Safety production management method and production safety management system based on safety helmet detection

Abstract

The invention discloses a safety production management method and a production safety management system based on safety helmet detection, and relates to the technical field of production safety management. Multi-modal data acquisition is realized by constructing an explosion-proof intelligent camera cluster and a wearable positioning terminal, an improved YOLOv7-Tiny model and an HRNet key point detection technology are adopted, illumination compensation, atomization elimination and data enhancement technologies are combined, and precise recognition of the wearing state, the angle deviation and the compliance of the safety helmet is completed; triggering hierarchical response in real time based on a five-level dynamic early warning rule, predicting a future risk probability of an operator through an optimized MLP prediction network in combination with historical violation data, and formulating a personalized scheme; according to the invention, a closed-loop management system of real-time monitoring, intelligent early warning, behavior prediction and active intervention is formed, the occurrence rate of safety accidents is significantly reduced, and the safety standardization of operators and the safety management efficiency of enterprises are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of environmental management, and in particular relates to a production safety management method and a production safety management system based on safety helmet detection. Background Art

[0002] In recent years, the rapid development of high-risk industries such as petrochemicals, mining, construction, and power operations has significantly increased the complexity of operating environments, and the risk of safety accidents continues to rise. Despite increasingly stringent national safety regulations, traditional safety management models continue to face severe challenges.

[0003] In existing technologies, data collection relies on single visible light data, which is easily affected by dynamic environmental interference and has low detection accuracy; helmet detection focuses more on wearing status identification, and lacks detailed analysis of wearing angle deviation and helmet compliance; a fixed threshold is used to trigger alarms, and the response strategy is not dynamically adjusted in combination with the risk level of the personnel position, resulting in the homogenization of risk treatment in high-risk areas and non-high-risk areas; it only relies on real-time monitoring, lacks in-depth mining of historical violation data and prediction of future violation probabilities, and cannot achieve proactive safety intervention. Summary of the Invention

[0004] (1) Technical problems solved

[0005] In response to the problems in the related art, the present invention provides a safety production management method based on safety helmet detection to overcome the above-mentioned technical problems existing in the existing related art.

[0006] (2) Technical solution

[0007] To solve the above technical problems, the present invention is achieved through the following technical solutions:

[0008] S1. Deploy explosion-proof smart cameras to obtain an explosion-proof smart camera cluster; collect and process production environment data through the explosion-proof smart camera cluster to obtain processed synchronized multimodal data; and set multi-level warning rules;

[0009] S2. Improve the YOLOv7-Tiny model to obtain an improved YOLOv7-Tiny model; use the improved YOLOv7-Tiny model, the HRNet key point detection model, and the enterprise-level hard hat feature database to analyze the processed synchronous multimodal data to obtain hard hat wearing classification results, worker location, real-time angle deviation, and compliance detection results;

[0010] S3. Executing multi-level warning rules based on the helmet wearing classification results, operator position, real-time angle deviation, and compliance detection results;

[0011] S4. Inputting the violation data of the operator to be predicted during the violation period into an optimized MLP operator violation probability prediction network obtained by combining historical warning data with an optimization algorithm to obtain the operator's future violation probability; formulating a training and work plan for the operator based on the future violation probability;

[0012] By integrating deep learning technology, multimodal data fusion and multi-level warning rules, the present invention achieves high-precision real-time detection of helmet wearing status, angle deviation and compliance in high-risk work scenarios, dynamically triggers graded warning responses, and combines the violation probability prediction model with the personalized solution mechanism to form a "real-time monitoring-intelligent warning-behavior prediction-active intervention" closed-loop management system, which significantly reduces the occurrence rate of safety accidents, improves the safety standardization of operators and the safety management efficiency of enterprises.

[0013] Preferably, the S1 comprises the following steps:

[0014] S11. Construct a set of key locations in the operation area; deploy explosion-proof smart cameras at key locations in the set of key locations in the operation area to obtain an explosion-proof smart camera cluster; the explosion-proof smart cameras integrate visible light cameras, thermal imaging sensors, and edge computing nodes;

[0015] S12. Equip the operator with a wearable positioning terminal that integrates an RFID chip, a UWB locator, and a vibration alarm; the RFID chip stores the operator's work number and position information; and the vibration alarm supports multi-level vibration feedback.

[0016] S13. Set the first-level warning rule to a vibration alarm that reminds workers to wear safety helmets correctly; set the second-level warning rule to a vibration alarm that vibrates and sounds to remind workers to wear safety helmets correctly; set the third-level warning rule to stop working equipment in high-risk areas, and the vibration alarm that vibrates and sounds to remind workers to wear safety helmets correctly; set the fourth-level warning rule to a vibration alarm that vibrates and sounds to remind workers to change into compliant safety helmets; set the fifth-level warning rule to stop working equipment in high-risk areas, and the vibration alarm that vibrates and sounds to remind workers to change into compliant safety helmets;

[0017] S14. Receive the video stream, thermal imaging data, and positioning signal in real time through the edge computing node, and obtain synchronized multimodal data using a spatiotemporal alignment algorithm; perform dynamic illumination compensation, fog elimination, and data enhancement on the synchronized multimodal data to obtain processed synchronized multimodal data;

[0018] The above steps build a multi-source perception network by deploying a cluster of explosion-proof smart cameras and wearable terminals that include visible light / thermal imaging / edge computing. Combined with five-level warning rules and edge computing data fusion, this network enables accurate identification and dynamic response to security risks in complex environments, significantly reducing missed detections and false detections, while also reducing unnecessary equipment downtime in high-risk areas and improving safety management efficiency and resource utilization.

