An industrial gas cylinder filling safety management system and method based on AI video

Abstract

The present invention discloses an industrial gas cylinder filling safety management system and method based on AI video, which belongs to the field of artificial intelligence technology. The present invention scans the QR code on the industrial gas cylinder, obtains the inspection information of the current cylinder, automatically accumulates the number of cylinders to be filled, generates the number of cylinders to be filled, and the number report; performs AI video analysis, detects and identifies the cylinders to be filled in the video data; compares the information of the cylinders to be filled obtained by AI video analysis with the inspection information; performs real-time detection and analysis of the video data during the filling process, and automatically calculates the total amount of gas currently required according to the number of cylinders to be filled; calculates the working time of the cryogenic pump according to the preset filling digital analysis model, combined with the total amount of gas currently required, starts the cryogenic pump, and performs the filling operation according to the calculated working time; automatically monitors and issues an alarm and stops the pump when abnormal.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and specifically to an industrial gas cylinder filling safety management system and method based on AI video. Background Art

[0002] In the industrial field, the safety management of gas cylinders, especially the safety during the filling process, has always been a critical issue. If industrial gas cylinders are not properly managed during transportation, storage and use, it is easy to cause safety hazards and even serious accidents such as explosions and leaks. In order to ensure the safe use of gas cylinders, it is usually required to conduct regular inspections of gas cylinders and strictly control the gas filling volume during the filling process.

[0003] In the existing gas cylinder filling management system, many links still rely on manual operation, such as checking the inspection information of the cylinder and monitoring the filling process. Manual operation is prone to negligence and errors, especially when there are many cylinders and the filling operation is frequent. These errors may lead to overfilling of gas or failure to fill as required, posing a safety hazard. The traditional system has limited real-time monitoring capabilities for the filling process of gas cylinders. In most cases, it only relies on simple sensors to detect the pressure and liquid level of the cylinders, and cannot fully monitor the dynamic changes in the filling process. This monitoring method lacks accuracy and it is difficult to detect abnormal conditions in the filling process in time, resulting in an increased risk of accidents. Most traditional filling systems rely on simple timing or quantitative control, and cannot be adjusted intelligently according to actual filling needs, which is prone to insufficient or excessive filling. Especially when filling multiple cylinders, it is difficult for traditional systems to accurately calculate the actual gas demand for each cylinder, resulting in low filling efficiency. Summary of the invention

[0004] The purpose of the present invention is to provide an industrial gas cylinder filling safety management system and method based on AI video to solve the problems raised in the above background technology.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0006] An industrial gas cylinder filling safety management method based on AI video, the method comprising the following steps:

[0007] S100, using a camera installed in the filling area to scan the QR code on the industrial gas cylinder, obtain the inspection information of the current cylinder, automatically accumulate the number of cylinders to be filled, determine whether the current cylinder is within the valid inspection period, and generate the number of cylinders to be filled; after all QR codes are scanned, generate a quantity report;

[0008] S200 uses the camera installed in the filling area to obtain the image and video data of the cylinders to be filled in real time, conducts AI video analysis, and detects and identifies the cylinders to be filled in the video data through convolutional neural network and YOLO algorithm; compares the information of the cylinders to be filled obtained through AI video analysis with the inspection information of the corresponding cylinders obtained by scanning the code; if the information is inconsistent, an alarm is issued and the operator is prompted to check and confirm;

[0009] S300 uses the YOLO algorithm and spatiotemporal convolutional neural network to perform real-time detection and analysis of video data during the filling process to ensure that the filling process of each gas cylinder meets safety standards; after receiving the filling instruction, the filling control pump relay automatically calculates the total amount of gas currently required based on the number of cylinders to be filled;

[0010] S400. Calculate the working time of the cryogenic pump based on the preset digital analysis model for filling and the total amount of gas currently required, start the cryogenic pump, and perform filling operations according to the calculated working time; when the filling reaches the preset time or it is detected that the gas filling amount exceeds the set safety range, automatically sound an alarm and stop the pump, and record the overfilling warning data into the database.

[0011] According to step S100, each cylinder is affixed with a unique QR code label, and the QR code label contains the inspection information of the current cylinder, and the inspection information includes the inspection record of the cylinder, the cylinder number, the production date, the last inspection date, the gas type and the maximum filling amount; the QR code on the industrial gas cylinder is scanned by a camera installed in the filling area, and the inspection information of the cylinder associated with the QR code is automatically extracted and stored in the database;

[0012] Verify the inspection information to determine whether the current cylinder is within the valid inspection period and record the inspection status; if the inspection information of a cylinder is unqualified or has exceeded the inspection validity period, a warning will be issued to remind the operator not to refill the cylinder; after each scan and acquisition of the inspection information, the qualified cylinders will be automatically counted into the total number of cylinders to be filled, and the number will be accumulated continuously to generate the current number of cylinders to be filled, and displayed on the operation interface for the operator to view in real time;

[0013] After the QR codes of all cylinders to be filled are scanned, a report on the quantity of cylinders to be filled is generated, which includes the inspection information and inspection status of each cylinder, as well as the accumulated total quantity.

[0014] According to step S200, the camera collects dynamic video data in the filling area in real time, and performs image preprocessing on the original video data transmitted by the camera, wherein the preprocessing includes denoising, image enhancement and white balance adjustment; for the preprocessed image, a convolutional neural network is used to perform target detection of the steel cylinder, and different features are detected by sliding a number of filters in the image, and the features are combined into a higher-level representation to form the outline of the steel cylinder;

[0015] After the cylinder detection is completed, the YOLO algorithm based on deep learning is used to further identify the specific information of the cylinder, including the shape, color and special mark of the cylinder. The outline of each cylinder is calculated based on the outline of the cylinder. The cylinder information in the video data is analyzed frame by frame, and each detected cylinder is uniquely identified, and the information of the cylinder to be filled detected in all frames is summarized.

[0016] The preprocessed image is input to the first layer of the convolutional neural network, which is a convolutional layer that slides over the image through a number of filters, scanning each area of ​​the image one by one. Each filter focuses on detecting a specific type of feature, including the edges, corners, and textures of the cylinder. As the filter slides across the image, it calculates each area and generates a feature map that represents the probability and location of the edge of the cylinder.

[0017] The convolutional neural network consists of several convolutional layers, each of which further extracts higher-level features based on the feature map of the previous layer to form the overall outline of the cylinder;

[0018] A pooling layer is applied after the convolution layer. The pooling layer downsamples the feature map to reduce the size of the feature map and retain the most important feature information in the overall contour of the cylinder. After several convolution and pooling layers, the convolutional neural network finally outputs a high-level feature representation. The high-level feature representation is composed of low-level features layer by layer to form the complete contour of the cylinder. The combined features can describe the shape, size and position of the cylinder.

