A method and system for proactive risk prevention and control in high-risk operations

Abstract

This invention relates to the field of industrial safety and automation control technology, specifically disclosing a method and system for risk pre-control and proactive management of high-risk operations. The method involves collecting work plans, environmental parameters, personnel physiological characteristics, and historical accident records to construct a risk memory represented by a causal relationship diagram and a probability matrix. Based on this memory, the work plan is subjected to counterfactual deduction to generate dynamic spatial boundaries, dynamic time windows, and preset trigger rules. During the operation, video, audio, and sensor data are integrated to measure deviations from the pre-control plan in real time. When deviations exceed the rules, proactive management instructions are generated. After the operation, the actual deviation trajectory is compared with the predicted deviation trajectory, and the risk memory is updated accordingly. This invention achieves closed-loop management of high-risk operations from passive alarm to proactive intervention and possesses self-evolution capabilities.

Description

Technical Field

[0001] This invention relates to the field of industrial safety and automation control technology, specifically to a method and system for risk pre-control and proactive management of high-risk operations. Background Technology

[0002] With the continuous improvement of industrial automation and informatization, high-risk industries such as power, petrochemicals, and mining have accumulated massive amounts of work permits, sensor monitoring data, personnel location records, and historical accident cases during production operations, forming a considerable industrial big data resource. However, current control measures for high-risk operations such as hot work, confined space work, and work at heights still mainly rely on static regulations and the experience-based judgment of on-site safety personnel. Existing industrial safety management systems typically only perform simple statistical displays or threshold alarm processing on this industrial big data, such as triggering audible and visual alarms when gas concentration exceeds a fixed upper limit, lacking the mining and utilization of deep semantic relationships between data.

[0003] The shortcomings of existing technologies addressed by this invention are: how to enable high-risk operation management systems to possess counterfactual reasoning and autonomous evolution capabilities based on industrial big data, that is, how to extract deep causal relationships between operational elements from static historical accident records and real-time monitoring data, and use these causal relationships to simulate and deduce "how the risk evolves after changing arbitrary conditions" of the proposed execution plan before the operation, thereby generating non-preset spatial boundaries and triggering rules that dynamically change with the operational environment. At the same time, the system can, after each operation, reverse-correct the strength of its stored causal relationships and the probability of temporal conflicts based on the difference between the actual deviation trajectory and the predicted deduction results, thereby breaking the technical limitations of traditional safety systems that "data cannot be converted into knowledge, rules cannot evolve, and intervention relies on manual intervention." Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for risk pre-control and proactive management of high-risk operations, so as to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions: A method for risk prevention and proactive management of high-risk operations includes the following steps: S1: Collect the text description of the proposed work plan, real-time environmental parameters of the work area, continuous physiological characteristics of the workers, and historical accident records to form the raw data set; S2: Transform the original dataset into a risk memory with semantic associations. The risk memory records the causal strength between work elements using a causal relationship diagram and records the probability of temporal conflicts between different work steps using a probability matrix. S3: Receive the execution plan and perform counterfactual deduction on the execution plan based on the risk memory, that is, simulate the change of risk evolution path after changing any operation conditions, and generate a pre-control plan that includes dynamic spatial boundaries, dynamic time windows and preset trigger rules. S4: During the operation, continuously receive the fused on-site video semantics, audio spectrum and sensor numerical stream, and measure the deviation between the fused multi-source features and the dynamic spatial boundary and time sequence rules in the pre-control scheme in real time. When the deviation measurement value exceeds the preset trigger rule, generate an active control command. S5: After the operation is completed, extract the deviation trajectory of the multi-source features in the actual records and compare it with the predicted deviation trajectory obtained by counterfactual inference in the pre-control scheme. Based on the difference value obtained by comparison, update the causal strength and temporal conflict probability in the risk memory in reverse.

[0006] As a further aspect of the present invention: S2 specifically includes: Extract the job location identifier, job category label, personnel qualification level, and environmental parameter values ​​from the original dataset as job element entities; Based on the order of co-occurrence of every two operational element entities in the historical accident record text, a directed causal edge is constructed from the antecedent to the consequence. The ratio of the number of times the directed causal edge appears in all records to the number of times the antecedent entity appears alone is used as the causal intensity. The proposed work plan is broken down into a sequence of work steps arranged in chronological order. The number of times the time windows of each pair of steps overlap in the historical record is counted. The ratio of the number of window overlaps to the total number of operations in the historical record is used as the probability of a time conflict between the two steps and is filled into the corresponding row and column positions of the probability matrix.

[0007] As a further aspect of the present invention: the calculation process for the causative intensity is as follows: For each historical accident record text, the event short sentence sequence is segmented according to the time sequence, and for each adjacent event short sentence pair, the subject that performs the action is marked as the cause entity and the object that receives the action is marked as the consequence entity, and directed candidate edges from the cause entity to the consequence entity are generated. The total number of times the same directed candidate edge is generated in all historical incident records is accumulated, and the independent number of times the same antecedent entity appears as the antecedent of any directed candidate edge is also accumulated. Dividing the total number of times by the number of independent times yields the causal strength of the directed candidate edge.

[0008] As a further aspect of the present invention: S3 specifically includes: Extract the start time, duration, work location, and required personnel qualifications for each work step from the proposed work plan, and use them as the set of work conditions to be simulated; One condition is selected from the set of operating conditions and its value is changed one by one, while the other conditions remain unchanged. Based on the causal strength and temporal conflict probability in the risk memory, the propagation path of the causal chain triggered by the change and the risk value changes of each node on the path are tracked to obtain at least one alternative risk evolution path. All alternative risk evolution paths are superimposed and compared with the original risk evolution paths when no conditions are changed. The spatiotemporal position where the risk value first exceeds the preset threshold on each alternative path is taken as the dynamic spatial boundary and dynamic time window. The combination of conditions that trigger the original risk is traced back from the spatiotemporal position as the preset triggering rule.

