Chemical vehicle intelligent AI loading and unloading risk management and control method, device, equipment and medium

Abstract

The application relates to a chemical vehicle intelligent AI loading and unloading risk management and control method, device, equipment and medium, and belongs to the technical field of chemical safety monitoring. The method comprises the following steps: determining whether the vehicle is allowed to enter according to the entry registration information of the vehicle; taking the first monitoring item in the sequence as a current monitoring item; acquiring video information of the vehicle in the loading and unloading operation area for the current monitoring item; analyzing the video information of the current monitoring item based on a deep learning model to obtain a corresponding recognition result; if the recognition result is standard, taking the next monitoring item in the sequence as a new current monitoring item; if the recognition result is non-standard, generating push information pushed to the management personnel, and automatically controlling the loading and unloading operation equipment to stop, re-recognizing until the recognition result is standard; and if all the monitoring items are standard, automatically controlling the loading and unloading operation equipment to start or continue to run. The application improves operation safety.

Description

Technical Field

[0001] This application relates to the technical field of chemical safety monitoring, and in particular to a method, device, equipment and medium for intelligent AI loading and unloading risk control of chemical vehicles. Background Technology

[0002] Loading and unloading operations at chemical plants involve flammable, explosive, toxic, and hazardous substances, making them a high-risk area for safety accidents. Therefore, ensuring the safety of loading and unloading operations is a crucial aspect of ensuring the safety of energy storage and transportation.

[0003] Currently, some chemical companies have deployed video surveillance and basic information registration systems for loading and unloading vehicle management, enabling functions such as vehicle entry and exit recording and area-wide video recording. However, these systems are only used for post-event follow-up and cannot automatically identify key steps in the loading and unloading process in real time. This results in potential safety hazards still existing in the loading and unloading process, making it difficult to intervene in a timely and accurate manner.

[0004] Therefore, there is an urgent need for a method for managing loading and unloading risks that can monitor loading and unloading vehicles in real time and provide automatic early warnings. Summary of the Invention

[0005] To improve operational safety, this application provides a method, device, equipment, and medium for intelligent AI-based loading and unloading risk management of chemical vehicles.

[0006] Firstly, this application provides an intelligent AI-based risk management method for loading and unloading chemical vehicles, employing the following technical solution: Determine whether a vehicle is allowed to enter based on its entry registration information; For vehicles allowed to enter, a sequence is obtained by sorting each monitoring item, and the first monitoring item in the sequence is taken as the current monitoring item; Acquire video information of the vehicle in the loading and unloading area for the current monitoring item; The video information of the current monitoring item is analyzed based on a deep learning model to obtain the corresponding recognition result; If the identification result is normal, then the next monitoring item in the sequence is taken as the new current monitoring item, and the steps of video information acquisition and video information analysis for the current monitoring item are continued. If the identification result is non-standard, a push message is generated and sent to the management personnel, and the loading and unloading equipment is automatically stopped. The video information acquisition and video information analysis steps are then repeated until the identification result is standard. If the current monitoring item is the last monitoring item in the sequence and the identification result is standard, then the automatic linkage control will start or continue to operate the loading and unloading equipment.

[0007] By adopting the above technical solution, real-time acquisition of vehicle entry information and video information of loading and unloading operation areas, and automatic identification of key monitoring items using deep learning models, non-standard behaviors in loading and unloading operations can be detected in a timely manner, and the loading and unloading equipment can be automatically stopped. At the same time, warning information is pushed to the management personnel, thereby effectively making up for the limitations of traditional systems in post-event follow-up. This achieves real-time, automatic, and precise safety intervention in the chemical loading and unloading operation process, significantly reducing operational risks and improving the overall safety management level.

[0008] Furthermore, before analyzing the video information based on the deep learning model, the method further includes adaptive neural network scheduling, including: The scene features of the current video information are analyzed in real time. The scene features include: image clarity, target occlusion ratio, light intensity, and the action type corresponding to the current monitoring item. The action type includes static targets and dynamic behaviors. Automatically select the target neural network model based on the scene characteristics: If the current monitoring item only needs to identify static targets, and the image clarity is higher than the preset clarity threshold and the image is unobstructed, then the image classification network is selected as the first neural network. If the current monitoring item needs to identify dynamic behavior, then a temporal convolutional network is selected as the second neural network; If the scene features meet the first condition, the arbitration network is activated. The arbitration network is a spatiotemporal joint model based on graph convolutional networks. The outputs of the first neural network and the second neural network are simultaneously input into the arbitration network for fusion recognition. The first condition includes: partial occlusion of the target, insufficient lighting, or joint judgment of static and dynamic behaviors.

[0009] By adopting the above technical solution, the use of neural network models is adaptively scheduled and optimized according to the real-time changing characteristics of video surveillance scenes. This avoids the problem of poor performance of a single model in all complex scenarios. It can flexibly call the most suitable recognition model for monitoring tasks and environmental conditions of different complexities. At the same time, the input frame rate is dynamically adjusted according to the complexity of the scene, which further optimizes the allocation of computing resources. This makes the entire risk management method better adapt to the complexity and variability of actual industrial scenarios, and improves the overall performance and reliability of intelligent AI loading and unloading risk management.

[0010] Further, if the scene features satisfy the first condition, the analysis of the video information of the current monitoring item based on the deep learning model to obtain the corresponding recognition result includes: The first neural network and the second neural network are run in parallel, wherein the first neural network extracts static semantic features related to the current monitoring item from a single frame image; and the second neural network extracts dynamic behavioral features related to the current monitoring item from multiple consecutive frames of images. When the identification results of the first neural network and the second neural network for the current monitoring item are consistent, a preliminary identification result for the current monitoring item is generated; The preliminary identification results are compared with the specifications of the corresponding monitoring items in the preset standard loading and unloading process sequence. The standard loading and unloading process sequence defines the legal order of each monitoring item and the specified time interval of each monitoring item. Output the identification result based on the comparison result; When the recognition results of the first neural network and the second neural network are inconsistent, the arbitration network is activated. The input is the fusion amount of the static semantic features and the dynamic behavioral features, and the output is the arbitrated recognition result. The arbitrated recognition result is then compared with the standard loading and unloading process sequence to obtain the final recognition result.

[0011] By adopting the above technical solution, when facing complex scenarios such as partial occlusion of the target, insufficient lighting, or the need for joint judgment of static and dynamic behaviors, the first and second neural networks run in parallel to comprehensively capture the features of the monitored items from static and dynamic dimensions, respectively. When the two identification results are consistent, the standardization can be quickly and accurately confirmed by comparing with the standard loading and unloading procedure sequence and the specified time interval. When the identification results are inconsistent, the third arbitration network based on graph convolutional network is activated. By integrating static semantic features and dynamic behavioral features, and combining spatiotemporal joint analysis capabilities, it can intelligently resolve conflicts and make more accurate judgments, effectively reducing the risk of false alarms and missed alarms. This ensures accurate identification and timely control of loading and unloading operation risks under various complex working conditions, thereby significantly improving operational safety.

[0012] Furthermore, before outputting the recognition result, the first neural network and / or the second neural network perform the following optimization on the extracted features: Calculate the radial distance between the scene features and each template vector in the preset standard template library; Multiply the radial distance by a preset scaling factor to obtain the centrifugal distance; The confidence level is determined based on the centrifugal distance: the smaller the centrifugal distance, the higher the confidence level; the larger the centrifugal distance, the lower the confidence level. If the centrifugal distance exceeds a preset distance threshold, the scene feature is determined to be an invalid feature and is not used for recognition; if the centrifugal distance does not exceed the preset distance threshold, the scene feature is retained and the recognition result with the highest confidence is output.