[0019] The last several layers of the convolutional neural network are fully connected layers, which are responsible for mapping high-level features to cylinders. By outputting a probability distribution, they determine whether there are cylinders in the image and determine the location of the cylinders. For each detected cylinder, a bounding box is generated to identify the specific location of the cylinder in the image and form the outline of the cylinder.

[0020] The YOLO algorithm receives a preprocessed image, which is cut into several grids, each of which is responsible for detecting the object features in its area; the YOLO model takes the entire image as input and detects several objects at the same time in a forward pass; the YOLO algorithm processes all grids in the image in parallel, and each grid generates a set of predictions, including the object category, the coordinates of the bounding box, the width and height of the box, and the confidence of each category, and the object category is set to cylinders;

[0021] The convolutional layer of the YOLO model extracts features from the image, including the edge, color distribution, and shape of the cylinder. The YOLO model uses these features to determine the specific information of the cylinder.

[0022] The YOLO model identifies the shape of the cylinder by analyzing the features within each grid, and identifies the shape of the cylinder based on edge detection and shape features; based on the shape features, the YOLO algorithm calculates the contour box of the cylinder, which includes the position coordinates and size of the cylinder in the image, and is used to identify the specific position of the cylinder; the YOLO algorithm extracts the color features of the cylinder, and identifies the main color and color mark of the cylinder by analyzing the color distribution within each grid. According to the color features, the model further subdivides the type of cylinder; different types of gas cylinders have different color logos, and the YOLO algorithm can automatically classify and identify different types of cylinders based on color features; special logos are printed on the cylinders, including QR codes, barcodes, labels and text, and the YOLO algorithm detects and identifies them by analyzing the texture and pattern features of each grid; parsing is performed to combine the relevant information with other identification information of the cylinder.

[0023] The YOLO algorithm processes each frame in the video data, identifying and marking the specific information of each cylinder; in each frame, the model will generate the outline box and related information of the cylinder.

[0024] According to step S300, the YOLO algorithm is used to detect key objects in the video frame in real time, the key objects include gas cylinders, filling equipment and operators, identify the key objects in each frame, and mark the bounding box; YOLO detects the position of the object and classifies it, distinguishes different types of gas cylinders, identifies the state of the filling connector, and distinguishes the posture and action of the operator, the state of the filling connector includes connected and unconnected;

[0025] Use a spatiotemporal convolutional neural network to analyze the behavior of detected objects in consecutive frames, analyze spatial information, and process changes in the time dimension. The spatial information represents the position of the object in each frame, and the changes in the time dimension represent the movement of the object between different frames, thereby identifying the specific actions of the operator; the spatiotemporal convolutional neural network is used to monitor the operating status of the filling equipment, and by analyzing the dynamic connection status of the filling connector, the filling progress of the gas cylinder, the liquid level and the pressure gauge readings, it is determined whether there are abnormal conditions or potential safety hazards;

[0026] The detection results of the YOLO algorithm and the motion recognition results of the spatiotemporal convolutional neural network are fused to determine whether the operator's actions are correct based on motion recognition, and whether the status of the equipment is consistent with expectations based on status monitoring; if the operating steps are not standardized or the equipment status is abnormal, an alarm is triggered.

[0027] The spatial information in the spatiotemporal convolutional neural network refers to the image content of each frame in the video, and the change in the time dimension refers to the movement and change of the object in several consecutive frames; a series of consecutive frames are extracted from the real-time video data, and a series of consecutive frames include the operator's actions and the status of the filling equipment; the consecutive frames are organized into a frame sequence in chronological order, and the frame sequence will be used as the input of the spatiotemporal convolutional neural network; in each frame, the spatiotemporal convolutional neural network extracts spatial features through the convolution layer, and the spatial features include the operator's posture, the position of the filling joint and the status of the gas cylinder; the specific position of each object in the frame is identified, and the corresponding feature map is generated, and the feature map is used to describe the distribution of the object in space;

[0028] The spatiotemporal convolutional neural network analyzes the changes of objects in consecutive frames by applying convolution operations in the time dimension, including the operator's hand movements, the connection and separation of the filling connector, the liquid level rise and pressure changes of the gas cylinder; through the convolution in the time dimension, the network can capture the movement pattern of the object between different frames and the dynamic connection state of the connector. The movement pattern represents the operator's action coherence;

[0029] Combining spatial and temporal features, the spatiotemporal convolutional neural network can identify the operator's specific actions, detect whether the operator has correctly connected the filling connector, whether the standardized operating procedures are followed, and whether there are potential misoperations; the network analyzes the operating status of the filling equipment, and determines whether the connector is firmly connected by monitoring the dynamic changes of the connector; by analyzing the changes in the liquid level and pressure gauge, it evaluates the progress of filling and whether there are any abnormal conditions; if the network detects that the operator's actions do not conform to the expected pattern, or the status of the filling equipment is abnormal, the recognition result will be marked as abnormal and an alarm will be triggered.

[0030] Assume that the input video data is a four-dimensional tensor X with dimensions , where T represents the number of time frames, H represents the height of the image, W represents the width of the image, and C represents the number of channels; for the convolution operation in the time dimension, a convolution kernel K is used with a dimension of ,in, represents the convolution kernel size in the time dimension, and Represent the convolution kernel size on the height and width of the image in the spatial dimension respectively;

[0031] The calculation formula for the output tensor Y of the time dimension convolution operation is as follows:

[0032] ;

[0033] Among them, t, h and w represent the index of the output tensor Y in the time, image height and image width dimensions respectively; i, j and k represent the offset of the convolution kernel in the time, image height and image width dimensions respectively; Represents the input channel index;

[0034] Through the convolution operation in the time dimension, the network can identify the operator's hand movement trajectory between several frames and determine whether the movement is coherent and standardized. The coherence of the hand movement is captured by the change of the hand position in adjacent frames. By analyzing the position and state changes of the filling connector in different frames, the network can detect whether the connector remains correctly connected or loose during the filling process.

[0035] For the capture of the operator’s hand movement pattern, the hand is moved at different time frames. The position on , then the motion pattern calculated by the network through the time dimension convolution for:

[0036] ;

[0037] For capturing the dynamic connection status of the filling joint, the joint is displayed at different time frames. The position and status on , then the network calculates the connection state mode through time dimension convolution for:

[0038] ;

[0039] in, is an indicator variable, indicating the connection status of the joint at time t, 1 for connected and 0 for unconnected;

[0040] Set the sport mode and connection state mode Used for anomaly detection, if the hand movement is incoherent or the joint connection status is abnormal in consecutive frames, an alarm is triggered.