[0009] As a further aspect of the present invention: the process of obtaining the alternative risk evolution path is as follows: The operational condition whose value has been changed is used as the propagation starting point, and an initial risk trigger mark is assigned to the propagation starting point. Find all directed causal edges in the risk memory with the current node as the antecedent entity, and propagate the initial risk trigger mark along each directed causal edge to the corresponding consequence entity. Each time it is propagated, multiply the current causal strength value by the risk value of the previous node to obtain the cumulative risk value of the next node. If the cumulative risk value of a node is lower than the preset propagation termination threshold, then the propagation from the node to the outside will stop; otherwise, it will continue to propagate downwards along its outgoing edge. All nodes that have been propagated to and whose cumulative risk value exceeds the preset propagation termination threshold are connected in the order they were visited, forming an alternative risk evolution path.

[0010] As a further aspect of the present invention: S4 specifically includes: Extract the three-dimensional spatial coordinates of the workers currently occupying the scene from the fused on-site video semantics, extract the matching confidence of the sound source azimuth angle and the preset warning word library from the audio spectrum, and extract the first-order rate of change of physical quantities from the sensor numerical stream. The system performs an internal and external inclusion judgment between the three-dimensional spatial coordinates and the dynamic spatial boundary, and performs an interval inclusion judgment between the time elapsed in the current operation step and the dynamic time window. At the same time, it compares the matching confidence and the first-order rate of change of the physical quantity with the corresponding thresholds in the preset trigger rules. When any judgment or comparison result exceeds the allowed range, the deviation event count is incremented once, and when the count reaches the upper limit of the number of deviations specified in the preset trigger rules, an active control instruction is generated.

[0011] As a further aspect of the present invention: the step of extracting the matching confidence level between the sound source azimuth angle and the preset warning word database from the audio spectrum specifically includes: The continuously acquired audio stream is divided into time frames with a fixed duration, and the frequency domain transformation is performed on each frame to obtain the frequency band energy distribution; Based on the time difference and energy difference of energy arriving at the two microphones in different frequency bands, the horizontal azimuth angle of the sound source relative to the acquisition center is calculated, and the output is the azimuth angle of the sound source. The frequency band energy distribution is compared frame by frame with the standard frequency band template of each warning word in the preset warning word library. The ratio of the number of time frames that match to the total number of continuous frames of the warning word is recorded as the matching confidence. The sound source azimuth sequence obtained within the current time window is time-aligned with the matching confidence sequence and then output.

[0012] As a further aspect of the present invention: S5 specifically includes: The actual recorded multi-source feature deviation trajectory is discretized into a deviation point sequence in chronological order, and the predicted deviation trajectory obtained from the counterfactual deduction in the pre-control scheme is discretized into the corresponding predicted deviation point sequence. Calculate the distance difference between the actual deviation point and the predicted deviation point at the same time location, and take the average of the distance differences for all time locations as the overall difference value; The overall difference value is compared with the first update threshold. If it exceeds the first update threshold, the causal strength related to the current deviation trajectory in the risk memory is reduced. If it is lower than the second update threshold, the causal strength is increased. The actual overlapping steps and their durations in this task are used to update the temporal conflict probability at the corresponding positions in the probability matrix.

[0013] As a further aspect of the present invention: the discretization of the predicted deviation from the trajectory obtained from counterfactual deduction in the pre-control scheme specifically includes: The start time of the work steps in the proposed work plan is selected as the time origin, and the entire work duration is divided into discrete time slices according to a fixed time step. For each time slice, the operation of changing the working conditions one by one and tracing the causal chain propagation path is repeated, and the spatial position offset of the risk value first exceeding the preset threshold is recorded on each alternative risk evolution path under the time slice. Arrange the spatial position offsets recorded in all time slices in chronological order to form a two-dimensional point column consisting of time values ​​and offset values, which serves as the discretized predicted deviation trajectory. Connect adjacent points in a two-dimensional point array with straight line segments to form a continuous track shape for subsequent difference comparison.

[0014] A high-risk operation risk pre-control and proactive management system includes: The data acquisition module is used to collect text descriptions of the proposed work plan, real-time environmental parameters of the work area, continuous physiological characteristics of the workers, and historical accident records to form a raw data set. The risk memory construction module transforms the original dataset into a risk memory with semantic associations. The risk memory records the causal strength between operational elements using a causal relationship diagram and records the probability of temporal conflicts between different operational steps using a probability matrix. The pre-control module is used to receive the execution operation plan and perform counterfactual deduction on the execution operation plan based on the risk memory, that is, to simulate the change of risk evolution path after changing any operation conditions, and generate a pre-control plan that includes dynamic spatial boundaries, dynamic time windows and preset trigger rules. The real-time deviation measurement module continuously receives the fused on-site video semantics, audio spectrum, and sensor data streams during the operation. It performs real-time deviation measurement between the fused multi-source features and the dynamic spatial boundaries and time sequence rules in the pre-control scheme. When the deviation measurement value exceeds the preset trigger rule, it generates an active control command. The feedback update module extracts the deviation trajectory of the multi-source features recorded in the actual operation and compares it with the predicted deviation trajectory obtained from the counterfactual inference in the pre-control scheme. Based on the difference value obtained from the comparison, it updates the causal strength and temporal conflict probability in the risk memory in reverse.

[0015] The beneficial effects of this invention are: (1) This invention, by constructing a risk memory and performing counterfactual reasoning, can automatically identify potential conflict risks and generate dynamic pre-control schemes before operations begin. During operations, through multimodal feature fusion and real-time deviation measurement, proactive control instructions can be directly issued to the on-site execution agency, achieving a rapid closed loop from perception to intervention without the need for manual confirmation at each level. This method elevates safety management from passive alarm to proactive intervention, reducing the risk of accident escalation due to human response delays.

