Aluminum processing deep well casting production line AI video violation snapshot monitoring method

Abstract

The invention provides an aluminum processing deep well casting production line AI video violation snapshot monitoring method, which is characterized in that an intelligent safety production early warning model is constructed based on a computer vision deep learning technology, embedded equipment is used as a data acquisition terminal, and the action of the data acquisition terminal is associated with the operation of a console of an aluminum processing deep well casting production line. When the console starts to execute a dangerous process, the data acquisition terminal is triggered, the data acquisition terminal carries out continuous data acquisition on a monitoring picture of the front-end monitoring equipment and sends the acquired data to the AI monitoring video server at the rear end, and an AI system of the AI monitoring video server carries out data analysis through an intelligent safety production early warning model. AI identification is carried out by an AI analysis module, and alarm response is carried out according to an analysis result; according to the invention, the operation behavior of the casting production field operation personnel can be monitored in real time, the unsafe behavior of the personnel is judged through AI visual analysis, an early warning is given to remind the personnel to correct the unsafe behavior, and the occurrence of safety accidents is avoided.

Description

Technical Field

[0001] The present invention relates to the field of monitoring technology, and in particular to an AI video violation capture monitoring method for an aluminum processing deep well casting production line. Background Art

[0002] The deep well casting production process is a key high-risk place in the production process of the aluminum processing industry. The explosion caused by the leakage of high-temperature aluminum liquid in water is the most serious safety hazard in the deep well casting process of aluminum processing, which is likely to cause mass deaths and injuries. According to statistics, in the past 20 years, there have been 14 explosion accidents caused by the leakage of high-temperature aluminum liquid in China (incomplete statistics), resulting in 64 deaths and direct economic losses of about 90 million yuan. For the 14 aluminum liquid explosion accidents, a comprehensive cause analysis was conducted from four aspects: unsafe behavior of people, unsafe state of objects, poor environment, and management defects. It was found that the rate of accidents caused by unsafe behavior of people was as high as 64%, including employees leaving their posts without authorization, resulting in the casting production site being unsupervised for a long time, and aluminum liquid leakage was not discovered in time; the operators neglected to observe and failed to discover aluminum leakage and emergency disposal in time; the operators violated the safety operating procedures and violated the regulations to cause accidents. It can be seen that unsafe behavior of people often plays a decisive role in the occurrence of accidents. According to Heinrich's Law, for every 330 accidents, 300 are caused by unsafe behavior. More than 90% of safety accidents are related to violations of regulations, and there are countless accidents that lead to death due to violations of regulations. Therefore, in order to prevent accidents from happening, it is necessary to strengthen the supervision and control of the safety behavior of on-site workers.

[0003] The "Standards for Determining Major Accident Hazards in Industry and Trade (2017 Edition)" lists "Article 7: The number of people in the casting workshop is not strictly controlled, and the entry of non-production personnel is not controlled" as a major safety hazard. Enterprise safety management needs to strictly control the number of people in the casting workshop to minimize the number of irrelevant personnel stranded at the production site. At present, most casting production line sites have not effectively supervised the number of people entering the production site and the on-the-job personnel of key processes. There is a possibility of too many people entering the site and the lack of on-the-job personnel of key processes. Safety management is mostly institutionalized manual management, lacking high-tech monitoring support. There is no real-time online monitoring process for potential risks in the operation process, and it cannot be perceived and handled in the first time. There are blind spots in manual safety supervision, and it is impossible to carry out targeted and precise management of some key areas, key time periods, and key processes, and it is impossible to respond to events efficiently and quickly. Summary of the invention

[0004] The present invention proposes an AI video violation capture monitoring method for an aluminum processing deep-well casting production line. Combined with the characteristics of the aluminum processing deep-well casting production site, visual AI analysis and safety early warning are carried out to monitor the working behavior of workers on the melting and casting production site in real time. By judging people's unsafe behavior through AI visual analysis, an early warning can be issued in the first place to remind people to stop and correct unsafe behavior, changing post-event management to pre-event early warning to avoid safety accidents.

[0005] The present invention adopts the following technical solutions.

[0006] A method for monitoring AI video violations of deep-well casting production lines for aluminum processing. The monitoring method constructs an intelligent production safety early warning model based on computer vision deep learning technology, uses an embedded device as a data acquisition terminal, and the action of the data acquisition terminal is associated with the console operation of the aluminum processing deep-well casting production line. When the console starts to execute a dangerous process, the data acquisition terminal is triggered. The data acquisition terminal continuously collects data from the monitoring screen of the front-end monitoring device and sends the collected data to the back-end AI monitoring video server. The AI ​​system of the AI ​​monitoring video server performs data analysis through the intelligent production safety early warning model, performs AI recognition with an AI analysis module, and responds to alarms based on the analysis results.

[0007] The monitoring method includes an AI video recognition and violation capture method for the converter area of ​​an aluminum processing deep well casting production line, and also includes an AI video recognition and violation capture method for the casting area of ​​an aluminum processing deep well casting production line; the data acquisition terminal collects data from the monitoring screen of the front-end monitoring equipment of the aluminum processing deep well casting production line through a camera.

[0008] The AI ​​video recognition and illegal capture method of the converter area specifically includes the following steps: Step A1, when the converter operation starts, the operator presses the converter discharge alarm button on the console, and the converter start signal is given to the AI ​​monitoring video server. After receiving the signal, the AI ​​monitoring video server starts data collection at the data collection terminal and starts to execute the illegal capture function.

