minimum value of the comprehensive minimum distance evolution curve, the reciprocal of the time when the minimum value is reached, and the curve descent rate before that time.

[0031] Based on the bidirectional interaction risk value and the work process status identifier, a violation determination result is generated using context-dependent risk weighting rules. These risk weighting rules assign different risk levels to the same bidirectional interaction risk value based on the work process status identifier. The violation determination result includes steps such as determining the urgency of the violation and the timing of the warning.

[0032] Based on the work process status identifier, extract the work stage type, work step number, and execution time;

[0033] The baseline risk weight coefficient corresponding to the work stage type is queried from the preset work process-risk mapping rule base. The baseline risk weight coefficient is set differently according to the safety distance requirements of the work stage type. When the work step number corresponds to a high-risk operation step, an amplification factor is applied to the baseline risk weight coefficient for adjustment. The fatigue accumulation factor is calculated based on the execution time. The adjusted risk weight coefficient is corrected according to the fatigue accumulation factor to obtain the final risk weight coefficient.

[0034] The two-way interactive risk value is multiplied by the final risk weight coefficient to obtain the weighted risk value, and the weighted risk value is divided into different risk levels according to the preset risk level threshold.

[0035] The urgency of the violation is determined based on the risk level; the moment when the two-way interactive risk value reaches its peak within a future time window is extracted as the warning timing; the urgency of the violation and the warning timing are combined as the violation determination result.

[0036] The steps of calculating the fatigue accumulation factor based on the executed duration, and correcting the adjusted risk weight coefficient according to the fatigue accumulation factor to obtain the final risk weight coefficient include:

[0037] The initial fatigue factor is calculated based on the ratio of the executed duration to the preset fatigue threshold duration.

[0038] Extract the motion intensity change curve within the executed duration from the joint position sequence, calculate the cumulative duration ratio of high-intensity motions, and adjust the initial fatigue factor based on the cumulative duration ratio to obtain a comprehensive fatigue factor.

[0039] The overall fatigue factor is attenuated and corrected according to the rest period in the work process status indicator; the correction method is selected according to the numerical range of the overall fatigue factor. When the overall fatigue factor is lower than the preset fatigue segment threshold, linear correction is used, and when the overall fatigue factor is higher than the preset fatigue segment threshold, nonlinear accelerated correction is used.

[0040] Determine whether there is a conflict between the baseline risk weight coefficient corresponding to the operation stage type and the risk weight coefficient after adjustment of the high-risk step amplification coefficient. When the baseline risk weight coefficient is already in the high-risk range, limit the amplification factor of the high-risk step amplification coefficient.

[0041] The adjusted risk weight coefficient is combined with the revised comprehensive fatigue factor to obtain the final risk weight coefficient.

[0042] A second aspect of the present invention provides an electronic device, comprising:

[0043] processor;

[0044] Memory used to store processor-executable instructions;

[0045] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0046] A third aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0047] This invention establishes a predicted motion envelope for workers and a predicted motion envelope for hazards, and calculates a two-way interactive risk value. This enables a quantitative assessment of the relative motion relationship between workers and dynamic hazards, overcoming the limitations of traditional methods that only focus on the unilateral actions of workers while ignoring the dynamic changes of hazards. It can accurately capture the interactive collision risk between people and dynamic hazards, improving the accuracy and reliability of violation detection in complex dynamic work scenarios. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating the method for detecting unauthorized actions of oilfield workers based on posture recognition, according to an embodiment of the present invention.

[0049] Figure 2 This is a flowchart for obtaining the subsequent attitude evolution trend based on the key point position sequence analysis. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0052] Figure 1 This is a flowchart illustrating the method for detecting unauthorized actions of oilfield workers based on posture recognition, as described in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0053] Acquire the sequence of key locations for workers, dynamic data of hazard sources, and status indicators of work processes;

[0054] Temporal evolution analysis is performed based on the joint position sequence to obtain the subsequent attitude evolution trend;

[0055] Based on the subsequent attitude evolution trend and the dynamic data of the hazard source, a predicted motion envelope for the worker and a predicted motion envelope for the hazard source are established. By calculating the minimum distance evolution curves of the predicted motion envelopes for the worker and the hazard source within a future time window, a two-way interactive risk value is obtained, which characterizes the probability of relative motion collision between the worker and the dynamic hazard source.

[0056] Based on the two-way interactive risk value and the work process status identifier, a violation judgment result is generated through context-dependent risk weight rules. The risk weight rules assign different risk levels to the same two-way interactive risk value according to the work process status identifier. The violation judgment result includes the urgency of the violation and the timing of the warning.

[0057] In one optional implementation, the step of performing temporal evolution analysis based on the joint position sequence to obtain the subsequent attitude evolution trend includes:

[0058] The actual posture is extracted from the joint position sequence, and an inertial posture is constructed assuming that active control has stopped. By comparing the actual posture with the inertial posture, the posture evolution is decomposed into intention-driven components and inertial constraint components.

[0059] The intent-driven components are mapped to a predefined job action semantic space to identify the currently executed action type and its completion degree, and the subsequent changes of the intent-driven components are predicted based on the evolution pattern of the action type.

[0060] Physical motion constraints and environmental contact constraints are applied to the inertial constraint components to generate predictions of subsequent changes in the inertial constraint components;

[0061] The fusion weights of the intention-driven component and the inertial constraint component are dynamically adjusted according to the prediction time length to obtain the joint position prediction sequence. The uncertainty range of each predicted position is marked based on the action semantic recognition confidence. The joint position prediction sequence and its uncertainty range are used as the subsequent attitude evolution trend.

[0062] Combination Figure 2 The following flowchart illustrates the subsequent posture evolution trend based on joint position sequence analysis. For example, actual posture data is extracted from the collected joint position sequence, which includes eight joints: left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, spine center, and head joint. Each joint contains three-dimensional spatial coordinate information. The extraction time window is set to 0.5 seconds, and the sampling frequency is 30 Hz, meaning 15 frames of joint position data are collected within 0.5 seconds. The collected three-dimensional coordinates of the joints are spatially normalized, and a three-dimensional coordinate system is established with the spine center point as the origin, referencing the natural standing direction of the human body.

[0063] A method for estimating inertial attitude after assuming the cessation of active control is constructed. This method calculates the motion velocity vector of each joint based on the joint position data of the first three frames, decomposing the velocity vector of each joint into three orthogonal directional components. For the right wrist joint, if the coordinates of the first three frames are (X1, Y1, Z1), (X2, Y2, Z2), and (X3, Y3, Z3), then the velocity vector from the second frame to the third frame is calculated as follows: X-direction velocity V_X = X3 - X2, Y-direction velocity V_Y = Y3 - Y2, and Z-direction velocity V_Z = Z3 - Z2. Based on this velocity vector, the influence of gravitational acceleration is applied, with a value of -9.8 m / s² acting in the negative Z-axis direction. Simultaneously, an air resistance attenuation coefficient of 0.95 is applied to the horizontal velocity component. The joint positions of subsequent frames are recursively calculated using this inertial estimation method to generate the inertial attitude trajectory.

[0064] The actual posture and the inertial posture are compared frame by frame, and the deviation vector between the actual position and the inertial predicted position of each joint in each frame is calculated. Taking the fifth frame of the right wrist joint as an example, the actual position coordinates are (0.45m, 0.32m, 1.10m), and the inertial predicted position is (0.42m, 0.28m, 1.05m), so the deviation vector is (0.03m, 0.04m, 0.05m). This deviation vector is the intention-driven component, representing the posture correction generated by the operator's active muscle control. The reference trajectory obtained by subtracting the intention-driven component from the actual posture is the inertial constraint component.

[0065] The extracted intention-driven components are mapped to a predefined semantic space for work actions, which includes twelve typical industrial work actions: grasping, placing, twisting, pushing, pulling, lifting, flipping, pressing, sliding, tapping, observing, and waiting. A feature descriptor is constructed to characterize the spatial geometric properties of the intention-driven components. This descriptor includes four dimensions: the rate of change of the relative distance between the hands, the change of the azimuth angle of the main operator's hand relative to the spine center, the curvature value of the main operator's movement trajectory, and the angle between the line connecting the shoulders and the direction of movement. For the grasping action, its typical characteristics are: the relative distance between the hands rapidly decreases from more than 0.6 meters to less than 0.1 meters; the azimuth angle of the right hand relative to the spine center remains within a 60-degree fan-shaped area in front; the curvature of the movement trajectory is less than 0.2, indicating near-linear movement; and the angle between the line connecting the shoulders and the direction of movement is nearly perpendicular.