[0016] (2) After each operation, this invention automatically extracts the difference between the actual deviation trajectory and the predicted deviation trajectory, and updates the causal strength and temporal conflict probability in the risk memory accordingly. As the number of operations accumulates, the causal relationships in the risk memory gradually approach the risk evolution law of real industrial scenarios, making the pre-control plan for subsequent operations more accurate. This mechanism realizes the continuous accumulation and optimization of experience knowledge, avoiding long-term dependence on fixed rules and expert experience. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system block diagram of the present invention. Detailed Implementation

[0019] 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.

[0020] Please see Figure 1 As shown, this invention is a method for risk pre-control and proactive management of high-risk operations, comprising the following steps: S1: Collect the text description of the proposed work plan, real-time environmental parameters of the work area, continuous physiological characteristics of the workers, and historical accident records to form the raw data set; S2: Transform the original dataset into a risk memory with semantic associations. The risk memory records the causal strength between work elements using a causal relationship diagram and records the probability of temporal conflicts between different work steps using a probability matrix. S3: Receive the execution plan and perform counterfactual deduction on the execution plan based on the risk memory, that is, simulate the change of risk evolution path after changing any operation conditions, and generate a pre-control plan that includes dynamic spatial boundaries, dynamic time windows and preset trigger rules. S4: During the operation, continuously receive the fused on-site video semantics, audio spectrum and sensor numerical stream, and measure the deviation between the fused multi-source features and the dynamic spatial boundary and time sequence rules in the pre-control scheme in real time. When the deviation measurement value exceeds the preset trigger rule, generate an active control command. S5: After the operation is completed, extract the deviation trajectory of the multi-source features in the actual records and compare it with the predicted deviation trajectory obtained by counterfactual inference in the pre-control scheme. Based on the difference value obtained by comparison, update the causal strength and temporal conflict probability in the risk memory in reverse.

[0021] In S1, the raw data set is composed of a text description of the proposed work plan, real-time environmental parameters of the work area, continuous physiological characteristics of the workers, and historical accident records. Specifically, it includes: Collect a text description of the proposed work plan. The text description is directly exported from the company's approved electronic work ticket system and should include at least the work content, start and end times, work location coordinates, work step sequence, number of personnel required and their qualification requirements, and a list of equipment to be used.

[0022] Real-time environmental parameters of the work area are collected. Industrial gas detectors (for detecting the concentration of combustible and toxic gases), temperature and humidity sensors, anemometers, and noise sensors are deployed near the work location. These sensors upload data to the on-site data collection terminal at a sampling frequency of once per second.

[0023] Continuous physiological data of workers are collected. Each person entering the work area is equipped with a smart wristband or smart safety helmet liner patch with functions including heart rate, body temperature, skin conductance, and accelerometer. These wearable devices record the personnel's physiological data at a frequency of once every two seconds and transmit it to the on-site data collection terminal via short-range wireless communication.

[0024] Collect historical accident records. Export all accident investigation reports, near miss records, and violation penalty records from the past three years in text format from the enterprise safety management database. Each record should include at least a description of the accident, an analysis of the cause of the accident, the type of work, environmental conditions, information on the personnel involved, and the severity of the consequences.

[0025] The collected text descriptions of the proposed work plan, real-time environmental parameters of the work area, continuous physiological characteristics of the workers, and historical accident records are all stored in the on-site data collection terminal to form a raw data set for use in subsequent steps.

[0026] In S2, the original dataset is transformed into a risk memory with semantic associations. The risk memory records the causal strength between operational elements using a causal relationship graph and records the probability of temporal conflicts between different operational steps using a probability matrix, specifically including: The original dataset is transformed into a risk memory with semantic associations. This risk memory consists of two parts: the first part is the causal strength between operational elements recorded in the form of a causal relationship graph, and the second part is the probability of temporal conflicts between different operational steps recorded in the form of a probability matrix.

[0027] The work element entities are extracted from the original dataset. Specifically, the work location identifier, work category label, and personnel qualification level are extracted from the text description of the proposed work plan; the environmental parameter values ​​corresponding to each sensor measurement point are extracted from the real-time environmental parameters of the work area; and the ranges of physiological indicators such as heart rate and body temperature are extracted from the continuous physiological characteristics of the workers. These extracted items are collectively referred to as work element entities, and each entity is accompanied by its value or status description.

[0028] Construct a causal relationship graph and calculate the causal strength. Based on the co-occurrence order of every two operational element entities in the historical accident record text, construct directed causal edges from antecedent to consequence. Specifically: traverse each historical accident record text, detect two operational element entities that appear sequentially, take the earlier-appearing entity as the antecedent and the later-appearing entity as the consequence, and generate a directed candidate edge from the antecedent to the consequence. After completing the above operation for all historical accident records, count the total number of times each directed candidate edge appears in all records, and simultaneously count the independent number of times each antecedent entity appears alone as the antecedent of any directed candidate edge. Divide the total number of occurrences by the independent number of occurrences, and the resulting ratio is the causal strength of the directed candidate edge. This ratio is between 0 and 1, representing the conditional probability that the occurrence of the antecedent leads to the occurrence of the consequence.