[0009] Step A2: During the entire converter operation process, when AI identifies that the operator at the converter process is not wearing a safety helmet, a protective face shield, or foot protection, an alarm sounds to remind the operator to correct the violation; Step A3: When the AI ​​system identifies that a person who is not on the converter operation whitelist is performing converter operation at the converter process, an alarm is issued to remind the operator to correct the problem; Step A4: When the AI ​​system identifies that the converter operator is off-duty at the converter process, an alarm is issued to remind the operator to correct the problem; Step A5, when the converter operation is completed, the operator turns off the converter discharge alarm through the console, and the console simultaneously sends the converter end signal to the AI ​​monitoring video server. After receiving the signal, the AI ​​monitoring video server executes the shutdown of the AI ​​recognition function.

[0010] The AI ​​video recognition and illegal capture method in the foundry area specifically includes the following steps: Step B1: When the operator presses the "casting preparation" button on the operation interface of the console, the casting start signal is sent to the AI ​​monitoring video server. After receiving the signal, the AI ​​monitoring video server starts data collection at the data collection terminal and starts to execute the function of identifying illegal shots; Step B2: During the entire casting operation, when AI identifies that the operator entering the casting safety warning area is not wearing any of the labor protection supplies such as foot protection, protective face shield with front and back neck gaiter, or protective clothing, an alarm will be sounded in time to remind the operator to correct the problem; Step B3: When the AI ​​system recognizes that the total number of people in the corresponding control area of ​​the foundry exceeds the set value, the system will immediately and automatically issue an alarm to remind people to evacuate to avoid overcrowding in the area; Step B4: When the AI ​​system recognizes that there is less than one person in the main operation room of the foundry and that a person is off duty, the system will immediately and automatically give an alarm to remind the operator to make corrections; Step B5: When the AI ​​system identifies that there are people sleeping on duty in the main operation room of the foundry, it can promptly issue an alarm to correct and remind the operator; Step B6: When the casting is finished, when the operator sends the relevant end instruction of the aluminum processing deep well casting production line on the console, the console synchronously sends a signal to the AI ​​monitoring video server. After receiving the signal, the AI ​​monitoring video server executes to shut down the AI ​​recognition function.

[0011] In step B3, an alarm is issued when the AI ​​system recognizes that the total number of people in the corresponding control area of ​​the foundry exceeds 8.

[0012] The intelligent production safety warning model includes an AI analysis module, which uses a convolutional neural network (CNN) structure and combines the target detection algorithm YOLOv5 for target detection and recognition. The efficient real-time detection performance of YOLOv5 is used to process the large-scale video stream of the front-end monitoring equipment monitoring screen of the aluminum processing deep well casting production line, and identify various types of objects and behavioral features in the video stream.

[0013] The intelligent production safety warning model is trained with a large number of video data sets of violation samples of aluminum processing deep well casting production lines, including target objects and violations in a variety of different scenarios. The video data sets are data-annotated and data-enhanced video data sets that cover a variety of samples under different lighting conditions and weather conditions. When annotating the video dataset, use annotation tools to annotate each image in the dataset, record the type of the target object, and annotate its location in the aluminum processing deep well casting production line with a bounding box; in addition, annotate the type of specific violations based on the characteristics of the violations; When performing data enhancement on the video dataset, the specific methods are as follows: rotation and scaling are used to enhance the recognition ability of the intelligent production safety warning model at different perspectives; random cropping and flipping are used to simulate target changes in the real scene of the aluminum processing deep well casting production line; and lighting adjustment is used to simulate different lighting conditions to improve the performance of the intelligent production safety warning model at night or in backlight conditions.

[0014] When training the intelligent production safety warning model, the YOLOv5 model, which has been pre-trained on the COCO public dataset and has good basic recognition capabilities, is used as the basis for fine-tuning on the target dataset to improve the recognition accuracy in specific scenarios; image enhancement techniques including rotation, scaling, and color adjustment are used during the training process to improve the robustness of the model; a video dataset containing a large number of violation samples of aluminum processing deep well casting production lines is used to fine-tune the trained model to improve the model's accuracy in identifying specific violations of aluminum processing deep well casting production lines; In terms of hyperparameter setting for model training, the following method is adopted: When setting the learning rate, set the initial learning rate to 0.001, and use the adaptive learning rate to adjust Reduce on Plateau during training. When the performance of the model on the validation set no longer improves, the learning rate is automatically reduced. When setting the batch size, set the batch size to 16 to balance the memory usage of the GPU device used for training and the training speed; When setting the optimizer, use the Adam optimizer combined with L2 regularization to prevent the model from overfitting; When setting the loss function, a multi-task loss function is used for YOLOv5, including classification loss, bounding box regression loss, and confidence loss, to ensure that the accuracy of the bounding box position is optimized while detecting.

[0015] The specific steps for training the intelligent production safety warning model are: Step S1: The model is trained on the training set for a total of 300 epochs, and the validation set is used to evaluate the model performance after each epoch.

[0016] Step S2: Introduce the Early Stopping strategy, and stop training when the accuracy on the validation set does not improve for 20 consecutive rounds to avoid overfitting.

[0017] Step S3: After training is completed, save the model parameters and perform performance evaluation to determine whether further optimization or adjustment of the model structure is needed.

[0018] In order to improve the AI ​​detection effect, the monitoring method adopts the following optimization strategies when training the intelligent production safety warning model: Anchor Box optimization: The K-means clustering algorithm is used to analyze the object sizes in the dataset and reset the AnchorBoxes according to the size distribution to ensure the accuracy of the model when detecting objects of different sizes. Model compression and quantization: After model training is completed, the model is quantized so that it can maintain a high detection speed when running on resource-constrained devices. At the same time, pruning technology is used to remove redundant neurons to further reduce the model size and improve computing efficiency.

