Safety production supervision system and method based on artificial intelligence and big data analysis

Abstract

The invention relates to a safety production supervision system and method based on artificial intelligence and big data analysis, and the system comprises a production equipment management system, a data and business overall planning server, a video analysis server and monitoring equipment. The video analysis server is used for forwarding and coordinating the service functions of the service terminals, docking with a fire fighting system, analyzing a security image corresponding to a field scene, and carrying out alarm reporting on an appearing scene picture conforming to security alarm; various monitoring problems in civil explosive production process application scenes are sensed, monitored, alarmed, evaluated and handled by using an artificial intelligence technology, and 24-hour uninterrupted unattended safety management is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of production safety supervision, and in particular to a production safety supervision system and method based on artificial intelligence and big data analysis. Background Art

[0002] In recent years, safety accidents have frequently occurred during the production of hazardous materials and chemicals. Existing video surveillance systems are essentially just video surveillance systems, serving only as an auxiliary tool for on-site control and comprehensive safety management. They are primarily used for post-event review and lack the ability to scientifically and systematically identify video footage, effectively failing to fully utilize the system's capabilities. On-site management personnel and safety managers can only correct some illegal operations, irregularities, and uncivilized production behaviors through on-site control. Although many key production sites for safety and quality have surveillance systems, they are unable to accurately assess and address on-site conditions in real time. This limited visibility makes accurate identification and timely response impossible.

[0003] Long-term, repetitive work can lead to complacency among on-site operators, leading to habitual and unforeseen violations of regulations and rules. This prevents on-site control personnel from exercising real-time safety oversight over specific operators. While existing production equipment has specific linkage measures, it lacks comprehensive control over the equipment and facilities, placing high demands on on-site operators. Therefore, fully leveraging AI technology to perceive, monitor, alert, assess, and address various monitoring issues in the civilian explosives production process is crucial to accelerating the safe and intelligent upgrade of civilian explosives production.

[0004] To address the above issues, this research team provides a production safety supervision system and method based on artificial intelligence and big data analysis. Summary of the Invention

[0005] The purpose of the present invention is to provide a production safety supervision system and method based on artificial intelligence and big data analysis.

[0006] On the one hand, the present invention provides a production safety supervision system based on artificial intelligence and big data analysis, including a production equipment management system, a data and business coordination server, a video analysis server and monitoring equipment.

[0007] The business coordination server coordinates the interaction of various data and the classification of different functions, forwards and coordinates the business functions of various business terminals, connects to the fire protection system, and provides background parameter control modules and business data coordination function modules;

[0008] The video analysis server is responsible for analyzing the security images corresponding to the on-site scenes, and reporting the scene images that meet the security alarm requirements, including a security scene function detection module.

[0009] Furthermore, a distributed flexible design model is adopted to highly integrate various functions in a modular manner, and corresponding module functions can be dynamically added according to actual scenario needs. In some embodiments, the security scenario function detection module includes:

[0010] Monitoring device status detection module: triggers an alarm when the camera is offline, blocked, displaced, the image is blurred, or blocked;

[0011] Item placement detection module: The system automatically identifies items that are placed outside the area or are abnormally retained, triggering an alarm;

[0012] Personnel quota detection module: When the number of personnel on the production line exceeds the preset value, an alarm is triggered;

[0013] Clothing detection module: Identifies the labor protection clothing of production line personnel and triggers an alarm when it finds personnel whose clothing does not meet the requirements;

[0014] Abnormal intrusion detection module: When a person enters an area that is not allowed to be entered during equipment operation, an alarm is triggered if a person is identified;

[0015] Matrix color detection module: When the color of the emulsion explosive matrix on the emulsifier changes abnormally, an alarm is triggered;

[0016] Screw conveyor stoppage detection module: triggers an alarm when it detects that the screw conveyor of the expanded explosives line has stopped working abnormally;

[0017] Detonator quantity detection module: When the electronic detonator laser coding machine codes the electronic detonators, if the number of detonators is insufficient, an alarm is triggered;

[0018] Charger drug quantity warning detection module: When the drug quantity in the drug hopper exceeds the warning position, an alarm is triggered;

[0019] Sensitizer overflow detection module: triggers an alarm when emulsion explosives overflow from the upper or lower layer of the sensitizer;

[0020] Sodium nitrate dropper abnormality detection module: When the sodium nitrate dropper does not swing normally to drop material, an alarm is triggered;

[0021] Belt stacking detection module: When the conveyor belt stops working abnormally, or when the transfer vehicle is not in place in time, or when the robot fails to malfunction and causes stacking of boxes, an alarm is triggered;

[0022] Spiral explosive leakage detection module: triggers an alarm when expanded explosives overflow the spiral hopper;

[0023] Expanded explosive spillage detection module: triggers an alarm when expanded explosives are detected to have spilled during transportation;

[0024] Leakage detection module: triggers an alarm when leakage or overflow occurs in the storage tank;

[0025] Cooling water interruption detection module: triggers an alarm when the emulsifier cooling water interruption is detected;

[0026] Personnel crossing detection module: When it is detected that someone crosses the prohibited crossing area on the production line, an alarm is triggered;

[0027] Loading robot stoppage detection module: When a box is brought in by the conveyor belt and the robot stops working, an alarm is triggered;

[0028] Unattended detection module: triggers an alarm when the monitoring duty room is unattended during the production process;

[0029] Rough loading and unloading detection module: triggers an alarm when it detects that the loading and unloading personnel are handling finished goods roughly;

[0030] Two-person double-lock detection module: triggers an alarm when one person is detected opening the warehouse door;

[0031] Fire passage obstruction detection module: When goods block the fire passage in the production workshop, an alarm is triggered.

[0032] Furthermore, the business data collaboration function module includes an MES equipment management system data synchronization function module, an alarm data upload and docking function module and a business overall management function module.

[0033] Furthermore, the personnel capacity detection function module and the detonator quantity detection function module adopt a convolutional neural network algorithm. Specifically, in the personnel capacity detection function module, the monitoring video stream is first captured in real time, and each frame of the image is input into a pre-trained convolutional neural network (CNN) model. The model is built based on a deep learning framework (such as TensorFlow or PyTorch), and uses multiple convolution layers and pooling layers to extract features in the image, and performs classification or counting operations through a fully connected layer. During the training process, a labeled personnel image dataset (including the number of people in different scenarios) is used for supervised learning, and the model parameters are optimized to minimize the error between the predicted number of people and the actual number of people. In actual applications, the model can analyze video frames in real time and output the number of people in the current scene, thereby realizing the capacity detection function.

[0034] The rough loading and unloading detection and two-person double-locking detection functional modules utilize a human behavior action learning algorithm. Specifically, in the rough loading and unloading detection functional module, a deep learning-based behavior recognition algorithm is used, combined with OpenPose human key point detection technology, to extract key point information of human actions in surveillance videos in real time. By constructing a spatiotemporal graph convolutional network (ST-GCN), the human action sequence is modeled and the feature representations of different action patterns are learned. After being trained on a large amount of labeled behavior data (such as normal loading and unloading and rough loading and unloading actions), the model can identify whether there is rough loading and unloading behavior and trigger an alarm in real time.