[0029] To ensure the accuracy of causal intensity calculation, each historical accident record text is segmented into event phrases. Specifically, each accident record text is divided into several event phrases using periods, semicolons, or line breaks as delimiters, and these phrases are arranged in chronological order. For each pair of adjacent event phrases (i.e., phrase 1 and phrase 2, phrase 2 and phrase 3, and so on), a dependency syntax-based annotation method is used to identify the noun phrases that perform the action as the antecedent entity and the noun phrases that receive the action as the consequent entity. If both the antecedent and consequent entities can be clearly identified in an adjacent event phrase pair, a directed candidate edge is generated from the antecedent entity to the consequent entity; otherwise, the phrase pair is skipped. After processing all historical accident record texts, a set of directed candidate edges is obtained. The total number of times the same directed candidate edge is generated across all records is summed. For example, the directed candidate edge "not wearing a helmet" pointing to "head injury" appears 15 times in all accident records. Simultaneously, the independent number of times the antecedent entity "not wearing a helmet" appears as an antecedent of any directed candidate edge is summed. For example, "not wearing a helmet" appears as an antecedent in 20 records (regardless of the consequence). Dividing the total number of 15 by the independent number of 20 yields a quotient of 0.75, which is the causal strength of "not wearing a helmet" pointing to "head injury".

[0030] Calculate the probability of timing conflicts. The proposed work plan is broken down into a sequence of chronologically ordered steps, such as steps A, B, and C, according to the steps described in its text description. The execution information of all historical operations is collected from historical accident records: for each historical operation, the actual start and end times of each step are recorded, forming a time window. For any two steps in the sequence (e.g., steps X and Y), all historical operation records are traversed, and the number of times the time windows of these two steps overlap is counted. The number of overlaps is divided by the total number of historical operations; the resulting ratio is the probability of a timing conflict between the two steps. This ratio is then entered into the corresponding row (step X) and column (step Y) of the probability matrix. If steps X and Y are the same, the probability value is recorded as 0. The probability of conflict for each pair of steps is calculated sequentially, resulting in a complete probability matrix, which serves as the second part of the risk memory.

[0031] In S3, the execution plan is received and counterfactually deduced based on the risk memory. This simulates the change in the risk evolution path after altering any operational conditions. Based on this, a pre-control plan is generated, which includes dynamic spatial boundaries, dynamic time windows, and preset triggering rules. Specifically, it includes: The system receives the proposed work plan and performs a counterfactual analysis of the plan based on the risk memory constructed by S2. This involves simulating the changes in the risk evolution path after any changes in work conditions, and generating a pre-control plan that includes dynamic spatial boundaries, dynamic time windows, and preset triggering rules.

[0032] First, extract the set of operational conditions to be simulated from the proposed work plan. Specifically, parse the start time, duration, three-dimensional coordinate range of the location, and personnel qualification level requirements for each operational step from the plan text description. Number these four pieces of information according to the order of the operational steps to form a set of operational conditions containing N steps, denoted as condition set C, where each condition contains four attributes: start time, duration, location coordinates, and qualification level.

[0033] Secondly, select one condition from the set of work conditions one by one and change its value, while keeping the other conditions unchanged. For example, for the i-th work step, delay its start time by 10 minutes, shift its location coordinates to one side by 1 meter, or lower the required personnel qualification level by one level. Each change generates a counterfactual scenario. Then, based on the causal strength and temporal conflict probability in the risk memory, trace the causal chain propagation path triggered by the change and the risk value changes of each node on the path to obtain at least one alternative risk evolution path. The specific propagation process is as follows: take the work condition whose value is changed as the propagation starting point, assign an initial risk trigger mark to the starting point, and set the initial value to 1. Search for all directed causal edges with the current node as the antecedent entity from the causal relationship graph of the risk memory. For each found directed causal edge, propagate the initial risk trigger mark along the edge to the corresponding consequence entity. During propagation, multiply the cumulative risk value of the current node by the causal strength corresponding to the above directed causal edge, and use the product as the cumulative risk value of the next node. That is, if the current node The cumulative risk value is ,from point to The causal strength of the directed causal edge is Then the cumulative risk value of node B Calculate using the following formula: ; in, The value range is from 0 to 1. The value range is from 0 to 1. Similarly, the value is between 0 and 1. After one propagation, the calculated result is assigned to the consequence entity node. Using this consequence entity node as the current node, the process of finding outgoing edges and multiplying the cumulative risk value is repeated until propagation can no longer continue or is terminated. The propagation termination criterion is a preset propagation termination threshold, fixed at 0.05. If the cumulative risk value calculated by the above formula for a node is lower than 0.05, propagation from that node stops; if the cumulative risk value is greater than or equal to 0.05, propagation continues downwards along its outgoing edges. All nodes reached by propagation and with a cumulative risk value greater than or equal to 0.05 are connected sequentially according to the order in which they were visited during propagation, forming an alternative risk evolution path. Since a propagation starting point may radiate outwards along multiple directed causal edges simultaneously, multiple alternative risk evolution paths can usually be obtained.

[0034] Then, all alternative risk evolution paths are superimposed and compared with the original risk evolution path without changing any conditions. The original risk evolution path is obtained by: without changing any operating conditions, taking each operating step in the operating plan as the starting point, traversing the causal relationship graph according to the same propagation rules as described above to obtain the original risk evolution path. For each alternative risk evolution path, the cumulative risk value of each node is checked sequentially along the path, and the position where the cumulative risk value on the path first exceeds a preset risk threshold is taken. The preset risk threshold is fixed at 0.8. The spatial coordinates and the time of occurrence at this position are recorded as the dynamic spatial boundary and the dynamic time window, respectively. For example, if the cumulative risk value of the 3rd node on a certain path is 0.75 (below 0.8) and the cumulative risk value of the 4th node is 0.85 (first exceeding 0.8), then the operating position coordinates corresponding to the 4th node are taken as the dynamic spatial boundary, and the time of occurrence of this node is taken as the start and end points of the dynamic time window. Multiple alternative paths may generate multiple different spatial boundaries and time windows. The smallest bounding cube of all boundaries is taken as the final dynamic spatial boundary, and the smallest intersection of all time windows is taken as the final dynamic time window. Simultaneously, backtracking begins from the node position where the threshold is first exceeded, sequentially tracing back its preceding nodes until the initial altered operational condition is reached. The condition combinations corresponding to all nodes on this backtracking chain are recorded as preset trigger rules. For example, a preset trigger rule might be: "Trigger high-risk intervention when the operator's position deviates from the original coordinates by more than 1 meter and the ambient temperature is above 40 degrees Celsius."