[0041] According to step S400, the filling digital analysis model is established based on historical data, gas cylinder type, gas properties, environmental conditions and performance parameters of the cryogenic pump, and can calculate the optimal working time of the cryogenic pump according to different types of gas cylinders and the required amount of gas to be filled, ensuring that the gas cylinders can be filled to the required amount of gas within a safe range; the input data of the model includes the current number of cylinders to be filled, the volume of each cylinder, the target filling amount, and the temperature and pressure conditions of the gas; the total gas demand of all cylinders is calculated based on the current number of cylinders to be filled and the target filling amount of each cylinder; during the calculation process, dynamic adjustments are made according to real-time environmental conditions and the status of the pump;

[0042] When building a digital analysis model for filling, different types of gas cylinders have different volumes, materials, design pressures and other characteristics, which are classified and recorded. The physical and chemical properties of different gases, such as density, expansion coefficient, boiling point, etc., have a great influence on the filling process, so it is necessary to accurately record the type of gas being filled and its key properties. External conditions such as ambient temperature, humidity, and atmospheric pressure will also affect the filling process, and these parameters are monitored and recorded in real time. Parameters such as the flow rate, working efficiency, and maximum working time of the cryogenic pump are also included in the input data of the model.

[0043] The input features of the model include the number of cylinders to be filled N, the volume of each cylinder V_i, the target filling volume Q_i, the temperature T_g and pressure P_g conditions of the gas, and the real-time environmental conditions T_e, H_e, P_e and the state parameters of the cryogenic pump, such as flow rate, power, etc. The output of the model is the optimal working time T_p of the cryogenic pump to ensure that each cylinder can be safely filled with the required amount of gas.

[0044] The relationship between input features and output variables is modeled using multivariate regression analysis. Through regression analysis, a preliminary mathematical model can be obtained to describe the impact of each input feature on the working time of the cryopump. Based on the preliminary regression model, the random forest algorithm is used for optimization to capture the complex nonlinear relationship between input features and output variables, thereby improving the prediction accuracy of the model. A dynamic adjustment mechanism is introduced into the model so that it can be adjusted according to real-time environmental conditions and the status of the pump. For example, when the ambient temperature is high, the model will extend the working time of the cryopump to compensate for the expansion effect of the gas.

[0045] The model is trained using the collected historical data. During the training process, the model automatically adjusts its parameters to minimize the prediction error. The model is validated using an independent dataset that was not involved in the training to ensure that it can accurately predict the optimal working time of the cryopump under different conditions. Based on the validation results, the model's parameter settings are further optimized to ensure that it has good generalization capabilities.

[0046] After calculating the working time, the cryogenic pump is automatically started to fill the cylinder with gas. The filling process enters the fully automatic control stage, and the working status of the pump and the filling progress are monitored in real time. During the filling process, the progress of the filling is monitored in real time through sensors and cameras, and the changes in liquid level and pressure are detected. If any abnormality is detected, the operating parameters will be adjusted and an alarm will be issued.

[0047] When the cryogenic pump reaches the preset filling time, the pump will automatically stop working to ensure that the filling volume does not exceed the set requirements; an overfilling detection mechanism is set up. If it is detected during the filling process that the gas filling volume exceeds the set safety range, an alarm will be issued and the cryogenic pump will automatically stop working.

[0048] An AI video-based industrial gas cylinder filling safety management system, including:

[0049] QR code scanning and quantity report generation module: including: QR code scanning unit, inspection information verification unit, quantity statistics unit and quantity report generation unit; wherein, the QR code scanning unit uses a camera installed in the filling area to scan the QR code on the industrial gas cylinder, the inspection information verification unit is responsible for obtaining the inspection information of the current cylinder, the quantity statistics unit is responsible for automatically accumulating the number of cylinders to be filled, judging whether the current cylinder is within the valid inspection period, and generating the number of cylinders to be filled, and the quantity report generation unit generates a quantity report after all QR codes are scanned;

[0050] AI video analysis and cylinder information comparison module: including: image and video data acquisition unit, cylinder detection and identification unit and information comparison unit; wherein, the image and video data acquisition unit uses the camera installed in the filling area to obtain the image and video data of the cylinder to be filled in real time, the cylinder detection and identification unit is responsible for AI video analysis, and detects and identifies the cylinder to be filled in the video data through convolutional neural network and YOLO algorithm, and the information comparison unit is responsible for comparing the information of the cylinder to be filled obtained through AI video analysis with the inspection information of the corresponding cylinder obtained by scanning the code. If the information is inconsistent, an alarm is issued and the operator is prompted to check and confirm;

[0051] The module for real-time detection and pump control calculation during the filling process includes: a real-time detection and analysis unit and a gas total amount calculation unit. The real-time detection and analysis unit uses the YOLO algorithm and spatiotemporal convolutional neural network to perform real-time detection and analysis on the video data during the filling process to ensure that the filling process of each gas cylinder meets safety standards. The gas total amount calculation unit is responsible for automatically calculating the current required gas total amount according to the number of cylinders to be filled after the filling control pump relay receives the filling instruction.

[0052] Filling control and safety monitoring module: including: filling time calculation unit, cryogenic pump control unit and safety monitoring unit; wherein, the filling time calculation unit is responsible for calculating the working time of the cryogenic pump according to the preset filling digital analysis model and the total amount of gas currently required; the cryogenic pump control unit is responsible for starting the cryogenic pump and performing filling operations according to the calculated working time; when the filling reaches the preset time or detects that the gas filling amount exceeds the set safety range, the safety monitoring unit automatically sounds an alarm and stops the pump, and records the overfilling warning data into the database.

[0053] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0054] The present invention uses AI video analysis technology to achieve highly automated and intelligent operation in the process of gas cylinder filling management. The camera scans the QR code to obtain the cylinder information, and uses AI technology to detect and analyze the video data of the filling process in real time, which greatly reduces the error rate of manual operation and improves the accuracy and safety of operation.

[0055] The present invention uses convolutional neural networks, YOLO algorithms and spatiotemporal convolutional neural networks to monitor the dynamic changes of gas cylinders in real time during the filling process, and comprehensively capture key parameters such as liquid level, pressure, and filling connector connection status. Through real-time analysis of these parameters, the system can promptly detect potential safety hazards and take immediate measures, such as sounding an alarm or stopping the filling operation, to ensure the safety of the entire filling process.