[0013] By adopting the above technical solution, the radial distance between the feature and each template vector in the preset standard template library is calculated and converted into centrifugal distance. This quantifies the deviation of the current feature from the standard specification pattern. By setting a preset distance threshold, invalid features that differ too much from the standard pattern are effectively filtered out. This avoids inputting low-quality or abnormal features into the recognition process, thereby significantly reducing the risk of false alarms and false negatives, and thus effectively improving the safety of chemical vehicle loading and unloading operations.

[0014] Furthermore, the standardized time interval is dynamically generated through the following steps: Collect time-series samples of each monitoring item during historical normal loading and unloading operations to form a time-series dataset; A dynamic time warping algorithm is used to map the start and end times of the same monitoring item in different loading and unloading operations to a common time axis, so as to flexibly align the time series dataset on the time axis. On the aligned common time axis, a kernel density estimation model is established to obtain the probability density function of the duration of each monitoring item; The highest density interval of the probability density function is used as the basic time interval, and the highest density interval is the longest connected interval formed by all duration points whose probability density function values ​​are greater than or equal to the maximum value. The complexity is determined based on the current vehicle model, the type of loading medium, and the experience level of the current operators. Statistics on all complexities in normal historical loading and unloading operations the median; Based on the aforementioned complexity Stretching the base time interval includes: If complexity If the value is higher than the median, the right boundary of the base time interval will be increased. This yields the normalized time interval; if the complexity... If the value is below the median, the left boundary of the base time interval will be reduced. This yields the standardized time interval; among which, The length of the base time interval, The first preset proportional coefficient, This is the second preset proportional coefficient, where 0 < < <0.5; After each loading and unloading operation is completed, the actual completion time is used as a new sample to incrementally update the kernel density estimation model.

[0015] By adopting the above technical solution, the standardized time intervals for each monitoring item can be dynamically generated and adjusted based on historical data and the actual situation of current operations. The kernel density estimation model can accurately capture the probability distribution of the duration of each monitoring item, thereby determining the basic time interval that best represents normal operations. The complexity c is then used to elastically stretch the basic time interval, so that the standardized time interval can fully consider factors such as the current vehicle model, the type of loading medium, and the experience level of the operators. This avoids the inaccuracies caused by fixed time intervals, effectively reduces false alarms and missed alarms, and improves the accuracy and adaptability of overall risk identification.

[0016] Furthermore, the method also includes: Pre-defined deterministic association rules based on process logic are used to define the triggering relationship that non-standard behavior of a preceding monitoring item will inevitably lead to non-standard behavior of a subsequent monitoring item. Constructing probabilistic association rules based on historical data includes: collecting historical non-standard records and confusion parameters; taking any two non-standard preceding detection items and non-standard subsequent detection items as temporary association pairs; performing hierarchical statistical analysis of the conditional probabilities corresponding to the temporary association pairs according to the confusion parameters; and obtaining the net association probability after correction. If the net association probability and the sample size of the temporary association pairs meet the sample threshold, then the temporary association pairs are determined as probabilistic association rules, and the association time window is recorded. When a non-standard monitoring item is detected, the deterministic association rule is matched. If a match is found, it is inferred that subsequent monitoring items will inevitably be non-standard, and the loading and unloading equipment is stopped in a coordinated manner. If no match is found, the probabilistic association rule is matched: the predicted probability of the subsequent monitoring item being non-standard within the association time window is calculated. If the predicted probability is greater than a preset probability threshold, the recognition frequency of the deep learning model is increased within the association time window, and a predictive warning is pushed.

[0017] By adopting the above technical solutions, the risk management of chemical vehicle loading and unloading is elevated from a passive response to a proactive prevention level. The pre-set deterministic association rules can make immediate and mandatory risk inferences and interventions on current non-standard behaviors based on strict process logic, effectively avoiding serious safety accidents caused by errors in process logic. The probabilistic association rules use historical data to predict possible non-standard behaviors in the future, realizing forward-looking risk management. This enables the system not only to respond to non-standard behaviors that have already occurred, but also to predict and prevent potential risks, thereby effectively improving the intelligence and safety of chemical vehicle loading and unloading operations.

[0018] Furthermore, the calculation of the predicted probability that the identification result of the current monitoring item will be non-standard within the associated time window after the current monitoring item's identification result is non-standard includes: From the probabilistic association rules, read the historical conditional probability of the occurrence of the subsequent non-standard monitoring item Y within the association time window after the occurrence of the preceding non-standard monitoring item X. Determine the time decay coefficient starting from the moment the aforementioned non-standard monitoring item X occurs; The performance coefficient is determined based on the operator's historical operation records; Based on the current work period, weather conditions, and whether the project is scheduled to be completed on time, the corresponding risk multiplier can be obtained from historical data. The historical conditional probability is multiplied sequentially by the time decay coefficient, the personnel performance coefficient, and the risk multiplier to obtain the product result, and the product result is restricted to the range of 0 to 1 to obtain the predicted probability.

[0019] By adopting the above technical solution, historical conditional probabilities can be read from probabilistic association rules as a basis, and a time decay coefficient can be introduced to dynamically reflect the immediacy of risk. The personnel performance coefficient is determined by statistically analyzing the relative error rate of the current operators. The risk multiplier is queried by combining external factors such as the current working period, weather conditions, and whether the project is on a tight schedule. These dynamic factors are multiplied with historical conditional probabilities and limited to the range of 0 to 1, thereby obtaining a comprehensive, dynamic, and highly accurate predictive probability. This allows for earlier and more accurate identification of potential risks, enabling more effective preventive measures to be taken before risks occur, thus ensuring operational safety.

[0020] Secondly, this application provides an intelligent AI-based risk management device for loading and unloading chemical vehicles, employing the following technical solution: The vehicle verification module is used to determine whether a vehicle is allowed to enter the site based on its entry registration information. The monitoring item sorting module is used to sort the vehicles allowed to enter the site based on each monitoring item to obtain a sequence, and to take the first monitoring item in the sequence as the current monitoring item; The video information acquisition module is used to acquire video information of the vehicle in the loading and unloading operation area for the current monitoring item; The monitoring item analysis module is used to analyze the video information of the current monitoring item based on a deep learning model to obtain the corresponding recognition results; The first identification result execution module is used to, when the identification result of the monitoring item analysis module is normal, take the next monitoring item in the sequence as the new current monitoring item and continue to execute the video information acquisition module and the monitoring item analysis module; The second identification result execution module is used to generate push information to the management personnel when the identification result of the monitoring item analysis module is non-standard, and automatically control the loading and unloading equipment to stop, and re-execute the video information acquisition module and the monitoring item analysis module until the identification result is standard. The third identification result execution module is used to automatically control the loading and unloading equipment to start or continue operating when the current monitoring item is the last monitoring item in the sequence and the identification result is standard.

[0021] Thirdly, this application provides an electronic device that adopts the following technical solution: An electronic device, comprising: At least one processor; Memory; At least one computer program, wherein the at least one computer program is stored in the memory and configured to be executed by the at least one processor, the at least one computer program being configured to: perform the method as described in any one of the first aspects.

[0022] Fourthly, this application provides a computer-readable storage medium, which adopts the following technical solution: A computer-readable storage medium storing a computer program that can be loaded by a processor and execute the method as described in any one of the first aspects.

[0023] For a detailed description of the second to fourth aspects of the present invention and their various implementations, please refer to the detailed description in the first aspect and its various implementations; and for a detailed description of the beneficial effects of the second to fourth aspects and their various implementations, please refer to the beneficial effect analysis in the first aspect and its various implementations, which will not be repeated here.

