Industrial field personnel behavior specification AI monitoring method and system

Abstract

The invention relates to the technical field of behavior monitoring, and provides an industrial field personnel behavior specification AI monitoring method and system. Performing human body key point detection and time sequence motion feature extraction on the behavior data flow of each station to obtain a behavior feature sequence of each person, and performing semantic similarity matching on the behavior feature sequence according to a behavior semantic type set to generate behavior semantic description data, and performing risk index quantitative evaluation and feature fusion on the behavior feature sequence according to the behavior semantic description data to obtain a risk feature set of each person, and performing illegal behavior identification and evidence fusion on the risk feature set according to a business behavior rule to generate an illegal event record of each person. According to the method, stable features are extracted from multi-source space-time alignment, and a traceable structured violation event is finally generated through semantic mapping and risk fusion, so that the accuracy and response speed of field safety management and control are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of behavior monitoring technology, and in particular to an AI monitoring method and system for regulating the behavior of personnel in industrial settings. Background Technology

[0002] Monitoring personnel behavior in industrial settings has strategic value in directly ensuring personal and equipment safety, reducing accident losses and downtime risks, and meeting legal compliance and insurance requirements. Through continuous monitoring and behavioral analysis of personnel operations, companies can not only promptly identify and address hazardous behaviors and abnormal situations, but also accumulate traceable management evidence, optimize work processes, and guide targeted training, thereby improving production continuity and management efficiency while protecting employees and property.

[0003] Currently, behavioral monitoring in industrial settings primarily relies on manual inspections and CCTV monitoring, supplemented by independent sensors or systems such as access control / RFID card swipe logs, simple video motion detection or threshold-based rule-based alarms, physiological parameter monitoring from wearable devices, and regular safety inspections and training assessments. Enterprises also frequently use industry software from single vendors for statistical reporting and post-event auditing; however, these methods are often fragmented, relying on manual interpretation or static rules, and lacking deep semantic understanding and automated evidence chain management.

[0004] Existing methods often rely on a single sensor source or manual verification, lacking real-time multi-source fusion and behavioral semantic mapping. This results in high false alarm / false negative rates when there is occlusion, changes in lighting, or positional errors. Furthermore, it is difficult to form a structured, traceable chain of evidence and a global security posture, making it impossible to achieve accurate and interpretable real-time early warning and post-event auditing. Summary of the Invention

[0005] In view of this, this application provides an AI monitoring method and system for regulating the behavior of personnel in industrial sites, in order to solve the problems of fragmented perceptual information and insufficient semantic understanding in personnel behavior monitoring.

[0006] The first aspect of this application provides an AI monitoring method for regulating personnel behavior in industrial settings, the method comprising: Human key point detection and temporal motion feature extraction are performed on the obtained behavioral data streams of each workstation to obtain the behavioral feature sequence of each person. Based on a preset set of behavioral semantic types, semantic similarity matching is performed on the behavioral feature sequence to generate behavioral semantic description data; Based on the behavioral semantic description data, the behavioral feature sequence is subjected to risk index quantitative assessment and feature fusion processing to obtain the risk feature set of each person. Based on preset business behavior rules, the risk feature set is used to identify violations and fuse evidence to generate violation event records for each person.

[0007] In an optional implementation, the step of performing human keypoint detection and temporal motion feature extraction on the obtained behavioral data streams of each workstation to obtain the behavioral feature sequence of each person includes: Based on the preset human joint feature data, the spatial coordinate sequence of each key point is extracted from the obtained behavioral data stream of each workstation; The spatial coordinate sequence is smoothed and filtered according to the time order to obtain the motion trajectory data of each key point; The kinematic parameters of each key point are calculated based on the motion trajectory data through a preset sliding timing window; Calculate the joint angles of each key point based on the relative coordinates of the limbs in the motion trajectory data; The kinematic parameters and joint angles are spliced ​​and normalized according to time sequence features to obtain the behavioral feature sequence of each person.

[0008] In an optional implementation, the step of performing semantic similarity matching on the behavioral feature sequence according to a preset set of behavioral semantic types to generate behavioral semantic description data includes: The similarity algorithm calculates the similarity between each behavioral semantic type in the preset behavioral semantic type set and each behavioral feature in the behavioral feature sequence based on a preset unit time period, and generates the probability distribution of behavioral semantic categories corresponding to each time period. Extract the behavioral semantic type corresponding to the maximum probability value from the behavioral semantic category probability distribution, and use it as the optimal semantic type for each time period; Based on the received location data streams for each person, extract the corresponding area scene data for each person from the preset industrial site area database; By using a preset semantic description method, time-series data is fused based on the optimal semantic type and the regional scene data to generate behavioral semantic description data for each person.

[0009] In an optional implementation, the step of performing risk indicator quantification and feature fusion processing on the behavioral feature sequence based on the behavioral semantic description data to obtain the risk feature set of each individual includes: Based on the optimal semantic type, extract the corresponding semantic risk coefficient set from the preset risk weight configuration table; The behavioral feature sequence is classified based on the regional scene data to obtain employee behavioral feature datasets for each regional scene. Then, a preset density clustering algorithm is used to perform cluster analysis on the employee behavioral feature datasets to calculate the behavioral anomaly value of each person. Based on the preset workstation and protective equipment data set and the positioning data stream, extract the personal protective equipment wearing status data of each person from the behavior data stream; Based on the preset dangerous area location dataset and the location data stream, area intrusion detection is performed to calculate the spatial relationship data between each person and the dangerous area; The risk feature set of each individual is obtained by weighted and fused calculation of the behavioral anomaly value, the wearing status data, and the spatial relationship data based on the semantic risk coefficient set.

[0010] In an optional implementation, the step of identifying violations and fusing evidence based on preset business behavior rules of the risk feature set to generate violation event records for each person includes: By using a preset membership function, the risk feature set is matched with a preset set of behaviors and violation types to obtain the probability distribution of violation types; Based on the preset membership threshold and the probability distribution of the violation type, a set of potential violation behavior features and the violation type corresponding to each behavior feature are extracted from the risk feature set; Based on each behavioral feature in the potential violation feature set, the corresponding violation evidence slice data is extracted from the behavioral data stream; Based on the violation type, a consistency analysis and verification is performed on the potential violation behavior feature set and the violation evidence slice data to obtain an evidence dataset; The comprehensive credibility score corresponding to the evidence dataset is calculated using a preset evidence evaluation algorithm; By using a preset evidence fusion method, the comprehensive credibility score and the evidence dataset are fused according to the violation type to generate violation event records for each person.

[0011] In an optional implementation, the method further includes: According to the preset alarm triggering strategy, the violation event records are classified and processed according to the severity of the violation, and graded alarm information is generated. Real-time alarms are triggered based on the real-time alarm information in the tiered alarm information, using a preset alarm method. Based on a preset time period, the hierarchical alarm information is analyzed to generate a list of key individuals to be monitored for violations.

