A method for monitoring the continuity of wearing based on timing characteristics

Abstract

This invention proposes a method for continuous monitoring of wearability compliance based on temporal features, comprising the following steps: synchronously collecting video data and environmental data of the monitoring area, and preprocessing the collected data; constructing a continuously sliding temporal window based on the collected video data, and extracting the temporal features of the worker's wearability status within each temporal window; identifying the current operational behavior stage of the worker based on the video data within each temporal window, and establishing a correlation mapping between each behavior stage and wearability compliance; dynamically judging the worker's wearability status by combining the temporal features of each temporal window and the corresponding correlation mapping; and outputting violation event information including the start and end times of the violation, duration, and type of violation based on the judgment results. This invention elevates the monitoring dimension from frame-level to temporal-level, effectively capturing the changing patterns of wearability status over time, accurately distinguishing between short-term abnormal behavior and genuine violations, and significantly reducing false alarm and false negative rates.

Description

Technical Field

[0001] This invention relates to the field of visual recognition technology, and in particular to a method for continuous monitoring of wearability specifications based on temporal features. Background Technology

[0002] The proper attire of personnel in industrial pilot zones directly determines the safety of production operations and the order of factory operations. The proper wearing of safety helmets, protective masks, work clothes, anti-impact shoes, protective gloves, and other items is a mandatory requirement for industrial safety production supervision, and scientific and efficient monitoring methods are the key to ensuring that the attire standards are implemented.

[0003] Currently, there are two main types of wearable monitoring methods in industrial pilot zones, both of which have significant technical shortcomings: One type is traditional manual inspection, which is affected by human factors and has problems such as incomplete coverage, delayed response, low efficiency, inability to monitor continuously for 24 hours and lack of traceability. It is difficult to achieve timely correction of violations and review of supervision. In addition, the labor cost is high and it is not suitable for large-scale and routine supervision needs. At the same time, some areas of the industrial pilot zone have high-risk operation risks, and manual inspection also poses safety hazards to the personnel themselves.

[0004] Another type is intelligent monitoring based on single-frame image recognition. Although it achieves automated monitoring and reduces reliance on manual labor, it can only complete frame-level instantaneous recognition, ignoring the dynamic temporal characteristics of staff operations. It has defects such as high false alarm and false alarm rates, difficulty in distinguishing between short-term anomalies and real violations, unstable recognition in occluded and complex environments, failure to combine behavior stage optimization judgment, and only being able to output frame-level results without forming a complete violation event.

[0005] In addition, existing wearable monitoring technologies are only suitable for general occupational safety scenarios. They have not been specifically optimized for special environments such as dust, oil, strong light, mechanical obstruction, and high-altitude operations in industrial test zones, as well as for the dynamic operation characteristics of workers, such as equipment operation, material handling, and welding. This further limits their promotion and application in various industrial test zone scenarios.

[0006] Therefore, developing a method that integrates temporal characteristics, enables continuous and accurate monitoring, and outputs complete violation events, based on the actual monitoring needs of industrial pilot zones, has become a pressing technical challenge in the field of industrial safety production supervision. Summary of the Invention

[0007] This invention proposes a continuous monitoring method for wearing compliance based on time-series features, which solves the problems of high false alarm and false negative rates and difficulty in distinguishing between short-term anomalies and real violations in existing industrial area wearing compliance monitoring methods.

[0008] The technical solution of this invention is implemented as follows: This invention provides a method for continuous monitoring of wearability compliance based on time-series features, comprising the following steps: Simultaneously collect video and environmental data of the monitoring area, and preprocess the collected data; A continuously sliding time-series window is constructed based on the collected video data, and the time-series features of the staff's clothing status within each time-series window are extracted. Based on the video data within each time window, the current operational behavior stage of the staff is identified, and a mapping between each behavior stage and the wearing specifications is established. By combining the temporal characteristics of each time window and the corresponding association mapping, the staff's wearing status is dynamically judged; Based on the judgment result, output violation event information including the start and end time of the violation, duration, and type of violation.

[0009] Specifically, the monitoring area includes at least one or more of the following: processing area, material handling area, high-altitude operation area, equipment operation area, and entrance / exit; the environmental data includes at least one or more of the following: temperature, humidity, dust concentration, light intensity, and oil concentration.

[0010] Specifically, the preprocessing method is: based on The criteria include outlier removal for environmental data, and the outlier removal formula is as follows: ; in, These are environmental parameter values ​​collected in a single session. This is the average value of environmental parameters within a preset time window. This represents the standard deviation of environmental parameters within the preset timing window.

[0011] Specifically, the length of the time-series window is adaptively adjusted based on the dust concentration in the environmental data, and the adjustment formula is as follows: ; in, This is the adjusted timing window length. For standard timing window length, This represents the current dust concentration. This is the baseline value for dust concentration. This represents the maximum threshold for dust concentration.

[0012] Specifically, the temporal characteristics include existence temporal characteristics, duration of state changes, and consistency temporal characteristics; The existence time-series feature is quantified by the following formula: ; Where Exist represents the existence-based temporal feature result. This is the current time series window length; For the first Confidence level for identifying wearable items in a frame. To identify the confidence threshold; The duration of the state change is calculated using the following formula: ; in, The duration of the state change. The frame number indicating the start of the state change. This is the frame number indicating the end of the state change. This refers to the video capture frequency. The consistency time series feature is calculated using the following formula: ; in, As a consistency indicator, The number of frames with the same wearing status within the timing window.