[0024] In summary, this application includes at least one of the following beneficial technical effects: 1. By acquiring real-time vehicle entry information and video information of loading and unloading operation areas, and using deep learning models to automatically identify key monitoring items, non-standard behaviors in loading and unloading operations can be detected in a timely manner, and the loading and unloading equipment can be automatically stopped. At the same time, warning information is pushed to the management personnel, thereby effectively making up for the limitations of traditional systems in post-event follow-up. This achieves real-time, automatic, and precise safety intervention in the chemical loading and unloading operation process, significantly reducing operational risks and improving the overall safety management level. 2. Based on the real-time changing characteristics of video surveillance scenes, the use of neural network models is adaptively scheduled and optimized, avoiding the problem of poor performance of a single model in all complex scenarios. It can flexibly call the most suitable recognition model for monitoring tasks and environmental conditions of different complexities. 3. By calculating the radial distance between the feature and each template vector in the preset standard template library and converting it into centrifugal distance, the deviation of the current feature from the standard specification pattern can be quantified. This effectively filters out invalid features that differ too much from the standard pattern, avoids inputting low-quality or abnormal features into the recognition process, and effectively improves the safety of chemical vehicle loading and unloading operations. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the intelligent AI loading and unloading risk management method for chemical vehicles in the embodiments of this application.

[0026] Figure 2 This is a structural block diagram of the intelligent AI loading and unloading risk control device for chemical vehicles in the embodiments of this application.

[0027] Figure 3 This is a structural block diagram of the electronic device in the embodiments of this application. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article, unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship.

[0030] This application discloses an intelligent AI-based risk management method for loading and unloading chemical vehicles. (Refer to...) Figure 1 This is performed by an electronic device, which can be a server or a terminal device. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. The terminal device can be a smartphone, tablet, desktop computer, etc., but is not limited to these. (Steps S101 to S107)

[0031] Step S101: Determine whether the vehicle is allowed to enter based on the vehicle's entry registration information.

[0032] Specifically, the electronic device acquires the vehicle's entry registration information, compares it with a preset database of allowed entry information, and determines whether the vehicle is allowed to enter based on the comparison results. The entry registration information includes license plate information, driver's name, driver's license number, and a photo of the vehicle's exterior.

[0033] Specifically, electronic devices can automatically identify license plate numbers through an image recognition system and match them with a list of license plates in a database. Driver information and vehicle appearance information are identified through pre-trained face recognition networks and vehicle classification networks, respectively. The network models have been trained before system deployment and extract information directly during runtime.

[0034] The system compares the entry registration information with a pre-set database of allowed entry vehicles, which stores information on the company's whitelist of vehicles and authorized drivers. If the comparison is successful, the vehicle is allowed to enter the loading and unloading area; otherwise, an alarm is triggered and entry is denied.

[0035] Step S102: For vehicles allowed to enter, sort them according to each monitoring item to obtain a sequence, and take the first monitoring item in the sequence as the current monitoring item.

[0036] The monitoring items include: verification of license plate consistency, vehicle parking location, handing over vehicle keys, placement of wheel chocks, static grounding, and connection of oil and gas recovery pipes.

[0037] Step S103: Acquire video information of the vehicle in the loading and unloading area for the current monitoring item.

[0038] Specifically, this application sets up several fixed-angle network cameras at key operational points for each monitoring item in the loading and unloading area, continuously recording video streams and transmitting them to a central storage device. Electronic devices obtain video information for the current monitoring item by communicating with the network camera of the current monitoring item and acquiring the video stream.

[0039] Step S104: Analyze the video information of the current monitoring item based on the deep learning model to obtain the corresponding recognition result.

[0040] Specifically, electronic devices can train a general object detection model based on convolutional neural networks (CNNs) that can identify all monitored items in a video frame and process the image in each frame or every few frames to output the location and category of the monitored items.

[0041] In another possible implementation, an independent model can be trained for each monitoring item. When it is necessary to identify a specific monitoring item, image segments of the relevant areas in the video are input into the corresponding model for judgment. Therefore, before analyzing video information based on deep learning models, the method also includes adaptive neural network scheduling, including (steps S11 to S12):

[0042] Step S11: Analyze the scene features of the current video information in real time. Scene features include: image clarity, target occlusion ratio, light intensity, and the action type corresponding to the current monitoring item. Action types include static targets and dynamic behaviors.

[0043] Specifically, image sharpness can be quantified by evaluating the edge sharpness, contrast, or frequency domain information of video frames; the target occlusion ratio can be obtained by detecting and segmenting the target object and calculating the proportion of the pixel area occluded by other objects to the total area; and illumination intensity can be determined by analyzing the average brightness value or histogram distribution of video frames. The action type corresponding to the current monitoring item is defined according to the preset monitoring task. For example, license plate recognition and vehicle parking position determination are usually static target recognition, while handing over vehicle keys, placing wheel chocks, static grounding, and connecting oil and gas recovery pipes may involve the recognition of dynamic behaviors.

[0044] Step S12: Automatically select the target neural network model based on scene features:

[0045] (1) If the current monitoring item only needs to identify static targets, and the image clarity is higher than the preset clarity threshold and the image is unobstructed, then the image classification network is selected as the first neural network.

[0046] Specifically, if the current scene conditions are ideal, then an image classification network is selected as the first neural network. This network can quickly and accurately identify static targets under favorable conditions. The preset sharpness can be 0.8.

[0047] (2) If the current monitoring item needs to identify dynamic behavior, then a temporal convolutional network is selected as the second neural network.

[0048] Specifically, temporal convolutional networks are designed to process continuous video frame sequences, capturing features and behavioral patterns in the temporal dimension, and can effectively identify complex dynamic actions.

[0049] (3) If the scene features meet the first condition, the arbitration network is started, and the outputs of the first neural network and the second neural network are simultaneously input into the arbitration network for fusion recognition. The first condition includes: partial occlusion of the target, insufficient lighting, or joint judgment of static and dynamic behaviors.

[0050] Specifically, when a target object is partially occluded, such as when the occlusion exceeds 30%, a single network may struggle to make an accurate judgment. When insufficient lighting degrades image quality, such as when the light intensity is less than 50 lux, different networks may exhibit varying degrees of robustness to noise. When a monitoring item encompasses both static and dynamic states, a comprehensive assessment of both types of information is necessary. The arbitration network can be a small fully connected network, an attention mechanism network, or a spatiotemporal joint model based on graph convolutional networks. It can learn how to weigh and integrate recognition results or feature representations from different neural networks, thereby achieving more robust and accurate final recognition results under complex conditions.

[0051] In this embodiment, the image classification network is a multi-task network that can simultaneously output the recognition results of all static monitoring items; the temporal convolutional network outputs the recognition results of all dynamic behaviors; and the arbitration network is shared and used to fuse the features of both.

[0052] Furthermore, for cases (1) and (2), the electronic device analyzes the video information based on the selected model, identifies the current monitoring item, and obtains the corresponding identification results:

[0053] License plate consistency verification: The first neural network is used to identify the license plate number of the current vehicle and compare it with the license plate information in the entry registration information to determine whether they are consistent. If they are consistent, the license plate is consistent with the one used at entry, and the result of the license plate consistency verification item is "standard". If they are inconsistent, the license plate may have been changed, and the result of the license plate consistency verification item is "non-standard".

[0054] Vehicle parking position: The first neural network is used to detect whether the vehicle's parking position in the loading and unloading area is accurate, such as whether it is within the designated marked area or beyond the guardrail. If the vehicle's parking position is accurate, the vehicle parking position item is considered correct; otherwise, it is considered incorrect.