[0019] Preferably, said S2 comprises the following steps:

[0020] S21. Improve the YOLOv7-Tiny model to obtain an improved YOLOv7-Tiny model;

[0021] S22. Perform preliminary detection on the processed synchronized multimodal data using the improved YOLO-v5 to obtain a helmet wearing classification result; detect that the operator is in a high-risk work area using the UWB locator, and obtain the operator's location;

[0022] S23, setting a set of key head position points; setting a safety deviation threshold;

[0023] If the helmet wearing classification result is wearing a helmet, the HRNet key point detection model is used to identify the key points of the person's head according to the head key position point set. Combined with the UWB positioning data, the angle deviation between the helmet and the head is calculated to obtain the real-time angle deviation.

[0024] S24. Setting compliance detection rules and establishing an enterprise-level safety helmet feature database. The compliance detection rules include color coding, enterprise logo recognition, and reflective strip detection.

[0025] Based on the enterprise-level safety helmet feature database and compliance detection rules, detect whether the safety helmet worn by the operator is compliant, and obtain a compliance detection result;

[0026] The above steps improve detection speed and accuracy through the improved YOLOv7-Tiny model, combined with HRNet key point detection and UWB positioning data to accurately calculate the angle deviation of the helmet wearing. At the same time, based on the enterprise-level feature library, multi-dimensional compliance verification is achieved, effectively solving the problems of missed detection, false detection, improper wearing and substandard helmets in traditional detection, reducing accident risks and improving the standardization of safety management.

[0027] Preferably, the S21 includes the following steps:

[0028] S211. Build a YOLOv7-Tiny model and replace the C3 module in the backbone network of the YOLOv7-Tiny model with the RepVGG module through structural reparameterization technology; use multi-branch convolution in the training phase and merge it into a single 3×3 convolution in the inference phase;

[0029] The GIoU loss function of the YOLOv7-Tiny model is replaced by the CIoU loss function through the geometric perception constraint method, introducing the center point distance penalty term and aspect ratio consistency constraint;

[0030] S212, obtain the improved YOLOv7-Tiny model through S211;

[0031] The above steps are used to obtain an improved YOLOv7-Tiny model through the YOLOv7-Tiny model. Compared with the YOLOv7-Tiny model, the improved YOLOv7-Tiny model enhances the feature extraction capability, achieves improved detection accuracy and inference speed, improves the small target detection accuracy, reduces the positioning error of the long strip helmet, and improves the training convergence efficiency.

[0032] Preferably, the step S3 includes the following steps:

[0033] S31. If the helmet wearing classification result is not worn and the worker is not in a high-risk work area, a second-level warning is executed. If the helmet wearing classification result is not worn and the worker is in a high-risk work area, a third-level warning is executed and the warning data is uploaded to the database. Otherwise, no warning is executed.

[0034] S32. When the real-time angle deviation is greater than or equal to the safety deviation threshold and the operator is not in a high-risk work area, a first-level warning is executed. When the real-time angle deviation is greater than or equal to the safety deviation threshold and the operator is in a high-risk work area, a third-level warning is executed and the warning data is uploaded to the database. Otherwise, no warning is executed.

[0035] S33. If the compliance test result is non-compliant and the operator is not in a high-risk work area, a level 4 warning is executed. If the compliance test result is non-compliant and the operator is in a high-risk work area, a level 5 warning is executed and the warning data is uploaded to the database. Otherwise, no warning is executed.

[0036] The above steps achieve precise early warning response through a dynamic graded early warning mechanism, combined with a composite judgment of the risk level of personnel positions and safety status (not wearing or angle deviation / non-compliance) in high-risk or non-high-risk areas: directly trigger equipment shutdown (level three warning) for problems of not wearing or angle deviation in high-risk areas, and use vibration / sound and light reminders (level one to level two) in non-high-risk areas to avoid excessive interference in production; for non-compliant safety helmets, replacement reminders (level four) or forced replacement (level five) are activated according to regional classification; through multi-dimensional risk weight assessment, unnecessary downtime is reduced while ensuring safety. At the same time, early warning data is uploaded in real time to build a violation database to provide support for subsequent prediction model training, forming a "risk identification-graded disposal-data sedimentation" closed loop, and improving the adaptive ability of safety management and resource utilization efficiency.