[0013] Specifically, methods for establishing the association mapping between each behavioral stage and wearing norms include: By analyzing the actions of staff within a time-series window using a behavior recognition model, the behavior stage to which they belong is identified. The confidence level of the behavior stage is calculated using the following formula: ; in, The current time sequence window belongs to the first Confidence level of the behavioral stage. For the first Frame action and the first Similarity of standard actions in similar behavior stages The current time series window length; select The behavior stage corresponding to the maximum value is used as the behavior stage of the current time sequence window; Based on the identified behavioral stage, the confidence threshold for wearable item recognition is adaptively adjusted using the following formula: ; in, For the first Class behavior stage The confidence threshold for identifying wearable items To identify the confidence level threshold, For the first Class behavior stage The visibility weight of wearable items; the behavioral phase includes at least one or more of the following: equipment operation phase, material handling phase, welding and processing phase, and mobile inspection phase.

[0014] Specifically, dynamically assessing the attire of staff includes: Basic Wearable Status Determination: Based on the existence and consistency time sequence characteristics, the wearable status of the current window is initially determined to be compliant, suspected of being non-compliant, or abnormal. The basic wearable status determination formula is as follows: ; Among them, State is the basic wearable state; Short-term violation filtering: When the duration of the anomaly does not exceed the preset short-term anomaly judgment threshold, and the wearing status of the next adjacent time window is compliant, the anomaly is judged as a short-term anomaly and filtered out. The short-term anomaly judgment formula is as follows: ; ShortAbnormal represents the result of short-term anomaly detection. The duration of the anomaly. This is the threshold for short-term anomaly detection. This serves as the base wear state for the next adjacent time window; Cumulative violation confirmation: When the number of consecutive abnormal state windows exceeds a preset value, or the duration of a single abnormal state exceeds the short-term abnormality judgment threshold and the state change trend shows no recovery, it is judged as a cumulative violation. The cumulative violation judgment formula is: ; in, To accumulate violation judgment results, The number of time series windows where state anomalies occur consecutively; Trend represents the state change trend determined based on the state change time series curve.

[0015] Preferably, after dynamically judging the staff's wearing status, the method further includes a time-series compensation inference step under occlusion scenarios: Occlusion type is identified based on the state change time series curve and human posture trajectory in the time series features. The occlusion type determination formula is as follows: ; in, It is an occlusion type. Duration of occlusion; For temporary occlusion, linear interpolation is used to compensate for the recognition confidence during the occlusion period. The compensation formula is as follows: ; in, For the period of time of occlusion Frame compensation and confidence level identification. To determine the recognition confidence of the last frame before occlusion, The recognition confidence score for the first frame after occlusion. The sequence number of the last frame before the occlusion. The sequence number of the first frame after occlusion; For long-term occlusion, the wearing status during the occlusion period is inferred based on the historical wearing status time sequence characteristics, combined with the current behavior stage and environmental data; if the duration of occlusion exceeds the short-term abnormality judgment threshold and there is no trend of occlusion removal, it is determined whether it is a violation of occlusion based on the behavior stage, and if so, it is judged as a cumulative violation. For complete occlusion, the current wearing status can be inferred by combining the wearing standards of other staff in the same area and the wearing requirements of the current stage of behavior; The compensation inference result is compared with the wear status recognition result of the subsequent time window to verify the accuracy of the compensation inference. The verification formula for the compensation result is as follows: ; in, To compensate for errors, To compensate for the length of the verification window, This represents the actual recognition confidence level within the subsequent time window.

[0016] Furthermore, if the violation is determined to be cumulative, the complete violation event information will be output. The formula for calculating the duration of the violation is as follows: ; in, Indicates the duration of the violation. To accumulate the number of consecutive violation time windows, The length of each time window, This refers to the video capture frequency. The warning system is tiered based on the duration of the violation, and the formula for determining the tiered warning is as follows: ; in, It is at the warning level.

[0017] Furthermore, the monitoring method also includes a linked step of retaining violation traceability data and subsequent processing: Retain video time-series clips, wear status time-series feature data, and compensation inference records corresponding to the violation incidents to form a violation trajectory tracing archive; Link with access control systems, attendance systems, or safety production assessment systems to restrict the work permissions or impose mandatory safety training on personnel who have repeatedly violated regulations and failed to rectify them.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention improves the monitoring dimension from frame level to time level, which can effectively capture the changing pattern of the wearer status in the time dimension, accurately distinguish short-term abnormal behavior from real violations, and significantly reduce the false alarm rate and false alarm rate. At the same time, by introducing an adaptive adjustment mechanism for environmental data and behavior stages, it can still maintain stable recognition performance in complex industrial environments such as dust, oil, strong light, and mechanical obstruction. The final output violation event information includes complete elements such as start and end time, duration, and violation type, providing a continuous monitoring solution for the entire process from data collection to event output for industrial safety production supervision. (2) This invention presents the state evolution pattern of wearable items within a time window from multiple dimensions by extracting existence time features, duration of state change and consistency time features; existence features determine whether wearable items exist stably, duration of state change quantifies the duration of anomalies or occlusions, and consistency features assess the stability of the wearable state; the fusion of the three types of features provides rich time information for subsequent dynamic judgment, fundamentally solving the problem of misjudgment caused by instantaneous occlusion, motion interference and other factors in single-frame recognition; (3) This invention establishes a three-tiered progressive judgment mechanism: basic wear status judgment, short-term violation filtering, and cumulative violation confirmation. The basic judgment initially identifies compliance, suspected violations, or abnormal states. The short-term violation filtering excludes brief abnormal behaviors such as wiping sweat and adjusting protective gear, avoiding invalid warnings. The cumulative violation confirmation accurately identifies real violations by combining the number of consecutive abnormal windows or the duration of abnormality with the trend of status changes. This effectively solves the industry pain point of difficulty in distinguishing between short-term abnormalities and real violations, and greatly improves regulatory efficiency. (4) The present invention designs differentiated compensation strategies for temporary occlusion, long-term occlusion and complete occlusion. For temporary occlusion, linear interpolation is used to perform smooth compensation based on the confidence level before and after occlusion. For long-term occlusion, the wearing status is inferred by combining historical time sequence features and behavioral stage information, and intentional violation of occlusion can be determined. For complete occlusion, the wearing standards of other staff in the same area are used to assist in the inference. The compensation results are optimized in a closed loop through error verification, which effectively makes up for the recognition loss caused by occlusion and further reduces the false alarm rate and false alarm rate. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1This is a flowchart illustrating the wearability compliance monitoring method based on time-series features according to the present invention. Detailed Implementation