[0055] Wheel chock placement: The first neural network is used to detect whether the wheel chocks have been correctly placed and positioned in front of and behind the wheels after the vehicle has stopped, to prevent the vehicle from rolling. If the wheel chocks are correctly placed and positioned in front of and behind the wheels, the wheel chock placement is considered correct; otherwise, it is considered incorrect.

[0056] Electrostatic grounding: A first neural network is used to detect whether the electrostatic grounding clamp is securely connected to the grounding stake, ensuring effective discharge of static electricity and preventing static electricity from igniting volatile oil or gas. If the electrostatic grounding is securely connected to the grounding stake, the electrostatic grounding item is up to standard; otherwise, it is not up to standard.

[0057] Fuel / gas recovery hose connection: A first neural network is used to detect whether the fuel / gas recovery hose is correctly connected to the vehicle's recovery interface and is not detached, meeting environmental protection and explosion-proof requirements. If the fuel / gas recovery hose is correctly connected to the vehicle's recovery interface and is not detached, the fuel / gas recovery hose connection is standard; otherwise, it is not standard.

[0058] Furthermore, if the scene features meet the first condition, such as the handing over of vehicle keys, it identifies whether the driver has placed the vehicle keys into the designated key box or handed them over as required, to prevent the vehicle from being accidentally started during loading and unloading. If the driver has handed over the keys as required, the handing over of vehicle keys is considered compliant; otherwise, it is not. The action of handing over vehicle keys includes the dynamic process of "moving the key to the box." A simple static image, such as the key ultimately being in the box, may not be sufficient to confirm whether the driver actively handed it over. Therefore, temporal behavior recognition is required, which meets the first condition that "static and dynamic behaviors must be jointly judged." The electronic device analyzes the video information of the current monitoring item based on a deep learning model to obtain the corresponding recognition results, including (steps S21 to S25):

[0059] Step S21: Run the first neural network and the second neural network in parallel. The first neural network extracts static semantic features related to the current monitoring item from a single frame image; the second neural network extracts dynamic behavioral features related to the current monitoring item from multiple consecutive frames of images.

[0060] Specifically, the first neural network identifies information such as the presence, location, and state of static targets from each frame of the video. For example, whether the key is ultimately inside the box. These features are determined based on the image content itself and do not depend on time-series information.

[0061] The second neural network analyzes temporal information in multiple consecutive frames of images to identify dynamic behaviors or actions. For example, it can determine whether a driver has handed over the vehicle keys. These features capture the target's movement patterns or behavioral sequences that change over time.

[0062] Step S22: When the identification results of the first neural network and the second neural network for the current monitoring item are consistent, a preliminary identification result for the current monitoring item is generated.

[0063] Specifically, when two independent and complementary neural networks arrive at the same conclusion regarding the current monitored item, it indicates that the recognition result has a high degree of reliability. For example, if the first neural network recognizes that the key is ultimately inside the box, and the second neural network recognizes the driver's action of handing over the key, then the current monitored item is preliminarily determined to be compliant.

[0064] Step S23: Compare the preliminary identification results with the specifications of the corresponding monitoring items in the preset standard loading and unloading process sequence to obtain the comparison results. The standard loading and unloading process sequence defines the legal order of each monitoring item and the specified time interval of each monitoring item.

[0065] For example, wheel chocks must be placed before electrostatic grounding can be performed. This sequence can be stored in a database and represented as structured data. Furthermore, for certain key monitoring items, in addition to the sequential requirements, completion must also be within a specific time frame. For instance, wheel chocks should be placed within X minutes after a vehicle parks. This time frame can be fixed or dynamically generated based on historical data. By comparing the initial identification results with these process sequences and time frames, electronic equipment can verify the rationality of the identification results from a higher-level business logic perspective, further improving the accuracy and reliability of the judgment.

[0066] Step S24: Output the recognition result based on the comparison result. If the comparison result matches, the output recognition result is standard; if the comparison result does not match, the output recognition result is non-standard.

[0067] The standardized time interval is dynamically generated through the following steps (steps Sa to Sh):

[0068] Step Sa: Collect time-series samples of each monitoring item during historical normal loading and unloading operations to form a time-series dataset.

[0069] Specifically, electronic devices collect a large amount of data from completed and deemed standard loading and unloading operations, recording the start and end times or durations of each monitoring item. These time-series samples form the basis for subsequent analysis and model training, ensuring that the model can learn "normal" patterns from actual operations.

[0070] Step Sb: Using a dynamic time warping algorithm, the start and end times of the same monitoring item in different loading and unloading operations are mapped to a common time axis to flexibly align the time series dataset on the time axis.

[0071] Specifically, the dynamic time warping algorithm can effectively handle differences in the duration or rhythm of the same monitoring item in different work batches, and can optimally match and align them in the time dimension even if the operation speed varies. Flexible alignment eliminates the stretching or compression of time series caused by factors such as operator habits and equipment status, enabling meaningful statistical analysis of the aligned data.

[0072] Step Sc: On the aligned common time axis, establish a kernel density estimation model to obtain the probability density function of the duration of each monitoring item.

[0073] Specifically, kernel density estimation is used to estimate the probability distribution of the duration of each monitoring item from the aligned time series data. It can produce a smooth probability density function curve, which clearly shows which durations are the most common and most compliant with the specifications. This avoids making pre-set assumptions about the data distribution and more accurately reflects the actual situation.

[0074] Step Sd: Use the highest density interval of the probability density function as the basic time interval. The highest density interval is the longest connected interval formed by all duration points whose probability density function values ​​are greater than or equal to the maximum value.

[0075] Specifically, the most core and typical duration range is extracted from the probability density function as a basis. By setting a preset percentage, such as 80% of the maximum probability density value, the most frequently occurring duration points that best represent normal operation can be identified, and the longest continuous interval can be selected from them.

[0076] Step Se: Determine the complexity based on the current vehicle model, the type of loading medium, and the experience level of the current operators. Statistics on all complexities in normal historical loading and unloading operations of the median.

[0077] Among them, determining the complexity The steps include (steps Se1 to Se5):

[0078] Step Se1: Map the current vehicle model, type of loading medium, and experience level of the current operator into numerical feature values.

[0079] Specifically, the vehicle model feature value is obtained from a preset vehicle model complexity mapping table based on parameters such as vehicle length, wheelbase, and tank volume. The value ranges from 0 to 1, with a larger value indicating a more complex vehicle model.

[0080] Loading medium characteristic value: It is obtained by weighted summation and normalization based on the medium's flash point, viscosity, volatility and other safety parameters. The value ranges from 0 to 1. The larger the value, the more dangerous the medium or the higher the operation requirements.

[0081] Operator experience level characteristic value: Based on the operator's cumulative loading and unloading times, number of violations in the past three months, and skills assessment scores, it is calculated using an exponential decay function, with a value range of 0 to 1. The larger the value, the more experienced the operator.

[0082] Step Se2: Weight the sum of the vehicle model characteristic value, the loading medium characteristic value, and the operator experience level characteristic value to obtain the initial complexity.

[0083] Step Se3: Calculate the average deviation rate between the actual time taken by the current operator to complete each monitoring item in the past preset number of loading and unloading operations and the base time interval predicted by the kernel density estimation model.

[0084] Specifically, for each monitoring item in each task, if the actual completion time is within the base time interval, the deviation is 0; if it is outside the interval, the deviation is (actual time - nearest boundary) / interval length. If the preset number of times is 20, the average deviation of all monitoring items and all tasks is taken.