[0035] The clothing and wear detection function module adopts a deep learning classification algorithm. Specifically, in the clothing and wear detection function module, a deep learning classification algorithm is adopted to perform real-time detection of the clothing of personnel based on a convolutional neural network (CNN). First, the image of the personnel in the monitoring area is obtained through an image acquisition device, and the image is preprocessed (such as cropping, normalization, etc.). Then, the image is input into a CNN model that has been trained with a large number of clothing samples. The model extracts features in the image through the convolution layer and the pooling layer, and classifies it through the fully connected layer to determine whether the personnel are wearing the necessary protective equipment (such as helmets, work clothes, gloves, etc.) as required. The model outputs the classification results. If a situation that does not comply with the dress code is detected, the system will issue a prompt in real time to remind the relevant personnel to make corrections.

[0036] The object location detection module adopts a deep learning recognition algorithm. Specifically, in the object location detection module, a deep learning recognition algorithm is adopted to perform real-time detection of the position and status of the object based on a convolutional neural network (CNN). Images of the working area are collected by a high-resolution camera, and the images are input into a CNN model that has been trained with a large number of object samples. The model can identify the characteristics of the object and determine whether it is placed in the specified location. During the training process, the model is supervised by using the labeled object location data, and the model parameters are optimized to improve the recognition accuracy. In actual applications, the system can analyze the image in real time to determine whether the object is in the correct position. If it is found that the object is displaced or not placed as required, an alarm will be issued immediately to remind relevant personnel to deal with it.

[0037] The screw conveyor stop working detection function module and the leakage detection function module adopt inter-frame difference and RGB→HSI color space change algorithm. Specifically, in the screw conveyor stop working detection function module, the inter-frame difference algorithm is combined with the RGB to HSI color space conversion technology. First, the video frames of the screw conveyor operation area are collected in real time by the monitoring camera, and the differential image between consecutive frames is calculated to detect motion changes. At the same time, the RGB image is converted into the HSI color space to enhance the color features of the image, which is convenient for identifying whether the screw conveyor has stopped running (such as color change or disappearance of the motion area). The model determines whether the screw conveyor is in a stopped state by analyzing the inter-frame differential image and HSI color features, and issues an alarm in real time.

[0038] The belt stacking detection function module adopts an unsupervised learning feature comparison algorithm. Specifically, in the belt stacking detection function module, an unsupervised learning feature comparison algorithm is adopted to extract image features through an autoencoder or a contrastive learning method. First, a high-resolution camera is used to capture an image of the belt conveying area, and the image is input into the autoencoder model. The autoencoder compresses the image into a low-dimensional feature representation through the encoder part, and then reconstructs the image through the decoder part. During the training process, the model learns the feature representation of the image by minimizing the error between the input image and the reconstructed image. In actual application, the system compares the features of the current image with the features of the normal state (no stacking). If the feature difference exceeds the set threshold, it is judged that there is a stacking phenomenon, and an alarm is issued in real time to remind relevant personnel to deal with it.

[0039] The monitoring device status detection function module first captures the monitoring video stream in real time and inputs each frame of the image into a pre-trained convolutional neural network model. The model is built based on a deep learning framework, uses multiple convolutional layers and pooling layers to extract image features, and performs classification or status judgment through a fully connected layer. During training, supervised learning is carried out using a labeled normal and abnormal state image dataset of monitoring equipment, and the model parameters are optimized to narrow the gap between the predicted device status and the actual status. The video frames are analyzed in real time and the results of whether the monitoring device is in a normal or abnormal state are output, thereby realizing the monitoring device status detection function;

[0040] The abnormal intrusion detection module utilizes a convolutional neural network algorithm to capture surveillance video streams in real time. Each frame is fed into a pre-trained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features, and a fully connected layer is used to determine whether an intrusion is abnormal. During training, supervised learning is performed using a dataset of images labeled with abnormal intrusion patterns. This optimizes model parameters and reduces the error between predicted and actual results. The module then analyzes video frames in real time and outputs whether an abnormal intrusion is present in the current scene, effectively achieving abnormal intrusion detection.

[0041] The substrate color detection module uses a convolutional neural network algorithm. It first acquires a surveillance video stream in real time and feeds each frame into a pre-trained convolutional neural network model. Based on a deep learning framework, it uses multiple convolutional and pooling layers to extract substrate color features, and then performs color classification and recognition through a fully connected layer. During training, supervised learning is carried out using a standard image dataset with substrate colors annotated. The model parameters are optimized to ensure that the predicted color is as close as possible to the actual color. The video frames are analyzed in real time and the color of the substrate in the current scene is output to realize the substrate color detection function.

[0042] The charge loader charge warning detection module uses a convolutional neural network algorithm to capture surveillance video streams in real time. Each frame is fed into a pre-trained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features related to the charge loader charge level. A fully connected layer is then used to determine whether the charge level has reached the warning range. During the training phase, supervised learning is performed using a dataset of charge loader images labeled with different charge levels. This optimizes model parameters and reduces the deviation between the predicted charge level and the actual level. The module then analyzes video frames in real time and outputs whether the charge loader's current charge level is in the warning state, completing the charge loader charge warning detection function.

[0043] The sensitizer overflow detection module, centered around a convolutional neural network algorithm, first captures surveillance video streams in real time. Each frame is fed into a pretrained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features indicating potential sensitizer overflow. A fully connected layer then determines whether an overflow has occurred. During training, supervised learning is performed using a dataset of images of sensitizer operating scenes, labeled with overflow patterns. This optimizes model parameters, narrowing the gap between predicted overflow states and actual conditions. The module then analyzes video frames in real time and outputs a signal indicating whether the sensitizer is currently overflowing, effectively implementing sensitizer overflow detection.

[0044] The sub-sodium dropper anomaly detection module utilizes a convolutional neural network algorithm. It first acquires a surveillance video stream in real time and feeds each frame into a pre-trained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features related to the sub-sodium dropper's operating status. A fully connected layer is then used to determine whether the dropper is abnormal. During training, supervised learning is performed using a dataset of images labeled with normal and abnormal dropper states. This optimizes model parameters and reduces the error between predicted results and actual conditions. The system then analyzes video frames in real time and outputs whether the sub-sodium dropper is currently in an abnormal state, achieving sub-sodium dropper anomaly detection.

[0045] The spiral drug counterfeiting detection module uses a convolutional neural network algorithm to capture surveillance video streams in real time. Each frame is fed into a pre-trained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features related to drug counterfeiting in the spiral. A fully connected layer is then used to determine whether drug counterfeiting is occurring. During training, supervised learning is performed using a dataset of images of spiral work scenes labeled with drug counterfeiting. Model parameters are optimized to ensure that predicted drug counterfeiting results are as consistent as possible with actual drug counterfeiting. Video frames are analyzed in real time to output whether drug counterfeiting is currently occurring in the spiral, thus enabling spiral drug counterfeiting detection.

[0046] The expanded explosive spill detection module utilizes a convolutional neural network algorithm. It first captures surveillance video streams in real time and feeds each frame into a pretrained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features related to explosive spillage. A fully connected layer then determines whether spillage has occurred. During the training phase, supervised learning is performed using a dataset of images labeled with the presence or absence of explosive spillage. This optimizes model parameters, narrowing the gap between predicted spillage and actual conditions. The system then analyzes video frames in real time and outputs a signal indicating whether explosive spillage has occurred in the current scene, completing the expanded explosive spill detection function.

[0047] The leak detection module utilizes a convolutional neural network algorithm to capture a real-time surveillance video stream. Each frame is fed into a pre-trained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features that indicate potential leaks. A fully connected layer then determines whether a leak is present. During training, supervised learning is performed using a dataset of images of equipment operating scenes labeled for leaks. This optimizes model parameters, reduces the error between predicted and actual leak states, and analyzes video frames in real time to output whether the current device is leaking, thus enabling leak detection.