[0021] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] Reference Figure 1 This invention provides a method for continuous monitoring of wearability specifications based on time-series features, comprising the following steps: Step S1, Multi-source time series data acquisition and preprocessing: Simultaneously collect video and environmental data of the monitoring area, and preprocess the collected data.

[0023] Specifically, high-definition cameras and environmental sensors are deployed in key monitoring areas of the industrial test zone, such as the processing area, material handling area, high-altitude operation area, equipment operation area, and entrances / exits, to achieve synchronous and continuous acquisition of multi-source time-series data, as detailed below: S11, Dynamic Video Timing Data Acquisition: Continuously acquire dynamic video streams of workers in the industrial test area using high-definition camera equipment, with an acquisition frequency set to 25-30 FPS to ensure complete capture of workers' operational actions and temporal changes; the video stream is divided into continuous video segments at fixed time intervals as the basis for subsequent timing analysis, while the original video stream is retained for subsequent violation tracking; for strong light and backlight scenes in the industrial test area, the camera equipment is equipped with wide dynamic range and backlight compensation functions to ensure image acquisition clarity.

[0024] S12, Environmental Data Acquisition: Environmental parameters such as temperature, humidity, dust concentration, light intensity, and oil concentration in the industrial test area are collected through environmental sensing devices. The acquisition frequency is synchronized with the video acquisition frequency to determine whether there are excessive dust or oil pollution or strong direct light causing image blurring. This provides data support for subsequent time series compensation inference and time series window adjustment, adapting to the complex environmental monitoring needs of the industrial test area.

[0025] S13, Data Preprocessing: The acquired video time-series data is processed for deblurring, illumination equalization, and dust and noise removal. To address image blurring caused by dust and oil, an adaptive filtering algorithm is employed to preserve the detailed features of wearable items. To address issues such as overexposure in strong light and underexposure in backlight, adjust image brightness and contrast to prevent feature extraction failure. based on The criteria include outlier removal for environmental data, and the outlier removal formula is as follows: ; in, These are the environmental parameter values ​​collected in a single instance (temperature, humidity, dust concentration, light intensity, oil concentration). It is the average value of environmental parameters within a preset time window (default 10 seconds), used to characterize the baseline level of environmental parameters during that period; It is the standard deviation of environmental parameters within a preset time window (default 10 seconds), used to characterize the fluctuation range of environmental parameters.

[0026] Value range: temperature 15-40℃, humidity Dust concentration: 30%-80% 0-3 mg / m³, light intensity Oil concentration: 100-1000 Lux The concentration is 0-1.5 mg / m³, suitable for the normal environment of the industrial test area, and consistent with the environmental parameter settings in subsequent embodiments.

[0027] Step S2, Construction of wearable state temporal window and extraction of temporal features: Based on the collected video data, a continuously sliding time-series window is constructed, and the time-series features of the workers' attire status within each time-series window are extracted. Specifically, the steps include: S21, Timing Window Construction: A sliding window mechanism is adopted to construct a fixed-length timing window for the wearable status. The window length is set to 30-60 frames (corresponding to a 1-2 second video segment), and the sliding step size is set to 10-15 frames to ensure the continuity and overlap of the timing window, achieving seamless monitoring of the wearable status. The length of the timing window can be adaptively adjusted according to the operating rhythm of the staff in the industrial test area and the complexity of the environment. For example, when the dust concentration is too high, the window length is shortened to 30 frames to improve the response speed, while avoiding too many blurry frames affecting the judgment. The adaptive adjustment formula for the timing window length is: ; in, The adjusted timing window length (unit: frames). The standard timing window length is 30 frames by default (corresponding to 1 second, adapting to the normal operation rhythm). The current dust concentration (mg / m³) is collected in real time by environmental sensing equipment; This is the baseline value for dust concentration, with a default value of 2 mg / m³. This is the maximum threshold for dust concentration, with a default value of 6 mg / m³ (the safe upper limit for dust concentration in industrial test areas).

[0028] When dust concentration hour, Use standard window length; when At the same time, the window length decreases linearly with the increase of dust concentration, with a minimum of 30 frames, to avoid incomplete feature extraction due to excessively short windows and to ensure the effectiveness of temporal features.

[0029] S22, Target Area Extraction: Within each time window, the target area of ​​the worker's human body is segmented using a target detection model to eliminate background interference from production equipment, materials, tooling fixtures, walls, etc. in the industrial test area; at the same time, key areas of the human body, such as the head (corresponding to a safety helmet), face (corresponding to a protective mask), torso (corresponding to work clothes), hands (corresponding to protective gloves), and feet (corresponding to anti-smashing shoes), are located for wearing status recognition to identify the monitoring object.