[0019] The present invention can combine the characteristics of the deep-well casting production site of aluminum processing to conduct visual AI analysis and safety warnings, monitor the working behaviors of workers on the melting and casting production site in real time, and determine people's unsafe behaviors through AI visual analysis. It can issue warnings in the first place to remind people to stop and correct unsafe behaviors, changing post-event management to pre-event warnings to avoid safety accidents.

[0020] The beneficial effects of the present invention also lie in: 1. Real-time monitoring: Promote the cloud platform of equipment and systems to realize all-round real-time monitoring of production.

[0021] 2. Rapid perception: Apply intelligent sensors, measuring instruments, edge computing and other technologies to realize the collection of all-factor information.

[0022] 3. System evaluation: Develop an evaluation model and tool set based on the Industrial Internet to achieve a comprehensive evaluation of the safety behavior of operators and a systematic evaluation of safe production.

[0023] 4. Joint response: Build and use emergency drills and emergency response plan libraries to improve the scientificity, accuracy and rapid response of emergency response.

[0024] 5. Advance warning: Build a risk and failure feature library and apply big data analysis to achieve accurate prediction and intelligent advance warning.

[0025] 6. Efficient and intelligent: Automatically identify and capture violations to reduce the workload of manual monitoring.

[0026] 7. Strong real-time performance: Able to process video streams in real time, identify and record violations.

[0027] 8. Wide applicability: The system is applicable to a variety of scenarios and can adapt to different environmental requirements through model adjustment.

[0028] 9. The system has a mature technical solution, strong feasibility and low risk. After the project is completed, it will be able to build a comprehensive management platform for personnel positioning, overcrowding alarm and violation supervision, to achieve strict implementation of the system, scientific management, safety compliance, and improve the company's safety information management level. It will effectively implement the actual use of information technology and means to manage safety, and has great safety value and comprehensive benefits.

[0029] 10. The completion of this system can significantly improve the safety management level of deep well casting process, effectively realize the accurate limit of the number of personnel in the foundry workshop, and timely issue alarms for the absence of personnel in important positions. It can effectively monitor general violations and the wearing of special labor protection products. Under the background that the Ministry of Emergency Management attaches great importance to the production safety of deep well casting, it has excellent promotion and application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments: Attached Figure 1 It is a schematic diagram of the equipment principle of the present invention. DETAILED DESCRIPTION

[0031] As shown in the figure, an AI video violation capture monitoring method for an aluminum processing deep well casting production line is provided. The monitoring method constructs an intelligent production safety early warning model based on computer vision deep learning technology, and uses an embedded device as a data acquisition terminal. The action of the data acquisition terminal is associated with the console operation of the aluminum processing deep well casting production line. When the console starts to execute a dangerous process, the data acquisition terminal is triggered. The data acquisition terminal continuously collects data from the monitoring screen of the front-end monitoring device and sends the collected data to the back-end AI monitoring video server. The AI ​​system of the AI ​​monitoring video server performs data analysis through the intelligent production safety early warning model, performs AI recognition with the AI ​​analysis module, and responds to alarms based on the analysis results.

[0032] The monitoring method includes an AI video recognition and violation capture method for the converter area of ​​an aluminum processing deep well casting production line, and also includes an AI video recognition and violation capture method for the casting area of ​​an aluminum processing deep well casting production line; the data acquisition terminal collects data from the monitoring screen of the front-end monitoring equipment of the aluminum processing deep well casting production line through a camera.

[0033] The AI ​​video recognition and illegal capture method of the converter area specifically includes the following steps: Step A1, when the converter operation starts, the operator presses the converter discharge alarm button on the console, and the converter start signal is given to the AI ​​monitoring video server. After receiving the signal, the AI ​​monitoring video server starts data collection at the data collection terminal and starts to execute the illegal capture function.

[0034] Step A2: During the entire converter operation process, when AI identifies that the operator at the converter process is not wearing a safety helmet, a protective face shield, or foot protection, an alarm sounds to remind the operator to correct the violation; Step A3: When the AI ​​system identifies that a person who is not on the converter operation whitelist is performing converter operation at the converter process, an alarm is issued to remind the operator to correct the problem; Step A4: When the AI ​​system identifies that the converter operator is off-duty at the converter process, an alarm is issued to remind the operator to correct the problem; Step A5, when the converter operation is completed, the operator turns off the converter discharge alarm through the console, and the console simultaneously sends the converter end signal to the AI ​​monitoring video server. After receiving the signal, the AI ​​monitoring video server executes the shutdown of the AI ​​recognition function.

[0035] The AI ​​video recognition and illegal capture method in the foundry area specifically includes the following steps: Step B1: When the operator presses the "casting preparation" button on the operation interface of the console, the casting start signal is sent to the AI ​​monitoring video server. After receiving the signal, the AI ​​monitoring video server starts data collection at the data collection terminal and starts to execute the function of identifying illegal shots; Step B2: During the entire casting operation, when AI identifies that the operator entering the casting safety warning area is not wearing any of the labor protection supplies such as foot protection, protective face shield with front and back neck gaiter, or protective clothing, an alarm will be sounded in time to remind the operator to correct the problem; Step B3: When the AI ​​system recognizes that the total number of people in the corresponding control area of ​​the foundry exceeds the set value, the system will immediately and automatically issue an alarm to remind people to evacuate to avoid overcrowding in the area; Step B4: When the AI ​​system recognizes that there is less than one person in the main operation room of the foundry and that a person is off duty, the system will immediately and automatically give an alarm to remind the operator to make corrections; Step B5: When the AI ​​system identifies that there are people sleeping on duty in the main operation room of the foundry, it can promptly issue an alarm to correct and remind the operator; Step B6: When the casting is finished, when the operator sends the relevant end instruction of the aluminum processing deep well casting production line on the console, the console synchronously sends a signal to the AI ​​monitoring video server. After receiving the signal, the AI ​​monitoring video server executes to shut down the AI ​​recognition function.