[0056] The present invention automatically compares the cylinder information obtained by AI video analysis with the inspection information scanned by the QR code, greatly improving the efficiency and accuracy of information comparison and reducing the errors that may be caused by manual comparison. The system can immediately issue an alarm when it identifies information mismatches, prompting the operator to check and confirm, avoiding safety hazards. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0058] Figure 1It is a schematic diagram of the steps of an industrial gas cylinder filling safety management method based on AI video of the present invention;

[0059] Figure 2 This is a system structure diagram of an industrial gas cylinder filling safety management system based on AI video in the present invention. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] See also Figure 1 and Figure 2 , the present invention provides a technical solution:

[0062] According to one embodiment of the present invention, Figure 1 A step diagram of an industrial gas cylinder filling safety management method based on AI video is shown in the figure. A method for industrial gas cylinder filling safety management based on AI video includes the following steps:

[0063] S100, using a camera installed in the filling area to scan the QR code on the industrial gas cylinder, obtain the inspection information of the current cylinder, automatically accumulate the number of cylinders to be filled, determine whether the current cylinder is within the valid inspection period, and generate the number of cylinders to be filled; after all QR codes are scanned, generate a quantity report;

[0064] S200 uses the camera installed in the filling area to obtain the image and video data of the cylinders to be filled in real time, conducts AI video analysis, and detects and identifies the cylinders to be filled in the video data through the convolutional neural network and YOLO algorithm; compares the information of the cylinders to be filled obtained through AI video analysis with the inspection information of the corresponding cylinders obtained by scanning the code; if the information is inconsistent, an alarm is issued and the operator is prompted to check and confirm;

[0065] S300 uses the YOLO algorithm and spatiotemporal convolutional neural network to perform real-time detection and analysis of video data during the filling process to ensure that the filling process of each gas cylinder meets safety standards; after receiving the filling instruction, the filling control pump relay automatically calculates the total amount of gas currently required based on the number of cylinders to be filled;

[0066] S400. Calculate the working time of the cryogenic pump based on the preset digital analysis model for filling and the total amount of gas currently required, start the cryogenic pump, and perform filling operations according to the calculated working time; when the filling reaches the preset time or it is detected that the gas filling amount exceeds the set safety range, automatically sound an alarm and stop the pump, and record the overfilling warning data into the database.

[0067] According to step S100, each cylinder is affixed with a unique QR code label, and the QR code label contains the inspection information of the current cylinder, and the inspection information includes the inspection record of the cylinder, the cylinder number, the production date, the last inspection date, the gas type and the maximum filling amount; the QR code on the industrial gas cylinder is scanned by a camera installed in the filling area, and the inspection information of the cylinder associated with the QR code is automatically extracted and stored in the database;

[0068] Verify the inspection information to determine whether the current cylinder is within the valid inspection period and record the inspection status; if the inspection information of a cylinder is unqualified or has exceeded the inspection validity period, a warning will be issued to remind the operator not to refill the cylinder; after each scan and acquisition of the inspection information, the qualified cylinders will be automatically counted into the total number of cylinders to be filled, and the number will be accumulated continuously to generate the current number of cylinders to be filled, and displayed on the operation interface for the operator to view in real time;

[0069] After the QR codes of all cylinders to be filled are scanned, a report on the quantity of cylinders to be filled is generated, which includes the inspection information and inspection status of each cylinder, as well as the accumulated total quantity.

[0070] According to step S200, the camera collects dynamic video data in the filling area in real time, and performs image preprocessing on the original video data transmitted by the camera, wherein the preprocessing includes denoising, image enhancement and white balance adjustment; for the preprocessed image, a convolutional neural network is used to perform target detection of the steel cylinder, and different features are detected by sliding a number of filters in the image, and the features are combined into a higher-level representation to form the outline of the steel cylinder;

[0071] In this example, a video dataset containing various types of industrial gas cylinders is used. The dataset contains filling scenes under different environmental conditions, with a video frame resolution of 1920x1080 and a frame rate of 30fps. The types of cylinders include oxygen cylinders, nitrogen cylinders, argon cylinders, etc., and each cylinder has a different color and logo.

[0072] After the cylinder detection is completed, the YOLO algorithm based on deep learning is used to further identify the specific information of the cylinder, including the shape, color and special mark of the cylinder. The outline of each cylinder is calculated based on the outline of the cylinder. The cylinder information in the video data is analyzed frame by frame, and each detected cylinder is uniquely identified, and the information of the cylinder to be filled detected in all frames is summarized.

[0073] The preprocessed image is input to the first layer of the convolutional neural network, which is a convolutional layer that slides over the image through a number of filters, scanning each area of ​​the image one by one. Each filter focuses on detecting a specific type of feature, including the edges, corners, and textures of the cylinder. As the filter slides across the image, it calculates each area and generates a feature map that represents the probability and location of the edge of the cylinder.

[0074] The convolutional neural network consists of several convolutional layers, each of which further extracts higher-level features based on the feature map of the previous layer to form the overall outline of the cylinder;

[0075] A pooling layer is applied after the convolution layer. The pooling layer downsamples the feature map to reduce the size of the feature map and retain the most important feature information in the overall contour of the cylinder. After several convolution and pooling layers, the convolutional neural network finally outputs a high-level feature representation. The high-level feature representation is composed of low-level features layer by layer to form the complete contour of the cylinder. The combined features can describe the shape, size and position of the cylinder.

[0076] The last several layers of the convolutional neural network are fully connected layers, which are responsible for mapping high-level features to cylinders. By outputting a probability distribution, they determine whether there are cylinders in the image and determine the location of the cylinders. For each detected cylinder, a bounding box is generated to identify the specific location of the cylinder in the image and form the outline of the cylinder.

[0077] The YOLO algorithm receives a preprocessed image, which is cut into several grids, each of which is responsible for detecting the object features in its area; the YOLO model takes the entire image as input and detects several objects at the same time in a forward pass; the YOLO algorithm processes all grids in the image in parallel, and each grid generates a set of predictions, including the object category, the coordinates of the bounding box, the width and height of the box, and the confidence of each category, and the object category is set to cylinders;

[0078] The convolutional layer of the YOLO model extracts features from the image, including the edge, color distribution, and shape of the cylinder. The YOLO model uses these features to determine the specific information of the cylinder.

[0079] The YOLO model identifies the shape of the cylinder by analyzing the features within each grid, and identifies the shape of the cylinder based on edge detection and shape features; based on the shape features, the YOLO algorithm calculates the contour box of the cylinder, which includes the position coordinates and size of the cylinder in the image, and is used to identify the specific position of the cylinder; the YOLO algorithm extracts the color features of the cylinder, and identifies the main color and color mark of the cylinder by analyzing the color distribution within each grid. According to the color features, the model further subdivides the type of cylinder; different types of gas cylinders have different color logos, and the YOLO algorithm can automatically classify and identify different types of cylinders based on color features; special logos are printed on the cylinders, including QR codes, barcodes, labels and text, and the YOLO algorithm detects and identifies them by analyzing the texture and pattern features of each grid; parsing is performed to combine the relevant information with other identification information of the cylinder.