[0030] S23, Temporal Feature Extraction: Analyze the key human body regions within each temporal window, extract three types of core temporal features, and construct a wearable state temporal feature vector to provide feature support for subsequent dynamic judgment. This includes the following steps: S231, Existence Temporal Features: For each type of wearable item (safety helmet, protective face mask, work clothes, protective gloves, anti-impact shoes), the recognition confidence sequence of the wearable item in consecutive frames within the time window is statistically analyzed. If the recognition confidence in consecutive frames is greater than the preset threshold (0.8-0.85), it is marked as "stable existence"; otherwise, it is marked as "abnormal existence". The existence time-series feature is quantified by the following formula: ; Where Exist represents the existence-based temporal feature result. This is the current time series window length; For the first The confidence level for identifying wearable items in the frame, with a value range of [0,1]. The higher the confidence level, the stronger the recognition accuracy. The default confidence threshold is 0.85, which can be adaptively adjusted according to the behavioral stage.

[0031] S232, Temporal characteristics of state changes: Analyze the state change patterns of wearable items within and between adjacent time windows, including the change time and duration of "wearing-removing", "removing-wearing", "covering-uncovering", and "uncovering-covering", and construct a state change temporal curve to capture the dynamic change trend of the wearable state, providing support for subsequent short-term anomaly filtering and cumulative violation confirmation. The duration of the state change is calculated using the following formula: ; in, The duration of the state change (in seconds) is used to determine the duration of occlusion or abnormal wearing. The frame number indicating the start of the state change. This is the frame number indicating the end of the state change. This is the video capture frequency, with a default of 30 frames per second.

[0032] S233, Consistency Timing Feature: Calculate the consistency index of the wearing state within the timing window, that is, the proportion of frames with the same wearing state in consecutive frames within the window. When the consistency index is ≥90%, the wearing state is judged to be stable; otherwise, the wearing state is judged to be unstable. This is used for subsequent violation judgment. The consistency time series feature is calculated using the following formula: ; in, It is a consistency index used to characterize the stability of the wear state within a time window; The number of frames with the same wearing status within the timing window.

[0033] Step S3, Identification of Worker Behavior Stages and Mapping of Wearing Standards in the Industrial Pilot Zone: Based on video data within each time window, the current operational behavior stage of the staff is identified, and a mapping between each behavior stage and the wearing guidelines is established. This includes the following steps: S31, Behavioral Stage Recognition: Through a behavioral recognition model, the human posture, movement trajectory, and temporal characteristics of workers within a time window are analyzed to divide the operational behaviors of workers in the industrial test zone into four core behavioral stages: equipment operation stage, material handling stage, welding processing stage, and mobile inspection stage. The equipment operation stage corresponds to the actions of workers operating machine tools, buttons, valves, etc.; the material handling stage corresponds to the actions of workers lifting, carrying, and stacking materials; the welding processing stage corresponds to the actions of workers holding welding torches and bending over to weld; and the mobile inspection stage corresponds to the actions of workers moving between different areas and inspecting equipment. The action characteristics of each behavioral stage closely match the actual operations in the industrial test zone, ensuring the rationality of the recognition scenario. The confidence level for the behavioral phase is calculated using the following formula: ; in, The current time sequence window belongs to the first Confidence level of the behavioral stage ( , which correspond to the equipment operation, material handling, welding processing, and mobile inspection stages respectively, and have a value range of [0,1]. For the first Frame action and the first Similarity of standard actions in the class behavior stage, with a value range of [0,1]; This is the current time series window length; Judgment criteria: Selection The behavior stage corresponding to the maximum value is used as the behavior stage of the current timing window; when the maximum value is less than 0.7, the behavior stage is determined to be unknown.

[0034] S32, Wearing Standard Mapping: Establish a mapping table linking various behavioral stages with wearing standards, clarifying the visibility weight and timing judgment thresholds for various wearable items under different behavioral stages; for example, in the welding process stage, when workers hold welding torches and their faces are close to the workpiece, protective masks are easily obscured, so the occlusion misjudgment threshold for this stage is reduced; in the material handling stage, when workers grasp materials with their hands, protective gloves are easily obscured, so the recognition weight of key hand areas is increased; in the mobile inspection stage, wearable items have high visibility, so standard judgment thresholds are used; the mapping logic fits the operational characteristics of various behavioral stages, specifically addressing the occlusion problem; Based on the identified behavioral stage, the confidence threshold for wearable item recognition is adaptively adjusted using the following formula: ; in, For the first Class behavior stage The confidence threshold for identifying wearable items; The default threshold for identifying the confidence level is 0.85. For the first Class behavior stage The visibility weight of wearable items ranges from 0.75 to 1.0. The smaller the weight, the lower the threshold, and the more suitable the scene is for the probability of occlusion.

[0035] In this embodiment, the recognition confidence thresholds for different types of wearable items at different stages are set as follows: Welding process stage ( ), protective face shield ( ), protective gloves ( Weight The corresponding threshold is 0.8; Material handling stage ( ), protective gloves ( ), anti-smashing shoes ( Weight The corresponding threshold is 0.75; Weighting of various wearable items during equipment operation and mobile inspection phases The corresponding standard threshold is 0.85.