[0066] Action recognition is performed by calculating the similarity between the current feature descriptor and the standard templates for each action type. The similarity calculation uses Euclidean distance as the metric; a smaller distance value indicates higher similarity. The feature parameters for the standard template of the grabbing action are set as follows: distance change rate -0.8 m / s, azimuth angle 45°, curvature 0.15, and included angle 85°. The currently observed feature parameters are: distance change rate -0.75 m / s, azimuth angle 48°, curvature 0.18, and included angle 88°. The Euclidean distance is then calculated as follows: The result was 0.06, which falls within the high similarity range. After traversing all twelve action templates, the action type with the smallest distance value was selected as the recognition result.

[0067] The Euclidean distance value D is converted into a recognition confidence level C using the formula: C = e^(-λD); where λ is an adjustment parameter with a value of 5. When the distance value D = 0.06, the confidence level C = e^(-5 × 0.06) = e^(-0.3) ≈ 0.74; when the distance value D = 0.02, the confidence level C = e^(-5 × 0.02) = e^(-0.1) ≈ 0.90. To improve discriminative power, the results of the above exponential function are linearly normalized, mapping the interval [0.74, 0.90] to the interval [0.75, 0.92]. The mapping formula is: C_norm = 0.75 + (C - 0.74) × (0.92 - 0.75) ÷ (0.90 - 0.74). Therefore, a distance value of 0.06 corresponds to a normalized confidence level of 0.75, and a distance value of 0.02 corresponds to a normalized confidence level of 0.92.

[0068] The completion rate of the grasping action is calculated by breaking it down into three temporal phases: the initiation phase, the execution phase, and the termination phase. The initiation phase of the grasping action is characterized by the hand beginning to move towards the target; the execution phase is characterized by the hand approaching the target and retracting the fingers; and the termination phase is characterized by the hand clenching and beginning to withdraw. By matching the current posture features with the standard features of each phase, the current phase and the percentage of progress within that phase are determined. If the current distance between the hands is 0.25m, the initiation distance of the grasping action is 0.6m, and the termination distance is 0.08m, then the completion rate is calculated as: (0.6-0.25)÷(0.6-0.08)=0.35÷0.52≈67%.

[0069] Based on the identified action type, an evolutionary pattern database for that action is retrieved. This database stores the standard evolutionary trajectory of each action type from different completion stages to completion. The standard subsequent evolution of the grasping action at 67% completion involves both hands continuing to approach, with the distance change rate maintained between -0.7 m / s and -0.9 m / s, the azimuth gradually converging towards the target object's direction, and the motion curvature increasing to 0.4 indicating a fine-tuning stage. This standard evolutionary pattern is used as the prediction benchmark for the intent-driven component, and combined with current velocity and acceleration information, a predicted trajectory for the intent-driven component within the next 0.8 seconds is generated.

[0070] Applying physical motion constraints to the subsequent evolution prediction of the inertial constraint components is the first step in the inertial constraint component prediction process. This includes joint angle constraints and inter-joint connection constraints. The shoulder joint flexion angle is constrained to the range of 0° to 180°, the abduction angle to the range of 0° to 170°, the elbow joint flexion angle to the range of 0° to 145°, and the wrist joint flexion-extension angle to the range of -70° to 80°. When predicting the evolution of the inertial constraint components, it is detected in real time whether the predicted trajectory violates the above angle constraints. If the predicted elbow joint angle of 152° exceeds the limit, the component of the joint velocity in the direction of the excess is zeroed out, and a reverse restoring force is applied to bring the joint angle back to within the 145° boundary.

[0071] The second step in the inertial constraint component prediction process involves applying environmental contact constraints to the predicted inertial constraint components. Three-dimensional point cloud data of the object's surface within the operating space is acquired using an environmental depth sensor and voxelized into an occupied grid map with a resolution of 0.02 meters. When predicting joint position evolution, it is determined whether the predicted position falls into an occupied grid. If the grid corresponding to the right wrist's predicted position coordinates (0.52m, 0.30m, 0.95m) is occupied, a collision is determined. Upon collision, the component of the joint velocity along the contact surface normal is attenuated to 10% of its original value, while retaining 85% of the tangential velocity component to simulate energy loss and friction effects during actual contact. For support contact, such as when the hand is supported on a worktable, contact is detected, and the joint point is constrained to the contact surface, allowing only tangential movement along the contact surface.

[0072] After applying physical and environmental constraints, a complete prediction sequence of subsequent changes in the inertial constraint components is generated. This step is the third step in the inertial constraint component prediction process. This prediction sequence covers a time range of 0.8 seconds, with a sampling interval of 0.033 seconds corresponding to a frequency of 30 Hz, totaling 24 prediction frames. Each prediction frame contains the three-dimensional position coordinates of all joints, as well as the velocity vector and angular velocity information for each joint. The prediction process employs an iterative recursive approach, with each iteration having a time step of 0.001 seconds, and 33 iterations are performed within a single sampling interval to improve numerical stability.

[0073] The fusion weights of the intent-driven and inertial constraint components are dynamically adjusted based on the prediction time length. Intent-driven components dominate in short-term predictions, while the effect of inertial constraints strengthens in long-term predictions. For predictions within 0.2 seconds, the intent-driven weight W_intent is set to 0.85, and the inertial constraint weight W_inertia is set to 0.15. For predictions between 0.2 and 0.5 seconds, the intent-driven weight linearly decreases to 0.60, while the inertial constraint weight increases to 0.40. For predictions between 0.5 and 0.8 seconds, the intent-driven weight further decreases to 0.35, while the inertial constraint weight increases to 0.65. The fusion formula for each prediction frame position is: P_fusion = W_intent × P_intent + W_inertia × P_inertia; where P_fusion is the fused keypoint position prediction, P_intent is the coordinate of the intent-driven component, and P_inertia is the coordinate of the inertial constraint component.

[0074] The uncertainty range is labeled for each predicted location based on the confidence score of action semantic recognition. The relationship between the uncertainty radius R and the confidence score C is mapped using an inverse proportional function: R = k ÷ (1 + C). α Where k is the scaling factor with a value of 0.12m, and α is the shape parameter with a value of 3. When the confidence level C = 0.92, the uncertainty radius R ≈ 0.02m. When the confidence level C = 0.75, the uncertainty radius R ≈ 0.02m. To reflect the impact of confidence level differences, the scaling factor k is adjusted to 0.24m. At this point, the uncertainty radius R for a confidence level of 0.92 is approximately 0.24 ÷ 7.09 ≈ 0.03m, and the uncertainty radius R for a confidence level of 0.75 is approximately 0.24 ÷ 5.36 ≈ 0.05m. Further considering the influence of prediction time, a time gain factor β = 1 + 0.5 × t is applied to the uncertainty radius, where t is the prediction time length (seconds). For the predicted right wrist position (0.38m, 0.28m, 1.08m) in the 10th frame (t=10×0.033=0.33 seconds), if the corresponding confidence level is 0.88, the basic uncertainty radius is: R_base=0.24÷(1+0.88)³≈0.036m; the uncertainty radius after applying time gain is: R_final=R_base×(1+0.5×0.33)≈0.04m; the uncertainty range of this position is a spherical region with a radius of 0.04 meters centered on the predicted coordinates. The key position prediction sequence and its corresponding uncertainty range data are combined and output to form a complete description of the subsequent attitude evolution trend, which is used by the subsequent collision prediction and trajectory planning modules.

[0075] This invention decomposes attitude evolution into two components: intention-driven and inertial constraint, enabling more accurate prediction of subsequent action trends of workers. By combining action semantic space mapping and physical constraint analysis, it achieves precision adjustment for predictions at different time scales and introduces uncertainty annotation, improving the reliability and timeliness of attitude prediction and providing a more comprehensive spatiotemporal information foundation for hazard source interaction assessment.

[0076] In one optional implementation, the step of decomposing attitude evolution into intention-driven components and inertial constraint components by comparing the actual attitude with the inertial attitude includes:

[0077] Extract the joint position difference between adjacent moments from the joint position sequence, and calculate the actual motion velocity and actual motion acceleration of each joint;

[0078] Based on the current motion velocity of each joint, a velocity decay coefficient is applied to simulate the motion decay after the active control stops. The joint position at the next moment is calculated based on the decayed velocity, and the inertial motion velocity of each joint under the inertial attitude is obtained. The inertial acceleration is calculated based on the inertial motion velocity.

[0079] The difference between the actual motion velocity and the inertial motion velocity of each joint is calculated as the velocity domain intention component, and the difference between the actual motion acceleration and the inertial acceleration is calculated as the acceleration domain intention component.

[0080] The velocity domain intention component and the acceleration domain intention component are smoothed to obtain the intention drive component.

[0081] Subtracting the position change corresponding to the intention-driven component from the joint position sequence yields the inertial constraint component.