[0080] The YOLO algorithm processes each frame in the video data, identifying and marking the specific information of each cylinder; in each frame, the model will generate the outline box and related information of the cylinder.

[0081] In the experimental data set, the convolutional neural network can successfully detect more than 95% of the cylinders. For scenes with complex environments, such as large changes in light or a lot of background interference, the detection accuracy drops slightly, but still remains above 90%. After using GPU acceleration, the processing time per frame is about 0.03 seconds, which can meet the needs of real-time processing. The YOLO algorithm can accurately identify the shape of the cylinder, and the shape recognition accuracy reaches 97%. In most cases, the algorithm can successfully distinguish different types of cylinders, such as long cylinders and short cylinders, and generate accurate contour boxes. The color recognition accuracy of the cylinder is 96%. The YOLO algorithm can accurately distinguish different types of cylinders based on color features, especially in identifying the main color of the cylinder and specific color logos. The recognition accuracy of special logos such as QR codes, barcodes, labels and text on the cylinder is 94%. The YOLO algorithm can effectively recognize these logos and combine them with other recognition information.

[0082] According to step S300, the YOLO algorithm is used to detect key objects in the video frame in real time, wherein the key objects include gas cylinders, filling equipment and operators, and the key objects are identified in each frame and the bounding boxes are marked; YOLO detects the position of the object and classifies it, distinguishes different types of gas cylinders, identifies the state of the filling connector, and identifies the posture and action of the operator, wherein the state of the filling connector includes connected and disconnected; in this embodiment, the experimental data set includes: 50 different types of gas cylinders, including oxygen cylinders, nitrogen cylinders and argon cylinders; 10 different operators perform filling operations, and the operation process includes connecting the filling connector, monitoring the liquid level and pressure gauge, and separating the connector after filling is completed; the filling equipment includes a liquid level sensor, a pressure gauge and a filling connector.

[0083] Use a spatiotemporal convolutional neural network to analyze the behavior of detected objects in consecutive frames, analyze spatial information, and process changes in the time dimension. The spatial information represents the position of the object in each frame, and the changes in the time dimension represent the movement of the object between different frames, thereby identifying the specific actions of the operator; the spatiotemporal convolutional neural network is used to monitor the operating status of the filling equipment, and by analyzing the dynamic connection status of the filling connector, the filling progress of the gas cylinder, the liquid level and the pressure gauge readings, it is determined whether there are abnormal conditions or potential safety hazards;

[0084] The detection results of the YOLO algorithm and the action recognition results of the spatiotemporal convolutional neural network are fused to determine whether the operator's actions are correct based on action recognition, and to verify whether the status of the equipment is consistent with expectations based on status monitoring; if the operation steps are not standardized or the equipment status is abnormal, an alarm is triggered. The accuracy of the YOLO algorithm in detecting gas cylinders, filling equipment, and operators is 96%. In highly complex scenes, such as light changes or background interference, the accuracy rate decreases, but it still remains above 93%. The processing time per frame is 0.02 seconds, which can meet real-time detection needs.

[0085] The spatial information in the spatiotemporal convolutional neural network refers to the image content of each frame in the video, and the change in the time dimension refers to the movement and change of the object in several consecutive frames; a series of consecutive frames are extracted from the real-time video data, and a series of consecutive frames include the operator's actions and the status of the filling equipment; the consecutive frames are organized into a frame sequence in chronological order, and the frame sequence will be used as the input of the spatiotemporal convolutional neural network; in each frame, the spatiotemporal convolutional neural network extracts spatial features through the convolution layer, and the spatial features include the operator's posture, the position of the filling joint and the status of the gas cylinder; the specific position of each object in the frame is identified, and the corresponding feature map is generated, and the feature map is used to describe the distribution of the object in space;

[0086] The spatiotemporal convolutional neural network analyzes the changes of objects in consecutive frames by applying convolution operations in the time dimension, including the operator's hand movements, the connection and separation of the filling connector, the liquid level rise and pressure changes of the gas cylinder; through the convolution in the time dimension, the network can capture the movement pattern of the object between different frames and the dynamic connection state of the connector. The movement pattern represents the operator's action coherence;

[0087] Combining spatial and temporal features, the spatiotemporal convolutional neural network can identify the operator's specific actions, detect whether the operator has correctly connected the filling connector, whether the standardized operating procedures are followed, and whether there are potential misoperations; the network analyzes the operating status of the filling equipment, and determines whether the connector is firmly connected by monitoring the dynamic changes of the connector; by analyzing the changes in the liquid level and pressure gauge, it evaluates the progress of filling and whether there are any abnormal conditions; if the network detects that the operator's actions do not conform to the expected pattern, or the status of the filling equipment is abnormal, the recognition result will be marked as abnormal and an alarm will be triggered.

[0088] Assume that the input video data is a four-dimensional tensor X with dimensions , where T represents the number of time frames, H represents the height of the image, W represents the width of the image, and C represents the number of channels; for the convolution operation in the time dimension, a convolution kernel K is used with a dimension of ,in, represents the convolution kernel size in the time dimension, and Represent the convolution kernel size on the height and width of the image in the spatial dimension respectively;

[0089] The calculation formula for the output tensor Y of the time dimension convolution operation is as follows:

[0090] ;

[0091] Among them, t, h and w represent the index of the output tensor Y in the time, image height and image width dimensions respectively; i, j and k represent the offset of the convolution kernel in the time, image height and image width dimensions respectively; Represents the input channel index;

[0092] Through the convolution operation in the time dimension, the network can identify the operator's hand movement trajectory between several frames and determine whether the movement is coherent and standardized. The coherence of the hand movement is captured by the change of the hand position in adjacent frames. By analyzing the position and state changes of the filling connector in different frames, the network can detect whether the connector remains correctly connected or loose during the filling process.

[0093] For the capture of the operator’s hand movement pattern, the hand is moved at different time frames. The position on , then the motion pattern calculated by the network through the time dimension convolution for:

[0094] ;

[0095] For capturing the dynamic connection status of the filling joint, the joint is displayed at different time frames. The position and status on , then the network calculates the connection state mode through time dimension convolution for:

[0096] ;

[0097] in, is an indicator variable, indicating the connection status of the joint at time t, 1 for connected and 0 for unconnected;

[0098] Set the sport mode and connection state mode Used for anomaly detection, if the hand movement is incoherent or the joint connection status is abnormal in consecutive frames, an alarm is triggered.