[0035] Through the above counterfactual deduction, a complete pre-control scheme is finally generated, which includes dynamic spatial boundaries, dynamic time windows, and preset triggering rules, for real-time deviation measurement in subsequent operations.

[0036] In S4, during the execution of the operation, the system continuously receives the fused on-site video semantics, audio spectrum, and sensor data streams. It then performs real-time deviation measurement between the fused multi-source features and the dynamic spatial boundaries and time-series rules in the pre-control scheme. When the deviation measurement value exceeds the preset trigger rules, an active control command is generated, specifically including: During the operation, the system continuously receives the fused on-site video semantics, audio spectrum, and sensor data streams. It then performs real-time deviation measurement between the fused multi-source features and the dynamic spatial boundaries and timing rules in the pre-control scheme generated by S3. When the deviation measurement value exceeds the preset triggering rule, an active control command is generated.

[0037] The 3D spatial coordinates of the workers are extracted from the semantics of the fused on-site video. Specifically, a binocular camera is used to capture real-time footage of the work area at a rate of 25 frames per second. For each frame, a contour detection-based method is used to identify the worker's center point. The 3D spatial coordinates (lateral distance, longitudinal distance, and depth distance) are then constructed using the center point's lateral and longitudinal positions in the camera coordinate system, as well as the depth distance calculated from the binocular parallax. These 3D spatial coordinates are updated every 40 milliseconds.

[0038] The sound source azimuth angle is extracted from the audio spectrum to determine the matching confidence level with a preset warning word library. Specifically, a pair of omnidirectional microphones spaced 15 cm apart are deployed at the work site, synchronously acquiring dual-channel audio streams at a sampling rate of 16,000 times per second. The continuously acquired audio stream is divided into time frames every 30 milliseconds, with adjacent frames overlapping by 15 milliseconds. For each time frame, a Fast Fourier Transform is applied to convert the time-domain signal into an energy distribution in the frequency domain, obtaining a sequence of energy values ​​for each frequency band from low to high frequencies. Based on the arrival time difference of low-frequency energy between the two microphone channels within the same time frame (i.e., the small time difference between the sound arriving at the left or right microphone first) and the energy attenuation difference in the mid-to-high frequency bands, the horizontal angle of the sound source relative to the center points of the two microphones is calculated. This horizontal angle ranges from -90 degrees to +90 degrees, where 0 degrees represents directly in front, negative values ​​represent the left, and positive values ​​represent the right. This horizontal angle value is output as the sound source azimuth angle. Simultaneously, a warning word library is pre-established, containing eight warning words such as "firefighting," "leakage," "collapse," and "fracture." Each warning word corresponds to a pre-extracted standard frequency band template, which records the typical distribution curve of the energy in each frequency band when the warning word is pronounced normally. The frequency band energy distribution of the current time frame is compared frame-by-frame with the standard frequency band template of each warning word in the warning word library: the starting position of the time frame is moved sample by sample point on the audio stream, and after each move, the sum of the absolute values ​​of the differences between the frequency band energy distribution of the current frame and the standard frequency band template is calculated. When this sum is lower than a preset matching threshold, the time frame is considered to match the warning word. The number of consecutive matching time frames is recorded, and this number is divided by the total number of standard continuous frames corresponding to the complete pronunciation of the warning word (e.g., the pronunciation of "firefighting" lasts approximately 60 frames). The resulting ratio is used as the matching confidence of the warning word. The maximum matching confidence among all warning words is taken as the output value of this time window. Align the sound source azimuth sequence (one angle value per frame) obtained within the current time window with the matching confidence sequence (one confidence value per frame) according to the time order of the frames, and output them to the subsequent judgment steps.

[0039] The first-order rate of change of a physical quantity is extracted from the sensor data stream. Specifically, the current physical quantity value (e.g., the volume percentage concentration of combustible gas) is continuously read from devices such as industrial gas detectors, temperature and humidity sensors, and anemometers deployed in the work area. Each time a value is read, the previous value is subtracted to obtain the difference, which is then divided by the time interval between the two readings (fixed to 1 second). The quotient is the first-order rate of change, representing how fast the physical quantity changes per second.

[0040] A deviation metric is performed. The extracted 3D spatial coordinates are compared with the dynamic spatial boundary generated by S3 (a 3D cube region defined by the minimum horizontal coordinate, maximum horizontal coordinate, minimum vertical coordinate, maximum vertical coordinate, minimum depth coordinate, and maximum depth coordinate) for inclusion / exclusion judgment: if all three components of the 3D spatial coordinates fall within the range of the six coordinate values ​​of the boundary cube, it is determined to be inside; otherwise, it is outside. Simultaneously, the duration consumed by the current operation step since its start time is compared with the dynamic time window generated by S3 (a time period defined by the start and end times) for interval inclusion judgment: if the current time is after the start time and before the end time of the time window, it is determined to be inside; otherwise, it is outside. Finally, the matching confidence is compared with the corresponding matching confidence threshold (fixed at 0.7) in the preset trigger rule: if the matching confidence is greater than or equal to 0.7, it is determined to be outside the allowable range; if it is less than 0.7, it is determined to be within the allowable range. At the same time, the first-order rate of change of the physical quantity is compared with the corresponding first-order rate of change threshold in the preset triggering rules (for example, the upper limit of the first-order rate of change of the volume percentage concentration of combustible gas is a positive percentage of 0.5 per second): if the first-order rate of change is greater than the upper limit, it is determined that it exceeds the allowable range.