[0037] Preferably, said S4 comprises the following steps:

[0038] S41. Collect historical warning data from a database to obtain historical safety helmet warning data, wherein the historical safety helmet warning data includes data on illegal workers, violation frequency data, violation area data, and violation time series data;

[0039] S42, constructing an MLP prediction network, and setting weight values ​​and bias values ​​of the MLP prediction network;

[0040] S43, setting a ratio of training data to test data according to the MLP prediction network; dividing the historical helmet warning data into training data and test data according to the ratio of training data to test data;

[0041] S44, setting the training accuracy threshold of the MLP prediction network to j1, the training accuracy to j2, and the maximum number of training iterations to k; training the MLP prediction network using the training data; adjusting the network structure of the MLP prediction network during the training process; when the training accuracy j2 ≥ the training accuracy threshold j1 or the maximum number of training iterations k is reached, obtaining the MLP operator violation probability prediction network;

[0042] S45, using the test data to test the MLP operator violation probability prediction network, after the test is completed, an optimized MLP operator violation probability prediction network is obtained;

[0043] S46. Set the violation period to t, collect the violation data of the operator to be predicted within the violation period t, input the violation data of the operator within the violation period into the optimized MLP operator violation probability prediction network, and obtain the future violation probability of the operator;

[0044] S47. Set a violation probability threshold. If the operator's future violation probability is greater than or equal to the violation probability threshold, provide the operator with VR safety training and develop a personalized inspection route plan to avoid dangerous areas until the operator's future violation probability is less than the violation probability threshold.

[0045] The above steps construct an optimized MLP prediction network, integrate historical violation data including violation frequency, regional distribution and time series with the hyena optimization algorithm parameter adjustment technology, and achieve accurate prediction of the future violation probability of operators; extract multi-dimensional violation feature data from the database and divide it into training set and test set, and ensure the generalization ability of the model by dynamically adjusting the network structure and iterative training; use the optimization algorithm to adaptively adjust the learning rate in the testing phase to further improve the prediction accuracy; feature encode the violation data within period t and input it into the optimization model. If the output probability exceeds the threshold, VR safety training and inspection path planning will be triggered, forming a "prediction-intervention-reassessment" closed loop; transform traditional passive supervision into active prevention, through data-driven mining of potential risk individuals, targeted correction of behavioral habits, reduced violation recurrence rate, reduced ineffective inspections, optimized human resource allocation, and the long-term accumulation of violation data can reversely optimize the enterprise safety management system, promoting the intelligent upgrade of safety management from "ex post accountability" to "ex ante control".

[0046] Preferably, the S45 includes the following steps:

[0047] S451, setting a test error threshold to p1; using the test data to test the MLP operator violation probability prediction network, obtaining a test error of p2;

[0048] S452. When the test error p2 is less than the test error threshold p1, the MLP operator violation probability prediction network is used as an optimized MLP operator violation probability prediction network; when the test error p2 is greater than or equal to the test error threshold p1, an optimization algorithm is used to find a learning rate for the MLP operator violation probability prediction network to obtain an optimal solution; the optimal solution is used as the learning rate of the MLP operator violation probability prediction network to obtain an optimized MLP operator violation probability prediction network;

[0049] The above steps dynamically determine model performance by testing the error threshold. If the threshold is not met, an optimization algorithm is used to adaptively adjust the learning rate to ensure that the MLP network prediction accuracy meets the requirements. This avoids the inefficiency of traditional manual parameter adjustment, improves the model's generalization ability and stability, ensures the high reliability of violation probability prediction results, and reduces the cost of ineffective intervention caused by misjudgment.

[0050] Preferably, in S452, the step of using an optimization algorithm to find the learning rate of the MLP operator violation probability prediction network to obtain the optimal solution includes the following steps:

[0051] S4521. Construct a hyena population. Set the size of the hyena population to r. Then the hyena population is represented by u = {u1, u2, ..., u i ,...,u r}, where u i Represents the i-th hyena in the hyena population; sets the maximum number of optimization iterations;

[0052] S4522, according to the weight value and bias value of the MLP operator violation probability prediction network, randomly set the initial position of the hyena population, and obtain the hyena population initial position set m = {(m 11 ,m 12 ),(m 21 ,m 22 ),...,(m i1 ,m i2 ),...,(m r1 ,m r2 )}; where m i1 、m i2 They represent the first dimension coordinate value and the second dimension coordinate value of the i-th hyena in the hyena population, respectively. The position of the hyena represents the distance from the prey;

[0053] S4523. Based on the test error threshold p1 and the training accuracy p2, define the fitness function of the hyena position in the hyena population. The fitness function formula is as follows:

[0054]

[0055] x represents the fitness function, z represents the bias;

[0056] S4524. Iterate the initial position set of the hyena population, wherein the higher the fitness value, the better the position. During each round of iteration, calculate the fitness value of each position in the initial position set of the hyena population according to the fitness function, and update the position of each hyena in the initial position set of the hyena population in descending order of fitness value. In each round of iteration, obtain the best individual hyena position in the hyena population and the global best hyena position.