[0036] In step B3, an alarm is issued when the AI ​​system recognizes that the total number of people in the corresponding control area of ​​the foundry exceeds 8.

[0037] The intelligent production safety warning model includes an AI analysis module, which uses a convolutional neural network (CNN) structure and combines the target detection algorithm YOLOv5 for target detection and recognition. The efficient real-time detection performance of YOLOv5 is used to process the large-scale video stream of the front-end monitoring equipment monitoring screen of the aluminum processing deep well casting production line, and identify various types of objects and behavioral features in the video stream.

[0038] The intelligent production safety warning model is trained with a large number of video data sets of violation samples of aluminum processing deep well casting production lines, including target objects and violations in a variety of different scenarios. The video data sets are data-annotated and data-enhanced video data sets that cover a variety of samples under different lighting conditions and weather conditions. When annotating the video dataset, use annotation tools to annotate each image in the dataset, record the type of the target object, and annotate its location in the aluminum processing deep well casting production line with a bounding box; in addition, annotate the type of specific violations based on the characteristics of the violations; When performing data enhancement on the video dataset, the specific methods are as follows: rotation and scaling are used to enhance the recognition ability of the intelligent production safety warning model at different perspectives; random cropping and flipping are used to simulate target changes in the real scene of the aluminum processing deep well casting production line; and lighting adjustment is used to simulate different lighting conditions to improve the performance of the intelligent production safety warning model at night or in backlight conditions.

[0039] When training the intelligent production safety warning model, the YOLOv5 model, which has been pre-trained on the COCO public dataset and has good basic recognition capabilities, is used as the basis for fine-tuning on the target dataset to improve the recognition accuracy in specific scenarios; image enhancement techniques including rotation, scaling, and color adjustment are used during the training process to improve the robustness of the model; a video dataset containing a large number of violation samples of aluminum processing deep well casting production lines is used to fine-tune the trained model to improve the model's accuracy in identifying specific violations of aluminum processing deep well casting production lines; In terms of hyperparameter setting for model training, the following method is adopted: When setting the learning rate, set the initial learning rate to 0.001, and use the adaptive learning rate to adjust Reduce on Plateau during training. When the performance of the model on the validation set no longer improves, the learning rate is automatically reduced. When setting the batch size, set the batch size to 16 to balance the memory usage of the GPU device used for training and the training speed; When setting the optimizer, use the Adam optimizer combined with L2 regularization to prevent the model from overfitting; When setting the loss function, a multi-task loss function is used for YOLOv5, including classification loss, bounding box regression loss, and confidence loss, to ensure that the accuracy of the bounding box position is optimized while detecting.

[0040] The specific steps for training the intelligent production safety warning model are: Step S1: The model is trained on the training set for a total of 300 epochs, and the validation set is used to evaluate the model performance after each epoch.

[0041] Step S2: Introduce the Early Stopping strategy, and stop training when the accuracy on the validation set does not improve for 20 consecutive rounds to avoid overfitting.

[0042] Step S3: After training is completed, save the model parameters and perform performance evaluation to determine whether further optimization or adjustment of the model structure is needed.

[0043] In order to improve the AI ​​detection effect, the monitoring method adopts the following optimization strategies when training the intelligent production safety warning model: Anchor Box optimization: The K-means clustering algorithm is used to analyze the object sizes in the dataset and reset the AnchorBoxes according to the size distribution to ensure the accuracy of the model when detecting objects of different sizes. Model compression and quantization: After model training is completed, the model is quantized so that it can maintain a high detection speed when running on resource-constrained devices. At the same time, pruning technology is used to remove redundant neurons to further reduce the model size and improve computing efficiency.

[0044] In this example, the monitoring screen of the front-end monitoring equipment covers the aluminum processing deep well casting production line and the monitoring angle remains basically unchanged. When organizing the training data set, violations are sorted and labeled from historical monitoring videos. The trained intelligent production safety warning model is used to identify the distribution of monitoring points used in historical monitoring videos and monitoring videos in daily production.

[0045] Embodiment 1: The AI ​​video violation capture monitoring method for the aluminum processing deep well casting production line proposed in this example is based on the AI ​​video violation capture monitoring system, which integrates video access storage and AI reasoning. It follows the personnel positioning alarm information system of the aluminum processing deep well casting production line and adopts an embedded design. It can be compatible with existing front-end monitoring equipment, and can perform uninterrupted analysis and judgment on the monitoring screen to achieve all-weather automated video monitoring. The edge end is only responsible for collecting data, and the collected raw data is uniformly uploaded to the data acquisition control center for caching and device control. The video data will be passed to the intelligent reasoning platform for processing, and alarm information will be generated, and then the alarm information will be pushed to the reasoning platform monitoring center. The reasoning platform monitoring center will visualize the received data and trigger alarm events in real time.

[0046] The system adopts B / S architecture. The overall computing service of the system is deployed in the intelligent reasoning platform server and is accessed and controlled using a browser client. Users do not need to install a client. It has the advantages of simple maintenance, easy use, and strong distribution.