[0085] Step Se4: If the average deviation rate is greater than the preset correction threshold, multiply the initial complexity by (1 + average deviation rate) to obtain the final complexity. , will increase complexity Limit to 1 or less.

[0086] Specifically, the preset correction threshold is 0.2, or 20%.

[0087] If the average deviation rate is greater than 0.2, the initial complexity is multiplied by (1 + average deviation rate), and the result is limited to no more than 1, resulting in the final complexity = min[1, initial complexity × (1 + average deviation rate)].

[0088] Step Se5: If the average deviation rate is not greater than the preset correction threshold, then the initial complexity is used as the final complexity. .

[0089] Step Sg: Based on complexity Stretching the base time interval includes: If complexity If the value is higher than the median, then the right boundary of the base time interval will be increased. This yields the normalized time interval; if the complexity... If the value is below the median, the left boundary of the base time interval will be reduced. This yields the standardized time interval; among which, The length of the base time interval, The first preset proportional coefficient, This is the second preset proportional coefficient, where 0 < < <0.5.

[0090] Specifically, when the task complexity *c* is higher than the historical median, it means the task may take longer. Therefore, the right boundary of the base time interval is expanded outwards, allowing for a longer completion time. Conversely, when the complexity *c* is lower, the task may be completed faster, so the left boundary is contracted inwards, requiring a faster completion speed. (Preset scaling factor) and And the relationship between them, 0 < < The value <0.5 ensures the rationality and safety of the adjustment, that is, it gives a greater time margin when the complexity is high, while it imposes stricter requirements on efficiency when the complexity is low.

[0091] Step Sh: After each loading and unloading operation is completed, the actual completion time of this operation is used as a new sample to incrementally update the kernel density estimation model.

[0092] Step S25: When the recognition results of the first neural network and the second neural network are inconsistent, the arbitration network is started. The input is static semantic features and dynamic behavioral features, and the output is the recognition result after arbitration. The recognition result after arbitration is compared with the standard loading and unloading process sequence again to obtain the final recognition result.

[0093] Specifically, when the recognition results of the first and second neural networks contradict each other, the arbitration network employs a graph convolutional network structure, which can effectively handle unstructured data relationships, such as complex logical connections, temporal dependencies, and spatial relationships between different monitoring items. Simultaneously, the arbitration network is a spatiotemporal joint model, capable of combining historical time-series information and spatial context information for comprehensive judgment. Deep fusion of static semantic features and dynamic behavioral features can be performed at the feature level, for example, by combining the two feature vectors through concatenation, attention mechanisms, or gating mechanisms, providing the arbitration network with a more comprehensive decision-making basis.

[0094] The arbitration network outputs a comprehensive assessment result. To ensure the operational validity of this arbitration result, it is compared again with a pre-defined standard loading and unloading procedure sequence. This second comparison ensures that even in complex scenarios, the conclusion reached through the advanced arbitration mechanism conforms to actual operational procedures, thus yielding a final, highly reliable identification result.

[0095] In practical applications, the extracted features may still be affected by environmental noise, subtle variations in the target, or limitations in the model's generalization ability, leading to insufficient effectiveness or reliability of the features. In such cases, directly identifying features based on these potentially flawed features may produce inaccurate or low-confidence identification results, thereby affecting the accuracy of risk management and the reliability of decision-making.

[0096] Therefore, before outputting the recognition result, the first neural network or the second neural network performs the following optimization on the extracted features (steps S31 to S34):

[0097] Step S31: Calculate the radial distance between the static semantic features or dynamic behavioral features and each template vector in the preset standard template library.

[0098] Specifically, the pre-set standard template library stores multiple template vectors, each representing a typical or ideal feature pattern for a specific monitoring item, such as standardized vehicle parking positions or correct wheel chock placement. The radial distance calculation aims to quantify the similarity or deviation between the currently extracted features and these standard templates. For example, various distance metrics such as Euclidean distance, Mahalanobis distance, or cosine distance can be used. Euclidean distance intuitively reflects the geometric distance between feature vectors in multidimensional space; the smaller the distance, the higher the similarity. The pre-set standard template library is built and continuously updated through feature extraction, cluster analysis, and expert review of a large amount of manually labeled and verified standardized operation video data.

[0099] Step S32: Multiply the radial distance by a preset scaling factor to obtain the centrifugal distance.

[0100] Specifically, in order to map the raw radial distance to a more discriminative and interpretable scale and to provide a sensitive input for subsequent confidence assessment, this application multiplies the calculated radial distance by a preset scaling factor, such as 10, to obtain the centrifugal distance.

[0101] The scaling factor can be adjusted according to the characteristics of feature distribution in the actual application scenario and the requirements for recognition confidence sensitivity. Through linear amplification, even small changes in radial distance can be more clearly reflected in centrifugal distance, thereby enhancing the sensitivity of subsequent judgments.

[0102] Step S33: Determine the confidence level based on the centrifugation distance: the smaller the centrifugation distance, the higher the confidence level; the larger the centrifugation distance, the lower the confidence level.

[0103] Specifically, a smaller centrifugation distance indicates that the extracted features are highly similar to the preset standard template, meaning that the features are valid and standardized, and therefore the corresponding recognition results have a high degree of confidence. Conversely, a larger centrifugation distance indicates that the extracted features deviate significantly from the standard template, and their validity or standardization is questionable, thus the corresponding recognition results have a lower degree of confidence.

[0104] The confidence level can be determined using a monotonically decreasing function. For example, a function can be designed to map the centrifugal distance to a confidence level range of 0 to 100%, where the confidence level is highest when the centrifugal distance is 0, and gradually decreases as the centrifugal distance increases.

[0105] Step S34: If the centrifugal distance exceeds the preset distance threshold, the static semantic feature or dynamic behavioral feature is determined to be an invalid feature and is not used for recognition; if the centrifugal distance does not exceed the preset distance threshold, the feature is retained and the recognition result with the highest confidence is output.

[0106] Specifically, to avoid inputting low-quality, abnormal, or significantly different features from the standard pattern into subsequent recognition processes, this application sets a preset distance threshold. When the calculated centrifugal distance exceeds the preset distance threshold, the system immediately determines that the currently extracted feature is invalid and will no longer be used in subsequent recognition processes, thereby effectively preventing potential misidentification. The preset distance threshold can be optimized by statistical analysis of a large amount of historical data, combined with expert experience, or through machine learning methods.

[0107] When the centrifugal distance does not exceed a preset distance threshold, it indicates that the extracted features have sufficient validity and normalization, and can be retained and used for subsequent recognition. At this point, the first or second neural network will perform recognition based on these valid features, and will typically output multiple possible recognition results and their corresponding confidence levels. The electronic device selects the recognition result with the highest confidence level as the final output.

[0108] Step S105: If the identification result is normal, then the next monitoring item in the sequence is taken as the new current monitoring item, and steps S103 and S104 are executed.

[0109] Step S106: If the identification result is non-standard or no identification result is obtained within the preset time threshold, a push message is generated and sent to the management personnel. The loading and unloading equipment is automatically stopped, and steps S103 and S104 are re-executed until the identification result is standard. The electronic device has a preset timeout for a single identification, for example, 2 minutes. If no valid identification result is obtained within 2 minutes, the identification timeout occurs.

[0110] Step S107: If the current monitoring item is the last monitoring item in the sequence and the identification result is standard, then the automatic linkage control of the loading and unloading equipment will start or continue to operate.

[0111] Furthermore, the above method also includes (steps S41 to S44):

[0112] Step S41: Preset deterministic association rules based on process logic to define the triggering relationship that non-standard behavior of the preceding monitoring item will inevitably lead to non-standard behavior of the subsequent monitoring item.