[0012] A second aspect of this application provides an AI monitoring device for monitoring personnel behavior in industrial settings, the device comprising: The feature extraction module is used to perform human key point detection and temporal motion feature extraction on the obtained behavioral data streams of each workstation to obtain the behavioral feature sequence of each person. The semantic description module is used to perform semantic similarity matching on the behavioral feature sequence according to a preset set of behavioral semantic types, and generate behavioral semantic description data. The risk assessment module is used to perform risk indicator quantification assessment and feature fusion processing on the behavioral feature sequence based on the behavioral semantic description data to obtain the risk feature set of each person. The violation identification module is used to identify violations and fuse evidence based on the preset business behavior rules of the risk feature set, and generate violation event records for each person.

[0013] A third aspect of this application provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the AI ​​monitoring method for industrial field personnel behavior norms as described above.

[0014] The fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the AI ​​monitoring method for regulating the behavior of personnel in industrial settings as described above.

[0015] In summary, this application includes at least the following beneficial technical effects: 1. By deploying multi-source spatiotemporal alignment and lightweight real-time detection at the edge, the system can quickly generate hierarchical alarms and push them in real time, thereby shortening the time for accident detection and handling and reducing safety losses.

[0016] 2. By fusing visual skeleton, RFID positioning, PPE detection, ontological semantics and Dempster-Shafer evidence, it can effectively resist occlusion, lighting changes and single sensor errors, and reduce false alarms and missed alarms.

[0017] 3. Violations are recorded in a standardized manner and evidence chain (including digital fingerprints), which, together with offline situational awareness mining and statistical reports, facilitates accountability, root cause analysis and training improvements, thereby enhancing management efficiency and compliance. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of an AI monitoring method for regulating the behavior of personnel in an industrial field, provided in an embodiment of this application. Figure 2 This is a functional module diagram of an AI monitoring device for regulating the behavior of personnel in an industrial field, provided in an embodiment of this application. Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] like Figure 1 The diagram shows a flowchart of an AI-based method for monitoring personnel behavior in industrial settings, as provided in an embodiment of this application. The AI-based method for monitoring personnel behavior in industrial settings, as provided in this embodiment, includes the following steps.

[0022] Step S1: Perform human key point detection and temporal motion feature extraction on the obtained behavioral data streams of each workstation to obtain the behavioral feature sequence of each person.

[0023] It should be understood that behavioral data streams are standardized behavioral data streams that have undergone rigorous preprocessing, and their generation process involves the fusion and alignment of multiple sensor sources. Specifically, behavioral data streams are formed by distributed timestamp synchronization processing of raw video streams collected by high-definition network camera arrays deployed in industrial sites. First, a network time protocol server is used to synchronize the internal clocks of all cameras at the microsecond level to ensure the time consistency of video frames across different viewpoints. Then, a background modeling algorithm based on an improved version of ViBe is applied to construct a background model for each pixel through random sampling to quickly adapt to changes in lighting and periodic motion interference. Morphological closure operations are used to fill internal holes in the detected moving target areas, and then connected component analysis is used to separate adhered targets to ensure that each person's area is complete and independent. Next, multi-scale spatiotemporal normalization operations are performed. In the time dimension, Lagrange interpolation is used to uniformly resample the variable frame rate to a fixed rate of 30fps. In the spatial dimension, bilinear normalization is used. Interpolation normalizes the resolution to 1280×720 pixels. A contrast-limited adaptive histogram equalization algorithm is used to process the luminance channel to eliminate shadows and reflections. To establish a unified geometric reference system, calibration reference objects of known dimensions are placed in the scene. The camera intrinsic and extrinsic parameter matrices are solved using the Zhang Zhengyou calibration method, constructing a projection transformation model from the pixel coordinate system to the world coordinate system. Simultaneously, a sliding window outlier detection algorithm is used to remove jump points from the raw positioning data collected by the RFID badge system. Quaternion interpolation is then used to achieve temporal alignment with the video data. Finally, an extended Kalman filter is used to fuse and smooth the multi-source trajectory data. This filter linearizes the nonlinear motion model using the Jacobian matrix, effectively suppressing sensor noise and measurement errors, thereby generating a high-quality behavioral data stream. This data stream provides a unified spatiotemporal reference for subsequent behavioral analysis, ensuring that the movement information of personnel at different workstations is comparable and computable in time and space. For example, on an assembly line, it can eliminate misjudgments of personnel position caused by differences in camera angles or lighting changes, thus accurately capturing workers' real-time actions.

[0024] After obtaining the behavioral data stream, the spatial coordinate sequence of each key point is first extracted based on the preset human joint feature data. This application employs a lightweight convolutional neural network, which uses depthwise separable convolution and attention mechanisms to balance computational efficiency and detection accuracy. The output includes two-dimensional coordinates and confidence scores for 17 key points, including the head and neck, shoulders, elbows, wrists, hips, knees, and ankles. This is significant because it transforms the pixel information in the original video frames into structured body part position data, laying the foundation for subsequent motion analysis. For example, in a welding station, accurately extracting the coordinates of the elbows and wrists can help identify whether safe operating distances are followed, avoiding detection omissions caused by occlusion or rapid movement. To eliminate instantaneous detection errors while preserving the true motion trend, the spatial coordinate sequence is then smoothed and filtered according to the time sequence. This application uses an exponentially weighted moving average algorithm, controlling the weight of historical data through a decay coefficient to obtain stable key point motion trajectory data. The above filtering operations reduce coordinate jitter caused by sensor noise or model fluctuations, ensuring the continuity of motion trajectory. For example, when carrying heavy objects, the smoothed motion trajectory data of each key point can more realistically reflect the worker's center of gravity changes, preventing misjudgment as a fall or imbalance.

[0025] Subsequently, kinematic parameters of each key point are calculated based on the motion trajectory data using a preset sliding time window. These kinematic parameters include displacement vector, motion velocity, acceleration, and motion trajectory curvature. In this application, the sliding time window size is set to 1 second (i.e., 30 frames) to balance real-time performance and feature integrity. It should be understood that the displacement vector is calculated using the coordinate difference between adjacent frames, reflecting the direction and distance of movement of the key point; the motion velocity is obtained by removing bits over time intervals, representing the speed of movement; the acceleration is calculated using the rate of change of velocity to capture abrupt changes in motion; and the motion trajectory curvature is quantified by the reciprocal of the radius of the arc formed by three consecutive points to determine the degree of path curvature. By calculating these kinematic parameters, the dynamic characteristics of personnel movement are quantified, thereby identifying abnormal patterns. For example, in a high-voltage equipment area, a sudden increase in acceleration may indicate running or slipping, while a high-curvature trajectory may suggest evasive maneuvers—all of which are closely related to safety regulations. Simultaneously, a pre-defined geometric algorithm calculates the joint angles of key points based on the relative coordinates of the limbs in the motion trajectory data. For example, the elbow joint angle is calculated using the vector dot product formula, which calculates the angle between the upper arm and forearm vectors. The formula is: θ=arccos[(AB×BC) / (|AB|×|BC|)], where AB and BC represent the vectors of adjacent bone segments, the dot product operator represents the vector dot product, and the modulus operator represents the vector length. The compliance of body posture is assessed through these calculations. For instance, in repetitive labor, a persistently large knee joint angle may indicate improper squatting posture, which could lead to overuse injuries in the long term. Real-time monitoring of these angles can promptly remind workers to adjust their posture.