[0047] The system combines the characteristics of the aluminum processing deep well casting production site, conducts visual AI analysis and safety warning, and builds a proprietary intelligent production safety warning model based on computer vision deep learning technology. It conducts all-round, full-factor, and full-coverage real-time perception, prediction, and warning of the safety behaviors of workers on the aluminum processing and casting production site, and catches them early to stop and correct unsafe behaviors. It realizes real-time perception of the safety situation, intelligent supervision of on-site operations, and dynamic warning of risks and hidden dangers, avoids the occurrence of safety accidents, and empowers enterprise safety production! The AI ​​video violation capture monitoring system uses the virtual electronic fence set up by the personnel positioning alarm information system to conduct real-time monitoring of key equipment (such as the casting area, converter area, material preparation area, lifting and unloading area, etc.) and key operation areas in the monitoring area. The system automatically tracks and identifies personnel entering the area. When the AI ​​system detects typical unsafe behavior of personnel in the monitoring area, the system promptly sends a warning signal to the monitoring center and the parties involved, and promptly reminds on-site personnel. The highlight of this system is that it can monitor and identify specific operations (converter operation period, casting operation period) and specific areas (within the electronic fence). The safety monitoring function of the AI ​​video violation capture monitoring system includes but is not limited to the following: 1. Safety helmet wearing monitoring: It can automatically identify the workers in the monitoring area. If it is detected that the workers are not wearing safety helmets, an alarm will be issued in time to correct and remind the workers. This algorithm scheme first uses the target detection algorithm to detect the head, locate the head position in the image, and obtain the detection frame containing the head. Based on this detection frame, the safety helmet classification algorithm is used to classify the detection frame content. The safety helmet confidence output by the classification is compared with the preset threshold. If it is lower than the threshold, an alarm message is output.

[0048] 2. Monitoring of protective equipment wearing in the converter area: It can automatically identify personnel in the melting and casting converter operation area. If it is detected that the operator is not wearing a safety helmet, protective face shield or foot protection, an alarm can be issued in time to correct and remind the operator. This algorithm scheme first uses the target detection algorithm to detect the human body, locate the human body in the image, and obtain the detection frame containing the human body. Then, based on this detection frame, the quality model is used for filtering to remove images with poor image quality. The pedestrian detection frame with good image quality is then classified for protective equipment wearing. The confidence of the protective equipment output by the classification is compared with the preset threshold. If it is lower than the threshold, an alarm message is output.

[0049] 3. Monitoring of protective equipment worn in the casting area: It can automatically identify personnel in the melting and casting operation area. If it is detected that the operator is not wearing a full set of labor protection equipment, including: foot protection, protective face shield with front and back neck guards, and protective clothing, when the operator does not wear any of these protective equipment, an alarm can be issued in time to correct and remind the operator. This algorithm solution is similar to the above-mentioned monitoring solution for wearing protective equipment in the converter area.

[0050] 4. Restricted area intrusion monitoring (area intrusion): It can automatically identify people in the monitoring area. If unauthorized people are detected entering the control area of ​​the melting and casting line, the system will automatically alarm and notify the management personnel. This algorithm solution first uses the target detection algorithm to detect the human body, locate the human body in the picture, and obtain the detection frame containing the human body. Then, based on this detection frame, it uses the quality model to filter and remove the human body area image with poor image quality. Then, based on the configured restricted area, it is judged whether there are people in the restricted area, and an alarm is issued if there are people in the area.

[0051] 5. Overcrowding monitoring: The smelting and casting production area is set as three levels of control areas, namely, level 1, level 2, and level 3, according to the degree of danger. If the total number of people in the corresponding control area exceeds the set value, the system will automatically alarm immediately to remind people to evacuate to avoid overcrowding in the area. This algorithm first uses the target detection algorithm to detect the human body, locate the human body in the picture, and obtain the detection frame containing the human body. Then, based on this detection frame, the quality model is used for filtering to remove images of the human body area with poor image quality. Then, based on the configured number of people, it is judged whether there is an excess of the limit in the area, and an alarm is issued when the number of people exceeds the limit.

[0052] 6. Personnel off-duty monitoring: For important places such as the deep well casting area and the main operation room, when the number of personnel is less than the set value, the system will automatically alarm immediately to remind the lack of personnel and avoid personnel off-duty in the area. This solution mainly uses the target detection algorithm to detect pedestrians in the working area, and uses the quality model to filter the results of the detection algorithm. If no personnel are detected in the specified area, the timer starts, and when the time exceeds the set time threshold, the alarm information is output.

[0053] 7. Regional personnel qualification monitoring: The system sets specific qualifications for converter operators, establishes a white list of converter operators (a list of operators with more than 3 months of converter operating experience), and conducts real-time monitoring of the entire converter operation process in the converter area. If it is detected that an unqualified person is performing converter operation, an alarm will be issued in time to correct and remind the operator.

[0054] 8. Smoking detection: It can automatically identify people in the monitoring area. If it detects that a worker is smoking in a non-smoking area, it can issue an alarm in time to correct and remind the worker. This algorithm first uses the target detection algorithm to detect the head, locate the head in the image, and obtain the detection frame containing the head. Based on this detection frame, the smoking classification algorithm is used to classify the detection frame content. The smoking confidence output by the classification is compared with the preset threshold. If it is higher than the threshold, an alarm message is output.

[0055] 9. Running monitoring: It can automatically identify people in the monitoring area. If it detects that a worker is running, it can issue an alarm in time to correct and remind the worker. This algorithm first uses the target detection algorithm to detect the head, locate the head in the picture, and obtain the detection frame containing the head. Then, based on this detection frame, it uses the running classification algorithm to classify the content of the detection frame. The running confidence output by the classification is compared with the preset threshold. If it is higher than the threshold, an alarm message is output.