[0048] The cooling water flow interruption detection module uses a convolutional neural network algorithm to capture the monitoring video stream in real time. Each image frame is input into a pre-trained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features related to the cooling water flow state. A fully connected layer is then used to determine whether a flow interruption has occurred. During training, supervised learning is performed using an image dataset labeled with normal cooling water flow and interruption states. Model parameters are optimized to ensure that the predicted flow interruption results are as close as possible to the actual cooling water state. Video frames are analyzed in real time to output whether the cooling water is currently in a flow interruption state, thus achieving the cooling water flow interruption detection function.

[0049] The human boundary crossing detection module, centered around a convolutional neural network algorithm, first captures surveillance video streams in real time and feeds each frame into a pre-trained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features related to the person's location and boundaries. Fully connected layers are then used to determine whether a person has crossed the boundary. During the training phase, supervised learning is performed using a dataset of images annotated with human boundary crossings and normal activity areas. This optimizes model parameters, narrowing the gap between predicted boundary crossings and actual activity. The system then analyzes video frames in real time and outputs whether a person has crossed the boundary, effectively implementing human boundary crossing detection.

[0050] The loading robot inactivity detection module utilizes a convolutional neural network algorithm. It first captures a surveillance video stream in real time and feeds each frame into a pretrained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features related to the loading robot's operating state. A fully connected layer is then used to determine whether the robot is inactive. During training, supervised learning is performed using a dataset of images labeled with the robot in both normal and inactive states. This optimizes model parameters and reduces the error between the predicted and actual robot states. The module then analyzes video frames in real time and outputs whether the loading robot is currently inactive, completing the loading robot inactivity detection function.

[0051] The unmanned detection module utilizes a convolutional neural network algorithm to capture surveillance video streams in real time. Each frame is fed into a pretrained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features related to personnel present in the scene. A fully connected layer is then used to determine whether an unmanned vehicle is present. During training, supervised learning is performed using a dataset of images of scenes labeled with both attended and unattended personnel. The model parameters are optimized to ensure that predicted unmanned vehicle conditions are as consistent as possible with actual conditions. The model then analyzes video frames in real time and outputs whether the scene is currently unmanned, thus implementing unmanned vehicle detection.

[0052] The fire escape obstruction detection module uses a convolutional neural network algorithm to capture surveillance video streams in real time. Each image frame is fed into a pre-trained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features related to whether the fire escape is blocked. A fully connected layer then determines whether the fire escape is unobstructed. During the training phase, supervised learning is performed using a dataset of images labeled with unobstructed and blocked fire escapes. This optimizes model parameters, narrowing the gap between predicted and actual fire escape status. The module then analyzes video frames in real time and outputs whether the fire escape is currently blocked, achieving fire escape obstruction detection.

[0053] Behavior recognition technology based on human skeletal keypoint detection and localization utilizes deep learning algorithms to identify the skeletal keypoints of each person in a video frame. These keypoints are then connected to obtain skeletal connectivity information, which in turn reflects the person's posture. Human behavior is a sequential process, with a beginning and an end. The combination of a single person's continuous skeletal postures over time constitutes a complete behavior and motion trajectory, such as falling, hitting someone, or stealing. Therefore, posture estimation based on skeletal keypoint localization serves as the core algorithm for behavior recognition. The algorithm detects 18 key points on a single person to describe their posture.

[0054] Furthermore, the monitoring device is a network camera, which is connected to a switch on site, and the video data is transmitted to a video analysis server.

[0055] Furthermore, the business data collaboration function module includes a production equipment management system data synchronization function, an alarm data upload and docking function, and a business overall management function.

[0056] Furthermore, the data and business coordination server web terminal displays alarm information, manages each video analysis server connected to the system, and implements basic parameter settings in the background parameter control module.

[0057] Another aspect of the present invention provides a safety production supervision method based on artificial intelligence and big data analysis, including the following steps: hierarchical push, generation of emergency plans, intervention of relevant personnel, problem solving, and archiving and evidence storage.

[0058] Tiered push: Data collection and risk assessment: Use artificial intelligence algorithms to collect and analyze real-time data related to production safety (such as equipment operating parameters, environmental indicators, personnel operation records, etc.), combine historical data and industry standards to assess the risk level of the current production safety situation, and divide it into three levels: low, medium, and high.

[0059] Determine push recipients: Based on the risk level, determine the appropriate push recipients. Low-risk information is pushed to grassroots safety managers; medium-risk information is pushed to department heads and heads of safety management departments; high-risk information is pushed to the company's principal, safety director, and relevant regulatory authorities.

[0060] Information push: Through the system's built-in message push module, risk information will be pushed to the corresponding recipients in a timely manner in the form of notifications, and unprocessed risk information will be highlighted on the system interface.

[0061] Generate emergency plans: Risk type judgment: When the system identifies medium or high-risk situations, the artificial intelligence algorithm quickly determines the type of risk (such as fire, explosion, hazardous material leakage, mechanical failure, etc.).

[0062] Plan template matching: According to the risk type, the corresponding emergency plan template is matched from the plan database. The template contains emergency handling procedures, division of responsibilities, resource allocation and other contents for different risks.

[0063] Personalized customization: Combine real-time data with actual on-site conditions to customize the matching emergency plan template. For example, you can adjust the rescue route based on the specific location of the accident or determine the emergency repair plan based on the equipment status.

[0064] Plan generation and review: A detailed emergency plan is generated and automatically reviewed by the system to ensure its completeness and feasibility. Once approved, it is sent to the relevant person in charge for confirmation and activation.

[0065] Human intervention: Task allocation: Based on the generated emergency plan, the system automatically assigns tasks to the corresponding emergency rescue personnel and departments. For example, it notifies firefighters to extinguish the fire, arranges technicians to repair equipment, and dispatches medical personnel to treat the injured.

[0066] Personnel notification: timely notify relevant personnel to intervene in emergency response work through various means (such as telephone, intercom, system messages, etc.), and inform them of their specific tasks and responsibilities.

[0067] Personnel arrival confirmation: After arriving at the scene, rescuers sign in through the system to confirm their arrival in real time. At the same time, the system can also urge and remind personnel who have not arrived in time.

[0068] Problem Solving: On-site Command and Coordination: The emergency command center provides unified command and coordination for rescue efforts based on on-site feedback and real-time data provided by the system. AI-assisted decision-making optimizes rescue plans and rationally allocates resources.

[0069] Real-time monitoring and adjustment: Utilizing various sensors and monitoring equipment installed on-site, we monitor the progress of incident handling and changes in the on-site environment in real time. We adjust rescue strategies and measures based on actual conditions to ensure the safety and effectiveness of rescue efforts.

[0070] Problem Resolution Assessment: Once the incident is initially under control, evaluate the resolution of the problem. Determine whether the safety hazard has been completely eliminated and whether normal production has been restored. If any issues remain, continue to implement appropriate measures to address them.

[0071] Archiving: Data collection and organization: Collect all data from the entire production safety supervision and emergency response process, including risk assessment reports, emergency plans, personnel task allocation records, on-site monitoring videos, various data from the rescue process, etc., and classify and organize them.

[0072] Document generation: Generate a detailed accident handling report based on the collated data, including the cause of the accident, handling process, lessons learned, improvement measures, etc.

[0073] Archive storage: Store the organized data and generated reports in the system's database for long-term preservation. At the same time, establish indexing and retrieval mechanisms to facilitate subsequent queries and analysis.

[0074] Regular review and update: Regularly review archived materials, update and improve relevant content based on new regulations and standards, technological developments and the actual situation of the enterprise, and continuously improve the effectiveness and adaptability of the production safety supervision system.