[0057] S4525, repeat S4524, when the maximum number of optimization iterations is reached, stop the iteration, and take the global best hyena position as the optimal solution;

[0058] The above steps adaptively adjust the MLP network learning rate by adopting the hyena optimization algorithm; construct a hyena population and randomly initialize its positions to map the network parameter space; define a fitness function, and integrate the test error threshold and training accuracy to quantify the quality of the parameters; iteratively update the population position, and the higher the fitness, the closer it is to the optimal solution, and finally output the global best parameters as the optimal learning rate; avoid falling into the limitation of local optimum, combine swarm intelligence to quickly converge to the global optimum, improve the model prediction stability, and avoid the subjectivity of manual parameter adjustment, ensuring the dual optimization of violation probability prediction accuracy and efficiency.

[0059] A production safety management system based on hard hat detection, used to implement the above-mentioned production safety management method based on hard hat detection, including an operation area monitoring and early warning system deployment module, a multimodal data processing and safety detection module, a multi-level early warning execution and data management module, and a violation prediction and behavior optimization module;

[0060] The operation area monitoring and early warning system deployment module is responsible for building an intelligent monitoring system for the operation area. By defining a set of key locations and deploying a cluster of explosion-proof intelligent cameras, multi-dimensional environmental perception is achieved. Workers are equipped with wearable terminals that integrate RFID chips, UWB locators, and multi-level vibration alarm functions. This, combined with five-level early warning rules, forms a dynamic risk response framework, ensuring full coverage from hardware deployment to early warning logic.

[0061] The multimodal data processing and safety detection module is based on the improved YOLOv7-Tiny model and the HRNet key point detection model, and integrates the multimodal data after spatiotemporal alignment to achieve helmet wearing status classification, head posture angle deviation calculation and compliance detection, and finally outputs the personnel position, helmet status and compliance results, providing high-precision data support for early warning decision-making;

[0062] The multi-level warning execution and data management module is used to dynamically trigger graded warnings based on real-time detection results: for the behavior of not wearing a safety helmet, a second or third level warning is executed depending on whether the person is in a high-risk area; if the safety helmet wearing angle deviation exceeds the limit, a first or third level vibration reminder is triggered; when the safety helmet is detected to be non-compliant, a fourth or fifth level warning is activated in combination with the regional risk level; all warning events are synchronously uploaded to the database to form a complete record of violations, providing a structured data source for subsequent analysis.

[0063] The violation prediction and behavior optimization module constructs an optimized MLP prediction network based on historical warning data, uses the hyena optimization algorithm to adjust the learning rate parameters, and predicts the probability of future violations by operators by analyzing the frequency, region and time series characteristics within the violation cycle. If the probability exceeds the threshold, VR safety training and personalized inspection route planning are initiated, forming a "prediction-intervention-feedback" closed loop until the violation probability drops below the safety threshold, realizing an intelligent upgrade from passive warning to active behavior correction.

[0064] (3) Beneficial effects

[0065] The present invention has the following beneficial effects:

[0066] This invention fuses multimodal data from complex environments, adopts an improved YOLOv7-Tiny model and HRNet key point detection technology to achieve multi-dimensional and accurate detection of helmet wearing status, angular deviation and compliance. It combines five-level dynamic warning rules with an optimized MLP prediction network to achieve full-chain management and control from real-time risk blocking to violation prediction, forming a "perception-warning-prediction-optimization" closed loop, significantly improving the safety protection efficiency and standardized management level in high-risk work scenarios.

[0067] This invention integrates multi-source sensors such as visible light and thermal imaging through an explosion-proof intelligent camera cluster, and combines illumination compensation, defogging, and data enhancement technologies to effectively solve the problems of uneven illumination, haze interference, and local occlusion. The spatiotemporal alignment algorithm achieves high-precision synchronization of multimodal data, significantly improving the stability of target detection in dynamic scenes.

[0068] This paper improves the YOLOv7-Tiny model, increasing the model inference speed and reducing system energy consumption while ensuring detection accuracy, making it suitable for long-term stable operation in various explosion-proof scenarios. Based on the improved YOLOv7-Tiny model and the HRNet key point detection model, it achieves multi-dimensional detection of helmet wearing status, head angle deviation, and compliance. It also combines the five-level warning rules and regional risk classification strategy to dynamically match the warning level.

[0069] The present invention constructs an optimized MLP prediction network combined with an optimization algorithm, integrates historical violation counts, regional distribution, and time series characteristics, and accurately predicts the probability of future violations by operators. Combined with VR safety training and personalized inspection route planning, targeted behavioral corrections are implemented for high-probability violators, effectively reducing the incidence of violation incidents, reducing training costs, and achieving a transformation from "post-event disposal" to "pre-event prevention."

[0070] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] In order to more clearly illustrate the technical solutions of the embodiments of the invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the invention. For ordinary technicians in this field, they can also obtain drawings based on these drawings without paying any creative work.

[0072] Figure 1 Schematic diagram of the process of the safety production management method based on helmet detection of the present invention;

[0073] Figure 2 This is a module diagram of the production safety management system based on helmet detection of the present invention. DETAILED DESCRIPTION

[0074] The following will clearly and completely describe the technical solutions in the embodiments of the invention in conjunction with the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0075] In the description of the present invention, it should be understood that the terms "opening", "upper", "lower", "top", "middle", "inside" and the like indicating orientation or positional relationship are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the components or elements referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore cannot be understood as limiting the invention.