[0056] 10. Sleeping on the job detection: It can automatically identify people in the monitoring area. If a person is detected sleeping on the table, an alarm can be issued in time to correct and remind the operator. This algorithm solution first uses the human target detection algorithm to locate the human body in the image, classify the background and the human body, and obtain the detection frame containing the human body. Then, it uses the configured monitoring area and working hours to make regional judgments to determine whether there are people in the area. When no one is on duty during working hours, an alarm is displayed. When there are people in the area, a classification algorithm is used to classify the status. The results output normal status and sleeping status. When there are sleeping personnel, an alarm message is output.

[0057] The overall construction plan of the AI ​​video violation capture monitoring system is based on the intelligent reasoning platform. It obtains data from the data collection terminal, transmits it to the data collection control center through the network cable, and then pushes the data to the intelligent reasoning platform through the network cable. It is combined with artificial intelligence software and hardware processing, and finally responds to the alarm based on the analysis results of the data, thus forming an integrated intelligent AI video violation capture monitoring system.

[0058] The specific workflow of the AI ​​video violation capture monitoring system includes: 1) Video acquisition: Video images of the monitoring scene are acquired in real time through high-definition cameras.

[0059] 2) Video processing: Pre-process the collected video images, such as denoising and enhancement, to improve the accuracy of subsequent analysis.

[0060] 3) Target detection and recognition: Use AI algorithms to detect and identify targets in video images.

[0061] 4) Behavior analysis: Based on the results of target detection and recognition, further analyze the target's behavior pattern, such as whether there are abnormal behaviors (such as smoking, running, sleeping on duty, etc.).

[0062] 5) Early warning and alarm: When abnormal behavior is detected, the system triggers the early warning mechanism to alert the monitoring center personnel through pop-up screens, sounds, etc. At the same time, the system can generate alarm information and send it to relevant personnel for processing.

[0063] 6) Data storage and management: Store video images, analysis results, alarm information and other data on a server or cloud storage platform for subsequent query and analysis.

[0064] 7) System configuration and maintenance: Users can set camera parameters, algorithm models, etc. through the system configuration interface. System maintenance personnel can regularly maintain and upgrade the system to ensure stable operation of the system.

[0065] The AI ​​video violation capture monitoring system is executing the capture function. The implementation method is as follows: 1) Video stream access: The AI ​​video surveillance system first captures the real-time video stream of the monitored area through the camera. These cameras usually use HD or UHD technology to ensure the clarity and details of the image. The video stream is connected to the system's video access module through a specific video transmission protocol.

[0066] 2) Snapshot trigger mechanism: The system can set up a variety of snapshot trigger mechanisms, such as motion detection, face detection, specific behavior detection, etc. When the camera captures a picture that meets the trigger conditions, the system will immediately start the snapshot function.

[0067] 3) Captured image processing: The captured images will go through the image preprocessing module, including denoising, enhancement, cropping and other steps to improve the image quality. The processed images will be saved to the designated storage location for subsequent analysis and use.

[0068] The recognition function of the AI ​​video violation capture monitoring system is implemented as follows: 1) Object detection: The system uses deep learning algorithms to detect objects in captured images and identify targets in the images. Object detection modules are usually based on deep learning models such as convolutional neural networks, which improve recognition accuracy by training large amounts of image data.

[0069] 2) Feature extraction: For the identified objects, the system will further extract their feature information, such as the wearing of protective equipment on various parts of the body, the body movements of the person, etc. These feature information will be used in the subsequent reasoning and recognition process.

[0070] The inference function of the AI ​​video violation capture monitoring system is implemented as follows: 1) Behavior analysis: The system will perform behavioral analysis on the identified objects, such as determining whether pedestrians have abnormal behaviors such as smoking, running, sleeping on duty, etc. The behavior analysis module is usually based on a behavior recognition algorithm, which identifies abnormal behaviors by training behavior patterns in video data.

[0071] 2) Contextual understanding: The system also makes inferences based on the contextual information of the monitored area, which helps to more accurately identify and understand complex scenes and events in the monitored images.

[0072] The intelligent reasoning platform processing method of the AI ​​video violation capture monitoring system is: 1) Data processing flow: The video data is first transmitted to the data processing module of the intelligent reasoning platform. The data processing module decodes, preprocesses and extracts features of the video data to provide basic data for subsequent reasoning analysis.

[0073] 2) Application of reasoning algorithms: The intelligent reasoning platform applies deep learning algorithms to perform reasoning analysis on processed video data. These algorithms include object detection, behavior recognition, scene understanding, etc., which can achieve comprehensive analysis and understanding of the surveillance images.

[0074] 3) Alarm information generation: When the reasoning analysis module detects abnormal behavior or events, it will immediately generate alarm information, which includes the type of abnormal behavior, location and time of occurrence, as well as related video clips or image evidence.

[0075] 4) Alarm information push: When the system detects typical unsafe behaviors of personnel in the monitoring area, the system will promptly send warning signals to the monitoring center and the parties involved, and promptly remind the on-site personnel. This helps to respond to and handle abnormal events in a timely manner and improve monitoring efficiency.

[0076] The method of associating and connecting the AI ​​video violation capture monitoring system with the personnel positioning system is as follows: In the user interface of the external personnel positioning alarm information system, a new [Alarm Management-AI Alarm] page is added to display the acquired AI alarm information. After receiving the AI ​​alarm information, a pop-up window prompts, click the [View] button to jump to the [Alarm Management-AI Alarm] page to filter out this alarm information, and you can view the specific alarm information in the Home Page-Daily Alarm Statistics.

[0077] The hardware and software integration of the AI ​​video violation capture monitoring system is as follows: The AI ​​video violation capture monitoring system usually includes hardware devices and software components such as cameras, video access modules, and intelligent reasoning platforms. These devices and components need to be tightly integrated and work together to ensure the stability and performance of the system.