[0099] The spatiotemporal convolutional neural network (STCNN) successfully identified the operator's main actions, such as connecting and disconnecting the filling connector, checking the liquid level and pressure gauge, etc. The accuracy of action recognition reached 94%, especially in detecting the consistency of the operator's hand movements. STCNN can accurately track the connection status of the filling connector and detect loose or improperly connected connectors. By analyzing the liquid level and pressure gauge readings, the network can identify abnormal filling progress with an accuracy of 92%. When STCNN detects that the operator's actions are incoherent or the equipment status is abnormal, the system can trigger an alarm in time. In 50 test videos, the system triggered a total of 12 alarms, of which 10 were real anomalies and 2 were false alarms. The false alarm rate was 4%, and the alarm response time was 0.1 seconds.

[0100] According to step S400, the filling digital analysis model is established based on historical data, gas cylinder type, gas properties, environmental conditions and performance parameters of the cryogenic pump, and can calculate the optimal working time of the cryogenic pump according to different types of gas cylinders and the required amount of gas to be filled, ensuring that the gas cylinders can be filled to the required amount of gas within a safe range; the input data of the model includes the current number of cylinders to be filled, the volume of each cylinder, the target filling amount, and the temperature and pressure conditions of the gas; the total gas demand of all cylinders is calculated based on the current number of cylinders to be filled and the target filling amount of each cylinder; during the calculation process, dynamic adjustments are made according to real-time environmental conditions and the status of the pump;

[0101] After calculating the working time, the cryogenic pump is automatically started to fill the cylinder with gas. The filling process enters the fully automatic control stage, and the working status of the pump and the filling progress are monitored in real time. During the filling process, the progress of the filling is monitored in real time through sensors and cameras, and the changes in liquid level and pressure are detected. If any abnormality is detected, the operating parameters will be adjusted and an alarm will be issued.

[0102] When the cryogenic pump reaches the preset filling time, the pump will automatically stop working to ensure that the filling volume does not exceed the set requirements; an overfilling detection mechanism is set up. If it is detected during the filling process that the gas filling volume exceeds the set safety range, an alarm will be issued and the cryogenic pump will automatically stop working.

[0103] According to another embodiment of the present invention, Figure 2 As shown in the system structure diagram of an industrial gas cylinder filling safety management system based on AI video, an industrial gas cylinder filling safety management system based on AI video includes:

[0104] QR code scanning and quantity report generation module: including: QR code scanning unit, inspection information verification unit, quantity statistics unit and quantity report generation unit; wherein, the QR code scanning unit uses a camera installed in the filling area to scan the QR code on the industrial gas cylinder, the inspection information verification unit is responsible for obtaining the inspection information of the current cylinder, the quantity statistics unit is responsible for automatically accumulating the number of cylinders to be filled, judging whether the current cylinder is within the valid inspection period, and generating the number of cylinders to be filled, and the quantity report generation unit generates a quantity report after all QR codes are scanned;

[0105] AI video analysis and cylinder information comparison module: including: image and video data acquisition unit, cylinder detection and identification unit and information comparison unit; wherein, the image and video data acquisition unit uses the camera installed in the filling area to obtain the image and video data of the cylinder to be filled in real time, the cylinder detection and identification unit is responsible for AI video analysis, and detects and identifies the cylinder to be filled in the video data through convolutional neural network and YOLO algorithm, and the information comparison unit is responsible for comparing the information of the cylinder to be filled obtained through AI video analysis with the inspection information of the corresponding cylinder obtained by scanning the code. If the information is inconsistent, an alarm is issued and the operator is prompted to check and confirm;

[0106] The module for real-time detection and pump control calculation during the filling process includes: a real-time detection and analysis unit and a gas total amount calculation unit. The real-time detection and analysis unit uses the YOLO algorithm and spatiotemporal convolutional neural network to perform real-time detection and analysis on the video data during the filling process to ensure that the filling process of each gas cylinder meets safety standards. The gas total amount calculation unit is responsible for automatically calculating the current required gas total amount according to the number of cylinders to be filled after the filling control pump relay receives the filling instruction.

[0107] Filling control and safety monitoring module: including: filling time calculation unit, cryogenic pump control unit and safety monitoring unit; wherein, the filling time calculation unit is responsible for calculating the working time of the cryogenic pump according to the preset filling digital analysis model and the total amount of gas currently required; the cryogenic pump control unit is responsible for starting the cryogenic pump and performing filling operations according to the calculated working time; when the filling reaches the preset time or detects that the gas filling amount exceeds the set safety range, the safety monitoring unit automatically sounds an alarm and stops the pump, and records the overfilling warning data into the database.

[0108] In this embodiment, the equipment is configured as follows: number of gas cylinders: 100 gas cylinders of different specifications, including small, medium and large gas cylinders; filling equipment: cryogenic pump, pressure gauge, liquid level sensor, safety valve, etc.; gas type: liquid oxygen, liquid nitrogen, liquid argon; experimental cycle: each filling operation lasts about 30 minutes, and a total of 50 filling operations were performed in the experiment.

[0109] The filling time calculation unit accurately calculated the working time of the cryogenic pump in 50 filling operations, so that the filling amount of each gas cylinder was within the error range of ±2%. The calculation of each filling time takes about 2 seconds, and it can respond quickly and start the filling operation. The cryogenic pump control unit showed high operating accuracy during the experiment. In 50 operations, the filling of all gas cylinders was completed as planned, and there was no problem of insufficient or excessive filling due to pump control problems. The system can control the pump to fill at the optimal speed based on the calculated working time, and the difference between the average filling time and the estimated time is less than 1%.

[0110] The safety monitoring unit successfully detected the anomaly and issued an alarm in time during the four simulated overfilling tests, with the alarm response time being less than 0.5 seconds. After the alarm was issued, the cryogenic pump stopped working immediately, and no actual overfilling incident occurred, ensuring the safety of the operation. All alarm events and filling operation data were successfully recorded in the database.