[0041] If any of the following four conditions occur: the spatial coordinates are outside the boundary, the current time is outside the time window, the matching confidence level reaches or exceeds 0.7, or the first-order rate of change exceeds the upper limit, then the deviation event count is incremented by one. The initial value of the deviation event count is 0, and the count increases by 1 each time any judgment exceeds the allowable range. When the accumulated deviation event count reaches the upper limit of the number of deviations specified in the preset trigger rules (fixed at 3 times), an active control command is immediately generated. This active control command is issued in the form of a binary signal and sent to the audible and visual alarms, the vibrator of the personnel's wearable equipment, and the emergency stop interface of the production control system on site via wired or wireless network.

[0042] In S5, after the operation is completed, the deviation trajectories of the actual recorded multi-source features are extracted and compared with the predicted deviation trajectories obtained from counterfactual inference in the pre-control scheme. Based on the difference values ​​obtained from the comparison, the causal strength and temporal conflict probability in the risk memory are updated in reverse, specifically including: After the operation is completed, the deviation trajectory of the multi-source features extracted from the actual records is compared with the predicted deviation trajectory obtained from the counterfactual inference in the pre-control scheme. Based on the difference value obtained from the comparison, the causal strength and temporal conflict probability in the risk memory are updated in reverse.

[0043] The actual recorded multi-source feature deviation trajectories are discretized into a sequence of deviation points in chronological order. Specifically, during the operation, step S4 records the time of each deviation event occurring once per second, along with information such as the operator's three-dimensional spatial coordinates, sound source azimuth, matching confidence, and first-order rate of change of physical quantities at that moment. This information is then arranged chronologically to form the actual deviation point sequence. Each point contains a time value and a comprehensive deviation feature value for that moment (the comprehensive deviation feature value is obtained by weighted summation of the aforementioned four deviation indicators, with weights of 0.4, 0.2, 0.2, and 0.2, respectively).

[0044] The predicted deviation from the pre-control scheme, derived from counterfactual deduction, is discretized into a sequence of predicted deviation points. The discretization process is as follows: The start time of the first operation step in the proposed operation scheme is selected as the time origin. The entire estimated operation duration (e.g., 120 minutes) is divided into discrete time slices according to a fixed time step (e.g., 30 seconds), resulting in 240 time slices. For each time slice (i.e., each discrete time point), the operation described in S3, which involves changing the operation conditions one by one and tracing the causal chain propagation path, is repeated. Specifically, the time corresponding to the time slice is used as the trigger time. The values ​​of each operation condition at that time are changed sequentially. Causal chain propagation is performed based on the causal strength in the risk memory. The spatial position of the node on each alternative risk evolution path where the accumulated risk value first exceeds the preset threshold (threshold is 0.8) is recorded. The distance offset of this spatial position relative to the preset position in the original operation scheme is calculated, and this offset is used as the spatial position offset for that time slice. Using the time value corresponding to each time slice as the x-axis and the spatial offset obtained under that time slice as the y-axis, a two-dimensional point column is formed by the time value and the offset value (for example, the first slice: time 0 seconds, offset 0 meters; the second slice: time 30 seconds, offset 0.2 meters; and so on). This two-dimensional point column is the discretized predicted deviation trajectory. In order to handle continuous time points during subsequent difference comparison, adjacent two points in the two-dimensional point column are connected by a straight line segment to form a continuous trajectory shape, so that the corresponding predicted offset can be obtained by interpolation at any time.

[0045] Calculate the overall difference value. Specifically, for each actual deviation point in the actual deviation point sequence, find the predicted offset at the same time on the continuous trajectory of the predicted deviation trajectory. Calculate the absolute value of the difference between the comprehensive deviation characteristic value of the actual deviation point at that time (this value was recorded in the first step) and the predicted offset, and use this as the distance difference at that time. Repeat the above calculation for all times with actual deviation points to obtain a set of distance difference values. Add these distance difference values ​​together and divide them by the total number of actual deviation points; the quotient obtained is the overall difference value.

[0046] The causal strength is updated based on the overall difference value. A first update threshold of 0.3 and a second update threshold of 0.05 are set. The calculated overall difference value is compared with the first update threshold: if the overall difference value is greater than 0.3, the actual risk deviation of this operation is determined to be much higher than the prediction of the pre-control plan, indicating that the causal strength related to this operation in the risk memory is underestimated. Therefore, the values ​​of these causal strengths are reduced by the difference between the overall difference value multiplied by 0.1 and the original causal strength. If the overall difference value is less than 0.05, the actual deviation is determined to be much lower than the prediction, indicating that the relevant causal strength is overestimated. Therefore, the values ​​of these causal strengths are increased by the sum of the overall difference value multiplied by 0.1 and the original causal strength. If the overall difference value is between 0.05 and 0.3 (inclusive), the original causal strength remains unchanged.

[0047] Update the probability of timing conflicts. During the actual execution of this task, record the actual start and end times of each task step, forming a time window for each step. For any two different steps in the proposed task plan (e.g., step A and step B), determine whether their time windows overlap (i.e., the start time of one step is earlier than the end time of the other step but later than its start time). If there is an overlap, record this pair of steps as an overlapping step pair and record the overlap duration (i.e., the number of minutes of the overlap). Use the newly generated overlapping step pairs and their overlap durations in this task to update the probability of timing conflicts at the corresponding positions in the probability matrix: in the original probability matrix, the probability value of the timing conflict corresponding to this step pair is updated to (original value multiplied by the total number of historical tasks plus the current overlap marker value of 1) divided by (total number of historical tasks plus 1); where the overlap duration is more than 5 minutes, it is marked as 1, and less than 5 minutes, it is marked as 0.5. Process all overlapping step pairs in sequence to complete the update of the probability matrix. The updated risk memory will be used for counterfactual inference in the next task.