[0026] Finally, the obtained kinematic parameters and joint angles are encapsulated. In this application, this process is divided into a splicing operation and a normalization operation. The splicing operation combines displacement vectors, motion velocity, acceleration, motion trajectory curvature, and joint angles in chronological order into a multi-dimensional feature vector. The normalization operation uses a minimum-maximum normalization method to map features of different dimensions to the [0,1] interval, eliminating the influence of individual body size and motion amplitude differences, thereby obtaining the behavioral feature sequence of each person. These operations create a unified and comparable feature representation, providing standardized input for subsequent behavioral semantic analysis. For example, on a multi-person collaborative assembly line, the normalized features allow the system to fairly evaluate the action efficiency of workers of different heights, avoiding bias caused by body differences, while ensuring stable convergence of the machine learning model, improving the robustness and accuracy of the overall monitoring system.

[0027] Through the above steps S1, from low-level coordinate extraction to high-level feature integration, the raw spatiotemporal data is transformed into semantically rich behavioral descriptions, laying the data foundation for safety monitoring in industrial sites and ensuring that any personnel action can be accurately quantified and tracked.

[0028] Step S2: Based on the preset set of behavioral semantic types, perform semantic similarity matching on the behavioral feature sequence to generate behavioral semantic description data.

[0029] Step S2 of this application is used to perform in-depth analysis on the behavioral feature sequence obtained through step S1 above, so as to map the numerical motion features to behavioral descriptions with clear semantic meaning, thereby providing understandable contextual information for subsequent risk assessment.

[0030] First, a pre-defined cosine similarity algorithm is used to calculate the similarity between the behavioral semantic type set and the behavioral feature sequence within a preset time period. The behavioral semantic type set is a predefined standardized behavior library containing typical industrial scene actions such as "normal walking," "running quickly," "bending down to pick up," "raising a hand to operate," and "climbing a railing." Each semantic type corresponds to a baseline feature vector learned from historical data. When calculating similarity, the feature vector at each time point in the behavioral feature sequence is first normalized to all baseline vectors in the semantic type set, transforming them into unit vectors. Then, the cosine similarity formula is applied to calculate their directional consistency. This formula is defined as the ratio of the dot product of two vectors to the product of their respective moduli, mathematically expressed as: Similarity = (A × B) / (||A|| × ||B||), where the dot operator represents the sum of the product of corresponding elements of the vectors, and the moduli operator represents the Euclidean length of the vector. The similarity calculation quantifies the degree of matching between actual actions and standard templates. For example, at an assembly station, if a worker's repetitive screw-tightening action shows a high similarity to the semantic type "standard installation," it indicates standardized operation; if it shows a high similarity to the "disorderly swaying" type, it may suggest fatigue or distraction. Through continuous rolling calculations over all time periods, this application's system generates a probability distribution for all semantic types within each time period. This distribution represents the weight of the probability of each behavior, providing a probabilistic basis for behavior recognition.

[0031] After obtaining the probability distribution of behavioral semantic categories through the above operations, the system of this application extracts the behavioral semantic type corresponding to the highest probability value as the optimal semantic type for that time period. Only the behavioral classification with the highest probability is retained to eliminate the uncertainty caused by fuzzy judgment and ensure the clarity of behavioral description. For example, in a chemical workshop, when the characteristics of a person's arm movement match both the semantic types of "adjusting a valve" and "waving and calling", the system selects the type with the higher probability as the final description. If the former has a probability of 0.8 and the latter has a probability of 0.2, it is determined to be a valve operation, avoiding safety misjudgments caused by multiple interpretations. At the same time, the system extracts regional scene data from a preset industrial site area database based on the received location data streams corresponding to each person. The location data streams come from the RFID and visual fusion trajectory smoothed by Kalman filtering in step S1. The industrial site area database in this application stores digital map information of the workshop, including the geographical boundaries and attribute labels of areas such as "high-voltage power distribution area", "robotic arm working range", "safety passage", and "material storage area". By matching real-time coordinates with geofences, the system dynamically obtains the scene context in which the personnel are located. For example, when the personnel's coordinates fall into a circular area with (X1,Y1) as the center and a radius of 5 meters, the system extracts the attribute of the area as "robot welding area" from the industrial site area database. The scene data of this area carries metadata such as environmental risk level and list of permitted behaviors.

[0032] Finally, this application employs a template-based semantic filling mechanism as the semantic description method, combining behavior types and scene attributes into a structured statement format. For example, when the optimal semantic type is "raising hand" and the scene is "high-voltage equipment area," the generated behavioral semantic description data may be recorded as "personnel raising hand in high-voltage equipment area," along with additional metadata such as timestamp, duration, and confidence level. The temporal data fusion operation in this application further considers the evolution of behavior over time. For instance, if three consecutive time periods are identified as "walking," "standing," and "bending over," the system will generate a composite behavioral description of "from walking to standing and then bending over." By capturing the coherent semantics of the action sequence through the above semantic fusion operation, the limitations of isolated judgments are avoided. Near a stamping machine, if the system detects that a person suddenly changes from "normal walking" to "rapid movement" and the scene is "equipment startup area," the generated semantic description will be labeled "accelerating movement next to running equipment." Such descriptions can be directly used to trigger safety warnings.

[0033] Step S2 described above, through progressive processing from probability calculation to context fusion, transforms low-level features into behavioral semantic data rich in scene information. This enables the system not only to identify "what people are doing," but also to understand "what they are doing, where they are doing it," providing a foundation for accurate and interpretable behavioral analysis for industrial safety monitoring. For example, on a high-altitude work platform, the description generated by combining the semantics of "climbing" with scene data of "missing guardrail areas" can accurately locate violations and guide safety officers to make targeted interventions, thereby effectively improving the semantic depth and practical value of behavioral monitoring.

[0034] Step S3: Based on the behavioral semantic description data, perform risk indicator quantification and feature fusion processing on the behavioral feature sequence to obtain the risk feature set of each person.