[0113] Deterministic association rules include: If electrostatic grounding is not up to standard, then the connection of the oil and gas recovery pipe is not allowed to be up to standard. If the handover of vehicle keys is not in accordance with regulations, then the placement of wheel chocks and static grounding are not allowed to be in accordance with regulations. If the vehicle is parked in an irregular position, the wheel chocks must not be placed in a standard position.

[0114] Specifically, pre-setting at least one deterministic association rule based on process logic refers to manually or through an expert system setting a series of mandatory association conditions according to the strict safety procedures and operational logic of chemical loading and unloading operations. That is, when a certain pre-monitoring item is non-compliant, subsequent monitoring items with a necessary logical relationship with it will also necessarily be non-compliant. For example, in the chemical loading and unloading process, if the electrostatic grounding operation is not standardized, the safety compliance of the oil and gas recovery pipe connection cannot be guaranteed. Therefore, a rule can be pre-set that "if electrostatic grounding is non-compliant, then the oil and gas recovery pipe connection is not allowed to be compliant." These rules are usually stored in the form of structured data, such as database tables or configuration files, so that the system can quickly query and match them during runtime.

[0115] Step S42: Construct probabilistic association rules based on historical data, including:

[0116] Collect historical non-standard records and obfuscated parameters. Each historical non-standard record includes the non-standard monitoring item and the timestamp of occurrence. Obfuscated parameters include the operator's identity, work period, and vehicle model at the time the historical non-standard record occurred.

[0117] Specifically, the electronic equipment constructs at least one probabilistic association rule based on historical data. By mining a large amount of historical loading and unloading operation data, it discovers non-necessary but statistically significant associations between different non-standard monitoring items. The electronic equipment first collects records of non-standard monitoring items from historical loading and unloading operations. These records include detailed information such as the name of the non-standard monitoring item, the timestamp of the occurrence, and the environmental and operational background information at the time, i.e., obfuscation parameters, such as operator identity, operation time (day / night, shift), vehicle model, etc.

[0118] Any two non-standard monitoring items X and Y are taken as temporary association pairs. Stratified statistics are performed according to the confusion parameter. The conditional probability of Y occurring within a preset time window after X occurs is calculated within the same stratum. The conditional probabilities of each stratum are summarized to obtain the corrected net association probability.

[0119] Specifically, to eliminate the influence of confounding parameters on correlation judgment, electronic devices perform stratified statistics according to these confounding parameters. For example, under the conditions of the same operator, the same working period, and the same vehicle model, the conditional probability of the subsequent non-standard monitoring item Y occurring within a preset time window after the occurrence of the preceding non-standard monitoring item X is calculated. By summing up the conditional probabilities calculated from each stratum, the corrected net correlation probability is obtained. The net correlation probability more accurately reflects the intrinsic correlation strength between X and Y.

[0120] When the net correlation probability is greater than the preset probability threshold and the number of samples with temporary correlation pairs in the non-standard monitoring item records reaches the sample threshold, the temporary correlation pairs are determined as probabilistic correlation rules, and the correlation time window is recorded.

[0121] Specifically, when the net association probability exceeds a preset probability threshold, such as 0.6 or 0.7, and the sample size of the non-standard monitoring items supporting the association pair reaches the sample threshold, thus meeting the statistical requirements (e.g., at least 50 records of the co-occurrence of X and Y), the temporary association pair is confirmed as a probabilistic association rule, and its corresponding association time window is recorded, such as within 5 minutes after X occurs. These rules are also stored in the database for subsequent real-time matching and prediction.

[0122] Step S43: When an irregularity is detected in the current monitored item, a deterministic association rule is matched: If a match is found, it is inferred that subsequent monitoring items must be non-compliant. If the loading and unloading equipment has not yet stopped due to the non-compliance of the current monitoring item, the loading and unloading equipment will be stopped in conjunction with the control.

[0123] In this situation, if the loading and unloading equipment has not yet stopped due to the non-compliance of the current monitoring item, the electronic equipment will immediately link up to control the loading and unloading equipment to stop, thereby forcibly interrupting the operation process that may have serious consequences and ensuring safety.

[0124] If the matching is inconsistent, a probabilistic association rule is used: calculate the predicted probability of the subsequent monitoring item being non-standard within the association time window after the current monitoring item is non-standard; if the predicted probability is greater than the preset probability threshold, the recognition frequency of the deep learning model is actively increased within the association time window, and a predictive warning is pushed.

[0125] Specifically, after the current monitoring item is identified as non-standard, the predicted probability that the subsequent monitoring item will also be identified as non-standard within the associated time window is calculated, including (steps S51 to S55):

[0126] Step S51: From the probabilistic association rules, read the historical conditional probability of the occurrence of the subsequent non-standard monitoring item Y within the association time window after the occurrence of the preceding non-standard monitoring item X.

[0127] Specifically, historical conditional probabilities are derived from statistical analysis of a large amount of historical loading and unloading operation data. They reflect the probability that a specific subsequent non-standard behavior Y will occur after a preceding non-standard behavior X has occurred, under normal circumstances. After identifying the preceding non-standard monitoring item X, the system queries the stored probabilistic association rule database to retrieve all associated subsequent non-standard monitoring items Y and their corresponding historical conditional probabilities.

[0128] Step S52: Starting from the moment when the previous non-standard monitoring item X occurs, determine the time decay coefficient.

[0129] The time decay coefficient is initially 1, and then decreases linearly over time until it reaches 0 at the end of the associated time window. If the same previous non-standard monitoring item X is detected again within the associated time window, the time decay coefficient is reset to the initial value.

[0130] Specifically, electronic devices maintain a timer or timestamp recording mechanism that records the occurrence time when a non-compliant monitoring item X is identified. Within the associated time window, the time decay coefficient is dynamically calculated based on the difference between the current time and the event occurrence time to reflect the immediacy of the risk impact.

[0131] Step S53: Determine the performance coefficient based on the operator's historical operation records.

[0132] The actual frequency of non-standard monitoring item Y in the past loading and unloading operations of the current on-duty operator is counted and divided by the average frequency of all personnel in the plant to obtain the relative error rate of the operator. If the relative error rate is greater than 1, the performance coefficient of the operator is determined to be greater than 1. If the relative error rate is not greater than 1, the performance coefficient of the operator is determined to be less than 1.

[0133] Specifically, the electronic equipment maintains an operator behavior database, recording the compliance or non-compliance of each operator in each monitoring item during loading and unloading operations. Within this database, multiple performance coefficients corresponding to relative error rates are set. By comparing the current operator's performance on specific non-compliant behaviors with the average level, the individual risk level of each operator can be quantified.

[0134] For example, the average frequency of the "oil and gas recovery pipe not connected" anomaly for all operators in the plant is 5%. The current operator "Zhang San" has experienced the "oil and gas recovery pipe not connected" anomaly twice in the past 20 loading and unloading operations.

[0135] Zhang San's actual frequency = 2 times / 20 times = 10% = 0.10; The average frequency of the entire plant = 5% = 0.05; The relative error rate is 0.10 / 0.05 = 2.0 (greater than 1).

[0136] The performance coefficient is set to 1.5 according to the preset mapping table, which means that Zhang San is more likely to make this type of mistake than the average person.

[0137] For example, Li Si encountered zero "oil and gas recovery pipe not connected" anomalies in the past 20 operations.

[0138] Actual frequency = 0%; The relative error rate = 0 / 0.05 = 0 (less than 1).

[0139] The performance coefficient is set to 0.5, indicating that Li Si is more reliable than the average person.

[0140] Step S54: Based on the current work period, weather conditions, and whether the project is scheduled to be completed, retrieve the corresponding risk multiplier from historical data.