[0082] For example, the temporal difference between two adjacent frames is extracted from the input joint position sequence. For each joint, the current 3D coordinates are subtracted from the previous 3D coordinates dimension by dimension. Assuming a shoulder joint has X coordinates of 120 pixels, Y coordinates of 80 pixels, and Z coordinates of 50 pixels in frame 10, and X coordinates of 115 pixels, Y coordinates of 78 pixels, and Z coordinates of 48 pixels in frame 9, then the position difference of this joint in the X direction is 5 pixels, in the Y direction is 2 pixels, and in the Z direction is 2 pixels. Dividing the position difference by the time interval between adjacent frames yields the actual motion velocity. When the video frame rate is 30 frames per second, the time interval is 0.033 seconds, so the actual motion velocity of the shoulder joint in the X direction is 5 pixels divided by 0.033 seconds, which is approximately 150 pixels per second. The same calculation is performed on the velocities of all joints in each direction to construct a complete actual motion velocity matrix.

[0083] The calculated actual motion velocities are subjected to time-domain difference processing to obtain the actual motion acceleration of each joint. The motion velocity of a joint at the current moment is subtracted from the motion velocity of the same joint at the previous moment, and then divided by the time interval. Continuing with the previous shoulder joint example, assuming that the velocity of the joint in the X direction is 150 pixels per second during the period from frame 9 to frame 10, and the velocity is 140 pixels per second during the period from frame 8 to frame 9, then the acceleration is the difference between the two velocities of 10 pixels per second divided by the time interval of 0.033 seconds, which gives approximately 303 pixels per second squared. This calculation process is repeated for all motion directions of all relevant nodes in the sequence to establish a complete data structure for the actual motion acceleration.

[0084] After obtaining the current motion velocity values ​​of each joint, a velocity decay coefficient is applied to simulate the natural deceleration process of the human body after ceasing active muscle control. The velocity decay coefficient typically ranges from 0.85 to 0.95, with the specific value adjusted according to the joint type and range of motion. A decay coefficient of 0.92 is set for large joints such as the shoulder, and 0.88 for small joints such as the wrist. The actual motion velocity at the current moment is multiplied by the corresponding decay coefficient to obtain the decayed velocity. If the current X-direction velocity of the shoulder joint is 150 pixels per second, applying a decay coefficient of 0.92 results in a decayed velocity of 138 pixels per second. The decayed velocity is multiplied by the time interval to obtain the position increment of the joint under inertia at the next moment. This position increment is added to the current joint position to calculate the three-dimensional coordinate position of the joint under inertial posture at the next moment.

[0085] Calculate the inertial velocity of each joint under the inertial posture. Subtract the actual joint position at the current moment from the calculated inertial posture joint position to obtain the position change under inertial drive. Divide this by the time interval to obtain the inertial velocity. Assuming the X-coordinate of the shoulder joint at the next moment under inertial action is calculated to be 124.6 pixels using decay velocity, and the current X-coordinate is 120 pixels, the position change is 4.6 pixels. Dividing this by the time interval of 0.033 seconds gives an inertial velocity of approximately 139 pixels per second. Perform a time-domain difference operation on the inertial velocity to calculate the inertial acceleration. Subtract the current inertial velocity from the previous inertial velocity and divide by the time interval. If the previous inertial velocity was 135 pixels per second and the current velocity is 139 pixels per second, then the inertial acceleration is 4 pixels per second divided by 0.033 seconds, approximately 121 pixels per second squared.

[0086] Subtracting the inertial velocity from the actual motion velocity yields a difference reflecting the velocity change caused by active control. Subtracting the inertial velocity of 139 pixels per second from the actual shoulder joint motion velocity of 150 pixels per second results in a velocity domain intention component of 11 pixels per second. This component characterizes the contribution of active muscle control to velocity. The same subtraction operation is performed on each joint in each motion direction in the sequence to form a complete matrix of velocity domain intention components. The same logic is used to calculate the acceleration domain intention components. Subtracting the inertial acceleration from the actual motion acceleration yields the acceleration change caused by active control. Subtracting the inertial acceleration of 121 pixels per second squared from the actual shoulder joint acceleration of 303 pixels per second squared results in an acceleration domain intention component of 182 pixels per second squared.

[0087] The velocity and acceleration intention components are smoothed to eliminate noise interference. A sliding window averaging method is used, with a window length of 5 frames. For the velocity intention component of a joint in a certain direction, the velocity intention component values ​​of that joint in the previous and next 2 frames (a total of 5 frames) are extracted, and the arithmetic mean is calculated as the smoothed velocity intention component. For example, if the velocity intention components of an elbow joint in the X direction in 5 consecutive frames are 8, 11, 9, 10, and 12 pixels per second, and the average value is 10 pixels per second, this is taken as the smoothed velocity intention component at that moment. The same sliding window averaging process is performed on the acceleration intention component. The smoothed velocity and acceleration intention components are then weighted and fused, with the velocity domain weight set to 0.6 and the acceleration domain weight set to 0.4. After smoothing a joint, the velocity domain intention component is 10 pixels per second, and the acceleration domain intention component is 150 pixels per second squared. Multiplying this by the time interval of 0.033 seconds converts it to approximately 5 pixels per second in velocity units. The weighted fusion result is 10 × 0.6 + 5 × 0.4 = 8 pixels per second. This value is defined as the intention-driven component of the joint in that direction.

[0088] The intention-driven component is converted into a change in the position domain. Multiplying the intention-driven component by the time interval yields the position increment generated by active control. The shoulder joint intention-driven component, at 8 pixels per second, multiplied by the time interval of 0.033 seconds, yields a position change of approximately 0.26 pixels. The position change corresponding to this intention-driven component is subtracted from the original 3D coordinates of each joint in the joint position sequence. The original shoulder joint X-coordinate, at 120 pixels, minus the intentional position change of 0.26 pixels, yields 119.74 pixels. This coordinate value represents the joint position generated solely by inertial constraints after removing the influence of active control, i.e., the inertial constraint component. This subtraction operation is performed on all coordinate dimensions of all joints at all times in the sequence to complete the extraction of the inertial constraint component. Through the above processing, the original posture evolution is decomposed into two independent motion representations: the intention-driven component reflecting subjective intention and the inertial constraint component reflecting physical inertia.

[0089] This invention introduces dual analysis of the velocity domain and acceleration domain, which enhances the ability to identify the active control intentions of personnel, while retaining the physical characteristics of inertial constraints, making the posture evolution prediction more in line with the laws of human movement.

[0090] In one optional implementation, the steps of establishing the predicted motion envelope of the worker and the predicted motion envelope of the hazard source based on the subsequent posture evolution trend and the dynamic data of the hazard source include:

[0091] The motion direction vectors and motion determinism levels of the joints in the subsequent posture evolution trend are extracted as personnel motion features, and the motion paths and spatial occupancy ranges of the hazard sources in the dynamic data of the hazard sources are extracted as hazard source motion features;

[0092] The relative approach velocity between the movement characteristics of the personnel and the movement characteristics of the hazard source is calculated, the spatiotemporal coupling strength is constructed to quantify the relative motion trend, and the interaction moment when the spatial distance reaches a local minimum is identified;

[0093] Based on the motion determinism level, the initial envelope radius of each joint is determined to form an initial envelope. When the spatiotemporal coupling strength is greater than a set value, the envelope boundary is expanded towards the hazard source. The expansion range is proportional to the spatiotemporal coupling strength, thus obtaining the predicted motion envelope of the operator.

[0094] Based on the motion characteristics of the hazard source, a spatial envelope of the hazard source is constructed, and the envelope size is adjusted based on motion predictability to obtain the predicted motion envelope of the hazard source;

[0095] At the interaction time, when the spatial distance between the predicted motion envelope of the operator and the predicted motion envelope of the hazard source is less than a set distance threshold, the boundary of the predicted motion envelope of the operator in the direction of the hazard source is expanded and the predicted motion envelope of the hazard source is extended.

[0096] For example, the three-dimensional spatial motion direction vector of each joint is extracted from the acquired subsequent posture evolution trend data. This direction vector is calculated by the spatial position difference of the joint within two consecutive time frames, and the difference result is normalized to form a unit direction vector. For a certain shoulder joint, the position coordinates at time T1 are X1=450 mm, Y1=320 mm, Z1=1200 mm, and the position coordinates at time T2 are X2=465 mm, Y2=318 mm, Z2=1195 mm. The differences in the directions of each coordinate axis are 15 mm, -2 mm, and -5 mm, respectively. The vector magnitude is calculated to be 15.9 mm. The normalized direction vector has an X-axis component of 0.943, a Y-axis component of -0.126, and a Z-axis component of -0.314.