[0078] The specific method of algorithm optimization and update of AI video violation capture monitoring system is as follows: Deep learning algorithm is the core of AI video violation capture monitoring system. In order to improve recognition accuracy and reasoning efficiency, it is necessary to continuously optimize and update the algorithm model. This includes adjusting model parameters, increasing training data, introducing new algorithm architecture, etc.

[0079] The data security and privacy protection method of the AI ​​video violation capture monitoring system is as follows: If the AI ​​video violation capture monitoring system involves a large amount of video data and sensitive information, strict data security measures and privacy protection mechanisms need to be adopted, including data encryption, access control, data backup and recovery, etc.

[0080] Embodiment 2: This example involves the implementation process of the AI ​​video violation capture monitoring function in the converter area and casting area of ​​the melting and casting line: 1. AI video recognition and illegal capture function in converter area: 1) When the converter starts, the operator presses the converter discharge alarm button, and the converter start signal is given to the AI ​​monitoring video server. After receiving the signal, the AI ​​server starts to execute the illegal capture function.

[0081] 2) During the entire converter process, operators must wear protective face shields and foot protectors in addition to regular labor protection supplies. When AI recognizes that an operator is not wearing a safety helmet, protective face shield or foot protectors, it can promptly sound an alarm to correct and remind the operator.

[0082] 3) When the AI ​​system identifies that a person who is not on the whitelist of converter operators is performing converter operations, it can promptly issue an alarm to correct and remind the operator.

[0083] 4) When the AI ​​system identifies that the converter operator is off-duty, it can promptly issue an alarm to correct and remind the operator.

[0084] 5) When the converter is finished, the operator turns off the converter discharge alarm, and the converter end signal is given to the AI ​​monitoring video server. After receiving the signal, the AI ​​server executes the shutdown of the AI ​​recognition function.

[0085] 2. AI video recognition and illegal capture function in the foundry area: 1) When the operator presses the "Casting Preparation" button on the operation interface, the casting start signal is given to the AI ​​monitoring video server. After receiving the signal, the AI ​​server starts to execute the illegal capture function.

[0086] 2) During the entire casting process, workers entering the red safety warning area of ​​the casting must wear a full set of labor protection equipment, including: foot protection, protective face shields with front and back neck guards, and protective clothing. When AI recognizes that a worker is not wearing one of the labor protection equipment, it can promptly sound an alarm to correct and remind the worker.

[0087] 3) When the AI ​​system recognizes that the total number of people in the corresponding control area of ​​the foundry exceeds the set value, such as when the number of people in the first-level control area exceeds the set value of 8, the system will immediately issue an automatic alarm to remind personnel to evacuate to avoid overcrowding in the area.

[0088] 4) When the AI ​​system recognizes that there is less than one person in the main operation room of the foundry and that someone is off duty, the system will immediately issue an automatic alarm to correct and remind the operator.

[0089] 5) When the AI ​​system identifies that there are people sleeping on duty in the main operation room of the foundry, it can issue an alarm in time to correct and remind the operators.

[0090] 6) When the casting is finished, execute the shutdown of AI recognition function.

[0091] The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made according to the scope of the patent application of the present invention should fall within the scope of the present invention.

Claims

1. An AI video illegal capture monitoring method for aluminum processing deep well casting production line, characterized by: The monitoring method constructs an intelligent production safety early warning model based on computer vision deep learning technology, uses an embedded device as a data acquisition terminal, and the action of the data acquisition terminal is associated with the console operation of the aluminum processing deep well casting production line. When the console starts to execute a dangerous process, the data acquisition terminal is triggered. The data acquisition terminal continuously collects data from the monitoring screen of the front-end monitoring device and sends the collected data to the back-end AI monitoring video server. The AI ​​system of the AI ​​monitoring video server performs data analysis through the intelligent production safety early warning model, performs AI recognition with an AI analysis module, and responds to alarms based on the analysis results.

2. The AI ​​video illegal capture monitoring method for aluminum processing deep well casting production line according to claim 1 is characterized by: The monitoring method includes an AI video recognition and violation capture method for the converter area of ​​an aluminum processing deep well casting production line, and also includes an AI video recognition and violation capture method for the casting area of ​​an aluminum processing deep well casting production line; the data acquisition terminal collects data from the monitoring screen of the front-end monitoring equipment of the aluminum processing deep well casting production line through a camera.

3. The AI ​​video illegal capture monitoring method for aluminum processing deep well casting production line according to claim 2 is characterized by: The AI ​​video recognition and illegal capture method for the converter area specifically includes the following steps: Step A1, when the converter operation starts, the operator presses the converter discharge alarm button on the console, and the converter start signal is given to the AI ​​monitoring video server. After receiving the signal, the AI ​​monitoring video server starts data collection at the data collection terminal and starts to execute the illegal capture function; Step A2: During the entire converter operation process, when AI identifies that the operator at the converter process is not wearing a safety helmet, a protective face shield, or foot protection, an alarm sounds to remind the operator to correct the violation; Step A3: When the AI ​​system identifies that a person who is not on the converter operation whitelist is performing converter operation at the converter process, an alarm is issued to remind the operator to correct the problem; Step A4: When the AI ​​system identifies that the converter operator is off-duty at the converter process, an alarm is issued to remind the operator to correct the problem; Step A5, when the converter operation is completed, the operator turns off the converter discharge alarm through the console, and the console simultaneously sends the converter end signal to the AI ​​monitoring video server. After receiving the signal, the AI ​​monitoring video server executes the shutdown of the AI ​​recognition function.