[0111] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0112] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for industrial gas cylinder filling safety management based on AI video, characterized in that: The method comprises the following steps: S100, using a camera installed in the filling area to scan the QR code on the industrial gas cylinder, obtain the inspection information of the current cylinder, automatically accumulate the number of cylinders to be filled, determine whether the current cylinder is within the valid inspection period, and generate the number of cylinders to be filled; after all QR codes are scanned, generate a quantity report; S200 uses the camera installed in the filling area to obtain the image and video data of the cylinders to be filled in real time, conducts AI video analysis, and detects and identifies the cylinders to be filled in the video data through convolutional neural network and YOLO algorithm; compares the information of the cylinders to be filled obtained through AI video analysis with the inspection information of the corresponding cylinders obtained by scanning the code; if the information is inconsistent, an alarm is issued and the operator is prompted to check and confirm; S300 uses the YOLO algorithm and spatiotemporal convolutional neural network to perform real-time detection and analysis of video data during the filling process to ensure that the filling process of each gas cylinder meets safety standards; after receiving the filling instruction, the filling control pump relay automatically calculates the total amount of gas currently required based on the number of cylinders to be filled; According to step S300, the YOLO algorithm is used to detect key objects in the video frame in real time, the key objects include gas cylinders, filling equipment and operators, identify the key objects in each frame, and mark the bounding box; YOLO detects the position of the object and classifies it, distinguishes different types of gas cylinders, identifies the state of the filling connector, and distinguishes the posture and action of the operator, the state of the filling connector includes connected and unconnected; Use a spatiotemporal convolutional neural network to analyze the behavior of detected objects in consecutive frames, analyze spatial information, and process changes in the time dimension. The spatial information represents the position of the object in each frame, and the changes in the time dimension represent the movement of the object between different frames, thereby identifying the specific actions of the operator; the spatiotemporal convolutional neural network is used to monitor the operating status of the filling equipment, and by analyzing the dynamic connection status of the filling connector, the filling progress of the gas cylinder, the liquid level and the pressure gauge readings, it is determined whether there are abnormal conditions or potential safety hazards; The detection results of the YOLO algorithm and the action recognition results of the spatiotemporal convolutional neural network are fused to determine whether the operator's actions are correct based on action recognition, and whether the status of the equipment is consistent with expectations based on status monitoring; if the operation steps are not standardized or the equipment status is abnormal, an alarm is triggered; The spatial information in the spatiotemporal convolutional neural network refers to the image content of each frame in the video, and the change in the time dimension refers to the movement and change of the object in several consecutive frames; a series of consecutive frames are extracted from the real-time video data, and a series of consecutive frames include the operator's actions and the status of the filling equipment; the consecutive frames are organized into a frame sequence in chronological order, and the frame sequence will be used as the input of the spatiotemporal convolutional neural network; in each frame, the spatiotemporal convolutional neural network extracts spatial features through the convolution layer, and the spatial features include the operator's posture, the position of the filling joint and the status of the gas cylinder; the specific position of each object in the frame is identified, and the corresponding feature map is generated, and the feature map is used to describe the distribution of the object in space; The spatiotemporal convolutional neural network analyzes the changes of objects in consecutive frames by applying convolution operations in the time dimension, including the operator's hand movements, the connection and separation of the filling connector, the liquid level rise and pressure changes of the gas cylinder; through the convolution in the time dimension, the network can capture the movement pattern of the object between different frames and the dynamic connection state of the connector. The movement pattern represents the operator's action coherence; Combining spatial and temporal features, the spatiotemporal convolutional neural network can identify the operator's specific actions, detect whether the operator correctly connects the filling connector, whether the standardized operating steps are followed, and whether there are potential misoperations; the network analyzes the operating status of the filling equipment and determines whether the connector is firmly connected by monitoring the dynamic changes of the connector; by analyzing the changes in the liquid level and pressure gauge, the progress of the filling and whether there are any abnormalities; if the network detects that the operator's actions do not conform to the expected pattern, or the status of the filling equipment is abnormal, the recognition result will be marked as abnormal and an alarm will be triggered; Assume that the input video data is a four-dimensional tensor X with dimensions , where T represents the number of time frames, H represents the height of the image, W represents the width of the image, and C represents the number of channels; for the convolution operation in the time dimension, a convolution kernel K is used with a dimension of ,in, represents the convolution kernel size in the time dimension, and Represent the convolution kernel size on the height and width of the image in the spatial dimension respectively; The calculation formula for the output tensor Y of the time dimension convolution operation is as follows: ; Among them, t, h and w represent the index of the output tensor Y in the time, image height and image width dimensions respectively; i, j and k represent the offset of the convolution kernel in the time, image height and image width dimensions respectively; Represents the input channel index; Through the convolution operation in the time dimension, the network can identify the operator's hand movement trajectory between several frames and determine whether the movement is coherent and standardized. The coherence of the hand movement is captured by the change of the hand position in adjacent frames. By analyzing the position and state changes of the filling connector in different frames, the network can detect whether the connector remains correctly connected or loose during the filling process. For the capture of the operator’s hand movement pattern, the hand is moved at different time frames. The position on , then the motion pattern calculated by the network through the time dimension convolution for: ; For capturing the dynamic connection status of the filling joint, the joint is displayed at different time frames. The position and status on , then the network calculates the connection state mode through time dimension convolution for: ; in, is an indicator variable, indicating the connection status of the joint at time t, 1 for connected and 0 for unconnected; Set the sport mode and connection state mode For anomaly detection, if the hand movement is incoherent or the joint connection status is abnormal in consecutive frames, an alarm is triggered; S400, according to the preset filling digital analysis model, combined with the total amount of gas currently required, calculate the working time of the cryogenic pump, start the cryogenic pump, and perform the filling operation according to the calculated working time; when the filling reaches the preset time or it is detected that the gas filling amount exceeds the set safety range, automatically sound an alarm and stop the pump, and record the overfilling warning data into the database; According to step S400, the filling digital analysis model is established based on historical data, gas cylinder type, gas properties, environmental conditions and performance parameters of the cryogenic pump, and can calculate the optimal working time of the cryogenic pump according to different types of gas cylinders and the required amount of gas to be filled, ensuring that the gas cylinders can be filled to the required amount of gas within a safe range; the input data of the model includes the current number of cylinders to be filled, the volume of each cylinder, the target filling amount, and the temperature and pressure conditions of the gas; the total gas demand of all cylinders is calculated based on the current number of cylinders to be filled and the target filling amount of each cylinder; during the calculation process, dynamic adjustments are made according to real-time environmental conditions and the status of the pump; After calculating the working time, the cryogenic pump is automatically started to fill the cylinder with gas. The filling process enters the fully automatic control stage, and the working status of the pump and the filling progress are monitored in real time. During the filling process, the progress of the filling is monitored in real time through sensors and cameras, and the changes in liquid level and pressure are detected. If any abnormality is detected, the operating parameters will be adjusted and an alarm will be issued. When the cryogenic pump reaches the preset filling time, the pump will automatically stop working to ensure that the filling volume does not exceed the set requirements; an overfilling detection mechanism is set up. If it is detected during the filling process that the gas filling volume exceeds the set safety range, an alarm will be issued and the cryogenic pump will automatically stop working.

2. According to claim 1, an AI video-based industrial gas cylinder filling safety management method is characterized by: According to step S100, each cylinder is affixed with a unique QR code label, and the QR code label contains the inspection information of the current cylinder, and the inspection information includes the inspection record of the cylinder, the cylinder number, the production date, the last inspection date, the gas type and the maximum filling amount; the QR code on the industrial gas cylinder is scanned by a camera installed in the filling area, and the inspection information of the cylinder associated with the QR code is automatically extracted and stored in the database; Verify the inspection information, determine whether the current cylinder is within the valid inspection period, and record the inspection status; If the inspection information of a cylinder fails or has exceeded the inspection validity period, a warning will be issued to remind the operator not to refill the cylinder; After each scan and acquisition of inspection information, the qualified cylinders are automatically counted into the total number of cylinders to be filled, and the number is continuously accumulated to generate the current number of cylinders to be filled, which is displayed on the operation interface for operators to view in real time; After the QR codes of all cylinders to be filled are scanned, a report on the quantity of cylinders to be filled is generated, which includes the inspection information and inspection status of each cylinder, as well as the accumulated total quantity.