[0036] Step S4, Dynamic determination of wearable status based on timing consistency: By combining the temporal characteristics of each time window and the corresponding association mapping, the staff's attire status is dynamically determined, specifically including the following steps: S41, Basic Wearing Status Determination: For each time window, combined with the judgment threshold corresponding to the behavior stage, the existence time sequence characteristics and consistency time sequence characteristics of various wearable items are analyzed to preliminarily determine the wearing status (compliant, abnormal, suspected violation) in the current window, providing a basis for subsequent dynamic judgment. The basic wearable status determination formula is: ; Among them, State is the basic wearable state.

[0037] S42, Short-term violation filtering: For each time window, combined with the judgment threshold corresponding to the behavior stage, the existence time characteristics and consistency time characteristics of various wearable items are analyzed to preliminarily determine the wearing status (compliant, abnormal, suspected violation) in the current window, providing a basis for subsequent dynamic judgment. The formula for judging short-term anomalies is: ; ShortAbnormal represents the result of short-term anomaly detection. The duration of the anomaly; The threshold for short-term anomaly detection is set to 4 seconds by default, but can be adjusted according to the operating rhythm of the industrial test zone (range: 3-5 seconds). This represents the basic wear state for the next adjacent time window.

[0038] S43, Cumulative Violation Confirmation: Set the criteria for cumulative violation judgment. If the same type of wearable item shows abnormal status in multiple consecutive time windows (≥3), or the duration of a single abnormal status exceeds the short-term abnormality judgment threshold, and the status change time curve shows no trend of recovery to compliance, it is judged as a cumulative violation (real violation). This condition can accurately capture real violations and avoid missed reports. The formula for determining cumulative violations is: ; in, To accumulate violation judgment results (i.e., actual violation judgment results); The number of consecutive time series windows with abnormal states is specified, with a default threshold of ≥3. Trend is the state change trend (no recovery / recovery) determined based on the state change time series curve, which is determined by the state change time series curve.

[0039] S44, Dynamic Judgment Iteration: The judgment result of the current time series window is correlated with the judgment result of the historical time series window, and the judgment result is iteratively optimized to avoid false alarms caused by misjudgment in a single window, ensure the continuity and accuracy of wearable status judgment, and further improve monitoring accuracy.

[0040] Step S5, temporal compensation inference in occluded scenes: This step addresses scenarios in industrial test zones where wearable items are temporarily obscured by production equipment or materials, or when the wearer is facing away from the device, or where dust, oil, or strong light cause incomplete wearable feature extraction. Based on the correlation of temporal features, it achieves wearable state compensation inference in obscured scenarios, further reducing the false negative and false positive rates. Specifically, it includes the following steps: S51, Occlusion Type Recognition: Occlusion types (temporary occlusion, long-term occlusion, complete occlusion, and partial occlusion) are identified through the state change time-series curve and human posture trajectory in the time-series features. Temporary occlusion refers to occlusion lasting ≤3 seconds and the occlusion is removed within the subsequent time-series window; long-term occlusion refers to occlusion lasting >3 seconds; complete occlusion means that the worn item is completely unrecognizable; partial occlusion means that some features of the worn item are recognizable. The occlusion type classification is consistent with the actual scenario of the industrial test zone and provides a basis for subsequent targeted compensation. The formula for determining the type of occlusion is as follows: ; in, It is an occlusion type. Duration of occlusion; Complete occlusion and partial occlusion are determined by the confidence level of a single frame. To completely block out, For partial occlusion ( (This is an adaptive threshold for the current behavior stage).

[0041] S51, Temporal Compensation Inference: Differentiated compensation strategies are adopted for different occlusion types to ensure accurate compensation results, as detailed below: S511, Temporary Obstruction Compensation: Based on the temporal characteristics of the wearing status before and after obstruction, a linear interpolation method is used to infer the wearing status during the obstruction period; for example, if the helmet remains stable before and after obstruction, it is inferred that the helmet is in a properly worn state during the obstruction period; the compensation method is tailored to the characteristics of temporary obstruction to ensure reasonable inference; the compensation formula is: ; in, For the period of time of occlusion Frame compensation and confidence level identification. To determine the recognition confidence of the last frame before occlusion, The recognition confidence score for the first frame after occlusion. The sequence number of the last frame before the occlusion. This is the sequence number of the first frame after the occlusion.

[0042] S512, Long-term occlusion compensation: Combining the current behavior stage and environmental data, the wear status during the occlusion period is inferred through a time-series analysis model based on the historical wear status time-series characteristics; at the same time, if the duration of occlusion exceeds the short-term anomaly judgment threshold and there is no trend of occlusion removal, it is determined whether it is a violation of occlusion (such as intentionally occluding worn items, or improper operation causing protective gear to fall off) based on the behavior stage. If it is a violation of occlusion, it is judged as a cumulative violation. S513, Complete Occlusion Compensation: If the worn items are completely unidentifiable, the current wearing status can be inferred by combining the wearing standards of other workers in the same area and the wearing requirements of the current behavior stage, thereby reducing the false negative rate; for example, if other workers in the welding processing area wear protective masks in accordance with regulations, it can be inferred that the workers in the area whose items are completely occluded should also wear them in accordance with regulations.

[0043] S514, Compensation Result Verification: Compare the compensation inference result with the wear status recognition result of the subsequent time window to verify the accuracy of the compensation inference. If there is a deviation, adjust the compensation algorithm parameters to improve the accuracy of subsequent compensation inference and ensure the optimizability of the compensation method. The formula for verifying the compensation result is: ; in, To compensate for errors, the accuracy of the compensation inference is evaluated; the smaller the error, the better the compensation effect. To compensate for the verification window length, the default is 30 frames, which is consistent with the standard timing window length; This represents the actual recognition confidence level within the subsequent time window.