[0141] Specifically, for example, the current work period can be either a day shift or a night shift, with the risk multiplier being relatively higher for night shifts; weather conditions such as sunny, rainy, or snowy weather also contribute to the higher risk multiplier; and the risk multiplier is relatively higher when there is a tight deadline. The electronic device maintains a database mapping environmental conditions and work conditions to risk multipliers. This database can be built through historical data analysis or expert experience. When making a prediction, the electronic device obtains information such as the current work period, weather conditions, and whether there is a tight deadline, and then queries this database to obtain the corresponding risk multiplier, taking into account the impact of external environment and work conditions on risk.

[0142] For example, the database uses day shift + sunny days + non-rush period as the baseline, with a historical anomaly rate of 5% and a risk multiple of 1; while the historical anomaly rate for night shift + rain / snow + rush period is 11%, so the risk multiple of night shift + rain / snow + rush period relative to the baseline is 11% / 5% = 2.2.

[0143] Step S55: Multiply the historical conditional probability by the time decay coefficient, the personnel performance coefficient, and the risk multiple in sequence to obtain the product result, and limit the product result to the range of 0 to 1 to obtain the predicted probability.

[0144] For example, electronic devices multiply the historical conditional probability by the time decay coefficient, the personnel performance coefficient, and the risk multiplier in sequence to obtain the product result. If the product result is greater than 1, it is taken as 1; if it is less than 0, it is taken as 0; otherwise, the original value is retained.

[0145] Step S44: If the predicted probability is greater than the preset probability threshold, the recognition frequency of the deep learning model will be actively increased to N times the preset normal frequency within the associated time window, where N≥3, and a predictive warning will be pushed.

[0146] If the calculated predicted probability exceeds a preset probability threshold (e.g., 0.5), indicating a high risk of non-compliance in subsequent monitoring items, the electronic device will immediately and proactively increase the recognition frequency of the deep learning model to more than three times the preset normal frequency within the associated time window. For example, from one frame per second to more than three frames per second, to achieve denser and more refined monitoring. Simultaneously, the electronic device will also push predictive warnings to management personnel, reminding them to pay attention to potential problems so that timely manual intervention or preventative measures can be taken.

[0147] To better implement the above method, this application also provides an intelligent AI-based risk management device for loading and unloading chemical vehicles, referring to... Figure 2 The 200-unit intelligent AI loading and unloading risk management device for chemical vehicles includes: The vehicle verification module 201 is used to determine whether a vehicle is allowed to enter the site based on the vehicle's entry registration information. The monitoring item sorting module 202 is used to sort the vehicles allowed to enter the site based on each monitoring item to obtain a sequence, and take the first monitoring item in the sequence as the current monitoring item; The video information acquisition module 203 is used to acquire video information of the vehicle in the loading and unloading operation area for the current monitoring item; The monitoring item analysis module 204 is used to analyze the video information of the current monitoring item based on a deep learning model to obtain the corresponding recognition results; The first identification result execution module 205 is used to take the next monitoring item in the sequence as the new current monitoring item when the identification result of the monitoring item analysis module is normal, and continue to execute the video information acquisition module 203 and the monitoring item analysis module 204. The second identification result execution module 206 is used to generate push information to the management personnel when the identification result of the monitoring item analysis module is non-standard or no identification result is obtained within the preset time threshold, and automatically link to control the loading and unloading equipment to stop, and re-execute the video information acquisition module 203 and the monitoring item analysis module 204 until the identification result is standard. The third identification result execution module 207 is used to automatically control the loading and unloading equipment to start or continue operating when the current monitoring item is the last monitoring item in the sequence and the identification result is standard.

[0148] The various variations and specific examples of the methods in the foregoing embodiments are also applicable to the intelligent AI loading and unloading risk control device for chemical vehicles in this embodiment. Through the foregoing detailed description of the intelligent AI loading and unloading risk control method for chemical vehicles, those skilled in the art can clearly understand the implementation method of the intelligent AI loading and unloading risk control device for chemical vehicles in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0149] To better implement the above methods, embodiments of this application provide an electronic device, referring to... Figure 3 The electronic device 300 includes a processor 301, a memory 303, and a display screen 305. The memory 303 and the display screen 305 are both connected to the processor 301, such as via a bus 302. Optionally, the electronic device 300 may also include a transceiver 304. It should be noted that in practical applications, the transceiver 304 is not limited to one type, and the structure of this electronic device 300 does not constitute a limitation on the embodiments of this application.

[0150] Processor 301 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 301 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0151] Bus 302 may include a pathway for transmitting information between the aforementioned components. Bus 302 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 302 may be divided into address bus, data bus, control bus, etc.

[0152] The memory 303 may be a ROM (Read Only Memory) or other type of static storage device capable of storing static information and instructions, RAM (Random Access Memory) or other type of dynamic storage device capable of storing information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0153] The memory 303 is used to store application code that executes the solution of this application, and its execution is controlled by the processor 301. The processor 301 is used to execute the application code stored in the memory 303 to implement the content shown in the foregoing method embodiments.

[0154] Figure 3 The electronic device 300 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0155] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the program implements the intelligent AI loading and unloading risk management method for chemical vehicles provided in the above embodiments. By acquiring real-time vehicle entry information and video information of the loading and unloading operation area, and using a deep learning model to automatically identify key monitoring items, it can promptly detect non-standard behaviors in loading and unloading operations and automatically control the loading and unloading equipment to stop. At the same time, it pushes warning information to management personnel, thereby effectively making up for the limitations of traditional systems in post-event follow-up. It realizes real-time, automatic, and precise safety intervention in the chemical loading and unloading operation process, significantly reduces operational risks, and improves the overall safety management level.

[0156] In this embodiment, the computer-readable storage medium can be a tangible device that holds and stores instructions used by an instruction execution device. The computer-readable storage medium can be, but is not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any combination thereof. Specifically, the computer-readable storage medium can be a portable computer disk, a hard disk, a USB flash drive, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory stick, floppy disk, optical disk, magnetic disk, mechanical encoding device, or any combination thereof.

[0157] The computer program in this embodiment includes program code for performing all the aforementioned methods. The program code may include instructions corresponding to the method steps provided in the above embodiments. The computer program can be downloaded from a computer-readable storage medium to various computing / processing devices, or downloaded to an external computer or external storage device via a network (e.g., the Internet, a local area network, a wide area network, and / or a wireless network). The computer program can be executed entirely on the user's computer as a standalone software package.

[0158] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

[0159] Additionally, it should be understood that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

Claims

1. A method for intelligent AI-based risk management and control of chemical vehicle loading and unloading, characterized in that, include: Determine whether a vehicle is allowed to enter based on its entry registration information; For vehicles allowed to enter, a sequence is obtained by sorting each monitoring item, and the first monitoring item in the sequence is taken as the current monitoring item; Acquire video information of the vehicle in the loading and unloading area for the current monitoring item; The video information of the current monitoring item is analyzed based on a deep learning model to obtain the corresponding recognition result; If the identification result is normal, then the next monitoring item in the sequence is taken as the new current monitoring item, and the steps of video information acquisition and video information analysis for the current monitoring item are continued. If the identification result is non-standard or no identification result is obtained within the preset time threshold, a push message is generated and sent to the management personnel, and the loading and unloading equipment is automatically stopped and the video information acquisition and video analysis steps are re-executed until the identification result is standard. If the current monitoring item is the last monitoring item in the sequence and the identification result is standard, then the automatic linkage control will start or continue to operate the loading and unloading equipment.