3. According to claim 1, an industrial gas cylinder filling safety management method based on AI video is characterized in that: According to step S200, the camera collects dynamic video data in the filling area in real time, and performs image preprocessing on the original video data transmitted by the camera, wherein the preprocessing includes denoising, image enhancement and white balance adjustment; for the preprocessed image, a convolutional neural network is used to perform target detection of the steel cylinder, and different features are detected by sliding a number of filters in the image, and the features are combined into a higher-level representation to form the outline of the steel cylinder; After the cylinder detection is completed, the YOLO algorithm based on deep learning is used to identify the specific information of the cylinder, which includes the shape, color and logo of the cylinder. The outline of each cylinder is calculated based on the outline of the cylinder. The cylinder information in the video data is analyzed frame by frame, and each detected cylinder is uniquely identified, and the information of the cylinders to be filled detected in all frames is summarized.

4. According to claim 3, an AI video-based industrial gas cylinder filling safety management method is characterized by: The preprocessed image is input to the first layer of the convolutional neural network, which is a convolutional layer. The convolutional layer slides over the image through a number of filters, scanning each area of ​​the image one by one. Each filter focuses on detecting features, including the edges, corners and textures of the cylinder. As the filter slides across the image, each area is calculated to generate a feature map, which represents the probability and location of the edge of the cylinder. The convolutional neural network consists of several convolutional layers, each of which extracts higher-level features based on the feature map of the previous layer to form the overall outline of the cylinder; A pooling layer is applied after the convolution layer. The pooling layer downsamples the feature map to reduce the size of the feature map and retain the most important feature information in the overall contour of the cylinder. After passing through several convolutional and pooling layers, the convolutional neural network finally outputs a high-level feature representation. The high-level feature representation is composed of low-level features layer by layer to form the complete contour of the cylinder. The high-level feature representation can describe the shape, size and position of the cylinder. The last several layers of the convolutional neural network are fully connected layers, which are responsible for mapping high-level features to cylinders. By outputting a probability distribution, they determine whether there are cylinders in the image and determine the location of the cylinders. For each detected cylinder, a bounding box is generated to identify the specific location of the cylinder in the image and form the outline of the cylinder.

5. According to claim 3, an industrial gas cylinder filling safety management method based on AI video is characterized in that: The YOLO algorithm receives a preprocessed image, which is cut into several grids, each of which is responsible for detecting the object features in its area; the YOLO model takes the entire image as input and detects several objects at the same time in a forward pass; the YOLO algorithm processes all grids in the image in parallel, and each grid generates a set of predictions, including the object category, the coordinates of the bounding box, the width and height of the box, and the confidence of each category, and the object category is set to cylinders; The convolutional layer of the YOLO model extracts features from the image, including the edge, color distribution, and shape of the cylinder. The YOLO model uses these features to determine the specific information of the cylinder. The YOLO model identifies the shape of the cylinder by analyzing the features within each grid, and identifies the shape of the cylinder based on edge detection and shape features; based on the shape features, the YOLO algorithm calculates the outline of the cylinder, which includes the position coordinates and size of the cylinder in the image, and is used to identify the specific location of the cylinder; the YOLO algorithm extracts the color features of the cylinder, and identifies the main color and color mark of the cylinder by analyzing the color distribution within each grid. According to the color features, the model subdivides the type of cylinder; different types of gas cylinders have different color logos, and the YOLO algorithm can automatically classify and identify different types of cylinders based on color features; logos are printed on the cylinders, including QR codes, barcodes, labels and text, and the YOLO algorithm detects and identifies them by analyzing the texture and pattern features of each grid; parsing is performed to combine the relevant information with other identification information of the cylinder; The YOLO algorithm processes each frame in the video data, identifying and marking the specific information of each cylinder; in each frame, the model will generate the outline box and related information of the cylinder.

6. An industrial gas cylinder filling safety management system based on AI video, using an industrial gas cylinder filling safety management method based on AI video according to any one of claims 1-5, characterized in that: include: QR code scanning and quantity report generation module: including: QR code scanning unit, inspection information verification unit, quantity statistics unit and quantity report generation unit; wherein, the QR code scanning unit uses a camera installed in the filling area to scan the QR code on the industrial gas cylinder, the inspection information verification unit is responsible for obtaining the inspection information of the current cylinder, the quantity statistics unit is responsible for automatically accumulating the number of cylinders to be filled, judging whether the current cylinder is within the valid inspection period, and generating the number of cylinders to be filled, and the quantity report generation unit generates a quantity report after all QR codes are scanned; AI video analysis and cylinder information comparison module: including: image and video data acquisition unit, cylinder detection and identification unit and information comparison unit; wherein, the image and video data acquisition unit uses the camera installed in the filling area to obtain the image and video data of the cylinder to be filled in real time, the cylinder detection and identification unit is responsible for AI video analysis, and detects and identifies the cylinder to be filled in the video data through convolutional neural network and YOLO algorithm, and the information comparison unit is responsible for comparing the information of the cylinder to be filled obtained through AI video analysis with the inspection information of the corresponding cylinder obtained by scanning the code. If the information is inconsistent, an alarm is issued and the operator is prompted to check and confirm; The module for real-time detection and pump control calculation during the filling process includes: a real-time detection and analysis unit and a gas total amount calculation unit. The real-time detection and analysis unit uses the YOLO algorithm and spatiotemporal convolutional neural network to perform real-time detection and analysis on the video data during the filling process to ensure that the filling process of each gas cylinder meets safety standards. The gas total amount calculation unit is responsible for automatically calculating the current required gas total amount according to the number of cylinders to be filled after the filling control pump relay receives the filling instruction. Filling control and safety monitoring module: including: filling time calculation unit, cryogenic pump control unit and safety monitoring unit; wherein, the filling time calculation unit is responsible for calculating the working time of the cryogenic pump according to the preset filling digital analysis model and the total amount of gas currently required; the cryogenic pump control unit is responsible for starting the cryogenic pump and performing filling operations according to the calculated working time; when the filling reaches the preset time or detects that the gas filling amount exceeds the set safety range, the safety monitoring unit automatically sounds an alarm and stops the pump, and records the overfilling warning data into the database.

Images