[0076] Example 1:

[0077] See also Figure 1 The present invention discloses a safety production management method based on safety helmet detection, comprising the following steps:

[0078] S1. Deploy explosion-proof smart cameras to obtain an explosion-proof smart camera cluster; collect and process production environment data through the explosion-proof smart camera cluster to obtain processed synchronized multimodal data; and set multi-level warning rules;

[0079] Said S1 comprises the following steps:

[0080] S11, construct the key position point set of the operation area a={a1,a2,...,a i ,...,a b}; where a irepresents the i-th key location point in the operation area, and b represents the total number of key location points in the operation area; explosion-proof smart cameras are deployed at key locations in the operation area to obtain an explosion-proof smart camera cluster; the key locations include high-risk operation areas, entrances and exits, and commanding heights; the explosion-proof smart cameras integrate visible light cameras, thermal imaging sensors, and edge computing nodes;

[0081] S12. Equip the operator with a wearable positioning terminal that integrates an RFID chip, a UWB locator, and a vibration alarm; the RFID chip stores the operator's work number and position information; and the vibration alarm supports multi-level vibration feedback.

[0082] S13. Set the first-level warning rule to a vibration alarm that reminds workers to wear safety helmets correctly; set the second-level warning rule to a vibration alarm that vibrates and sounds to remind workers to wear safety helmets correctly; set the third-level warning rule to stop working equipment in high-risk areas, and the vibration alarm that vibrates and sounds to remind workers to wear safety helmets correctly; set the fourth-level warning rule to a vibration alarm that vibrates and sounds to remind workers to change into compliant safety helmets; set the fifth-level warning rule to stop working equipment in high-risk areas, and the vibration alarm that vibrates and sounds to remind workers to change into compliant safety helmets;

[0083] S14. Receive video streams, thermal imaging data, and positioning signals in real time through the edge computing node, and obtain synchronized multimodal data using a spatiotemporal alignment algorithm; perform dynamic illumination compensation, fog removal, and data enhancement on the synchronized multimodal data using Retinex-based ambient lighting modeling, a DehazeNet dehazing algorithm, and random occlusion simulation multi-angle rotation expansion to obtain processed synchronized multimodal data;

[0084] S2. Improve the YOLOv7-Tiny model to obtain an improved YOLOv7-Tiny model; use the improved YOLOv7-Tiny model, the HRNet key point detection model, and the enterprise-level hard hat feature database to analyze the processed synchronous multimodal data to obtain hard hat wearing classification results, worker location, real-time angle deviation, and compliance detection results;

[0085] The S2 comprises the following steps:

[0086] S21. Improve the YOLOv7-Tiny model to obtain an improved YOLOv7-Tiny model;

[0087] The S21 includes the following steps:

[0088] S211. Build a YOLOv7-Tiny model and replace the C3 module in the backbone network of the YOLOv7-Tiny model with the RepVGG module through structural reparameterization technology; use multi-branch convolution in the training phase and merge it into a single 3×3 convolution in the inference phase;

[0089] The GIoU loss function of the YOLOv7-Tiny model is replaced by the CIoU loss function through the geometric perception constraint method, introducing the center point distance penalty term and aspect ratio consistency constraint;

[0090] S212, obtain the improved YOLOv7-Tiny model through S211;

[0091] S22. Perform preliminary detection on the processed synchronized multimodal data using the improved YOLO-v5 to obtain a helmet wearing classification result; detect that the operator is in a high-risk work area using the UWB locator, and obtain the operator's location;

[0092] S23, set the key position point set of the head c={c1,c2,...,c i ,...,c d}; where c i represents the i-th key position point of the head, and d represents the total number of key position points of the head; the key position points of the head are such as the top of the head, the temple, and the back of the head; set a safety deviation threshold;

[0093] If the helmet wearing classification result is wearing a helmet, the HRNet key point detection model is used to identify the key points of the person's head according to the head key position point set. Combined with the UWB positioning data, the angle deviation between the helmet and the head is calculated to obtain the real-time angle deviation.

[0094] S24. Setting compliance detection rules and establishing an enterprise-level safety helmet feature database. The compliance detection rules include color coding using HSV color space matching, enterprise logo recognition using Siamese network-based feature comparison, and reflective strip detection using thermal imaging data reflective characteristics analysis.

[0095] Based on the enterprise-level safety helmet feature database and compliance detection rules, detect whether the safety helmet worn by the operator is compliant, and obtain a compliance detection result;

[0096] S3. Executing multi-level warning rules based on the helmet wearing classification results, operator position, real-time angle deviation, and compliance detection results;

[0097] The S3 includes the following steps:

[0098] S31. If the helmet wearing classification result is not worn and the worker is not in a high-risk work area, a second-level warning is executed. If the helmet wearing classification result is not worn and the worker is in a high-risk work area, a third-level warning is executed and the warning data is uploaded to the database. Otherwise, no warning is executed.