[0048] Please see Figure 2 As shown, a high-risk operation risk pre-control and proactive management system includes: The data acquisition module is used to collect text descriptions of the proposed work plan, real-time environmental parameters of the work area, continuous physiological characteristics of the workers, and historical accident records to form a raw data set. The risk memory construction module transforms the original dataset into a risk memory with semantic associations. The risk memory records the causal strength between operational elements using a causal relationship diagram and records the probability of temporal conflicts between different operational steps using a probability matrix. The pre-control module is used to receive the execution operation plan and perform counterfactual deduction on the execution operation plan based on the risk memory, that is, to simulate the change of risk evolution path after changing any operation conditions, and generate a pre-control plan that includes dynamic spatial boundaries, dynamic time windows and preset trigger rules. The real-time deviation measurement module continuously receives the fused on-site video semantics, audio spectrum, and sensor data streams during the operation. It performs real-time deviation measurement between the fused multi-source features and the dynamic spatial boundaries and time sequence rules in the pre-control scheme. When the deviation measurement value exceeds the preset trigger rule, it generates an active control command. The feedback update module extracts the deviation trajectory of the multi-source features recorded in the actual operation and compares it with the predicted deviation trajectory obtained from the counterfactual inference in the pre-control scheme. Based on the difference value obtained from the comparison, it updates the causal strength and temporal conflict probability in the risk memory in reverse.

[0049] The working principle of this invention is as follows: A raw data set is formed by collecting textual descriptions of the proposed work plan, real-time environmental parameters of the work area, continuous physiological characteristics of the workers, and historical accident records. This raw data set is then transformed into a semantically related risk memory. This risk memory records the causal strength between work elements using a causal relationship graph and the probability of temporal conflicts between different work steps using a probability matrix. Next, the work plan is received, and counterfactual analysis is performed on the work plan based on the risk memory to simulate changes in the risk evolution path after altering any work conditions. Based on this, a pre-control plan containing dynamic spatial boundaries, dynamic time windows, and preset triggering rules is generated. During the work execution, fused on-site video semantics, audio spectrum, and sensor data streams are continuously received. The multi-source features obtained from the fusion are compared with the dynamic spatial boundaries and temporal rules in the pre-control plan in real time to measure deviations. When the deviation measurement value exceeds the preset triggering rules, an active control command is generated. Finally, after the work is completed, the actual recorded multi-source feature deviation trajectory is extracted and compared with the predicted deviation trajectory obtained from the counterfactual analysis in the pre-control plan. The causal strength and temporal conflict probability in the risk memory are updated in reverse based on the comparison difference value.

[0050] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for risk pre-control and proactive management of high-risk operations, characterized in that, Includes the following steps: S1: Collect the text description of the proposed work plan, real-time environmental parameters of the work area, continuous physiological characteristics of the workers, and historical accident records to form the raw data set; S2: Transform the original dataset into a risk memory with semantic associations. The risk memory records the causal strength between work elements using a causal relationship diagram and records the probability of temporal conflicts between different work steps using a probability matrix. S3: Receive the execution plan and perform counterfactual deduction on the execution plan based on the risk memory, that is, simulate the change of risk evolution path after changing any operation conditions, and generate a pre-control plan that includes dynamic spatial boundaries, dynamic time windows and preset trigger rules. S4: During the operation, continuously receive the fused on-site video semantics, audio spectrum and sensor numerical stream, and measure the deviation between the fused multi-source features and the dynamic spatial boundary and time sequence rules in the pre-control scheme in real time. When the deviation measurement value exceeds the preset trigger rule, generate an active control command. S5: After the operation is completed, extract the deviation trajectory of the multi-source features in the actual records and compare it with the predicted deviation trajectory obtained by counterfactual inference in the pre-control scheme. Based on the difference value obtained by comparison, update the causal strength and temporal conflict probability in the risk memory in reverse.

2. The method for risk pre-control and proactive management of high-risk operations according to claim 1, characterized in that, S2 specifically includes: Extract the job location identifier, job category label, personnel qualification level, and environmental parameter values ​​from the original dataset as job element entities; Based on the order of co-occurrence of every two operational element entities in the historical accident record text, a directed causal edge is constructed from the antecedent to the consequence. The ratio of the number of times the directed causal edge appears in all records to the number of times the antecedent entity appears alone is used as the causal intensity. The proposed work plan is broken down into a sequence of work steps arranged in chronological order. The number of times the time windows of each pair of steps overlap in the historical record is counted. The ratio of the number of window overlaps to the total number of operations in the historical record is used as the probability of a time conflict between the two steps and is filled into the corresponding row and column positions of the probability matrix.

3. The method for risk pre-control and proactive management of high-risk operations according to claim 2, characterized in that, The calculation process for the causative intensity is as follows: For each historical accident record text, the event short sentence sequence is segmented according to the time sequence, and for each adjacent event short sentence pair, the subject that performs the action is marked as the cause entity and the object that receives the action is marked as the consequence entity, and directed candidate edges from the cause entity to the consequence entity are generated. The total number of times the same directed candidate edge is generated in all historical incident records is accumulated, and the independent number of times the same antecedent entity appears as the antecedent of any directed candidate edge is also accumulated. Dividing the total number of times by the number of independent times yields the causal strength of the directed candidate edge.

4. The method for risk pre-control and proactive management of high-risk operations according to claim 1, characterized in that, S3 specifically includes: Extract the start time, duration, work location, and required personnel qualifications for each work step from the proposed work plan, and use them as the set of work conditions to be simulated; One condition is selected from the set of operating conditions and its value is changed one by one, while the other conditions remain unchanged. Based on the causal strength and temporal conflict probability in the risk memory, the propagation path of the causal chain triggered by the change and the risk value changes of each node on the path are tracked to obtain at least one alternative risk evolution path. All alternative risk evolution paths are superimposed and compared with the original risk evolution paths when no conditions are changed. The spatiotemporal position where the risk value first exceeds the preset threshold on each alternative path is taken as the dynamic spatial boundary and dynamic time window. The combination of conditions that trigger the original risk is traced back from the spatiotemporal position as the preset triggering rule.