[0075] The production equipment management system transmits production equipment and production data to the data and business coordination server. The data and business coordination server transmits the business control data required by the video analysis server to the video analysis server. The monitoring equipment transmits the data acquired through monitoring to the video analysis server. The video analysis server performs a security analysis of the on-site scene based on the received data, and reports the detected scene images that meet the security alarm requirements to the data and business coordination server. The data and business coordination server uploads the detection alarm data to the production equipment management system. The video analysis server receives the basic data of the production equipment, production status data, business control data, and image data input by the IP network camera, and collaboratively detects abnormal problems in the scene based on the corresponding data and specific algorithms, iteratively trains the detection results, and detects alarms.

[0076] The beneficial effects of the present invention are:

[0077] 1. The present invention adopts a distributed flexible design mode, and a system that highly integrates various functions in a modular manner. The corresponding module functions can be dynamically added according to the actual scenario needs, and the existing equipment management system can be connected to obtain the specific status of the production equipment. Various algorithms use the corresponding scenario data standards for processing and analysis according to the specific status of the production equipment. Compared with traditional production safety management that mostly uses manual on-site patrols or setting fixed workstations for detection, it solves the real-time supervision problems of production safety, safe and civilized production supervision, violations of regulations, etc., and realizes 24-hour uninterrupted unmanned safety management.

[0078] 2. The present invention uses intelligent image analysis technologies such as image analysis technology and face recognition technology to collaborate with the Internet of Things, big data and the Internet to conduct real-time intelligent data analysis of various safety alarm scenarios at different time periods and in different working forms, including people, production equipment, production products and production processes, to provide early warnings and timely feedback on various possible safety issues.

[0079] 3. The deep learning target detection algorithm adopted in this application adds operations such as inter-layer fusion on the basis of commonly used network layers such as convolutional layers and pooling layers, making the network more adaptable to the size of the target and the detection results more accurate; the cost function designed in the algorithm integrates detection and multi-target classification, making the training of the network model more efficient, supporting target detection of more categories, filtering layer by layer, locating targets from coarse to fine, filtering out a large number of candidates each cascade, reducing a large amount of calculation, and making the entire algorithm fast and efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 This is a system network structure diagram of Example 1 of the present invention;

[0081] Figure 2 This is a network structure diagram of a production safety supervision system based on artificial intelligence and big data analysis according to Example 1 of the present invention;

[0082] Figure 3 This is the data flow of the system in Example 2 of the present invention; Figure 4 This is a flow chart of the video analysis function of Example 2 of the present invention. DETAILED DESCRIPTION

[0083] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The exemplary embodiments and descriptions of the present invention are used to explain the present invention but are not intended to limit the present invention.

[0084] Example 1

[0085] See Figure 1 - Figure 2 , a production safety supervision system based on artificial intelligence and big data analysis, including a production equipment management system, a data and business coordination server, a video analysis server and a monitoring device. In this embodiment, the monitoring device is a network camera, which is connected to a switch on site, and the video data is transmitted to the video analysis server;

[0086] The business coordination server coordinates the interaction of various data and organizes and categorizes different functions, forwards and coordinates the business functions of various business terminals, connects to the fire protection system, and provides background parameter control modules and business data coordination function modules;

[0087] The video analysis server is responsible for analyzing the security images corresponding to the on-site scenes, and reporting alarms for scene images that meet the security alarm requirements, including a security scene function detection module.

[0088] The business data collaboration function module includes a production equipment management system data synchronization function, an alarm data upload and docking function, and a business overall management function.

[0089] The data and business coordination server web terminal displays alarm information, manages each video analysis server connected to the system, and implements basic parameter settings in the background parameter control module.

[0090] The business data collaboration function module includes an MES equipment management system data synchronization function module, an alarm data upload and docking function module, and a business overall management function module.

[0091] A distributed flexible design model is adopted to highly integrate various functions in a modular manner. Corresponding module functions can be dynamically added according to actual scenario needs. The security scenario function detection module described in this embodiment includes:

[0092] Monitoring equipment status detection module: When the camera is offline, blocked, displaced, the image is blurred, or blocked, an alarm is triggered; deployed for equipment status monitoring throughout the entire factory.

[0093] Item Positioning Detection Module: The system automatically identifies items that are placed outside the designated area or are abnormally retained, triggering an alarm. This module is deployed throughout the detonator branch, the explosives branch, and the storage area.

[0094] Personnel quota detection module: When the number of personnel on the production line exceeds the preset value, an alarm is triggered. It is deployed in the entire detonator branch, the entire explosives branch, and the loading and unloading operation area of the general warehouse;

[0095] Clothing detection module: This module identifies the labor protection clothing worn by production line personnel and triggers an alarm when any personnel's clothing does not meet the requirements. This module is deployed in the entire detonator branch, the entire explosives branch, and the general warehouse area.

[0096] Abnormal intrusion detection module: When a person enters an area where personnel are not allowed to enter during equipment operation, an alarm is triggered. This module is deployed in all human-machine isolation workstations in the detonator branch and the explosives branch.

[0097] Matrix color detection module: When the color of the emulsion explosive matrix on the emulsifier changes abnormally, an alarm is triggered. It is deployed on emulsion explosive lines 1 and 2;

[0098] Screw conveyor stoppage detection module: When it detects that the screw conveyor of the expanded explosives line has stopped working abnormally, an alarm is triggered. It is deployed on the expanded explosives line;

[0099] Detonator quantity detection module: When the electronic detonator laser coding machine codes the electronic detonators, if the number of detonators is insufficient, an alarm is triggered and deployed on the detonator line;

[0100] Charger charge quantity warning detection module: When the charge quantity in the charge hopper exceeds the warning position, an alarm is triggered. It is deployed on emulsion explosive lines 1 and 2.

[0101] Sensitizer overflow detection module: When the emulsion explosives overflow from the upper or lower layer of the sensitizer, an alarm is triggered. It is deployed on the emulsion explosives production line;

[0102] Sodium hypochlorite dropper anomaly detection module: When the sodium hypochlorite dropper does not swing normally to drip material, an alarm is triggered. It is deployed in the emulsion explosives production line;

[0103] Belt stacking detection module: When the conveyor belt stops working abnormally, or when the transfer vehicle is not in place in time, or when the robot fails to malfunction and causes stacking, an alarm is triggered. This module is deployed on the finished product conveyor belt, loading platform, and packaging line.

[0104] Spiral explosive leakage detection module: When expanded explosives overflow the spiral hopper, an alarm is triggered. It is deployed on the explosive drying spiral and spiral hopper;

[0105] Expanded explosive spill detection module: When expanded explosive spillage is detected during transportation, an alarm is triggered. It is deployed on the expanded explosive production line.

[0106] Leakage detection module: When leakage or overflow occurs in a storage tank, an alarm is triggered. This module is deployed in the outdoor ammonium nitrate tank area and the indoor water-oil phase storage room.

[0107] Cooling water interruption detection module: When the emulsifier cooling water interruption is detected, an alarm is triggered. It is deployed on emulsion explosive lines 1 and 2;

[0108] Personnel crossing detection module: When it detects that someone crosses the prohibited crossing area on the production line, it triggers an alarm and is deployed on the production line;

[0109] Loading robot stoppage detection module: When a conveyor belt carries a box and the robot stops working, an alarm is triggered. This module is deployed at the loading platform of the explosives line.

[0110] Unattended detection module: When the monitoring duty room is unattended during the production process, an alarm is triggered. It is deployed in all production lines of the explosives branch, the detonator branch, and all video surveillance rooms within the factory.