[0099] S32. When the real-time angle deviation is greater than or equal to the safety deviation threshold and the operator is not in a high-risk work area, a first-level warning is executed. When the real-time angle deviation is greater than or equal to the safety deviation threshold and the operator is in a high-risk work area, a third-level warning is executed and the warning data is uploaded to the database. Otherwise, no warning is executed.

[0100] S33. If the compliance test result is non-compliant and the operator is not in a high-risk work area, a level 4 warning is executed. If the compliance test result is non-compliant and the operator is in a high-risk work area, a level 5 warning is executed and the warning data is uploaded to the database. Otherwise, no warning is executed.

[0101] S4. Inputting the violation data of the operator to be predicted during the violation period into an optimized MLP operator violation probability prediction network obtained by combining historical warning data with an optimization algorithm to obtain the operator's future violation probability; formulating a training and work plan for the operator based on the future violation probability;

[0102] The S4 comprises the following steps:

[0103] S41. Collect historical warning data from a database to obtain historical safety helmet warning data, wherein the historical safety helmet warning data includes data on illegal workers, violation frequency data, violation area data, and violation time series data;

[0104] S42, constructing an MLP prediction network, and setting weight values ​​and bias values ​​of the MLP prediction network;

[0105] S43, setting a ratio of training data to test data according to the MLP prediction network; dividing the historical helmet warning data into training data and test data according to the ratio of training data to test data;

[0106] S44, setting the training accuracy threshold of the MLP prediction network to j1, the training accuracy to j2, and the maximum number of training iterations to k; training the MLP prediction network using the training data; adjusting the network structure of the MLP prediction network during the training process; when the training accuracy j2 ≥ the training accuracy threshold j1 or the maximum number of training iterations k is reached, obtaining the MLP operator violation probability prediction network;

[0107] S45, using the test data to test the MLP operator violation probability prediction network, after the test is completed, an optimized MLP operator violation probability prediction network is obtained;

[0108] The S45 includes the following steps:

[0109] S451, setting a test error threshold to p1; using the test data to test the MLP operator violation probability prediction network, obtaining a test error of p2;

[0110] S452. When the test error p2 is less than the test error threshold p1, the MLP operator violation probability prediction network is used as an optimized MLP operator violation probability prediction network; when the test error p2 is greater than or equal to the test error threshold p1, an optimization algorithm is used to find a learning rate for the MLP operator violation probability prediction network to obtain an optimal solution; the optimal solution is used as the learning rate of the MLP operator violation probability prediction network to obtain an optimized MLP operator violation probability prediction network;

[0111] In S452, the optimization algorithm is used to find the learning rate of the MLP operator violation probability prediction network, and obtaining the optimal solution includes the following steps:

[0112] S4521. Construct a hyena population. Set the size of the hyena population to r. Then the hyena population is represented by u = {u1, u2, ..., u i ,...,u r}, where u i Represents the i-th hyena in the hyena population; sets the maximum number of optimization iterations;

[0113] S4522, according to the weight value and bias value of the MLP operator violation probability prediction network, randomly set the initial position of the hyena population, and obtain the hyena population initial position set m = {(m 11 ,m 12 ),(m 21 ,m 22 ),...,(m i1 ,m i2 ),...,(m r1 ,m r2 )}; where m i1 、m i2 They represent the first dimension coordinate value and the second dimension coordinate value of the i-th hyena in the hyena population, respectively. The position of the hyena represents the distance from the prey;

[0114] S4523. Based on the test error threshold p1 and the training accuracy p2, define the fitness function of the hyena position in the hyena population. The fitness function formula is as follows:

[0115]

[0116] x represents the fitness function, z represents the bias;

[0117] S4524. Iterate the initial position set of the hyena population, wherein the higher the fitness value, the better the position. During each round of iteration, calculate the fitness value of each position in the initial position set of the hyena population according to the fitness function, and update the position of each hyena in the initial position set of the hyena population in descending order of fitness value. In each round of iteration, obtain the best individual hyena position in the hyena population and the global best hyena position.

[0118] S4525, repeat S4524, when the maximum number of optimization iterations is reached, stop the iteration, and take the global best hyena position as the optimal solution;

[0119] S46. Set the violation period to t, collect the violation data of the operator to be predicted within the violation period t, input the violation data of the operator within the violation period into the optimized MLP operator violation probability prediction network, and obtain the future violation probability of the operator;

[0120] S47. Set a violation probability threshold. If the operator's future violation probability is greater than or equal to the violation probability threshold, conduct VR safety course training for the operator and develop a personalized inspection route plan to avoid dangerous areas until the operator's future violation probability is less than the violation probability threshold.

[0121] Example 2:

[0122] See also Figure 2 , a production safety management system based on hard hat detection, used to implement the above-mentioned production safety management method based on hard hat detection, including an operation area monitoring and early warning system deployment module, a multimodal data processing and safety detection module, a multi-level early warning execution and data management module, and a violation prediction and behavior optimization module;

[0123] The operation area monitoring and early warning system deployment module is responsible for building an intelligent monitoring system for the operation area. By defining a set of key locations and deploying a cluster of explosion-proof intelligent cameras, multi-dimensional environmental perception is achieved. Workers are equipped with wearable terminals that integrate RFID chips, UWB locators, and multi-level vibration alarm functions. This, combined with five-level early warning rules, forms a dynamic risk response framework, ensuring full coverage from hardware deployment to early warning logic.