[0044] Step S6, Violation event determination and output: Based on the judgment result, output violation event information including the start and end times of the violation, duration, and type of violation, specifically including the following steps: S61, Violation Event Determination: If it is determined to be a cumulative violation, it constitutes a complete violation event. The core information of the violation event must be clearly defined, including: the start and end time of the violation event, the duration of the violation, the type of violation (such as not wearing a safety helmet, not wearing a protective mask, not wearing protective gloves, etc.), the violating personnel, and the violating area; the core information must be complete to meet the needs of regulatory review and liability determination. The formula for calculating the duration of the violation is: ; in, Indicates the duration of the violation, used for tiered early warning systems; To accumulate the number of consecutive violation time windows, The length of each time window, This is the video capture frequency, with a default of 30 frames per second.

[0045] S62, Violation Incident Output: The core information of the violation incident is organized into structured data and output through the monitoring terminal. At the same time, a graded warning is triggered (the warning level is set according to the duration and type of violation). The warning information is pushed to the safety management personnel terminal and the on-site warning screen in the industrial pilot zone, so as to facilitate timely handling by management personnel and immediate rectification by on-site personnel. The graded warning logic is clear and meets the needs of industrial safety production supervision. The formula for determining the graded early warning is: ; in, The levels are designated as warning levels (Level 1 / Level 2 / Level 3). The higher the level, the more serious the violation and the greater the safety risk.

[0046] In this embodiment, the first-level warning only provides a local reminder and a light prompt on the on-site warning screen; the second-level warning pushes the information to the management personnel's APP and provides on-site voice reminders; and the third-level warning is linked to the on-site emergency stop device and access control system to forcibly stop operations in the illegal area, restrict the operating permissions of the illegal personnel, and strengthen the rectification of violations.

[0047] S63, Data Retention: Retain video time-series segments (10 seconds before the violation to 10 seconds after the violation is rectified), wearable status time-series feature data, compensation inference records, early warning processing records, etc., corresponding to the violation incidents, to form a violation trajectory tracing archive for subsequent regulatory review, responsibility determination, safety training and monitoring algorithm optimization.

[0048] S64, Follow-up Processing Linkage: It can be linked with the access control system, attendance system, and safety production assessment system of the industrial pilot zone to restrict the work permissions of personnel who have repeatedly violated regulations without rectification (such as restricting access to high-risk work areas), link violations to the safety production assessment and performance of staff, provide mandatory safety training for personnel who violate regulations multiple times, strengthen the enforcement of wearing regulations, and further improve the compliance rate of wearing regulations in the industrial pilot zone.

[0049] To further verify the feasibility and technical effectiveness of the present invention, based on the aforementioned full-process technical solution of "data acquisition - temporal modeling - behavior recognition - dynamic judgment - violation output", a typical mechanical processing industrial test area scenario was selected. The specific implementation process and verification results are as follows: (I) Overview of Implementation Scenarios Select 80m 2The mechanical processing industrial test zone (adapting to the promotion needs of various industrial test zone scenarios in the technical solution) covers four core monitoring areas: equipment operation area, material handling area, welding processing area, and entrance / exit. It has 12 staff members on duty, operating for 12 hours daily. During peak hours (8:30-11:30, 13:30-16:30), dust and oil concentrations are high. Staff frequently perform equipment operation, material handling, and welding processing, resulting in scenarios where personal protective equipment (PPE) is frequently obscured by production equipment and materials. This zone perfectly meets the monitoring needs of complex environments and dynamic operations in the technical solution. The monitoring targets are five types of PPE: safety helmets, face shields, work clothes, protective gloves, and anti-impact shoes. The core requirements are 24-hour continuous monitoring, short-term anomaly filtering, and accurate identification and tracing of violations, ensuring that the selected scenario aligns with the application requirements of the technical solution.

[0050] (II) System Deployment

[0051] One 1080P high-definition infrared explosion-proof camera was deployed in each of the four key monitoring areas of the industrial test zone, with a sampling frequency of 30FPS. The cameras in the welding processing area and equipment operation area were equipped with dust-proof and oil-proof protective covers to prevent dust and oil from affecting image acquisition. Three multi-parameter environmental sensors were deployed to simultaneously collect environmental parameters such as temperature, humidity, dust concentration, light intensity, and oil concentration (the sampling frequency is consistent with the video acquisition frequency). One edge computing node was deployed for core algorithm calculations such as temporal feature extraction, behavior stage recognition, and violation judgment. Two monitoring terminals were deployed (one in the safety management room and one in the on-site operation room) for early warning information output and violation data viewing. Two LED warning screens were also deployed on-site for on-site early warning prompts. The entire system does not require large-scale modification of the industrial test zone's production environment, is easy to deploy, and simple to operate and maintain. Compared with existing industrial-grade monitoring systems, the deployment cost is reduced by 32%.