[0097] The extraction of motion determinism level is based on the stability analysis of the motion trajectory of joints over consecutive frames. The motion direction vectors of the joints in the past ten time frames are selected, and the cosine values ​​of the angles between adjacent vectors are calculated. The stability coefficient is obtained by averaging the ten cosine values. A stability coefficient greater than 0.95 is defined as high determinism level A, corresponding to a value of 5; a stability coefficient between 0.85 and 0.95 is defined as medium determinism level B, corresponding to a value of 3; and a stability coefficient less than 0.85 is defined as low determinism level C, corresponding to a value of 1. For a specific elbow joint, the calculated ten cosine values ​​are 0.98, 0.96, 0.97, 0.95, 0.96, 0.98, 0.97, 0.96, 0.95, and 0.97, with an average value of 0.965. This joint is assigned a high determinism level A, with a value of 5.

[0098] The extraction of the motion path from the dynamic data of the hazard source is achieved by continuously tracking the centroid position of the hazard source. For a mobile work platform, its centroid coordinate sequence is recorded during a continuous monitoring period. At time T1, the centroid position is X=2000 mm, Y=1500 mm, Z=800 mm; at time T2, the centroid position is X=2035 mm, Y=1520 mm, Z=800 mm; and at time T3, the centroid position is X=2070 mm, Y=1540 mm, Z=800 mm. Connecting the centroid coordinates at each time point forms a spatial motion path curve. The spatial occupancy is determined by the three-dimensional outer contour dimensions of the hazard source. The work platform is 1200 mm long, 800 mm wide, and 600 mm high. The centroid is used as the center point to expand in all directions to form a cuboid envelope, expanding ±600 mm in the X-axis direction, ±400 mm in the Y-axis direction, and ±300 mm in the Z-axis direction.

[0099] The calculation of relative approach velocity requires considering the motion states of both the personnel's joints and the hazard source to quantify their relative motion trends in space. A specific hand joint of the personnel is selected, with its velocity vector having an X-axis component of 50 mm / s, a Y-axis component of 30 mm / s, and a Z-axis component of -20 mm / s. The hazard source's velocity vector has an X-axis component of 35 mm / s, a Y-axis component of 25 mm / s, and a Z-axis component of 0 mm / s. The velocity difference between the two is 15 mm / s on the X-axis, 5 mm / s on the Y-axis, and -20 mm / s on the Z-axis. The spatial vector from the hand joint to the hazard source's centroid is calculated; this vector has an X-axis component of 1500 mm, a Y-axis component of 1180 mm, and a Z-axis component of -400 mm. After normalization, a pointing vector is obtained. Projecting the velocity difference vector onto this pointing vector, the projected length is the relative approach velocity, calculated to be 26.3 mm / s. This relative approach velocity reflects the speed at which the hand joints move along the direction of approaching the hazard source, providing a velocity basis for subsequent spatiotemporal coupling analysis.

[0100] The spatiotemporal coupling strength was constructed by integrating relative approach velocity and spatial distance information to comprehensively assess the risk level of dynamic interaction between personnel and hazard sources. The shortest spatial distance between the hand joint and the hazard source boundary was calculated to be 850 mm. Dividing the relative approach velocity by this spatial distance yielded a preliminary coupling value of 0.0309 seconds per second. This coupling value reflects the approach rate per unit distance; a higher value indicates a stronger approach trend at the current distance. A time decay factor was introduced to adjust the coupling strength at different future moments: 1.0 for the next 1 second, 0.8 for the next 2 seconds, and 0.6 for the next 3 seconds. At the next 1 second, the spatiotemporal coupling strength remained at 0.0309 seconds per second; at the next 2 seconds, it decreased to 0.0247 seconds per second; and at the next 3 seconds, it decreased to 0.0185 seconds per second.

[0101] The identification of local minimum spatial distances is achieved by analyzing distance change curves over future time periods. The moment corresponding to this local minimum is the closest interaction moment between the person and the hazard. Based on the current motion trend, the distance change between the hand joint and the center of mass of the hazard is predicted over the next 5 seconds. The distance is 850 mm at time T0, 824 mm at time T0+1, 795 mm at time T0+2, 770 mm at time T0+3, 785 mm at time T0+4, and 810 mm at time T0+5. The distance sequence detection reveals that the minimum distance value of 770 mm is found at time T0+3, and this moment is identified as the interaction moment. The significance of identifying this interaction moment lies in determining the key time point where collision risk needs to be assessed, providing a time benchmark for envelope spatial distance judgment.

[0102] The initial envelope radius is directly related to the motion certainty level. For joints with a certainty level of A, an initial envelope radius of 80 mm is assigned; for joints with a certainty level of B, an initial envelope radius of 120 mm is assigned; and for joints with a certainty level of C, an initial envelope radius of 180 mm is assigned. For a given worker, the shoulder joint has a certainty level of A and an initial envelope radius of 80 mm; the elbow joint has a certainty level of B and an initial envelope radius of 120 mm; and the wrist joint has a certainty level of C and an initial envelope radius of 180 mm. The initial envelope radii of each joint are combined to form the initial envelope of the human body.

[0103] When the spatiotemporal coupling strength exceeds a set threshold of 0.025 seconds per second, the envelope boundary expansion mechanism is triggered. For the wrist joint, the current spatiotemporal coupling strength is 0.0309 seconds per second, exceeding the set threshold, requiring the envelope boundary to expand towards the hazard source. The direction vector from the wrist joint to the hazard source centroid is calculated. The expansion magnitude in this direction is obtained by subtracting the threshold from the spatiotemporal coupling strength and multiplying by an expansion coefficient of 8000 mm. Specifically, the value is (0.0309 - 0.025) × 8000 mm, resulting in an expansion magnitude of 47.2 mm. The original envelope radius of 180 mm increases by 47.2 mm in the hazard source direction, expanding the envelope boundary in this direction to 227.2 mm, forming an asymmetric ellipsoidal envelope. This expansion mechanism ensures that the envelope can adaptively adjust according to actual movement trends, reserving more sufficient safety space in high-risk directions.

[0104] The construction of the hazard source prediction motion envelope is based on the initial envelope setting of the spatial occupancy range of the hazard source. For the aforementioned mobile work platform, its spatial envelope is a cuboid with a length of 1200 mm, a width of 800 mm, and a height of 600 mm. The linearity of the hazard source's movement path is analyzed to assess motion predictability. Twenty centroid position points recorded during a 10-second observation period are fitted to a straight line, and the deviation distance from each point to the fitted straight line is calculated. The maximum deviation is 15 mm, and the average deviation is 6 mm. The motion predictability coefficient is defined as 0.92. The envelope size is adjusted according to this coefficient. When the predictability coefficient is greater than 0.9, the envelope size is expanded by 10% of the original size in each direction: the length expands to 1320 mm, the width expands to 880 mm, and the height expands to 660 mm. Hazards with higher motion predictability have smaller trajectory deviations, thus requiring a relatively smaller envelope expansion. Conversely, hazard sources with lower predictability require a larger expansion to cover the range of motion offset.

[0105] The spatial distance of the envelope is determined at the interaction moment. At the interaction moment T0+3 seconds of the aforementioned identification, the shortest spatial distance between the extended envelope boundary of the wrist joint and the extended envelope boundary of the hazard source is calculated to be 285 mm. The set distance threshold is 300 mm. The current distance of 285 mm is less than this threshold, triggering a secondary expansion mechanism. The boundary of the worker's predicted motion envelope in the direction of the hazard source is increased by an additional 60 mm, and the predicted motion envelope of the hazard source is simultaneously expanded by 40 mm in all directions, ensuring sufficient safety buffer space for subsequent risk assessment and early warning response. After the secondary expansion, the total envelope radius of the wrist joint in the direction of the hazard source reaches 287.2 mm, and the total envelope size of the hazard source reaches 1400 mm in length, 960 mm in width, and 740 mm in height, further ensuring the effective safety gap between the two.

[0106] This invention constructs a dynamic predictive motion envelope that considers the spatiotemporal coupling strength. It can adaptively adjust the envelope boundary to reflect the actual risk. By identifying the interaction moment with the minimum spatial distance and selectively expanding the envelope range of high-risk areas, it achieves early perception of potential collision risks.

[0107] In one optional implementation, the step of calculating the minimum distance evolution curves of the predicted motion envelope of the worker and the predicted motion envelope of the hazard source within a future time window to obtain the bidirectional interactive risk value includes:

[0108] The spatial locations of personnel joints are extracted from the predicted motion envelope of the workers as key monitoring points, and hazardous boundary surfaces are extracted from the predicted motion envelope of the hazard sources. The minimum geometric distance between each key monitoring point and the hazardous boundary surface is calculated to obtain multiple location distance sequences.

[0109] Based on the movement characteristics of the personnel and the movement characteristics of the hazard source, a relative movement direction vector is calculated. When the relative movement direction vector points close together, a directional attenuation coefficient is applied to the distance sequence of the body parts; when the relative movement direction vector points far apart, a directional gain coefficient is applied to the distance sequence of the body parts.