[0035] First, based on the optimal semantic type extracted from the behavioral semantic description data, the corresponding semantic risk coefficient set is retrieved from a pre-defined risk weight configuration table. This risk weight configuration table is a mapping table built based on historical accident data and safety expert experience, establishing a correlation between different behavioral semantic types and risk coefficients. For example, "climbing a guardrail" is mapped to a high-risk coefficient of 0.9, "walking normally" to a low-risk coefficient of 0.1, and "running" is mapped to a variable coefficient ranging from 0.3 to 0.7 depending on the area. This mapping operation transforms qualitative behavioral descriptions into quantifiable risk indicators, enabling the system to conduct preliminary risk assessments based on the inherent danger level of the behavioral type itself. For example, in a chemical plant area, when the semantic type "handling chemicals" is identified, the system automatically assigns a baseline risk coefficient of 0.8, while "operating control panel" is assigned only a coefficient of 0.3. Thus, even in the absence of other abnormal signals, the system can establish a basic risk perception based on the essential attributes of the behavior.

[0036] While acquiring semantic risk coefficients, this application's system performs spatial classification processing on behavioral feature sequences based on the obtained regional scene data, aggregating the behavioral features of personnel in different regions to form employee behavioral feature datasets specific to each regional scene. Through the above operations, a regional behavioral benchmark model is established, enabling risk assessment to consider the influence of environmental context. For example, in an assembly workshop, clustering the behavioral features of the "conveyor belt area" and the "quality inspection area" separately can reveal differences in normal behavioral patterns between different work units. Based on the employee behavioral feature datasets obtained by regionalization, this application's system uses a density-based clustering algorithm to analyze employee behavioral patterns within each regional scene. The density clustering algorithm identifies abnormal behavior by calculating the local density of data points in the feature space. Specifically, the algorithm first determines the number of other points contained in the ε-neighborhood of each data point, and then marks objects with significantly lower density than surrounding points as outliers. The abnormality value of a person's behavior is calculated using the standardized distance to the nearest cluster center, expressed as: Abnormality = (d - μ) d ) / σ d Where d represents the Euclidean distance to the nearest cluster center, and μ d and σ d These represent the mean and standard deviation of the distances between all personnel in the area, respectively. In the welding work area, if a worker's arm movement trajectory differs from the standard pattern of most workers by more than 2 standard deviations, their behavioral abnormality score will reach 0.8 or higher, indicating that non-standard operating behavior may exist.

[0037] Meanwhile, this application performs clothing recognition on industrial site personnel based on a preset set of workstation and protective equipment data and a location data stream to extract personal protective equipment (PPE) wearing status data from the behavioral data stream. The workstation and PPE data set used in this application defines the types of protective equipment required for different work areas, such as welding masks and protective gloves for the "welding area" and chemical protective suits and goggles for the "chemical area." Specifically, by combining personnel location information, the system activates the equipment detection module for the corresponding area and uses a target detection network based on the YOLOv5 architecture to perform real-time analysis of the video stream, outputting the detection confidence and location coordinates of equipment such as safety helmets, protective glasses, and insulating gloves associated with the personnel in the video. The output data is combined into wearing status data in the form of a multi-dimensional vector, where each dimension corresponds to the wearing status value of a type of equipment (0 indicates not worn, 1 indicates correctly worn). In the high-voltage operation area, if the system detects that a person is not wearing insulating gloves (status value 0), this data will be used as an important risk indicator in subsequent fusion calculations.

[0038] Meanwhile, this application employs a pre-defined convex hull and ray detection method to perform precise area intrusion detection based on the hazardous area location dataset and personnel location data stream. The hazardous area location dataset used in this application stores the geometric boundaries of each hazardous area in the form of a sequence of polygon vertices; for example, the working range of a robotic arm is defined as a convex polygon enclosed by multiple three-dimensional coordinate points. The convex hull and ray detection method used in this application specifically includes convex hull detection and ray detection. Convex hull detection is used to determine the spatial relationship between a personnel location and a hazardous area. By calculating the positional relationship between a point and the minimum boundary convex set of the polygon, it determines whether an intrusion has occurred. Ray detection, on the other hand, emits virtual rays from the personnel location in any direction and calculates the number of intersections with the area boundary to determine the inside-outside relationship (an odd number of intersections indicates the point is inside the polygon, while an even number indicates it is outside). Finally, the obtained spatial relationship data for measuring the spatial relationship between personnel and hazardous areas includes parameters such as minimum distance, intrusion depth, and dwell time. In automated warehousing areas, when the minimum distance between a personnel and a running AGV is less than the safety threshold of 1.5 meters, the distance parameter in the spatial relationship data will trigger a risk escalation.

[0039] Subsequently, this application employs a linear weighted model to perform weighted fusion of the behavioral anomaly values, wearing status data, and spatial relationship data obtained from the above operations based on the semantic risk coefficient set. The weights of each dimension are determined using the analytic hierarchy process (AHP), comprehensively considering the impact of each factor on the overall risk. Before performing the above fusion calculation process, this application also needs to convert the wearing status data into an equipment violation index. Specifically, the system of this application performs a weighted summation calculation based on the wearing status values ​​of each necessary piece of equipment in the wearing status data to obtain the equipment violation index. The weight coefficients corresponding to each wearing status value are determined based on the safety importance of the equipment, using the AHP and expert scoring. For example, in high-voltage operating areas, the weight of insulating gloves (0.4) is much higher than the weight of reflective vests (0.1). All weight coefficients are normalized to ensure their sum is 1. In chemical handling areas, if personnel are not wearing protective suits (weight 0.5) and goggles (weight 0.3), but are correctly wearing safety helmets (weight 0.2), then the equipment violation index = 0.5 × 1 + 0.3 × 1 + 0.2 × 0 = 0.8. The above-mentioned equipment violation index calculation process differentiates the severity of violations by different equipment, avoiding treating all violations the same. Furthermore, before performing the aforementioned fusion calculation process, this application also needs to convert the spatial relationship coefficient into a spatial risk index. The spatial risk index is a composite indicator that comprehensively considers the relative position and movement state of personnel and the hazardous area. The calculation formula can be expressed as: Spatial Risk Index = α × f(d) + β × g(t) + γ × h(v). Here, d represents the minimum distance between the personnel and the boundary of the hazardous area, t represents the time the personnel spend within the hazardous area, and v represents the speed at which the personnel move towards the hazardous area. f(d) is a distance function, defined as f(d) = e (-k×d) Where k is the attenuation coefficient, meaning that the closer the distance, the greater the risk; g(t) is a time function, defined as g(t) = e^(-k / t). (-λ×t) This indicates that as the dwell time increases, the risk gradually approaches its maximum value; h(v) is the velocity function, defined as h(v) = min(v / v) max ,1), indicating that the faster the speed towards the danger zone, the higher the risk. α, β, and γ are weighting coefficients that sum to 1 and are adjusted according to the characteristics of different areas. In automated storage areas, when personnel approach the path of a running AGV at a relatively high speed and the distance is less than the safety threshold of 1.5 meters, the spatial risk index will increase rapidly. The above spatial risk index calculation dynamically reflects the evolution of spatial risk, considering not only static position but also the potential danger brought about by movement trends.