[0052] (III) Parameter Settings

[0053] Environmental data processing: using 3 σ The criteria for removing outliers are based on a preset time window of 10 seconds, using the baseline average values ​​of temperature, humidity, dust concentration, light intensity, and oil concentration. The concentrations were set at 25℃, 60%, 1.5 mg / m³, 500 Lux, and 0.8 mg / m³, respectively. Timing window: Standard window length =30 frames, the sliding step size is set to 12 frames, and the window length is adjusted according to the real-time dust concentration based on the time window adaptive adjustment formula, with a minimum of 30 frames. Identification threshold: Standard threshold for identification confidence =0.85, the adaptive threshold for the behavior stage is set according to different wearing categories for different behavior stages (weights for protective masks and protective gloves in the welding process stage). =0.94, corresponding to a threshold of 0.8; weight of protective gloves and anti-impact shoes during material handling. =0.88, corresponding to a threshold of 0.75; weights for equipment operation and mobile inspection stages. =1.0, corresponding to the standard threshold); Violation Detection: Short-Term Anomaly Detection Threshold =4 seconds, the cumulative violation judgment condition is that the same type of wearable abnormality occurs in ≥3 consecutive time windows; Occlusion Compensation and Early Warning: Temporary occlusion is compensated using linear interpolation, with the compensation verification window length adjusted accordingly. =30 frames, error adjustment standard is Optimize parameters in real time; classify warnings into three levels based on the duration of violations: Level 1 warning is a local reminder + a light prompt on the warning screen; Level 2 warning is an app push notification for management personnel + on-site voice reminder; Level 3 warning is a linkage to the on-site emergency stop device + work permission restriction.

[0054] (iv) Specific implementation steps and timeliness verification

[0055] Step S1 (Multi-source time-series data acquisition): The camera continuously acquires dynamic video streams of the staff, and the environmental sensor synchronously acquires environmental parameters. In the preprocessing stage, an adaptive filtering algorithm is used to remove image blur caused by dust and oil. Illumination equalization and backlight compensation processing solve the problems of strong light and backlight. The accuracy rate of environmental data outlier removal reaches 99.8%, ensuring the accuracy of basic data. Step S2 (Construction of Temporal Window and Extraction of Temporal Features): Peak dust concentration is 2.2-3.0 mg / m³ (exceeding the baseline value). =2mg / m³), the temporal window length was adjusted to 30-28 frames according to the adaptive formula, and the three types of temporal features, existence, state change and consistency, were successfully extracted with a feature extraction accuracy of 99%; Step S3 (Behavioral Stage Recognition and Wearing Specification Mapping): Through the behavioral recognition model, four behavioral stages—equipment operation, material handling, welding processing, and mobile inspection—were successfully identified, and the confidence level was determined. All values ​​are ≥0.75, with a recognition accuracy of ≥95%, and the recognition threshold for wearable items is adaptively adjusted according to the behavioral stage to meet the occlusion requirements of different action scenarios. Step S4 (Dynamic Judgment): Short-term anomaly filtering was successfully achieved. Short-term (≤4 seconds) actions such as wiping sweat and adjusting protective gear, such as removing and putting on protective masks and helmets, were accurately filtered out, with no invalid false alarms. Wearing anomalies occurring at 3 or more consecutive windows (such as staff not wearing helmets for more than 4 seconds or not wearing protective gloves when handling materials) were accurately judged as cumulative violations, with a judgment accuracy rate of 99.2%. Step S5 (Obscuration Compensation Inference): For scenarios such as obstruction by protective masks and gloves during welding, protective gloves and anti-impact shoes during material handling, and obstruction by machine tools during equipment operation, temporary obstruction compensation is completed using linear interpolation, and long-term obstruction compensation is completed by combining historical time series characteristics, compensating for errors. ≤0.1, effectively reducing the false negative rate, and keeping the false negative rate below 0.9% in complex environments; Step S6 (Output of Violation Events): Violation events can be output with complete core information such as start and end time, duration, violation type, violating personnel and area. Violation trajectory video and time sequence data are completely retained, enabling full traceability, timely push of graded early warnings, response time of safety management personnel ≤30 seconds, and on-site violation rectification completion time ≤2 minutes on average, meeting the needs of industrial safety production supervision.

[0056] The results of this implementation and testing are as follows: the accuracy rate of wearable item recognition is ≥98.3%, the false alarm rate is ≤1.3%, the false negative rate is ≤0.9%, the short-term anomaly filtering accuracy rate is ≥99%, the effective recognition rate in complex environments is ≥92%, and the compliance rate of staff wearing the items has increased from 72% before implementation to 98.2%, which fully verifies the feasibility, practicality and accuracy of the technical solution of this invention.

[0057] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for continuous monitoring of wearability compliance based on temporal features, characterized in that, Includes the following steps: Simultaneously collect video and environmental data of the monitoring area, and preprocess the collected data; A continuously sliding time-series window is constructed based on the collected video data, and the time-series features of the staff's clothing status within each time-series window are extracted. Based on the video data within each time window, the current operational behavior stage of the staff is identified, and a mapping between each behavior stage and the wearing specifications is established. By combining the temporal characteristics of each time window and the corresponding association mapping, the staff's wearing status is dynamically judged; Based on the judgment result, output violation event information including the start and end time of the violation, duration, and type of violation.

2. The method for continuous monitoring of wearability compliance based on time-series features as described in claim 1, characterized in that, The monitoring area includes at least one or more of the following: processing area, material handling area, high-altitude operation area, equipment operation area, and entrance / exit; the environmental data includes at least one or more of the following: temperature, humidity, dust concentration, light intensity, and oil concentration.

3. The method for continuous monitoring of wearability specifications based on time-series features as described in claim 1, characterized in that, The preprocessing method is: based on The criteria include outlier removal for environmental data, and the outlier removal formula is as follows: ； in, These are environmental parameter values ​​collected in a single session. This is the average value of environmental parameters within a preset time window. This represents the standard deviation of environmental parameters within the preset timing window.