[0111] Rough handling detection module: When it detects rough handling of finished goods by loading and unloading personnel, it triggers an alarm and is deployed in the general warehouse;

[0112] Two-person, two-lock detection module: When one person is detected opening the warehouse door, an alarm is triggered and deployed in the warehouse area;

[0113] Fire passage obstruction detection module: When goods block the fire passage in the production workshop, an alarm is triggered. It is deployed in the fire passages of the production workshop and office building.

[0114] The personnel capacity detection function module and the detonator quantity detection function module adopt a convolutional neural network algorithm. Specifically, the convolutional neural network (CNN) extracts and classifies images through convolution layers, pooling layers and fully connected layers.

[0115] The personnel capacity detection function module and the detonator quantity detection function module adopt a convolutional neural network algorithm. Specifically, in the personnel capacity detection function module, the monitoring video stream is first captured in real time, and each frame of the image is input into a pre-trained convolutional neural network (CNN) model. The model is built based on a deep learning framework (such as TensorFlow or PyTorch), and uses multiple convolution layers and pooling layers to extract features in the image, and performs classification or counting operations through a fully connected layer. During the training process, a labeled personnel image dataset (including the number of people in different scenes) is used for supervised learning, and the model parameters are optimized to minimize the error between the predicted number of people and the actual number of people. In actual applications, the model can analyze video frames in real time and output the number of people in the current scene, thereby realizing the capacity detection function.

[0116] The rough loading and unloading detection and two-person double-locking detection functional modules utilize a human behavior action learning algorithm. Specifically, in the rough loading and unloading detection functional module, a deep learning-based behavior recognition algorithm is used, combined with OpenPose human key point detection technology, to extract key point information of human actions in surveillance videos in real time. By constructing a spatiotemporal graph convolutional network (ST-GCN), the human action sequence is modeled and the feature representations of different action patterns are learned. After being trained on a large amount of labeled behavior data (such as normal loading and unloading and rough loading and unloading actions), the model can identify whether there is rough loading and unloading behavior and trigger an alarm in real time.

[0117] The clothing and wear detection function module adopts a deep learning classification algorithm. Specifically, in the clothing and wear detection function module, a deep learning classification algorithm is adopted to perform real-time detection of the clothing of personnel based on a convolutional neural network (CNN). First, the image of the personnel in the monitoring area is obtained through an image acquisition device, and the image is preprocessed (such as cropping, normalization, etc.). Then, the image is input into a CNN model that has been trained with a large number of clothing samples. The model extracts features in the image through the convolution layer and the pooling layer, and classifies it through the fully connected layer to determine whether the personnel are wearing the necessary protective equipment (such as helmets, work clothes, gloves, etc.) as required. The model outputs the classification results. If a situation that does not comply with the dress code is detected, the system will issue a prompt in real time to remind the relevant personnel to make corrections.

[0118] The object location detection module adopts a deep learning recognition algorithm. Specifically, in the object location detection module, a deep learning recognition algorithm is adopted to perform real-time detection of the position and status of the object based on a convolutional neural network (CNN). Images of the working area are collected by a high-resolution camera, and the images are input into a CNN model that has been trained with a large number of object samples. The model can identify the characteristics of the object and determine whether it is placed in the specified location. During the training process, the model is supervised by using the labeled object location data, and the model parameters are optimized to improve the recognition accuracy. In actual applications, the system can analyze the image in real time to determine whether the object is in the correct position. If it is found that the object is displaced or not placed as required, an alarm will be issued immediately to remind relevant personnel to deal with it.

[0119] The screw conveyor stop working detection function module and the leakage detection function module adopt inter-frame difference and RGB→HSI color space change algorithm. Specifically, in the screw conveyor stop working detection function module, the inter-frame difference algorithm is combined with the RGB to HSI color space conversion technology. First, the video frames of the screw conveyor operation area are collected in real time by the monitoring camera, and the differential image between consecutive frames is calculated to detect motion changes. At the same time, the RGB image is converted into the HSI color space to enhance the color features of the image, which is convenient for identifying whether the screw conveyor has stopped running (such as color change or disappearance of the motion area). The model determines whether the screw conveyor is in a stopped state by analyzing the inter-frame differential image and HSI color features, and issues an alarm in real time.

[0120] The belt stacking detection function module adopts an unsupervised learning feature comparison algorithm. Specifically, in the belt stacking detection function module, an unsupervised learning feature comparison algorithm is adopted to extract image features through an autoencoder or a contrastive learning method. First, a high-resolution camera is used to capture an image of the belt conveying area, and the image is input into the autoencoder model. The autoencoder compresses the image into a low-dimensional feature representation through the encoder part, and then reconstructs the image through the decoder part. During the training process, the model learns the feature representation of the image by minimizing the error between the input image and the reconstructed image. In actual application, the system compares the features of the current image with the features of the normal state (no stacking). If the feature difference exceeds the set threshold, it is judged that there is a stacking phenomenon, and an alarm is issued in real time to remind relevant personnel to deal with it.

[0121] The monitoring device status detection function module first captures the monitoring video stream in real time and inputs each frame of the image into a pre-trained convolutional neural network model. The model is built based on a deep learning framework, uses multiple convolutional layers and pooling layers to extract image features, and performs classification or status judgment through a fully connected layer. During training, supervised learning is carried out using a labeled normal and abnormal state image dataset of monitoring equipment, and the model parameters are optimized to narrow the gap between the predicted device status and the actual status. The video frames are analyzed in real time and the results of whether the monitoring device is in a normal or abnormal state are output, thereby realizing the monitoring device status detection function;

[0122] The abnormal intrusion detection module utilizes a convolutional neural network algorithm to capture surveillance video streams in real time. Each frame is fed into a pre-trained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features, and a fully connected layer is used to determine whether an intrusion is abnormal. During training, supervised learning is performed using a dataset of images labeled with abnormal intrusion patterns. This optimizes model parameters and reduces the error between predicted and actual results. The module then analyzes video frames in real time and outputs whether an abnormal intrusion is present in the current scene, effectively achieving abnormal intrusion detection.

[0123] The substrate color detection module uses a convolutional neural network algorithm. It first acquires a surveillance video stream in real time and feeds each frame into a pre-trained convolutional neural network model. Based on a deep learning framework, it uses multiple convolutional and pooling layers to extract substrate color features, and then performs color classification and recognition through a fully connected layer. During training, supervised learning is carried out using a standard image dataset with substrate colors annotated. The model parameters are optimized to ensure that the predicted color is as close as possible to the actual color. The video frames are analyzed in real time and the color of the substrate in the current scene is output to realize the substrate color detection function.

[0124] The charge loader charge warning detection module uses a convolutional neural network algorithm to capture surveillance video streams in real time. Each frame is fed into a pre-trained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features related to the charge loader charge level. A fully connected layer is then used to determine whether the charge level has reached the warning range. During the training phase, supervised learning is performed using a dataset of charge loader images labeled with different charge levels. This optimizes model parameters and reduces the deviation between the predicted charge level and the actual level. The module then analyzes video frames in real time and outputs whether the charge loader's current charge level is in the warning state, completing the charge loader charge warning detection function.

[0125] The sensitizer overflow detection module, centered around a convolutional neural network algorithm, first captures surveillance video streams in real time. Each frame is fed into a pretrained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features indicating potential sensitizer overflow. A fully connected layer then determines whether an overflow has occurred. During training, supervised learning is performed using a dataset of images of sensitizer operating scenes, labeled with overflow patterns. This optimizes model parameters, narrowing the gap between predicted overflow states and actual conditions. The module then analyzes video frames in real time and outputs a signal indicating whether the sensitizer is currently overflowing, effectively implementing sensitizer overflow detection.