[0040] Based on the above operations, the weighted fusion formula of this application can be expressed as: Comprehensive Risk Value = w1 × Semantic Risk Coefficient + w2 × Behavioral Anomaly Degree + w3 × Equipment Violation Index + w4 × Spatial Risk Index, where w1-w4 are the weight coefficients of each dimension, and w1+w2+w3+w4=1. This multi-dimensional weighted fusion operation overcomes the limitations of a single risk assessment indicator. By integrating multi-source information such as behavioral semantics, action patterns, equipment status, and spatial location, a comprehensive and accurate personnel risk profile is formed, providing precise quantitative basis for subsequent identification of violations, ultimately achieving precise prevention and control of industrial site safety risks. Finally, the generated risk feature set is a structured dataset containing the following core components: the original values ​​of the basic risk indicator dimensions, including the semantic risk coefficient, behavioral anomaly degree value, equipment violation index, and spatial risk index; the comprehensive risk value obtained from the above weighted fusion; a risk feature vector, a numerical vector formed by arranging the indicators of each dimension in a fixed order, used as input to the machine learning model; timestamps and location context information, recording the specific time and spatial location of the risk assessment; and confidence scores, reflecting the reliability of the calculation of each indicator.

[0041] For example, in a large assembly workshop, the risk feature set of a person who is not wearing a safety helmet and whose movement pattern is abnormal in the robot work area will include basic indicators such as semantic risk coefficient (0.6), behavioral abnormality degree (0.7), equipment violation index (0.8) and spatial risk index (0.9), as well as a comprehensive risk value (0.78) obtained by weight fusion. At the same time, the assessment time, work station number and the calculated confidence level of each indicator are recorded to generate the risk feature set of the employee.

[0042] Step S4: Based on the preset business behavior rules, identify violations and fuse evidence in the risk feature set to generate violation event records for each person.

[0043] It should be understood that business behavior rules include sets of behaviors and violation types, as well as membership thresholds. The set of behaviors and violation types is a predefined set of mapping relationships that establishes a correspondence between specific combinations of behavioral patterns and violation types. For example, "not wearing a safety helmet" and "entering a high-voltage area" maps to "serious safety violation," while "running quickly" and "approaching equipment" maps to "general behavioral violation." The membership threshold is a numerical boundary determined through statistical analysis and expert experience, used to distinguish between a violation and a potential violation. Business behavior rules transform abstract safety specifications into calculable and executable judgment criteria, enabling the system to trigger violation identification based on explicit conditions. For example, in a welding work area, the rule base defines a membership threshold of 0.8 for "not wearing safety goggles" and "performing welding operations." When the system detects the relevant combination of behavioral characteristics, it can determine a violation based on this threshold.

[0044] First, this application uses a Gaussian membership function to match the risk feature set with the behavior and violation type set, obtaining the violation type probability distribution. The membership function can be expressed as: μ(x) = exp[-(xc)]. 2 / (2σ 2 [ ], where x represents the risk characteristic value, c represents the typical characteristic value of the violation type, and σ controls the width of the function. The system in this application inputs each indicator from the risk characteristic set into the membership function to calculate the matching probability of each violation type. The matching probabilities of all violation types form the violation type probability distribution, which characterizes the likelihood that the current behavior belongs to any type of violation. In the chemical storage area, when abnormal personnel movement speed (characteristic value 0.7) and missing protective equipment (characteristic value 0.9) are detected, the system calculates the distribution results of "running violation" probability 0.6 and "missing protective equipment violation" probability 0.8 through the membership function. The significance of this probabilistic processing lies in quantifying the uncertainty of violation judgment and providing richer reference information for subsequent decision-making.

[0045] Furthermore, this application's system extracts a potential violation behavior feature set and the violation type corresponding to each behavior feature from the risk feature set based on the membership threshold and the probability distribution of violation types. The membership threshold in this application is typically set in the range of 0.5-0.8, with the specific value adjusted according to the severity of the violation type. When the matching probability of a certain violation type exceeds its corresponding membership threshold, the system marks that violation type as a potential violation and simultaneously extracts a relevant feature subset from the original risk feature set to form a potential violation behavior feature set. This feature set not only contains the main features that trigger the violation but also retains relevant auxiliary feature information. In an assembly line scenario, if the probability of "operation speed violation" reaches 0.75 (exceeding the threshold of 0.7), the system will extract features such as movement speed, acceleration, and operation frequency to form a feature subset and label the corresponding violation type as "operational specification violation." The significance of this screening process lies in focusing on key risk signals, reducing the amount of data processed subsequently, and improving system operating efficiency.

[0046] Based on a feature set of potential violations, the system extracts corresponding violation evidence slices from the behavioral data stream. The behavioral data stream, serving as the system's original data source, includes complete video sequences, positioning trajectories, and equipment status records. This application employs a time window backtracking mechanism for evidence slice extraction, extracting data segments extending forward and backward from the point of violation occurrence. The obtained violation evidence slices contain keyframe sequences several seconds before and after the violation, positioning evidence slices record the personnel's movement trajectory coordinates during the violation, and equipment status evidence slices preserve the historical operating parameters of the relevant equipment. In the robot's work area, when a "safety distance violation" is detected, the system automatically extracts a 5-second video clip, personnel positioning trajectory, and robotic arm operating status data before and after the violation as evidence slices. The significance of this evidence preservation mechanism lies in providing intuitive and verifiable original evidence for violation determination, ensuring the traceability of subsequent processing.

[0047] Subsequently, this application uses a pre-defined Dempster combination rule to perform consistency analysis and verification on the potential violation feature set and violation evidence slice data. The Dempster combination rule is a core algorithm in evidence theory, used to fuse the support of multiple independent evidence sources. Its mathematical expression is m1⊕m2(A)=Σm1(B)m2(C) / (1-K), where A, B, and C are hypothesis sets, and K represents the evidence conflict factor. In this application's system, the feature analysis results and the original evidence are treated as independent evidence sources. The consistency of the judgment is verified by calculating their support for the same violation hypothesis. In high-temperature work areas, when feature analysis shows a high probability of "fatigue operation violation," the system combines visual features such as personnel posture and operating rhythm from video evidence for evidence fusion. If the support of both is higher than 0.7 and the conflict factor is lower than 0.3, the violation is confirmed. The significance of this cross-validation of evidence lies in improving the reliability of the judgment and reducing the risk of false alarms.