5. A method for risk pre-control and proactive management of high-risk operations according to claim 4, characterized in that, The process of obtaining the alternative risk evolution path is as follows: The operational condition whose value has been changed is used as the propagation starting point, and an initial risk trigger mark is assigned to the propagation starting point. Find all directed causal edges in the risk memory with the current node as the antecedent entity, and propagate the initial risk trigger mark along each directed causal edge to the corresponding consequence entity. Each time it is propagated, multiply the current causal strength value by the risk value of the previous node to obtain the cumulative risk value of the next node. If the cumulative risk value of a node is lower than the preset propagation termination threshold, then the propagation from the node to the outside will stop; otherwise, it will continue to propagate downwards along its outgoing edge. All nodes that have been propagated to and whose cumulative risk value exceeds the preset propagation termination threshold are connected in the order they were visited, forming an alternative risk evolution path.

6. The method for risk pre-control and proactive management of high-risk operations according to claim 1, characterized in that, S4 specifically includes: Extract the three-dimensional spatial coordinates of the workers currently occupying the scene from the fused on-site video semantics, extract the matching confidence of the sound source azimuth angle and the preset warning word library from the audio spectrum, and extract the first-order rate of change of physical quantities from the sensor numerical stream. The system performs an internal and external inclusion judgment between the three-dimensional spatial coordinates and the dynamic spatial boundary, and performs an interval inclusion judgment between the time elapsed in the current operation step and the dynamic time window. At the same time, it compares the matching confidence and the first-order rate of change of the physical quantity with the corresponding thresholds in the preset trigger rules. When any judgment or comparison result exceeds the allowed range, the deviation event count is incremented once, and when the count reaches the upper limit of the number of deviations specified in the preset trigger rules, an active control instruction is generated.

7. A method for risk pre-control and proactive management of high-risk operations according to claim 6, characterized in that, The step of extracting the sound source azimuth angle from the audio spectrum and the matching confidence score with the preset warning word library specifically includes: The continuously acquired audio stream is divided into time frames with a fixed duration, and the frequency domain transformation is performed on each frame to obtain the frequency band energy distribution; Based on the time difference and energy difference of energy arriving at the two microphones in different frequency bands, the horizontal azimuth angle of the sound source relative to the acquisition center is calculated, and the output is the azimuth angle of the sound source. The frequency band energy distribution is compared frame by frame with the standard frequency band template of each warning word in the preset warning word library. The ratio of the number of time frames that match to the total number of continuous frames of the warning word is recorded as the matching confidence. The sound source azimuth sequence obtained within the current time window is time-aligned with the matching confidence sequence and then output.

8. The method for risk pre-control and proactive management of high-risk operations according to claim 1, characterized in that, S5 specifically includes: The actual recorded multi-source feature deviation trajectory is discretized into a deviation point sequence in chronological order, and the predicted deviation trajectory obtained from the counterfactual deduction in the pre-control scheme is discretized into the corresponding predicted deviation point sequence. Calculate the distance difference between the actual deviation point and the predicted deviation point at the same time location, and take the average of the distance differences for all time locations as the overall difference value; The overall difference value is compared with the first update threshold. If it exceeds the first update threshold, the causal strength related to the current deviation trajectory in the risk memory is reduced. If it is lower than the second update threshold, the causal strength is increased. The actual overlapping steps and their durations in this task are used to update the temporal conflict probability at the corresponding positions in the probability matrix.

9. A method for risk pre-control and proactive management of high-risk operations according to claim 8, characterized in that, The discretization of the predicted deviation from the trajectory obtained from counterfactual deduction in the pre-control scheme specifically includes: The start time of the work steps in the proposed work plan is selected as the time origin, and the entire work duration is divided into discrete time slices according to a fixed time step. For each time slice, the operation of changing the working conditions one by one and tracing the causal chain propagation path is repeated, and the spatial position offset of the risk value first exceeding the preset threshold is recorded on each alternative risk evolution path under the time slice. Arrange the spatial position offsets recorded in all time slices in chronological order to form a two-dimensional point column consisting of time values ​​and offset values, which serves as the discretized predicted deviation trajectory. Connect adjacent points in a two-dimensional point array with straight line segments to form a continuous track shape for subsequent difference comparison.

10. A high-risk operation risk pre-control and proactive management system, characterized in that, A method for implementing a high-risk operation risk pre-control and proactive management system as described in any one of claims 1-9, comprising: The data acquisition module is used to collect text descriptions of the proposed work plan, real-time environmental parameters of the work area, continuous physiological characteristics of the workers, and historical accident records to form a raw data set. The risk memory construction module transforms the original dataset into a risk memory with semantic associations. The risk memory records the causal strength between operational elements using a causal relationship diagram and records the probability of temporal conflicts between different operational steps using a probability matrix. The pre-control module is used to receive the execution operation plan and perform counterfactual deduction on the execution operation plan based on the risk memory, that is, to simulate the change of risk evolution path after changing any operation conditions, and generate a pre-control plan that includes dynamic spatial boundaries, dynamic time windows and preset trigger rules. The real-time deviation measurement module continuously receives the fused on-site video semantics, audio spectrum, and sensor data streams during the operation. It performs real-time deviation measurement between the fused multi-source features and the dynamic spatial boundaries and time sequence rules in the pre-control scheme. When the deviation measurement value exceeds the preset trigger rule, it generates an active control command. The feedback update module extracts the deviation trajectory of the multi-source features recorded in the actual operation and compares it with the predicted deviation trajectory obtained from the counterfactual inference in the pre-control scheme. Based on the difference value obtained from the comparison, it updates the causal strength and temporal conflict probability in the risk memory in reverse.

Images