4. The method for continuous monitoring of wearability compliance based on time-series features as described in claim 1, characterized in that, The length of the time-series window is adaptively adjusted based on the dust concentration in the environmental data, and the adjustment formula is as follows: ； in, This is the adjusted timing window length. For standard timing window length, This represents the current dust concentration. This is the baseline value for dust concentration. This represents the maximum threshold for dust concentration.

5. The method for continuous monitoring of wearability specifications based on time-series features as described in claim 1, characterized in that, The temporal characteristics include existence temporal characteristics, duration of state changes, and consistency temporal characteristics; The existence time-series feature is quantified by the following formula: ； Where Exist represents the existence-based temporal feature result. This is the current time series window length; For the first Confidence level for identifying wearable items in a frame. To identify the confidence threshold; The duration of the state change is calculated using the following formula: ； in, The duration of the state change. The frame number indicating the start of the state change. This is the frame number indicating the end of the state change. This refers to the video capture frequency. The consistency time series feature is calculated using the following formula: ； in, As a consistency indicator, The number of frames with the same wearing status within the timing window.

6. The method for continuous monitoring of wearability specifications based on time-series features as described in claim 1, characterized in that, Methods for establishing a mapping between each behavioral stage and wearing guidelines include: By analyzing the actions of staff within a time-series window using a behavior recognition model, the behavior stage to which they belong is identified. The confidence level of the behavior stage is calculated using the following formula: ； in, The current time sequence window belongs to the first Confidence level of the behavioral stage. For the first Frame action and the first Similarity of standard actions in similar behavior stages The current time series window length; select The behavior stage corresponding to the maximum value is used as the behavior stage of the current time sequence window; Based on the identified behavioral stage, the confidence threshold for wearable item recognition is adaptively adjusted using the following formula: ； in, For the first Class behavior stage The confidence threshold for identifying wearable items To identify the confidence level threshold, For the first Class behavior stage The visibility weight of wearable items; the behavioral phase includes at least one or more of the following: equipment operation phase, material handling phase, welding and processing phase, and mobile inspection phase.

7. The method for continuous monitoring of wearability compliance based on time-series features as described in claim 5, characterized in that, Dynamically assessing the attire of staff includes: Basic Wearable Status Determination: Based on the existence and consistency time sequence characteristics, the wearable status of the current window is initially determined to be compliant, suspected of being non-compliant, or abnormal. The basic wearable status determination formula is as follows: ； Among them, State is the basic wearable state; Short-term violation filtering: When the duration of the anomaly does not exceed the preset short-term anomaly judgment threshold, and the wearing status of the next adjacent time window is compliant, the anomaly is judged as a short-term anomaly and filtered out. The short-term anomaly judgment formula is as follows: ； ShortAbnormal represents the result of short-term anomaly detection. The duration of the anomaly. This is the threshold for short-term anomaly detection. This serves as the base wear state for the next adjacent time window; Cumulative violation confirmation: When the number of consecutive abnormal state windows exceeds a preset value, or the duration of a single abnormal state exceeds the short-term abnormality judgment threshold and the state change trend shows no recovery, it is judged as a cumulative violation. The cumulative violation judgment formula is: ； in, To accumulate violation judgment results, The number of time series windows where state anomalies occur consecutively; Trend represents the state change trend determined based on the state change time series curve.

8. The method for continuous monitoring of wearability specifications based on time-series features as described in claim 1, characterized in that, After dynamically determining the staff's attire status, the process also includes a timing compensation inference step for occlusion scenarios: Occlusion type is identified based on the state change time series curve and human posture trajectory in the time series features. The occlusion type determination formula is as follows: ； in, It is an occlusion type. Duration of occlusion; For temporary occlusion, linear interpolation is used to compensate for the recognition confidence during the occlusion period. The compensation formula is as follows: ； in, For the period of time of occlusion Frame compensation and confidence level identification. To determine the recognition confidence of the last frame before occlusion, The recognition confidence score for the first frame after occlusion. The sequence number of the last frame before the occlusion. The sequence number of the first frame after occlusion; For long-term occlusion, the wearing status during the occlusion period is inferred based on the historical wearing status time sequence characteristics, combined with the current behavior stage and environmental data; if the duration of occlusion exceeds the short-term abnormality judgment threshold and there is no trend of occlusion removal, it is determined whether it is a violation of occlusion based on the behavior stage, and if so, it is judged as a cumulative violation. For complete occlusion, the current wearing status can be inferred by combining the wearing standards of other staff in the same area and the wearing requirements of the current stage of behavior; The compensation inference result is compared with the wear status recognition result of the subsequent time window to verify the accuracy of the compensation inference. The verification formula for the compensation result is as follows: ； in, To compensate for errors, To compensate for the length of the verification window, This represents the actual recognition confidence level within the subsequent time window.

9. The method for continuous monitoring of wearability specifications based on time-series features as described in claim 7, characterized in that, If the violation is determined to be cumulative, output the complete violation event information; The formula for calculating the duration of the violation is as follows: ； in, Indicates the duration of the violation. To accumulate the number of consecutive violation time windows, The length of each time window, This refers to the video capture frequency. The warning system is tiered based on the duration of the violation, and the formula for determining the tiered warning is as follows: ； in, It is at the warning level.

10. The method for continuous monitoring of wearability specifications based on time-series features as described in claim 1, characterized in that, It also includes the linkage between the retention of violation tracking data and subsequent processing: Retain video time-series clips, wear status time-series feature data, and compensation inference records corresponding to the violation incidents to form a violation trajectory tracing archive; Link with access control systems, attendance systems, or safety production assessment systems to restrict the work permissions or impose mandatory safety training on personnel who have repeatedly violated regulations and failed to rectify them.

Images