[0048] Based on the evidence dataset selected through the above verification, this application calculates the comprehensive credibility score corresponding to the evidence dataset using a pre-defined evidence evaluation algorithm. The evidence evaluation algorithm used in this application is a weighted confidence model, comprehensively considering multiple dimensions of credibility indicators such as feature matching degree, evidence clarity, and temporal consistency. Each evidence source is assigned a different weight coefficient, and the weight allocation is determined based on the importance of the evidence type; for example, video evidence typically has a higher weight than trajectory evidence. The comprehensive credibility score is calculated using a linear weighting method, with a value range of [0,1]. A higher value indicates a more reliable violation determination. On a high-altitude work platform, when the system simultaneously obtains clear video evidence of not wearing a safety belt (weight 0.6, credibility 0.9) and evidence of location exceeding boundaries (weight 0.4, credibility 0.7), the comprehensive credibility score can reach 0.82. The significance of this quantitative evaluation lies in providing a basis for classifying the severity of violations and supporting differentiated processing strategies.

[0049] Finally, this application's system integrates the comprehensive credibility score and evidence dataset based on the violation type to generate violation event records for each person. The evidence fusion in this application employs a template-based structured assembly method, integrating key information such as violation type, occurrence time, spatial location, comprehensive credibility, and evidence index according to a predetermined format. Violation event records are stored in a standardized JSON format, containing complete fields such as header information, violation details, evidence links, and handling suggestions. In the electrical control room, when a "violation of unlicensed operation" is confirmed, the system-generated record will include detailed information such as operator information, violation time, equipment number, certificate status, and video evidence timestamps, while automatically marking the handling priority based on the credibility score. The significance of this structured record lies in forming a complete and standardized violation file, providing standardized data support for subsequent safety management, accountability, and statistical analysis.

[0050] The operation in step S4 above, through a progressive process from rule matching to evidence fusion, transforms risk characteristics into legally valid violation records, thereby achieving the standardization, evidence-based approach, and traceability of industrial site safety supervision, and providing technical support for improving safety management.

[0051] After generating violation event records, this application's system performs early warning classification based on the severity of the violations according to a preset alarm triggering strategy. The alarm triggering strategy is a classification rule system based on a multi-dimensional risk assessment matrix, which comprehensively considers factors such as violation type, frequency of occurrence, environmental risk level, and the severity of potential consequences. The early warning classification operation used in this application first analyzes key fields in the violation event records, including violation type code, comprehensive credibility score, spatial location attribute, and time duration index. A weighted assessment model is then used to calculate the severity index of each violation event. Specifically, a linear weighted method is used: Severity Index = α × Type Weight + β × Credibility Score + γ × Regional Risk Coefficient + δ × Duration Factor, where α, β, γ, and δ are weight coefficients determined through the analytic hierarchy process and satisfy normalization conditions. Based on the calculation results, the system classifies violations into three warning levels: red (urgent), orange (serious), and yellow (general). For example, in a chemical plant area, a violation involving the lack of protective equipment for handling hazardous chemicals would be automatically classified as red, while excessive walking speed in a general area would be classified as yellow. The significance of this tiered approach lies in enabling differentiated emergency responses, ensuring that limited safety management resources prioritize high-risk events while avoiding overreactions to low-risk events.

[0052] Based on the completed early warning classification, this application's system employs a multi-level, multi-channel hybrid notification mechanism to accurately alert to real-time alarm information within the tiered alarm information. The alarm methods used in this application's embodiments include on-site audible and visual alarms, mobile terminal push notifications, monitoring center pop-ups, and SMS reminders to responsible personnel. Different early warning levels trigger different alarm combinations: red-level alarms activate all alarm channels simultaneously, orange-level alarms enable notifications from both mobile terminals and the monitoring center, and yellow-level alarms only trigger monitoring center recordings. The on-site audible and visual alarm device is specifically designed for industrial environments; the high-frequency alarm sound combined with the rotating red light is suitable for high-noise workshop environments. Mobile terminal push notifications use a structured message format, containing key information such as the type of violation, location of occurrence, and suggested handling measures. In large assembly workshops, when the system detects an intrusion violation in the robotic arm's work area and determines it to be at the red level, the on-site warning lights immediately flash, a warning voice is broadcast in the area, and simultaneously, the safety officer receives detailed alarm information, including on-site video screenshots, on a handheld terminal. The significance of this differentiated alarm mechanism lies in ensuring that alarm information can be accurately and promptly conveyed to relevant responsible personnel and providing sufficient on-site context support, creating the necessary conditions for rapid and effective emergency response.

[0053] Meanwhile, the system in this application performs group behavior analysis on graded alarm information according to a preset time period. The time period in this embodiment is set as a configurable parameter, typically using daily, weekly, and monthly statistical analysis periods for short-term early warning, medium-term trend analysis, and long-term management decision-making, respectively. The group behavior analysis employs a multi-indicator comprehensive evaluation method. First, based on the alarm record database, data is aggregated by personnel identifier to calculate individual violation frequency index, severity index, and risk trend coefficient. The violation frequency index is calculated using standardized calculations based on the number of alarms per unit time, the severity index is based on the weighted average of historical alarm levels, and the risk trend coefficient is used to predict future risk changes through time series analysis. Based on this, the system applies a clustering algorithm to identify groups of personnel with similar violation characteristics, discovering potential systemic safety issues. The obtained list of key personnel for violation monitoring is sorted in descending order according to the comprehensive risk score. Each personnel entry contains structured content such as basic information, main violation patterns, risk rating, and improvement suggestions. For example, in a weekly analysis of a welding workshop, the system may find that a certain work group has widespread violations regarding the wearing of protective masks, thus listing this group as a key monitoring group and recommending strengthened management of protective equipment in that area. Through the aforementioned periodic analysis, patterns of safety hazards are identified at both the individual and group levels, providing data support for the development of targeted safety management measures and enabling a shift from post-event handling to pre-event prevention.

[0054] This application is applied to the field of behavior monitoring technology. It obtains behavioral feature sequences for each person by detecting key human points and extracting time-series motion features from behavioral data streams at each workstation. Semantic similarity matching is then performed on these behavioral feature sequences based on a set of behavioral semantic types to generate behavioral semantic description data. Risk indicator quantification and feature fusion are then performed on the behavioral feature sequences based on the behavioral semantic description data to obtain a risk feature set for each person. Finally, violation identification and evidence fusion are performed on the risk feature set according to business behavior rules to generate violation event records for each person. This application extracts stable features from multi-source spatiotemporal alignment, and through semantic mapping and risk fusion, ultimately generates traceable structured violation events, thereby significantly improving the accuracy and response speed of on-site safety management.

[0055] like Figure 2 The diagram shown is a functional block diagram of an AI monitoring device for monitoring the behavior of personnel in an industrial field, provided in an embodiment of this application.