[0126] The sub-sodium dropper anomaly detection module utilizes a convolutional neural network algorithm. It first acquires a surveillance video stream in real time and feeds each frame into a pre-trained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features related to the sub-sodium dropper's operating status. A fully connected layer is then used to determine whether the dropper is abnormal. During training, supervised learning is performed using a dataset of images labeled with normal and abnormal dropper states. This optimizes model parameters and reduces the error between predicted results and actual conditions. The system then analyzes video frames in real time and outputs whether the sub-sodium dropper is currently in an abnormal state, achieving sub-sodium dropper anomaly detection.

[0127] The spiral drug counterfeiting detection module uses a convolutional neural network algorithm to capture surveillance video streams in real time. Each frame is fed into a pre-trained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features related to drug counterfeiting in the spiral. A fully connected layer is then used to determine whether drug counterfeiting is occurring. During training, supervised learning is performed using a dataset of images of spiral work scenes labeled with drug counterfeiting. Model parameters are optimized to ensure that predicted drug counterfeiting results are as consistent as possible with actual drug counterfeiting. Video frames are analyzed in real time to output whether drug counterfeiting is currently occurring in the spiral, thus enabling spiral drug counterfeiting detection.

[0128] The expanded explosive spill detection module utilizes a convolutional neural network algorithm. It first captures surveillance video streams in real time and feeds each frame into a pretrained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features related to explosive spillage. A fully connected layer then determines whether spillage has occurred. During the training phase, supervised learning is performed using a dataset of images labeled with the presence or absence of explosive spillage. This optimizes model parameters, narrowing the gap between predicted spillage and actual conditions. The system then analyzes video frames in real time and outputs a signal indicating whether explosive spillage has occurred in the current scene, completing the expanded explosive spill detection function.

[0129] The leak detection module utilizes a convolutional neural network algorithm to capture a real-time surveillance video stream. Each frame is fed into a pre-trained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features that indicate potential leaks. A fully connected layer then determines whether a leak is present. During training, supervised learning is performed using a dataset of images of equipment operating scenes labeled for leaks. This optimizes model parameters, reduces the error between predicted and actual leak states, and analyzes video frames in real time to output whether the current device is leaking, thus enabling leak detection.

[0130] The cooling water flow interruption detection module uses a convolutional neural network algorithm to capture the monitoring video stream in real time. Each image frame is input into a pre-trained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features related to the cooling water flow state. A fully connected layer is then used to determine whether a flow interruption has occurred. During training, supervised learning is performed using an image dataset labeled with normal cooling water flow and interruption states. Model parameters are optimized to ensure that the predicted flow interruption results are as close as possible to the actual cooling water state. Video frames are analyzed in real time to output whether the cooling water is currently in a flow interruption state, thus achieving the cooling water flow interruption detection function.

[0131] The human boundary crossing detection module, centered around a convolutional neural network algorithm, first captures surveillance video streams in real time and feeds each frame into a pre-trained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features related to the person's location and boundaries. Fully connected layers are then used to determine whether a person has crossed the boundary. During the training phase, supervised learning is performed using a dataset of images annotated with human boundary crossings and normal activity areas. This optimizes model parameters, narrowing the gap between predicted boundary crossings and actual activity. The system then analyzes video frames in real time and outputs whether a person has crossed the boundary, effectively implementing human boundary crossing detection.

[0132] The loading robot inactivity detection module utilizes a convolutional neural network algorithm. It first captures a surveillance video stream in real time and feeds each frame into a pretrained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features related to the loading robot's operating state. A fully connected layer is then used to determine whether the robot is inactive. During training, supervised learning is performed using a dataset of images labeled with the robot in both normal and inactive states. This optimizes model parameters and reduces the error between the predicted and actual robot states. The module then analyzes video frames in real time and outputs whether the loading robot is currently inactive, completing the loading robot inactivity detection function.

[0133] The unmanned detection module utilizes a convolutional neural network algorithm to capture surveillance video streams in real time. Each frame is fed into a pretrained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features related to personnel present in the scene. A fully connected layer is then used to determine whether an unmanned vehicle is present. During training, supervised learning is performed using a dataset of images of scenes labeled with both attended and unattended personnel. The model parameters are optimized to ensure that predicted unmanned vehicle conditions are as consistent as possible with actual conditions. The model then analyzes video frames in real time and outputs whether the scene is currently unmanned, thus implementing unmanned vehicle detection.

[0134] The fire escape obstruction detection module uses a convolutional neural network algorithm to capture surveillance video streams in real time. Each image frame is fed into a pre-trained convolutional neural network model. Multiple convolutional and pooling layers are used to extract image features related to whether the fire escape is blocked. A fully connected layer then determines whether the fire escape is unobstructed. During the training phase, supervised learning is performed using a dataset of images labeled with unobstructed and blocked fire escapes. This optimizes model parameters, narrowing the gap between predicted and actual fire escape status. The module then analyzes video frames in real time and outputs whether the fire escape is currently blocked, achieving fire escape obstruction detection.

[0135] Example 2

[0136] See Figure 3 - Figure 4 , adopts a production safety supervision method based on artificial intelligence and big data analysis. The production equipment management system is the MES production management system, and the monitoring equipment is the IP network camera.

[0137] It includes the following steps: hierarchical push, generation of emergency plan, relevant personnel intervention, problem solving, and archiving and evidence storage.

[0138] Tiered push: Data collection and risk assessment: Use artificial intelligence algorithms to collect and analyze real-time data related to production safety (such as equipment operating parameters, environmental indicators, personnel operation records, etc.), combine historical data and industry standards to assess the risk level of the current production safety situation, and divide it into three levels: low, medium, and high.

[0139] Determine push recipients: Based on the risk level, determine the appropriate push recipients. Low-risk information is pushed to grassroots safety managers; medium-risk information is pushed to department heads and heads of safety management departments; high-risk information is pushed to the company's principal, safety director, and relevant regulatory authorities.

[0140] Information push: Through the system's built-in message push module, risk information will be pushed to the corresponding recipients in a timely manner in the form of notifications, and unprocessed risk information will be highlighted on the system interface.

[0141] Generate emergency plans: Risk type judgment: When the system identifies medium or high-risk situations, the artificial intelligence algorithm quickly determines the type of risk (such as fire, explosion, hazardous material leakage, mechanical failure, etc.).

[0142] Plan template matching: According to the risk type, the corresponding emergency plan template is matched from the plan database. The template contains emergency handling procedures, division of responsibilities, resource allocation and other contents for different risks.

[0143] Personalized customization: Combine real-time data with actual on-site conditions to customize the matching emergency plan template. For example, you can adjust the rescue route based on the specific location of the accident or determine the emergency repair plan based on the equipment status.

[0144] Plan generation and review: A detailed emergency plan is generated and automatically reviewed by the system to ensure its completeness and feasibility. Once approved, it is sent to the relevant person in charge for confirmation and activation.

[0145] Human intervention: Task allocation: Based on the generated emergency plan, the system automatically assigns tasks to the corresponding emergency rescue personnel and departments. For example, it notifies firefighters to extinguish the fire, arranges technicians to repair equipment, and dispatches medical personnel to treat the injured.

[0146] Personnel notification: timely notify relevant personnel to intervene in emergency response work through various means (such as telephone, intercom, system messages, etc.), and inform them of their specific tasks and responsibilities.