[0110] Differential weights are assigned to the location distance sequences of different key monitoring points based on the spatiotemporal coupling strength, with higher weights for the corresponding location distance sequences corresponding to greater spatiotemporal coupling strength.

[0111] The distance sequences of each part after directional adjustment and weight allocation are weighted and fused to obtain the comprehensive minimum distance evolution curve; the bidirectional interaction risk value is calculated by weighting the reciprocal of the global minimum value of the comprehensive minimum distance evolution curve, the reciprocal of the time when the minimum value is reached, and the curve descent rate before that time.

[0112] For example, in practical applications, joints with large ranges of motion and directly related to the task are selected and marked in the human skeletal model. For assembly operations, the palm, wrist, elbow, shoulder, head center, and waist center constitute a set of key monitoring points. Each joint is represented in three-dimensional space by X-axis, Y-axis, and Z-axis coordinates. At time T0, the position of a worker's right palm is X=1350 mm, Y=980 mm, Z=1150 mm, and the position of the right wrist joint is X=1420 mm, Y=1020 mm, Z=1180 mm. Corresponding boundary surface representation methods are used for hazards of different geometric shapes. The hazardous boundary surface of a cuboid-shaped mobile work platform consists of six planes, each defined by coefficients of a plane equation. The hazardous boundary surfaces of the work platform at time T0 include the front surface at X=2700 mm, the rear surface at X=1500 mm, the left side at Y=1100 mm, the right side at Y=300 mm, the top surface at Z=1460 mm, and the bottom surface at Z=800 mm. For cylindrical hazards such as rotating shafts, the hazardous boundary surfaces are defined by the position of the cylinder axis, radius, and height range.

[0113] A perpendicular distance algorithm from a point to a plane or a point to a curved surface is used to solve for the geometric minimum distance. For the distance calculation from a point to a plane, the coordinates of the joint point are substituted into the plane equation to solve for the absolute value of the perpendicular distance. The distance from the right palm to the front surface of the work platform is calculated using the X-coordinate difference, resulting in 2700 - 1350 = 1350 mm. For the distance calculation from a point to a cylindrical surface, the perpendicular distance from the joint point to the cylinder axis is first solved, and then the cylinder radius is subtracted to obtain the minimum distance. For each joint point, all dangerous boundary surfaces are traversed, and the minimum distance value is selected as the location distance of that joint point at the current moment. The distances from the right palm to the six boundary surfaces are 1350 mm, 150 mm, 120 mm, 680 mm, 310 mm, and 350 mm, respectively; the minimum value of 120 mm is taken as the location distance of the right palm. The future time window is set to 5 seconds, and the time step is set to 0.1 seconds. Based on the movement characteristics of personnel and hazards, the positions of key points and hazardous boundaries at the next 50 time points are predicted. The distance calculation process is repeated to obtain the location distance sequence for each key point. The location distance sequence for the right palm point is 120 mm at time T0, 115 mm at time T0+0.1 seconds, 109 mm at time T0+0.2 seconds, and so on, until it reaches 285 mm at time T0+5 seconds, forming a sequence containing 50 data points.

[0114] The velocity vectors of the personnel's joints and the hazard source are extracted, and their difference yields the relative velocity vector. The velocity vector of the right palm point has an X-axis component of 45 mm / s, a Y-axis component of -8 mm / s, and a Z-axis component of 12 mm / s. The velocity vector of the hazard source has an X-axis component of -20 mm / s, a Y-axis component of 5 mm / s, and a Z-axis component of 0 mm / s. The relative velocity vector is 65 mm / s on the X-axis, -13 mm / s on the Y-axis, and 12 mm / s on the Z-axis. The spatial vector from the joint point to the nearest point on the nearest hazard boundary is calculated as the distance direction vector. The distance direction vector from the right palm point to the nearest boundary is in the negative Y-axis direction, and after normalization, it has an X-axis component of 0, a Y-axis component of -1, and a Z-axis component of 0. The relative velocity vector and the distance direction vector are then multiplied to obtain the relative motion direction coefficient. The dot product result is 65×0+(-13)×(-1)+12×0=13. A value greater than 0 indicates the relative motion direction vector is pointing closer, while a value less than 0 indicates it is pointing further away. When pointing closer, the attenuation coefficient is adjusted according to the magnitude of the relative motion direction coefficient; the larger the relative motion direction coefficient, the more significant the attenuation. The attenuation coefficient is calculated as 1 - relative motion direction coefficient / 100. The attenuation coefficient for the right palm point is 1 - 13 / 100 = 0.87. Each distance value in the location distance sequence is multiplied by the attenuation coefficient 0.87. The distance at time T0 (120 mm) is adjusted to 104.4 mm, and the distance at time T0+0.1 seconds (115 mm) is adjusted to 100.05 mm. When pointing further away, the gain coefficient is 1 + absolute value of the relative motion direction coefficient / 100. If the relative motion direction coefficient of a certain elbow joint point is -8, the gain coefficient is 1 + 8 / 100 = 1.08. All distance values ​​in the location distance sequence are multiplied by 1.08 for adjustment.

[0115] The spatiotemporal coupling strength is calculated by dividing the relative approach velocity by the spatial distance. In a specific implementation case, the relative approach velocity of the right palm point is 13 mm / s, and the spatial distance to the dangerous boundary surface is 120 mm, so the calculated spatiotemporal coupling strength is 13 ÷ 120 ≈ 0.108 / s. A normalization threshold for the spatiotemporal coupling strength is set to 0.2 / s. The spatiotemporal coupling strength of each key monitoring point is divided by this threshold, with the maximum value limited to 1, to obtain the normalized coupling strength coefficient. The coupling strength coefficient of the right palm point is 0.108 ÷ 0.2 = 0.54, the coupling strength coefficient of the right wrist joint is 0.32, the coupling strength coefficient of the elbow joint is 0.18, and the coupling strength coefficients of the other joints are 0.12, 0.08, and 0.06, respectively. The coupling strength coefficient is used as the initial weight, and then a joint importance coefficient is introduced for correction. For hand joints, the importance coefficient was set to 1.2; for wrist and elbow joints, the importance coefficient was set to 1.0; and for shoulder and torso joints, the importance coefficient was set to 0.8. The final weight of the right palm joint was 0.54 × 1.2 = 0.648, and the final weight of the right wrist joint was 0.32 × 1.0 = 0.32. The final weights of all joints were normalized so that the sum of the weights was 1. The total weight of the six key monitoring points was 1.464. After normalization, the weight of the right palm joint was 0.648 / 1.464 = 0.443, the weight of the right wrist joint was 0.219, and the weights of the remaining joints were 0.123, 0.082, 0.055, and 0.041, respectively.

[0116] At each time point, the adjusted distance values ​​of each joint are multiplied by their corresponding weights and then summed. At time T0, the adjusted distance of the right palm joint is 104.4 mm × weight 0.443 = 46.2 mm; the adjusted distance of the right wrist joint is 135 mm × weight 0.219 = 29.6 mm; the adjusted distance of the elbow joint is 280 mm × weight 0.123 = 34.4 mm; the adjusted distance of the shoulder joint is 450 mm × weight 0.082 = 36.9 mm; the adjusted distance of the head center point is 520 mm × weight 0.055 = 28.6 mm; and the adjusted distance of the waist center point is 380 mm × weight 0.041 = 15.6 mm. The weighted sum is 191.3 mm. This calculation process is repeated for each of the 50 time points within the future time window to obtain the comprehensive minimum distance evolution curve. The curve has a distance of 191.3 mm at time T0, 185.7 mm at time T0+0.1 seconds, 179.8 mm at time T0+0.2 seconds, and reaches a global minimum of 152.4 mm at time T0+1.5 seconds. After that, the distance gradually increases to 208.6 mm at time T0+5 seconds.

[0117] The global minimum value of the curve, 152.4 mm, is extracted, and its reciprocal is calculated to be 0.00656 / mm. The moment the minimum value is reached is extracted as 1.5 seconds, and its reciprocal is calculated to be 0.667 / second. When calculating the curve's descent rate before this moment, the distance change from time T0 to time T0+1.5 seconds is selected. The initial distance of 191.3 mm minus the minimum distance of 152.4 mm equals 38.9 mm, which, divided by the time interval of 1.5 seconds, yields a descent rate of 25.9 mm / second. Weighting coefficients are set for the three indicators: the coefficient for the reciprocal of the global minimum is 0.5, the coefficient for the reciprocal of the arrival time is 0.3, and the coefficient for the descent rate is 0.2. The two-way interaction risk value is calculated as 0.5 × 0.00656 + 0.3 × 0.667 + 0.2 × 25.9 / 100 = 0.255. The risk value is expressed in mm / s; a higher value indicates a higher interaction risk. A risk level classification threshold is set: a risk value less than 0.1 mm / s is considered low risk, 0.1 to 0.3 mm / s is considered medium risk, and greater than 0.3 mm / s is considered high risk. The currently calculated risk value of 0.255 mm / s falls within the medium risk level, requiring the triggering of an early warning mechanism and advising workers to adjust their movement trajectory or speed to reduce the probability of collision.