[0056] In some embodiments, the AI ​​monitoring device 2 for industrial site personnel behavior may include multiple functional modules composed of computer program segments. The computer programs for each program segment in the AI ​​monitoring device 2 may be stored in the server's memory and executed by at least one processor to perform (see details). Figure 1(Description) Functions of AI monitoring methods for regulating the behavior of personnel in industrial settings.

[0057] In this embodiment, the industrial site personnel behavior monitoring device 2 can be divided into multiple functional modules according to its functions. These functional modules may include: a feature extraction module 21, a semantic description module 22, a risk assessment module 23, a violation identification module 24, and a graded alarm module 25. The module referred to in this invention is a series of computer program segments that can be executed by at least one processor and perform a fixed function, stored in memory. In this embodiment, the functions of each module will be detailed in subsequent embodiments.

[0058] Feature extraction module 21 is used to perform human key point detection and temporal motion feature extraction on the obtained behavioral data stream of each workstation to obtain the behavioral feature sequence of each person. The semantic description module 22 is used to perform semantic similarity matching on the behavioral feature sequence according to a preset set of behavioral semantic types to generate behavioral semantic description data; Risk assessment module 23 is used to perform risk indicator quantification assessment and feature fusion processing on the behavioral feature sequence based on the behavioral semantic description data to obtain the risk feature set of each person; The violation identification module 24 is used to identify violations and fuse evidence based on the preset business behavior rules of the risk feature set, and generate violation event records for each person.

[0059] In an optional implementation, the feature extraction module 21 is specifically used for: Based on the preset human joint feature data, the spatial coordinate sequence of each key point is extracted from the obtained behavioral data stream of each workstation; The spatial coordinate sequence is smoothed and filtered according to the time order to obtain the motion trajectory data of each key point; The kinematic parameters of each key point are calculated based on the motion trajectory data through a preset sliding timing window; Calculate the joint angles of each key point based on the relative coordinates of the limbs in the motion trajectory data; The kinematic parameters and joint angles are spliced ​​and normalized according to time sequence features to obtain the behavioral feature sequence of each person.

[0060] In an optional implementation, the semantic description module 22 is specifically used for: The similarity algorithm calculates the similarity between each behavioral semantic type in the preset behavioral semantic type set and each behavioral feature in the behavioral feature sequence based on a preset unit time period, and generates the probability distribution of behavioral semantic categories corresponding to each time period. Extract the behavioral semantic type corresponding to the maximum probability value from the behavioral semantic category probability distribution, and use it as the optimal semantic type for each time period; Based on the received location data streams for each person, extract the corresponding area scene data for each person from the preset industrial site area database; By using a preset semantic description method, time-series data is fused based on the optimal semantic type and the regional scene data to generate behavioral semantic description data for each person.

[0061] In an optional implementation, the risk assessment module 23 is specifically used for: Based on the optimal semantic type, extract the corresponding semantic risk coefficient set from the preset risk weight configuration table; The behavioral feature sequence is classified based on the regional scene data to obtain employee behavioral feature datasets for each regional scene. Then, a preset density clustering algorithm is used to perform cluster analysis on the employee behavioral feature datasets to calculate the behavioral anomaly value of each person. Based on the preset workstation and protective equipment data set and the positioning data stream, extract the personal protective equipment wearing status data of each person from the behavior data stream; Based on the preset dangerous area location dataset and the location data stream, area intrusion detection is performed to calculate the spatial relationship data between each person and the dangerous area; The risk feature set of each individual is obtained by weighted and fused calculation of the behavioral anomaly value, the wearing status data, and the spatial relationship data based on the semantic risk coefficient set.

[0062] In an optional implementation, the violation identification module 24 is specifically used for: By using a preset membership function, the risk feature set is matched with a preset set of behaviors and violation types to obtain the probability distribution of violation types; Based on the preset membership threshold and the probability distribution of the violation type, a set of potential violation behavior features and the violation type corresponding to each behavior feature are extracted from the risk feature set; Based on each behavioral feature in the potential violation feature set, the corresponding violation evidence slice data is extracted from the behavioral data stream; Based on the violation type, a consistency analysis and verification is performed on the potential violation behavior feature set and the violation evidence slice data to obtain an evidence dataset; The comprehensive credibility score corresponding to the evidence dataset is calculated using a preset evidence evaluation algorithm; By using a preset evidence fusion method, the comprehensive credibility score and the evidence dataset are fused according to the violation type to generate violation event records for each person.

[0063] In an optional implementation, the industrial site personnel behavior regulation AI monitoring device 2 further includes a hierarchical alarm module 25, which is specifically used for: According to the preset alarm triggering strategy, the violation event records are classified and processed according to the severity of the violation, and graded alarm information is generated. Real-time alarms are triggered based on the real-time alarm information in the tiered alarm information, using a preset alarm method. Based on a preset time period, the hierarchical alarm information is analyzed to generate a list of key individuals to be monitored for violations.

[0064] It should be understood that the various variations and specific embodiments of the methods provided in the above embodiments are also applicable to the AI ​​monitoring device for industrial site personnel behavior norms in this embodiment. Through the foregoing detailed description of the AI ​​monitoring method for industrial site personnel behavior norms, those skilled in the art can clearly understand the implementation method of the AI ​​monitoring device for industrial site personnel behavior norms in this embodiment. For the sake of brevity, it will not be described in detail here.

[0065] like Figure 3 The diagram shown is a structural schematic of an electronic device provided in an embodiment of this application.

[0066] In a preferred embodiment of the present invention, the electronic device 3 may include, but is not limited to, a memory 31, at least one processor 32, and at least one communication bus 33.

[0067] Those skilled in the art should understand that Figure 3 The structure of the electronic device 3 shown does not constitute a limitation of the embodiments of the present invention. The electronic device 3 may also include more or fewer other hardware or software than shown, or different component arrangements.

[0068] In some embodiments, the electronic device 3 is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits, programmable gate arrays, digital processors, and embedded devices.

[0069] It should be noted that the electronic device 3 is merely an example. Other existing or future electronic products that are suitable for this application should also be included within the scope of protection of this application and are incorporated herein by reference.

[0070] In some embodiments, the memory 31 stores a computer program that, when executed by the at least one processor 32, implements all or part of the steps in the AI ​​monitoring method for industrial field personnel behavior norms as described. The memory 31 includes a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electronically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data. Further, the computer-readable storage medium may primarily include a program storage area and a data storage area, wherein the program storage area may store an operating system, at least one application program required for a function, etc.

[0071] In some embodiments, the at least one processor 32 is the control unit of the electronic device 3, connecting various components of the electronic device 3 via various interfaces and lines. It executes programs or modules stored in the memory 31 and calls data stored in the memory 31 to perform various functions and process data. For example, when the at least one processor 32 executes the computer program stored in the memory 31, it implements all or part of the steps of the AI ​​monitoring method for industrial field personnel behavior norms described in this application embodiment; or it implements all or part of the functions of the AI ​​monitoring device for industrial field personnel behavior norms. The at least one processor 32 may be composed of integrated circuits, such as a single-packaged integrated circuit or multiple integrated circuits with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips.