2. The method according to claim 1, characterized in that, Before analyzing the video information based on the deep learning model, the method further includes adaptive neural network scheduling, including: The scene features of the current video information are analyzed in real time. The scene features include: image clarity, target occlusion ratio, light intensity, and the action type corresponding to the current monitoring item. The action type includes static targets and dynamic behaviors. Automatically select the target neural network model based on the scene characteristics: If the current monitoring item only needs to identify static targets, and the image clarity is higher than the preset clarity threshold and the image is unobstructed, then the image classification network is selected as the first neural network. If the current monitoring item needs to identify dynamic behavior, then a temporal convolutional network is selected as the second neural network; If the scene features meet the first condition, the arbitration network is activated. The arbitration network is a spatiotemporal joint model based on graph convolutional networks. The outputs of the first neural network and the second neural network are simultaneously input into the arbitration network for fusion recognition. The first condition includes: partial occlusion of the target, insufficient lighting, or joint judgment of static and dynamic behaviors.

3. The method according to claim 2, characterized in that, If the scene features satisfy the first condition, the analysis of the video information of the current monitoring item based on the deep learning model to obtain the corresponding recognition result includes: The first neural network and the second neural network are run in parallel, wherein the first neural network extracts static semantic features related to the current monitoring item from a single frame image; and the second neural network extracts dynamic behavioral features related to the current monitoring item from multiple consecutive frames of images. When the identification results of the first neural network and the second neural network for the current monitoring item are consistent, a preliminary identification result for the current monitoring item is generated; The preliminary identification results are compared with the specifications of the corresponding monitoring items in the preset standard loading and unloading process sequence to obtain the comparison results. The standard loading and unloading process sequence defines the legal order of each monitoring item and the specified time interval of each monitoring item. Output the identification result based on the comparison result; When the recognition results of the first neural network and the second neural network are inconsistent, the arbitration network is activated. The input is the fusion amount of the static semantic features and the dynamic behavioral features, and the output is the arbitrated recognition result. The arbitrated recognition result is then compared with the standard loading and unloading process sequence to obtain the final recognition result.

4. The method according to claim 2, characterized in that, Before the first neural network or the second neural network outputs the recognition result, the method further includes: Calculate the radial distance between the static semantic feature or the dynamic behavioral feature and each template vector in the preset standard template library; Multiply the radial distance by a preset scaling factor to obtain the centrifugal distance; The confidence level is determined based on the centrifugal distance: the smaller the centrifugal distance, the higher the confidence level; the larger the centrifugal distance, the lower the confidence level. If the centrifugal distance exceeds a preset distance threshold, the static semantic feature or the dynamic behavioral feature is determined to be an invalid feature and is not used for identification; if the centrifugal distance does not exceed the preset distance threshold, the static semantic feature or the dynamic behavioral feature is retained, and the identification result with the highest confidence is output.

5. The method according to claim 3, characterized in that, The standardized time interval is dynamically generated through the following steps: Collect time-series samples of each monitoring item during historical normal loading and unloading operations to form a time-series dataset; A dynamic time warping algorithm is used to map the start and end times of the same monitoring item in different loading and unloading operations to a common time axis, so as to flexibly align the time series dataset on the time axis. On the aligned common time axis, a kernel density estimation model is established to obtain the probability density function of the duration of each monitoring item; The highest density interval of the probability density function is used as the basic time interval, and the highest density interval is the longest connected interval formed by all duration points whose probability density function values ​​are greater than or equal to the maximum value. The complexity is determined based on the current vehicle model, the type of loading medium, and the experience level of the current operators. Statistics on all complexities in normal historical loading and unloading operations the median; Based on the aforementioned complexity Stretching the base time interval includes: If complexity If the value is higher than the median, the right boundary of the base time interval will be increased. This yields the normalized time interval; if the complexity... If the value is below the median, the left boundary of the base time interval will be reduced. This yields the standardized time interval; among which, The length of the base time interval, The first preset proportional coefficient, This is the second preset proportional coefficient, where 0 < < <0.5; After each loading and unloading operation is completed, the actual completion time is used as a new sample to incrementally update the kernel density estimation model.

6. The method according to claim 1, characterized in that, The method further includes: Pre-defined deterministic association rules based on process logic are used to define the triggering relationship that non-standard behavior of a preceding monitoring item will inevitably lead to non-standard behavior of a subsequent monitoring item. Constructing probabilistic association rules based on historical data includes: collecting historical non-standard records and confusion parameters; taking any two non-standard preceding detection items and non-standard subsequent detection items as temporary association pairs; performing hierarchical statistical analysis of the conditional probabilities corresponding to the temporary association pairs according to the confusion parameters; and obtaining the net association probability after correction. If the net association probability and the sample size of the temporary association pairs reach the sample threshold, the temporary association pairs are determined as probabilistic association rules, and the association time window is recorded. When a non-standard monitoring item is detected, the deterministic association rule is matched. If a match is found, it is inferred that subsequent monitoring items will inevitably be non-standard, and the loading and unloading equipment is stopped in a coordinated manner. If no match is found, the probabilistic association rule is matched: the predicted probability of the subsequent monitoring item being non-standard within the association time window is calculated. If the predicted probability is greater than a preset probability threshold, the recognition frequency of the deep learning model is increased within the association time window, and a predictive warning is pushed.

7. The method according to claim 6, characterized in that, The calculation of the predicted probability that the identification result of the current monitored item is non-standard, and the identification result of the subsequent monitored item will be non-standard within the associated time window, includes: From the probabilistic association rules, read the historical conditional probability of the occurrence of the subsequent non-standard monitoring item Y within the association time window after the occurrence of the preceding non-standard monitoring item X. Determine the time decay coefficient starting from the moment the aforementioned non-standard monitoring item X occurs; The performance coefficient is determined based on the operator's historical operation records; Based on the current work period, weather conditions, and whether the project is scheduled to be completed on time, the corresponding risk multiplier can be obtained from historical data. The historical conditional probability is multiplied sequentially by the time decay coefficient, the personnel performance coefficient, and the risk multiplier to obtain the product result, and the product result is restricted to the range of 0 to 1 to obtain the predicted probability.

8. A smart AI-powered risk management device for loading and unloading chemical vehicles, characterized in that, include: The vehicle verification module is used to determine whether a vehicle is allowed to enter the site based on its entry registration information. The monitoring item sorting module is used to sort the vehicles allowed to enter the site based on each monitoring item to obtain a sequence, and to take the first monitoring item in the sequence as the current monitoring item; The video information acquisition module is used to acquire video information of the vehicle in the loading and unloading operation area for the current monitoring item; The monitoring item analysis module is used to analyze the video information of the current monitoring item based on a deep learning model to obtain the corresponding recognition results; The first identification result execution module is used to, when the identification result of the monitoring item analysis module is normal, take the next monitoring item in the sequence as the new current monitoring item and continue to execute the video information acquisition module and the monitoring item analysis module; The second identification result execution module is used to generate push information to the management personnel when the identification result of the monitoring item analysis module is non-standard or no identification result is obtained within a preset time threshold, and automatically link to control the loading and unloading equipment to stop, and re-execute the video information acquisition module and the monitoring item analysis module until the identification result is standard. The third identification result execution module is used to automatically control the loading and unloading equipment to start or continue operating when the current monitoring item is the last monitoring item in the sequence and the identification result is standard.

9. An electronic device, characterized in that, include: At least one processor; Memory; At least one computer program, wherein the at least one computer program is stored in the memory and configured to be executed by the at least one processor, the at least one computer program being configured to perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, A computer program is stored that can be loaded by a processor and used in parallel with the method as described in any one of claims 1 to 7.

Images