[0118] This invention achieves dynamic focusing on key parts through directional adjustment and coupling strength weight allocation, integrates spatial and temporal dimension information by combining minimum distance evolution curves, and quantitatively assesses collision risk through bidirectional interactive risk values.

[0119] In one optional implementation, a violation determination result is generated based on the two-way interaction risk value and the work process status identifier using a context-dependent risk weighting rule. The risk weighting rule assigns different risk levels to the same two-way interaction risk value based on the work process status identifier. The violation determination result includes the steps of determining the urgency of the violation and the timing of the warning.

[0120] Based on the work process status identifier, extract the work stage type, work step number, and execution time;

[0121] The baseline risk weight coefficient corresponding to the work stage type is queried from the preset work process-risk mapping rule base. The baseline risk weight coefficient is set differently according to the safety distance requirements of the work stage type. When the work step number corresponds to a high-risk operation step, an amplification factor is applied to the baseline risk weight coefficient for adjustment. The fatigue accumulation factor is calculated based on the execution time. The adjusted risk weight coefficient is corrected according to the fatigue accumulation factor to obtain the final risk weight coefficient.

[0122] The two-way interactive risk value is multiplied by the final risk weight coefficient to obtain the weighted risk value, and the weighted risk value is divided into different risk levels according to the preset risk level threshold.

[0123] The urgency of the violation is determined based on the risk level; the moment when the two-way interactive risk value reaches its peak within a future time window is extracted as the warning timing; the urgency of the violation and the warning timing are combined as the violation determination result.

[0124] For example, the job phase type is represented by an enumeration encoding method: preparation phase is coded as 01, execution phase as 02, inspection phase as 03, and completion phase as 04. The job step number is the sequential number within the current job phase, starting from 1 and incrementing by 1 for each completed step. The execution time records the cumulative time from the start of the current job phase to the current moment, measured in seconds. In a certain assembly job scenario, the current job flow status indicator shows that the job phase type is execution phase code 02, the job step number is step 5, and the execution time is 1820 seconds. This data is obtained by parsing the status messages pushed in real time by the job management system. The message format is a fixed-length string, where the first two digits represent the phase type, the middle three digits represent the step number, and the last four digits represent the execution time.

[0125] The pre-defined work process risk mapping rule base uses a key-value pair structure for storage. The key is the work stage type code, and the value is the corresponding baseline risk weight coefficient. The baseline risk weight coefficient for the preparation stage is 0.8, for the execution stage it is 1.2, for the inspection stage it is 1.0, and for the closing stage it is 0.9. These coefficients reflect the differentiated safety distance requirements for different work stages. The execution stage, due to frequent personnel movement and equipment operation, has stricter safety distance requirements and is therefore assigned a higher weight coefficient. The query process matches the current work stage type code as the key in the rule base. For the execution stage code 02, the matched baseline risk weight coefficient is 1.2.

[0126] High-risk operation steps are identified using a pre-marked step risk level table, which associates each operation step number with its risk level. Risk levels are divided into two categories: routine operations and high-risk operations. Routine operations do not trigger amplification adjustments, while high-risk operations trigger amplification factor adjustments. High-risk operation steps include using power tools, working near rotating parts, and working at heights. For the assembly process, steps 3, 5, and 8 are marked as high-risk operation steps. The current operation step number is 5. Checking the step risk level table confirms that this step is a high-risk operation, triggering an amplification factor adjustment. The amplification factor is set to 1.3. Multiplying the baseline risk weight factor of 1.2 by the amplification factor of 1.3 yields an adjusted risk weight factor of 1.56.

[0127] The fatigue accumulation factor is calculated based on a comparison between the executed duration and a preset fatigue threshold time. The fatigue threshold time is set at 1800 seconds. When the executed duration does not exceed this threshold, the fatigue accumulation factor is 1.0. After exceeding the threshold, the fatigue accumulation factor increases proportionally to the excess duration. The calculation formula is: fatigue accumulation factor = 1.0 + the portion of the executed duration exceeding the fatigue threshold time divided by the fatigue threshold time, then multiplied by the growth rate coefficient 0.15. The current executed duration is 1820 seconds, exceeding the fatigue threshold time by 20 seconds. The excess proportion is 20 ÷ 1800 ≈ 0.011, and the fatigue accumulation factor is calculated as 1.0 + 0.011 × 0.15 = 1.0017. Multiplying the adjusted risk weight coefficient 1.56 by the fatigue accumulation factor 1.0017 yields the final risk weight coefficient 1.563.

[0128] Assuming the calculated two-way interactive risk value is 0.255 per millimeter, multiplied by the final risk weighting coefficient of 1.563, the weighted risk value is 0.399 per millimeter. The preset risk level threshold classification rule is as follows: a weighted risk value less than 0.15 per millimeter is defined as a safe level, 0.15 to 0.30 per millimeter as a warning level, 0.30 to 0.50 per millimeter as a dangerous level, and greater than 0.50 per millimeter as an emergency level. The current weighted risk value of 0.399 per millimeter falls within the 0.30 to 0.50 range, therefore the risk level is determined to be dangerous.

[0129] The urgency level of a violation is directly mapped to the risk level. A safety level corresponds to a violation urgency level of zero, indicating no warning is needed; a warning level corresponds to a violation urgency level of one, indicating that operators need to be alerted; a danger level corresponds to a violation urgency level of two, indicating that an immediate audible and visual warning is required and operation should be suspended; an emergency level corresponds to a violation urgency level of three, indicating that work must be forcibly stopped and an emergency response triggered. The current risk level is danger, and the violation urgency level is determined to be level two.

[0130] The evolution curve of the comprehensive minimum distance obtained from the previous steps is used to derive the evolution curve of the two-way interaction risk value. This curve shows a risk value of 0.255 mm at time T0, 0.288 mm at time T0+0.5 seconds, 0.334 mm at time T0+1.0 seconds, a peak risk value of 0.382 mm at time T0+1.5 seconds, and a drop back to 0.310 mm at time T0+2.0 seconds. By iterating through all time points within the next 5-second time window, the moment with the highest risk value is identified as the peak moment. The current peak moment is time T0+1.5 seconds, with a time offset of 1.5 seconds relative to the current moment. This moment is designated as the warning timing. The physical meaning of the warning timing is that 1.5 seconds after the current moment, the interaction risk between personnel and the hazard source will reach its highest level, and the system must complete the push of warning information and the generation of operational suggestions before this moment arrives.

[0131] The violation urgency level (Level 2) is combined with the 1.5-second warning timing to generate structured violation determination data. This data includes violation identifier, urgency level, warning timing, and suggested action fields. The violation identifier field is set to true, indicating that a violation has been detected; the urgency level field is assigned a value of 2, corresponding to the danger level; the warning timing field is assigned a value of 1.5 seconds, indicating the remaining time until the risk peak; the suggested action field is automatically filled with "Immediately issue an audible and visual warning and suggest pausing operations, adjusting personnel movement trajectories, or reducing movement speed" based on the Level 2 urgency. The violation determination result is pushed to the warning display terminal and the operation management platform via a message queue, triggering corresponding warning actions and recording. In cases with a short fatigue duration, assuming an execution time of 900 seconds, a fatigue accumulation factor of 1.0, a final risk weight coefficient of 1.56, and a weighted risk value of 0.398 per millimeter, the violation remains at the danger level, maintaining Level 2 urgency. However, the warning response can be appropriately delayed until the user completes the current micro-operation to avoid the risk of misoperation caused by sudden warnings.

[0132] This invention achieves differentiated assessment of the same risk value at different operational stages through context-dependent risk weighting rules. It combines fatigue accumulation and dynamic adjustment of weights for high-risk steps to improve the accuracy and timeliness of violation determination, providing a reliable basis for graded early warning and precise intervention.

[0133] In one optional implementation, the steps of calculating a fatigue accumulation factor based on the executed duration, and correcting the adjusted risk weight coefficient according to the fatigue accumulation factor to obtain the final risk weight coefficient include:

[0134] The initial fatigue factor is calculated based on the ratio of the executed duration to the preset fatigue threshold duration.

[0135] Extract the motion intensity change curve within the executed duration from the joint position sequence, calculate the cumulative duration ratio of high-intensity motions, and adjust the initial fatigue factor based on the cumulative duration ratio to obtain a comprehensive fatigue factor.