[0072] In some embodiments, the at least one communication bus 33 is configured to enable communication between the memory 31 and the at least one processor 32, etc. Although not shown, the electronic device 3 may also include a power supply (e.g., a battery) to power the various components. Preferably, the power supply can be logically connected to the at least one processor 32 via a power management device, thereby enabling functions such as charging, discharging, and power consumption management. The power supply may also include one or more DC or AC power supplies, recharging devices, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components. The electronic device 3 may also include various sensors, Bluetooth modules, Wi-Fi modules, etc., which will not be described in detail here.

[0073] The integrated unit implemented as a software functional module described above can be stored in a computer-readable storage medium. This software functional module, stored in a storage medium, includes several instructions to cause an electronic device (which may be a personal computer, electronic device, or network device, etc.) or processor to execute portions of the methods described in the various embodiments of this application.

[0074] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0075] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0076] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. An industrial site personnel behavior regulation AI monitoring method, characterized in that, The method comprises: obtaining the behavior data stream of each station and performing human key point detection and time sequence motion feature extraction to obtain the behavior feature sequence of each person; according to the preset behavior semantic type set, performing semantic similarity matching on the behavior feature sequence to generate behavior semantic description data; according to the behavior semantic description data, performing risk index quantitative evaluation and feature fusion processing on the behavior feature sequence to obtain the risk feature set of each person; according to the preset business behavior rule, performing rule violation behavior identification and evidence fusion on the risk feature set to generate the rule violation event record of each person.

2. The industrial field personnel behavioral norm AI monitoring method of claim 1, wherein, The human key point detection and time sequence motion feature extraction on the obtained behavior data stream of each station to obtain the behavior feature sequence of each person comprises: according to the preset human joint feature data, extracting the spatial coordinate sequence of each key point from the obtained behavior data stream of each station; according to the time sequence, performing smoothing filtering processing on the spatial coordinate sequence to obtain the motion trajectory data of each key point; according to the motion trajectory data, calculating the kinematic parameters of each key point through a preset sliding time sequence window; according to the relative coordinates of the limbs in the motion trajectory data, calculating the joint angles of each key point; performing time sequence feature splicing and normalization processing on the kinematic parameters and the joint angles to obtain the behavior feature sequence of each person.

3. The industrial field personnel behavioral norm AI monitoring method of claim 1, wherein, The semantic similarity matching on the behavior feature sequence according to the preset behavior semantic type set to generate behavior semantic description data comprises: according to the preset similarity algorithm, calculating the similarity between each behavior semantic type in the preset behavior semantic type set and each behavior feature in the behavior feature sequence according to the preset unit time period to generate the behavior semantic category probability distribution corresponding to each time period; extracting the behavior semantic type corresponding to the maximum probability value from the behavior semantic category probability distribution as the optimal semantic type corresponding to each time period; according to the received positioning data stream corresponding to each person, extracting the region scene data corresponding to each person from the preset industrial field region database; according to the optimal semantic type and the region scene data, performing time sequence data fusion through a preset semantic description method to generate the behavior semantic description data of each person.

4. The industrial field personnel behavioral norm AI monitoring method of claim 3, wherein, The risk index quantitative evaluation and feature fusion processing on the behavior feature sequence according to the behavior semantic description data to obtain the risk feature set of each person comprises: according to the optimal semantic type, extracting the corresponding semantic risk coefficient set from the preset risk weight configuration table; according to the region scene data, performing classification processing on the behavior feature sequence to obtain the employee behavior feature data set of each region scene, so as to perform clustering analysis on the employee behavior feature data set through a preset density clustering algorithm to calculate the behavior abnormality value of each person; according to the preset station and protective equipment data group set and the positioning data stream, extracting the wearing state data of the personal protective equipment of each person from the behavior data stream; According to the preset dangerous area positioning data set and the positioning data stream, area intrusion detection is performed to calculate spatial relationship data between each person and the dangerous area; According to the semantic risk coefficient set, the behavior anomaly degree value, the wearing state data, and the spatial relationship data are weighted and fused to obtain a risk feature set of each person.

5. The industrial field personnel behavior regulation AI monitoring method of claim 1, the business behavior rules comprising a behavior and violation type group set and a membership threshold, characterized in that, The rule violation identification and evidence fusion of the risk feature set according to the preset business behavior rule to generate a rule violation event record of each person include: Through a preset membership function, the risk feature set is matched with a preset behavior and rule type group set to obtain a rule type probability distribution; According to a preset membership threshold and the rule type probability distribution, a potential rule violation behavior feature set and a rule type corresponding to each behavior feature are extracted from the risk feature set; According to each behavior feature in the potential rule violation behavior feature set, corresponding rule violation evidence slice data is extracted from the behavior data stream; According to the rule type, consistency analysis and verification are performed on the potential rule violation behavior feature set and the rule violation evidence slice data to obtain an evidence data set; Through a preset evidence evaluation algorithm, a comprehensive credibility score corresponding to the evidence data set is calculated; According to the rule type, the comprehensive credibility score and the evidence data set are evidence fused through a preset evidence fusion mode to generate a rule violation event record of each person.

6. The industrial field personnel behavioral compliance AI monitoring method of claim 1, wherein, The method further includes: According to a preset alarm triggering strategy, a rule violation severity early warning classification process is performed on the rule violation event record to generate a hierarchical alarm information; According to real-time alarm information in the hierarchical alarm information, real-time alarm is performed through a preset alarm mode; According to a preset time period, a personnel group behavior analysis is performed on the hierarchical alarm information to generate a rule violation key focus personnel information list.

7. An industrial site personnel behavior regulation AI monitoring device applied to the industrial site personnel behavior regulation AI monitoring method of claim 1, characterized in that, The device includes: A feature extraction module for human key point detection and time sequence motion feature extraction on the obtained behavior data stream of each station to obtain a behavior feature sequence of each person; A semantic description module for semantic similarity matching of the behavior feature sequence according to a preset behavior semantic type set to generate behavior semantic description data; A risk assessment module for risk index quantitative evaluation and feature fusion processing of the behavior feature sequence according to the behavior semantic description data to obtain a risk feature set of each person; A rule violation identification module for rule violation identification and evidence fusion of the risk feature set according to a preset business behavior rule to generate a rule violation event record of each person.

8. An electronic device, comprising: The electronic device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor implements the steps of the industrial site personnel behavior specification AI monitoring method according to any one of claims 1 to 6 when executing the computer program.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the industrial site personnel behavior specification AI monitoring method according to any one of claims 1 to 6.

Images