[0124] The multimodal data processing and safety detection module is based on the improved YOLOv7-Tiny model and the HRNet key point detection model, and integrates the multimodal data after spatiotemporal alignment to achieve helmet wearing status classification, head posture angle deviation calculation and compliance detection, and finally outputs the personnel position, helmet status and compliance results, providing high-precision data support for early warning decision-making;

[0125] The multi-level warning execution and data management module is used to dynamically trigger graded warnings based on real-time detection results: for the behavior of not wearing a safety helmet, a second or third level warning is executed depending on whether the person is in a high-risk area; if the safety helmet wearing angle deviation exceeds the limit, a first or third level vibration reminder is triggered; when the safety helmet is detected to be non-compliant, a fourth or fifth level warning is activated in combination with the regional risk level; all warning events are synchronously uploaded to the database to form a complete record of violations, providing a structured data source for subsequent analysis.

[0126] The violation prediction and behavior optimization module constructs an optimized MLP prediction network based on historical warning data, uses the hyena optimization algorithm to adjust the learning rate parameters, and predicts the probability of future violations by operators by analyzing the frequency, region and time series characteristics within the violation cycle. If the probability exceeds the threshold, VR safety training and personalized inspection route planning are initiated, forming a "prediction-intervention-feedback" closed loop until the violation probability drops below the safety threshold, realizing an intelligent upgrade from passive warning to active behavior correction.

[0127] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the invention. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0128] The preferred embodiments of the invention disclosed above are intended only to help illustrate the invention. These preferred embodiments do not exhaust all details, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the content of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A safety production management method based on helmet detection, characterized in that: The following steps are involved: S1. Deploy explosion-proof smart cameras to obtain an explosion-proof smart camera cluster. Collect and process production environment data through explosion-proof smart camera clusters to obtain processed synchronous multimodal data; set multi-level early warning rules; S2. Improve the YOLOv7-Tiny model to obtain an improved YOLOv7-Tiny model; use the improved YOLOv7-Tiny model, the HRNet key point detection model, and the enterprise-level hard hat feature database to analyze the processed synchronous multimodal data to obtain hard hat wearing classification results, worker location, real-time angle deviation, and compliance detection results; S3. Executing multi-level warning rules based on the helmet wearing classification results, operator position, real-time angle deviation, and compliance detection results; S4. Input the violation data of the operator to be predicted within the violation period into the optimized MLP operator violation probability prediction network obtained by combining historical warning data with the optimization algorithm to obtain the future violation probability of the operator; and formulate a training and work plan for the operator based on the future violation probability.

2. The safety production management method based on helmet detection according to claim 1 is characterized in that: Said S1 comprises the following steps: S11. Construct a set of key locations in the operation area; deploy explosion-proof smart cameras at key locations in the set of key locations in the operation area to obtain an explosion-proof smart camera cluster; the explosion-proof smart cameras integrate visible light cameras, thermal imaging sensors, and edge computing nodes; S12. Equip the operator with a wearable positioning terminal that integrates an RFID chip, a UWB locator, and a vibration alarm; the RFID chip stores the operator's work number and position information; and the vibration alarm supports multi-level vibration feedback. S13. Set the first-level warning rule to a vibration alarm that reminds workers to wear safety helmets correctly; set the second-level warning rule to a vibration alarm that vibrates and sounds to remind workers to wear safety helmets correctly; set the third-level warning rule to stop working equipment in high-risk areas, and the vibration alarm that vibrates and sounds to remind workers to wear safety helmets correctly; set the fourth-level warning rule to a vibration alarm that vibrates and sounds to remind workers to change into compliant safety helmets; set the fifth-level warning rule to stop working equipment in high-risk areas, and the vibration alarm that vibrates and sounds to remind workers to change into compliant safety helmets; S14. Receive video streams, thermal imaging data, and positioning signals in real time through edge points of an explosion-proof intelligent camera cluster, and obtain synchronized multimodal data using a spatiotemporal alignment algorithm; perform dynamic illumination compensation, fog elimination, and data enhancement on the synchronized multimodal data to obtain processed synchronized multimodal data.

3. The safety production management method based on helmet detection according to claim 1 is characterized in that: The S2 comprises the following steps: S21. Improve the YOLOv7-Tiny model to obtain an improved YOLOv7-Tiny model; S22. Perform preliminary detection on the processed synchronized multimodal data using the improved YOLO-v5 to obtain a helmet wearing classification result; detect that the operator is in a high-risk work area using the UWB locator, and obtain the operator's location; S23, setting a set of key head position points; setting a safety deviation threshold; If the helmet wearing classification result is wearing a helmet, the HRNet key point detection model is used to identify the key points of the person's head according to the head key position point set. Combined with the UWB positioning data, the angle deviation between the helmet and the head is calculated to obtain the real-time angle deviation. S24. Setting compliance detection rules and establishing an enterprise-level safety helmet feature database. The compliance detection rules include color coding, enterprise logo recognition, and reflective strip detection. According to the enterprise-level safety helmet feature database combined with compliance detection rules, whether the safety helmet worn by the operator is compliant is detected to obtain a compliance detection result.