[0147] Personnel arrival confirmation: After arriving at the scene, rescuers sign in through the system to confirm their arrival in real time. At the same time, the system can also urge and remind personnel who have not arrived in time.

[0148] Problem Solving: On-site Command and Coordination: The emergency command center provides unified command and coordination for rescue efforts based on on-site feedback and real-time data provided by the system. AI-assisted decision-making optimizes rescue plans and rationally allocates resources.

[0149] Real-time monitoring and adjustment: Utilizing various sensors and monitoring equipment installed on-site, we monitor the progress of incident handling and changes in the on-site environment in real time. We adjust rescue strategies and measures based on actual conditions to ensure the safety and effectiveness of rescue efforts.

[0150] Problem Resolution Assessment: Once the incident is initially under control, evaluate the resolution of the problem. Determine whether the safety hazard has been completely eliminated and whether normal production has been restored. If any issues remain, continue to implement appropriate measures to address them.

[0151] Archiving: Data collection and organization: Collect all data from the entire production safety supervision and emergency response process, including risk assessment reports, emergency plans, personnel task allocation records, on-site monitoring videos, various data from the rescue process, etc., and classify and organize them.

[0152] Document generation: Generate a detailed accident handling report based on the collated data, including the cause of the accident, handling process, lessons learned, improvement measures, etc.

[0153] Archive storage: Store the organized data and generated reports in the system's database for long-term preservation. At the same time, establish indexing and retrieval mechanisms to facilitate subsequent queries and analysis.

[0154] Regular review and update: Regularly review archived materials, update and improve relevant content based on new regulations and standards, technological developments and the actual situation of the enterprise, and continuously improve the effectiveness and adaptability of the production safety supervision system.

[0155] The MES production management system transmits production equipment and production data to the data and business coordination server. The data and business coordination server transmits the business control data required by the video analysis server to the video analysis server. The monitoring equipment transmits the data obtained from monitoring to the video analysis server. The video analysis server performs a security analysis of the on-site scene based on the received data, and reports the detected scene images that meet the security alarm to the data and business coordination server. The data and business coordination server uploads the detection alarm data to the MES production management system. The video analysis server receives the basic data of production equipment, production status data, business control data and image data input by the IP network camera, and collaboratively detects abnormal problems in the scene based on the corresponding data and specific algorithms, iteratively trains the detection results, and detects alarms.

[0156] Although the present invention has been described in detail above using general descriptions and specific embodiments, it will be apparent to those skilled in the art that modifications and improvements may be made based on the present invention. Therefore, such modifications and improvements, which do not depart from the spirit of the present invention, are intended to be within the scope of protection claimed herein.

Claims

1. A production safety supervision system based on artificial intelligence and big data analysis, characterized by: Including production equipment management system, data and business coordination server, video analysis server and monitoring equipment; The production equipment management system manages production equipment and production data, uploads production equipment and production data to the data and business coordination server, and receives alarm data sent back by the data and business coordination server; The data and business coordination server coordinates the interaction of various types of data and the classification of different functions, forwards and coordinates the business functions of various business terminals, connects to the fire protection system, and provides background parameter control modules and business data coordination function modules; The video analysis server is responsible for analyzing the security images corresponding to the scene, and reporting the scene images that meet the security alarm requirements, including a security scene function detection module; The monitoring equipment monitors the on-site production situation.

2. The production safety supervision system based on artificial intelligence and big data analysis according to claim 1 is characterized in that: The safety scenario function detection module includes: Monitoring device status detection module: triggers an alarm when the camera is offline, blocked, displaced, the image is blurred, or falls; Item placement detection module: The system automatically identifies items that are placed outside the area or are abnormally retained, triggering an alarm; Personnel quota detection module: When the number of personnel on the production line exceeds the preset value, an alarm is triggered; Clothing detection module: Identifies the labor protection clothing of production line personnel and triggers an alarm when it finds personnel whose clothing does not meet the requirements; Abnormal intrusion detection module: When a person enters an area that is not allowed to be entered during equipment operation, an alarm is triggered if a person is identified; Matrix color detection module: When the color of the emulsion explosive matrix on the emulsifier changes abnormally, an alarm is triggered; Screw conveyor stoppage detection module: triggers an alarm when it detects that the screw conveyor of the expanded explosives line has stopped working abnormally; Detonator quantity detection module: When the electronic detonator laser coding machine codes the electronic detonators, if the number of detonators is insufficient, an alarm is triggered; Charger drug quantity warning detection module: When the drug quantity in the drug hopper exceeds the warning position, an alarm is triggered; Sensitizer overflow detection module: triggers an alarm when emulsion explosives overflow from the upper or lower layer of the sensitizer; Sodium nitrate dropper abnormality detection module: When the sodium nitrate dropper does not swing normally to drop material, an alarm is triggered; Belt stacking detection module: When the conveyor stops working due to abnormality, or when the car transfer is not timely positioned, or when the robot fails to malfunction and causes stacking of boxes, an alarm is triggered; Spiral explosive leakage detection module: triggers an alarm when expanded explosives overflow the spiral hopper; Expanded explosive spillage detection module: triggers an alarm when expanded explosives are detected to have spilled during transportation; Leakage detection module: triggers an alarm when leakage or overflow occurs in the storage tank; Cooling water interruption detection module: triggers an alarm when the emulsifier cooling water interruption is detected; Personnel crossing detection module: When it is detected that someone crosses the prohibited crossing area on the production line, an alarm is triggered; Loading robot stoppage detection module: When a box is brought in by the conveyor belt and the robot stops working, an alarm is triggered; Unattended detection module: triggers an alarm when the monitoring duty room is unattended during the production process; Rough loading and unloading detection module: triggers an alarm when it detects that the loading and unloading personnel are handling finished goods roughly; Two-person double-lock detection module: triggers an alarm when one person is detected opening the warehouse door; Fire passage obstruction detection module: When goods block the fire passage in the production workshop, an alarm is triggered.

3. The production safety supervision system based on artificial intelligence and big data analysis according to claim 1 is characterized in that: The business data collaboration function module includes an MES equipment management system data synchronization function module, an alarm data upload and docking function module, and a business overall management function module.