[0136] The overall fatigue factor is attenuated and corrected according to the rest period in the work process status indicator; the correction method is selected according to the numerical range of the overall fatigue factor. When the overall fatigue factor is lower than the preset fatigue segment threshold, linear correction is used, and when the overall fatigue factor is higher than the preset fatigue segment threshold, nonlinear accelerated correction is used.

[0137] Determine whether there is a conflict between the baseline risk weight coefficient corresponding to the operation stage type and the risk weight coefficient after adjustment of the high-risk step amplification coefficient. When the baseline risk weight coefficient is already in the high-risk range, limit the amplification factor of the high-risk step amplification coefficient.

[0138] The adjusted risk weight coefficient is combined with the revised comprehensive fatigue factor to obtain the final risk weight coefficient.

[0139] For example, the initial fatigue factor is calculated by comparing the executed time extracted from the work process status identifier with a preset fatigue threshold time. The fatigue threshold time is set according to the work type: 3600 seconds for assembly work, 2700 seconds for handling work, and 4500 seconds for maintenance work. The initial fatigue factor is calculated by dividing the executed time by the fatigue threshold time, with the quotient rounded to three decimal places. In a certain assembly work scenario, the executed time is 2580 seconds, and the fatigue threshold time is 3600 seconds. The ratio is calculated as 2580 ÷ 3600, resulting in an initial fatigue factor of 0.717. When the executed time exceeds the fatigue threshold time, the ratio exceeds 1.0, indicating that the worker has entered a state of fatigue. If the executed time is 4320 seconds, the initial fatigue factor is calculated as 4320 ÷ 3600, resulting in 1.200. The portion exceeding the baseline value of 1.0 reflects the degree of accumulated fatigue due to overtime.

[0140] The joint position sequence contains three-dimensional coordinate data of the worker's shoulder, elbow, wrist, hip, knee, and ankle joints, sampled at a frequency of 30 frames per second. The distance change rate between joints is extracted from all frames from the start of the sequence to the end of the current execution time to construct a motion intensity change curve. Motion intensity is quantified by calculating the sum of the Euclidean distances of joint displacements between adjacent frames, in millimeters per frame. A threshold for high-intensity motion is set at a total joint displacement exceeding 150 millimeters in a single frame. All frames within the entire time window are traversed, and the number of frames meeting this threshold condition is counted. Within the 2580-second execution time, the total number of frames is 2580 × 30 = 77400, of which 18600 frames are considered high-intensity motion, representing a cumulative duration percentage of 18600 ÷ 77400 ≈ 0.240. The initial fatigue factor is adjusted for intensity based on the cumulative duration percentage. The adjustment coefficient is set to 1.0, plus the cumulative duration percentage multiplied by the intensity influence weight of 0.35, resulting in a adjustment coefficient of 1.0 + 0.240 × 0.35 = 1.084. Multiplying the initial fatigue factor of 0.717 by the adjustment coefficient of 1.084 yields a comprehensive fatigue factor of 0.777.

[0141] The work process status identifier records the start and end timestamps of rest periods. A rest period is defined as the time when a worker leaves the work area for more than 180 seconds. The rest flag field in the status identifier is parsed to identify all rest periods occurring within the executed time range, and the total rest duration is calculated. Assuming that within the executed time of 2580 seconds, a rest period from time 900 to time 1200 is detected, with a rest duration of 300 seconds, the attenuation correction uses an exponential decay model with an attenuation coefficient set to 0.0008 per second. The overall fatigue factor decreases after the rest period ends by multiplying the attenuation coefficient by the rest duration. The attenuation amount is calculated as the overall fatigue factor multiplied by the attenuation coefficient and then by the rest duration, resulting in 0.777 × 0.0008 × 300 ≈ 0.186. Subtracting the attenuation amount of 0.186 from the overall fatigue factor of 0.777 yields the attenuated overall fatigue factor of 0.591. When there are no rest periods, the overall fatigue factor remains unchanged.

[0142] The preset fatigue segmentation threshold is set to 0.800 to distinguish between the linear correction interval and the nonlinear accelerated correction interval. When the comprehensive fatigue factor is below 0.800, linear correction is used, and the corrected comprehensive fatigue factor equals the original comprehensive fatigue factor multiplied by the linear correction factor of 1.15. When the comprehensive fatigue factor is equal to or higher than 0.800, nonlinear accelerated correction is used, and the corrected comprehensive fatigue factor equals the original comprehensive fatigue factor raised to the power of 1.8 multiplied by the accelerated correction base of 1.10. In the current case, the comprehensive fatigue factor is 0.591, which is lower than the fatigue segmentation threshold of 0.800, triggering linear correction. The corrected comprehensive fatigue factor is 0.591 × 1.15 ≈ 0.680. In the comparison case, assuming the comprehensive fatigue factor is 0.950, which is higher than the fatigue segmentation threshold, nonlinear accelerated correction is triggered. Calculating 0.950 raised to the power of 1.8 is approximately 0.912, and multiplying by 1.10 yields a corrected comprehensive fatigue factor of 1.003. A value exceeding 1.0 indicates that the fatigue state has reached a high-risk level.

[0143] The conflict judgment between the baseline risk weight coefficient corresponding to the work phase type and the adjusted risk weight coefficient of the high-risk step amplification factor is achieved by setting a high-risk interval boundary. The high-risk interval is defined as the range where the risk weight coefficient is greater than or equal to 1.50. When the baseline risk weight coefficient is already in this interval, the amplification factor of the high-risk step amplification factor is limited. In the preceding calculation, the baseline risk weight coefficient is 1.20, the high-risk step amplification factor is 1.30, and the adjusted risk weight coefficient is 1.20 × 1.30 = 1.56, which is in the high-risk interval. The conflict judgment module detects that although the baseline risk weight coefficient of 1.20 has not reached the high-risk interval, the adjusted risk weight coefficient of 1.56 has exceeded 1.50, triggering the amplification factor limitation mechanism. The limitation rule is that when the adjusted risk weight coefficient exceeds 1.50, the high-risk step amplification factor is reduced to the minimum multiple that makes the adjusted risk weight coefficient equal to 1.50. The calculated limited amplification factor is 1.50 ÷ 1.20 = 1.25. This replaces the original high-risk step amplification factor of 1.30, and the adjusted risk weight factor is corrected to 1.50. When the baseline risk weight factor of 1.60 is already in the high-risk range, the high-risk step amplification factor is directly disabled, and the adjusted risk weight factor remains at the baseline value of 1.60.

[0144] The adjusted risk weight coefficient is fused with the corrected comprehensive fatigue factor using a weighted product model. The final risk weight coefficient equals the adjusted risk weight coefficient multiplied by the corrected amplification factor of the comprehensive fatigue factor. The corrected amplification factor of the comprehensive fatigue factor is calculated as follows: subtract the baseline fatigue factor of 0.500 from the corrected comprehensive fatigue factor, multiply the difference by the fatigue influence coefficient of 0.40, and then add 1.0 to the product to obtain the amplification factor. In the current case, the corrected comprehensive fatigue factor is 0.680, the baseline fatigue factor is 0.500, the difference is 0.180, and the amplification factor is 1.0 + 0.180 × 0.40 = 1.072. The adjusted risk weight coefficient is 1.50, multiplied by the amplification factor of 1.072, resulting in a final risk weight coefficient of 1.608. In a low-fatigue scenario, assuming the corrected comprehensive fatigue factor is 0.450, lower than the baseline fatigue factor of 0.500, the difference is -0.050. The amplification factor is 1.0 + (-0.050) × 0.40 = 0.980, and the final risk weight coefficient is 1.50 × 0.980 = 1.470, indicating that the low-fatigue state has a slight reducing effect on the risk weight. When the corrected comprehensive fatigue factor exceeds 1.20, the difference is 1.20 minus 0.500 equals 0.700. The amplification factor is calculated as 1.0 + 0.700 × 0.40 = 1.280, and the adjusted risk weight coefficient is 1.50. Multiplying by the amplification factor of 1.280, the final risk weight coefficient is 1.920, significantly increasing the risk assessment weight and reflecting the enhanced sensitivity of violation judgment under extreme fatigue conditions.

[0145] A numerical boundary protection mechanism is implemented during the fusion calculation process, limiting the final risk weight coefficient to between 0.50 and 2.50. When the calculated result is below 0.50, it is forcibly corrected to 0.50; when the calculated result is above 2.50, it is forcibly corrected to 2.50. This prevents distortion of risk assessment caused by abnormal fatigue states or incorrect configuration parameters. In actual deployment, the adjusted risk weight coefficient, the corrected comprehensive fatigue factor, and the final risk weight coefficient are all recorded in the work process log database. Each record includes a timestamp, operator identifier, work stage type, step number, values ​​of coefficients at each level, and fusion parameter configuration, supporting post-event auditing and parameter optimization analysis.