4. The safety production management method based on helmet detection according to claim 2 is characterized in that: The S21 includes the following steps: The C3 module in the backbone network of the YOLOv7-Tiny model was replaced with the RepVGG module through structural reparameterization technology; multi-branch convolution was used in the training phase and merged into a single 3×3 convolution in the inference phase; The GIoU loss function of the YOLOv7-Tiny model is replaced by the CIoU loss function through the geometric perception constraint method, and the center point distance penalty term and aspect ratio consistency constraint are introduced.

5. The safety production management method based on helmet detection according to claim 1 is characterized in that: The S3 includes the following steps: S31. When the helmet wearing classification result is not worn, the operator is not in a high-risk work area and a level 2 warning is executed. If the operator is in a high-risk work area, a level 3 warning is executed and the warning data is uploaded to the database. Otherwise, no warning is executed. S32. When the real-time angle deviation is greater than or equal to the safety deviation threshold, and the operator is not in a high-risk work area, a first-level warning is executed; when the operator is in a high-risk work area, a third-level warning is executed, and the warning data is uploaded to the database; otherwise, no warning is executed; S33. If the compliance test result is non-compliant, when the operator is not in a high-risk work area, a level 4 warning will be executed; when the operator is in a high-risk work area, a level 5 warning will be executed, and the warning data will be uploaded to the database. Otherwise, no warning will be executed.

6. The safety production management method based on helmet detection according to claim 1 is characterized in that: The S4 comprises the following steps: S41. Collect historical warning data from a database to obtain historical safety helmet warning data, wherein the historical safety helmet warning data includes data on illegal workers, violation frequency data, violation area data, and violation time series data; S42, constructing an MLP prediction network, and setting weight values ​​and bias values ​​of the MLP prediction network; S43, setting a ratio of training data to test data according to the MLP prediction network; dividing the historical helmet warning data into training data and test data according to the ratio of training data to test data; S44, setting a training accuracy threshold of the MLP prediction network to j, a training accuracy to j2, and a maximum number of training iterations; training the MLP prediction network using the training data; adjusting the network structure of the MLP prediction network during the training process; when the training accuracy is greater than or equal to the training accuracy threshold or the maximum number of training iterations is reached, obtaining an MLP operator violation probability prediction network; S45, using the test data to test the MLP operator violation probability prediction network, after the test is completed, an optimized MLP operator violation probability prediction network is obtained; S46. Set a violation period, collect violation data of the operator to be predicted within the violation period, input the violation data of the operator within the violation period into the optimized MLP operator violation probability prediction network, and obtain the future violation probability of the operator; S47. Set a violation probability threshold. If the operator's future violation probability is greater than or equal to the violation probability threshold, conduct VR safety course training for the operator and develop a personalized inspection route plan to avoid dangerous areas until the operator's future violation probability is less than the violation probability threshold.

7. The safety production management method based on helmet detection according to claim 6 is characterized in that: The S45 includes the following steps: S451, setting a test error threshold; using the test data to test the MLP operator violation probability prediction network to obtain a test error; S452. When the test error is less than the test error threshold, the MLP operator violation probability prediction network is used as the optimized MLP operator violation probability prediction network; when the test error is greater than or equal to the test error threshold, an optimization algorithm is used to find the learning rate of the MLP operator violation probability prediction network to obtain the optimal solution; the optimal solution is used as the learning rate of the MLP operator violation probability prediction network to obtain the optimized MLP operator violation probability prediction network.

8. The safety production management method based on helmet detection according to claim 7 is characterized in that: In S452, the optimization algorithm is used to find the learning rate of the MLP operator violation probability prediction network, and obtaining the optimal solution includes the following steps: S4521. Construct a hyena population; set the maximum number of optimization iterations; S4522. Randomly set the initial position of the hyena population based on the weight value and bias value of the MLP operator violation probability prediction network to obtain an initial position set of the hyena population; S4523. Defining a fitness function of the hyena position in the hyena population according to the test error threshold and the training accuracy; S4524, performing an iterative operation on the initial position set of the hyena population; during each round of iteration, calculating the fitness value of each position in the initial position set of the hyena population according to the fitness function, updating the position of each hyena in the initial position set of the hyena population in descending order of fitness value, and obtaining the best individual hyena position in the hyena population and the global best hyena position during each round of iteration; S4525. Repeat S4524. When the maximum number of optimization iterations is reached, stop the iteration and take the global best hyena position as the optimal solution.

9. A safety and environmental protection system based on multi-pollutant coordinated treatment, characterized in that: Implement the safe production management method based on safety helmet detection as described in any one of claims 1-8.

10. A storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, the safe production management method based on safety helmet detection as described in any one of claims 1 to 8 is implemented.

Images