4. The production safety supervision system based on artificial intelligence and big data analysis according to claim 2 is characterized in that: The personnel quota detection function module, the detonator quantity detection function module, the monitoring equipment status detection function module, the abnormal intrusion detection function module, the matrix color detection function module, the charging machine charge warning detection function module, the sensitizer overflow detection function module, the sub-sodium dropper abnormality detection function module, the spiral leakage detection function module, the expanded explosive spillage detection function module, the leakage detection function module, the cooling water interruption detection function module, the personnel crossing the boundary detection function module, the loading robot stop working detection function module, the unattended detection function module, and the fire passage obstruction detection function module adopt a convolutional neural network algorithm. Specifically, first, the monitoring video stream is captured in real time, and each frame of the image is input into a pre-trained convolutional neural network model. The model is built based on a deep learning framework, uses multiple convolution layers and pooling layers to extract features in the image, and performs classification or counting operations through a fully connected layer. During the training process, a labeled personnel image data set is used for supervised learning, the model parameters are optimized to minimize the error between the prediction and the actual, the video frames are analyzed in real time, and the current scene results are output, thereby realizing the above-mentioned detection function; The rough loading and unloading detection and two-person double-locking detection modules utilize a human behavior learning algorithm. Specifically, they utilize a deep learning-based behavior recognition algorithm, combined with OpenPose human key point detection technology, to extract key point information of human actions in surveillance videos in real time. They then construct a spatiotemporal graph convolutional network (ST-GCN) to model human action sequences and learn the feature representations of different action patterns. The model is trained with a large amount of labeled behavior data to identify whether there is rough loading and unloading behavior and trigger an alarm in real time. The clothing detection function module adopts a deep learning classification algorithm. Specifically, it uses a deep learning classification algorithm to perform real-time detection of personnel clothing based on a convolutional neural network (CNN). First, an image of a person in the monitoring area is acquired through an image acquisition device and the image is preprocessed. The image is then input into a CNN model trained with a large number of clothing samples. The model extracts features from the image through convolutional layers and pooling layers, and classifies it through a fully connected layer to determine whether the person is wearing protective equipment as required. The model outputs the classification result. If a situation that does not comply with the dress code is detected, the system will issue a prompt in real time to remind the person to make corrections. The object location detection module uses a deep learning recognition algorithm. Specifically, it uses a deep learning recognition algorithm to perform real-time detection of the position and status of objects based on a convolutional neural network (CNN). A high-resolution camera captures images of the working area and inputs the images into a CNN model trained with a large number of object samples. The model recognizes the features of the objects and determines whether they are placed in the specified location. During the training process, the model is supervised by using labeled object location data, and the model parameters are optimized to improve recognition accuracy. In actual application, the system analyzes images in real time to determine whether the objects are in the correct location. If an object is found to be displaced or not placed as required, an alarm will be immediately issued to remind personnel to take action. The screw conveyor stoppage detection function module and the leakage detection function module adopt inter-frame difference and RGB→HSI color space change algorithms. Specifically, the inter-frame difference algorithm is combined with RGB to HSI color space conversion technology. First, the video frames of the screw conveyor operation area are collected in real time by the monitoring camera, and the difference image between consecutive frames is calculated to detect motion changes. At the same time, the RGB image is converted into the HSI color space to enhance the color features of the image and identify whether the screw conveyor has stopped. By analyzing the inter-frame difference image and HSI color features, the model determines in real time whether the screw conveyor is in a stopped state and issues an alarm if it has stopped. The belt stacking detection function module adopts an unsupervised learning feature comparison algorithm. Specifically, image features are extracted through an autoencoder or a contrastive learning method. First, a high-resolution camera is used to capture images of the belt conveyor area, and the images are input into the autoencoder model. The autoencoder compresses the image into a low-dimensional feature representation through the encoder part, and then reconstructs the image through the decoder part. During the training process, the model learns the feature representation of the image by minimizing the error between the input image and the reconstructed image. In actual application, the system compares the features of the current image with the features of the normal non-stacked state. If the feature difference exceeds the set threshold, it is judged that there is a stacking phenomenon, and an alarm is issued in real time to remind relevant personnel to deal with it.

5. The production safety supervision system based on artificial intelligence and big data analysis according to claim 1 is characterized in that: The monitoring device is a network camera, which is connected to a switch on site and transmits video data to a video analysis server.

6. The production safety supervision system based on artificial intelligence and big data analysis according to claim 1 is characterized in that: The business data collaboration function module includes a production equipment management system data synchronization function, an alarm data upload and docking function, and a business overall management function.

7. The production safety supervision system based on artificial intelligence and big data analysis according to claim 1 is characterized in that: The data and business coordination server web terminal displays alarm information, manages each video analysis server connected to the system, and implements basic parameter settings in the background parameter control module.

8. A production safety supervision method based on artificial intelligence and big data analysis, characterized in that: The method is implemented using the production safety supervision system based on artificial intelligence and big data analysis as described in any one of claims 1 to 7, specifically comprising the following steps: S10 tiered push, specifically, S11 Data Collection and Risk Assessment: Utilizes artificial intelligence algorithms to collect and analyze real-time data on production safety, and combines historical data with industry standards to assess the risk level of the current production safety situation, categorizing it into three levels: low, medium, and high. S12 Determine push recipients: Determine the corresponding push recipients based on risk levels. Low-risk information is pushed to grassroots safety managers, medium-risk information is pushed to department heads and heads of safety management departments, and high-risk information is pushed to the company's principal, safety director, and relevant regulatory authorities. S13 Information Push: Through the system's built-in message push module, risk information is promptly pushed to the corresponding recipients in the form of notifications, and unprocessed risk information is highlighted on the system interface; S20 generates an emergency plan, specifically, S21 Risk Type Determination: When the system identifies medium or high risk situations, the AI algorithm determines the risk type; S22 Plan Template Matching: Based on the risk type, the corresponding emergency plan template is matched from the plan database. The template includes the emergency response process, division of responsibilities, and resource allocation for different risks; S23 Personalized Customization: Combine real-time data with actual on-site conditions to personalize the matching emergency plan templates; S24 Plan Generation and Review: Generate a detailed emergency plan, which the system automatically conducts a preliminary review to ensure its completeness and feasibility. Once approved, it is pushed to the relevant person in charge for confirmation and activation. S30 personnel intervention, specifically, S31 Task Allocation: Based on the generated emergency plan, the system automatically allocates tasks to the corresponding emergency rescue personnel and departments; S32 Personnel Notification: Notify relevant personnel to participate in emergency response work in a timely manner through telephone, intercom, and system messages, and inform them of their specific tasks and responsibilities; S33 Personnel arrival confirmation: After arriving at the scene, rescuers sign in and confirm through the system, so as to obtain real-time information on personnel arrival. The system will urge and remind personnel who have not arrived in time. S40 solves the problem, specifically, S41 On-site Command and Coordination: The emergency command center conducts unified command and coordination of rescue work based on on-site feedback and real-time data provided by the system. It uses artificial intelligence to assist in decision-making, optimize rescue plans, and rationally allocate resources. S42 Real-time Monitoring and Adjustment: Utilizes various sensors and monitoring equipment installed on-site to monitor the progress of accident handling and changes in the on-site environment in real time, and promptly adjusts rescue strategies and measures based on actual conditions; S43 Problem Solving Assessment: After the accident is initially under control, the problem solving situation will be assessed to determine whether the safety hazard has been completely eliminated and whether normal production order has been restored. If problems still exist, appropriate measures will be taken to address them. S50 archive archive, specifically, S51 Data Collection and Collation: Collect all data from the entire production safety supervision and emergency response process, including risk assessment reports, emergency plans, personnel task assignment records, on-site monitoring videos, and various data from the rescue process, and classify and collate them; S52 document generation: Generate a detailed accident handling report based on the collated data, including the cause of the accident, handling process, lessons learned, and improvement measures; S53 Archiving Storage: Store the organized data and generated reports in the system's database for long-term preservation, and establish indexing and retrieval mechanisms to facilitate subsequent query and analysis; S54 Regular review and update: Regularly review archived materials and update and improve relevant content based on new regulations and standards, technological developments and the actual situation of the enterprise.

9. The method for production safety supervision based on artificial intelligence and big data analysis according to claim 8 is characterized in that: Specifically include: The production equipment management system transmits production equipment and production data to the data and business coordination server. The data and business coordination server transmits the business control data required by the video analysis server to the video analysis server. The monitoring equipment transmits the acquired data to the video analysis server. The video analysis server performs a security analysis of the on-site scene based on the received data, and reports the detected scene images that meet the security alarm requirements to the data and business coordination server. The data and business coordination server uploads the detection alarm data to the production equipment management system. The video analysis server receives basic data of production equipment, production status data, business control data and image data input by IP network cameras. It collaboratively detects abnormal problems in the scene based on the corresponding data and specific algorithms, iteratively trains the detection results, and detects alarms.

Images