[0146] This invention achieves refined adjustment of risk weight coefficients by using multi-dimensional fatigue accumulation factor calculation and segmented correction strategies, combined with high-risk step conflict judgment and dynamic fusion mechanism, thereby improving the sensitivity and accuracy of violation judgment under fatigue state.

[0147] A second aspect of the present invention provides an electronic device, comprising:

[0148] processor;

[0149] Memory used to store processor-executable instructions;

[0150] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0151] A third aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0152] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

Claims

1. A method for detecting unauthorized actions of workers in oilfield scenarios based on posture recognition, characterized in that, include: Acquire the sequence of key locations for workers, dynamic data of hazard sources, and status indicators of work processes; Temporal evolution analysis is performed based on the joint position sequence to obtain the subsequent attitude evolution trend; Based on the subsequent attitude evolution trend and the dynamic data of the hazard source, a predicted motion envelope for the worker and a predicted motion envelope for the hazard source are established. By calculating the minimum distance evolution curves of the predicted motion envelopes for the worker and the hazard source within a future time window, a two-way interactive risk value is obtained, which characterizes the probability of relative motion collision between the worker and the dynamic hazard source. Based on the two-way interactive risk value and the work process status identifier, a violation judgment result is generated through context-dependent risk weight rules. The risk weight rules assign different risk levels to the same two-way interactive risk value according to the work process status identifier. The violation judgment result includes the urgency of the violation and the timing of the warning.

2. The method according to claim 1, characterized in that, The steps for performing temporal evolution analysis based on the joint position sequence to obtain the subsequent attitude evolution trend include: The actual posture is extracted from the joint position sequence, and an inertial posture is constructed assuming that active control has stopped. By comparing the actual posture with the inertial posture, the posture evolution is decomposed into intention-driven components and inertial constraint components. The intent-driven components are mapped to a predefined job action semantic space to identify the currently executed action type and its completion degree, and the subsequent changes of the intent-driven components are predicted based on the evolution pattern of the action type. Physical motion constraints and environmental contact constraints are applied to the inertial constraint components to generate predictions of subsequent changes in the inertial constraint components; The fusion weights of the intention-driven component and the inertial constraint component are dynamically adjusted according to the prediction time length to obtain the joint position prediction sequence. The uncertainty range of each predicted position is marked based on the action semantic recognition confidence. The joint position prediction sequence and its uncertainty range are used as the subsequent attitude evolution trend.

3. The method according to claim 2, characterized in that, The step of decomposing attitude evolution into intention-driven components and inertial constraint components by comparing the actual attitude with the inertial attitude includes: Extract the joint position difference between adjacent moments from the joint position sequence, and calculate the actual motion velocity and actual motion acceleration of each joint; Based on the current motion velocity of each joint, a velocity decay coefficient is applied to simulate the motion decay after the active control stops. The joint position at the next moment is calculated based on the decayed velocity, and the inertial motion velocity of each joint under the inertial attitude is obtained. The inertial acceleration is calculated based on the inertial motion velocity. The difference between the actual motion velocity and the inertial motion velocity of each joint is calculated as the velocity domain intention component, and the difference between the actual motion acceleration and the inertial acceleration is calculated as the acceleration domain intention component. The velocity domain intention component and the acceleration domain intention component are smoothed to obtain the intention drive component. Subtracting the position change corresponding to the intention-driven component from the joint position sequence yields the inertial constraint component.

4. The method according to claim 1, characterized in that, Based on the subsequent posture evolution trend and the dynamic data of the hazard source, the steps for establishing the predicted motion envelope of the worker and the predicted motion envelope of the hazard source include: The motion direction vectors and motion determinism levels of the joints in the subsequent posture evolution trend are extracted as personnel motion features, and the motion paths and spatial occupancy ranges of the hazard sources in the dynamic data of the hazard sources are extracted as hazard source motion features; The relative approach velocity between the movement characteristics of the personnel and the movement characteristics of the hazard source is calculated, the spatiotemporal coupling strength is constructed to quantify the relative motion trend, and the interaction moment when the spatial distance reaches a local minimum is identified; Based on the motion determinism level, the initial envelope radius of each joint is determined to form an initial envelope. When the spatiotemporal coupling strength is greater than a set value, the envelope boundary is expanded towards the hazard source. The expansion range is proportional to the spatiotemporal coupling strength, thus obtaining the predicted motion envelope of the operator. Based on the motion characteristics of the hazard source, a spatial envelope of the hazard source is constructed, and the envelope size is adjusted based on motion predictability to obtain the predicted motion envelope of the hazard source; At the interaction time, when the spatial distance between the predicted motion envelope of the operator and the predicted motion envelope of the hazard source is less than a set distance threshold, the boundary of the predicted motion envelope of the operator in the direction of the hazard source is expanded and the predicted motion envelope of the hazard source is extended.

5. The method according to claim 4, characterized in that, The steps for calculating the minimum distance evolution curves of the predicted motion envelope of the worker and the predicted motion envelope of the hazard source within a future time window to obtain the bidirectional interactive risk value include: The spatial locations of personnel joints are extracted from the predicted motion envelope of the workers as key monitoring points, and hazardous boundary surfaces are extracted from the predicted motion envelope of the hazard sources. The minimum geometric distance between each key monitoring point and the hazardous boundary surface is calculated to obtain multiple location distance sequences. Based on the movement characteristics of the personnel and the movement characteristics of the hazard source, a relative movement direction vector is calculated. When the relative movement direction vector points close together, a directional attenuation coefficient is applied to the distance sequence of the body parts; when the relative movement direction vector points far apart, a directional gain coefficient is applied to the distance sequence of the body parts. Differential weights are assigned to the location distance sequences of different key monitoring points based on the spatiotemporal coupling strength, with higher weights for the corresponding location distance sequences corresponding to greater spatiotemporal coupling strength. The distance sequences of each part after directional adjustment and weight allocation are weighted and fused to obtain the comprehensive minimum distance evolution curve; the bidirectional interaction risk value is calculated by weighting the reciprocal of the global minimum value of the comprehensive minimum distance evolution curve, the reciprocal of the time when the minimum value is reached, and the curve descent rate before that time.

6. The method according to claim 1, characterized in that, Based on the bidirectional interaction risk value and the work process status identifier, a violation determination result is generated using context-dependent risk weighting rules. These risk weighting rules assign different risk levels to the same bidirectional interaction risk value based on the work process status identifier. The violation determination result includes steps such as determining the urgency of the violation and the timing of the warning. Based on the work process status identifier, extract the work stage type, work step number, and execution time; The baseline risk weight coefficient corresponding to the work stage type is queried from the preset work process-risk mapping rule base. The baseline risk weight coefficient is set differently according to the safety distance requirements of the work stage type. When the work step number corresponds to a high-risk operation step, an amplification factor is applied to the baseline risk weight coefficient for adjustment. The fatigue accumulation factor is calculated based on the execution time. The adjusted risk weight coefficient is corrected according to the fatigue accumulation factor to obtain the final risk weight coefficient. The two-way interactive risk value is multiplied by the final risk weight coefficient to obtain the weighted risk value, and the weighted risk value is divided into different risk levels according to the preset risk level threshold. The urgency of the violation is determined based on the risk level; the moment when the two-way interactive risk value reaches its peak within a future time window is extracted as the warning timing; the urgency of the violation and the warning timing are combined as the violation determination result.

7. The method according to claim 6, characterized in that, The steps of calculating the fatigue accumulation factor based on the executed duration, and correcting the adjusted risk weight coefficient according to the fatigue accumulation factor to obtain the final risk weight coefficient include: The initial fatigue factor is calculated based on the ratio of the executed duration to the preset fatigue threshold duration. Extract the motion intensity change curve within the executed duration from the joint position sequence, calculate the cumulative duration ratio of high-intensity motions, and adjust the initial fatigue factor based on the cumulative duration ratio to obtain a comprehensive fatigue factor. The overall fatigue factor is attenuated and corrected according to the rest period in the work process status indicator; the correction method is selected according to the numerical range of the overall fatigue factor. When the overall fatigue factor is lower than the preset fatigue segment threshold, linear correction is used, and when the overall fatigue factor is higher than the preset fatigue segment threshold, nonlinear accelerated correction is used. Determine whether there is a conflict between the baseline risk weight coefficient corresponding to the operation stage type and the risk weight coefficient after adjustment of the high-risk step amplification coefficient. When the baseline risk weight coefficient is already in the high-risk range, limit the amplification factor of the high-risk step amplification coefficient. The adjusted risk weight coefficient is combined with the revised comprehensive fatigue factor to obtain the final risk weight coefficient.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images