4. The AI ​​video illegal capture monitoring method for aluminum processing deep well casting production line according to claim 2 is characterized by: The AI ​​video recognition and illegal capture method in the foundry area specifically includes the following steps: Step B1: When the operator presses the "casting preparation" button on the operation interface of the console, the casting start signal is sent to the AI ​​monitoring video server. After receiving the signal, the AI ​​monitoring video server starts the data collection of the data collection terminal and starts to execute the function of identifying illegal snapshots; Step B2: During the entire casting operation, when AI identifies that the operator entering the casting safety warning area is not wearing any of the labor protection supplies such as foot protection, protective face shield with front and back neck gaiter, or protective clothing, an alarm will be sounded in time to remind the operator to correct the problem; Step B3: When the AI ​​system recognizes that the total number of people in the corresponding control area of ​​the foundry exceeds the set value, the system will immediately and automatically issue an alarm to remind personnel to evacuate to avoid overcrowding in the area; Step B4: When the AI ​​system recognizes that there is less than one person in the main operation room of the foundry and that a person is off duty, the system will immediately and automatically give an alarm to remind the operator to make corrections; Step B5: When the AI ​​system identifies that there are people sleeping on duty in the main operation room of the foundry, it can promptly issue an alarm to correct and remind the operator; Step B6: When the casting is finished, when the operator sends the relevant end instruction of the aluminum processing deep well casting production line on the console, the console synchronously sends a signal to the AI ​​monitoring video server. After receiving the signal, the AI ​​monitoring video server executes to shut down the AI ​​recognition function.

5. The AI ​​video illegal capture monitoring method for aluminum processing deep well casting production line according to claim 4 is characterized by: In step B3, an alarm is issued when the AI ​​system recognizes that the total number of people in the corresponding control area of ​​the foundry exceeds 8.

6. The AI ​​video illegal capture monitoring method for aluminum processing deep well casting production line according to claim 2 is characterized by: The intelligent production safety warning model includes an AI analysis module, which uses a convolutional neural network (CNN) structure and combines the target detection algorithm YOLOv5 to perform target detection and recognition; YOLOv5's efficient real-time detection performance is used to process large-scale video streams of the front-end monitoring equipment of the aluminum processing deep-well casting production line, and identify various types of objects and behavioral features in the video streams.

7. The AI ​​video illegal capture monitoring method for aluminum processing deep well casting production line according to claim 6 is characterized by: The intelligent production safety warning model is trained with a large number of video data sets of violation samples of aluminum processing deep well casting production lines, including target objects and violations in a variety of different scenarios. The video data sets are data-annotated and data-enhanced video data sets that cover a variety of samples under different lighting conditions and weather conditions. When annotating the video dataset, use annotation tools to annotate each image in the dataset, record the type of the target object, and annotate its location in the aluminum processing deep well casting production line with a bounding box; in addition, annotate the type of specific violations based on the characteristics of the violations; When performing data enhancement on the video dataset, the specific methods are as follows: rotation and scaling are used to enhance the recognition ability of the intelligent production safety warning model at different perspectives; random cropping and flipping are used to simulate target changes in the real scene of the aluminum processing deep well casting production line; and lighting adjustment is used to simulate different lighting conditions to improve the performance of the intelligent production safety warning model at night or in backlight conditions.

8. The AI ​​video illegal capture monitoring method for aluminum processing deep well casting production line according to claim 7 is characterized by: When training the intelligent production safety warning model, the YOLOv5 model, which has been pre-trained on the COCO public dataset and has good basic recognition capabilities, is used as the basis for fine-tuning on the target dataset to improve the recognition accuracy in specific scenarios; Image enhancement techniques including rotation, scaling, and color adjustment are used during the training process to improve the robustness of the model. Video datasets containing a large number of violation samples of aluminum processing deep well casting production lines are used to fine-tune the trained model to improve the model's accuracy in identifying specific violations of aluminum processing deep well casting production lines. In terms of hyperparameter setting for model training, the following method is adopted: When setting the learning rate, set the initial learning rate to 0.001, and use adaptive learning rate adjustment Reduce on Plateau during training. When the performance of the model on the validation set no longer improves, the learning rate is automatically reduced. When setting the batch size, set the batch size to 16 to balance the memory usage of the GPU device used for training and the training speed; When setting the optimizer, use the Adam optimizer combined with L2 regularization to prevent the model from overfitting; When setting the loss function, a multi-task loss function is used for YOLOv5, including classification loss, bounding box regression loss, and confidence loss, to ensure that the accuracy of the bounding box position is optimized while detecting.

9. The AI ​​video illegal capture monitoring method for aluminum processing deep well casting production line according to claim 8 is characterized by: The specific steps for training the intelligent production safety warning model are: Step S1: The model is trained on the training set for a total of 300 epochs, and the validation set is used to evaluate the model performance after each epoch; Step S2: Introduce the Early Stopping strategy, which stops training when the accuracy on the validation set does not improve for 20 consecutive rounds to avoid overfitting; Step S3: After training is completed, save the model parameters and perform performance evaluation to determine whether further optimization or adjustment of the model structure is needed.

10. The AI ​​video illegal capture monitoring method for aluminum processing deep well casting production line according to claim 9 is characterized in that: In order to improve the AI ​​detection effect, the monitoring method adopts the following optimization strategies when training the intelligent production safety warning model: Anchor Box optimization: The K-means clustering algorithm is used to analyze the object sizes in the dataset and reset the anchor boxes based on the size distribution to ensure the accuracy of the model when detecting objects of different sizes. Model compression and quantization: After model training is completed, the model is quantized so that it can maintain a high detection speed when running on resource-constrained devices; At the same time, pruning technology is used to remove redundant neurons to further reduce the model size and improve computing efficiency.

Images