A hazardous waste warehouse monitoring and identifying method based on video image analysis
By constructing a risk event status identification model for hazardous waste warehouses through video image analysis, and combining fuzzy clustering and multi-factor decomposition, accurate monitoring and early identification of risks in hazardous waste warehouses are achieved. This solves the problems of delayed risk identification and false alarms/missed alarms in traditional methods, and improves the real-time performance and accuracy of early warning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SUZHOU LUZE ENVIRONMENTAL PROTECTION TECHNOLOGY CO LTD
- Filing Date
- 2026-01-07
- Publication Date
- 2026-06-05
AI Technical Summary
Traditional hazardous waste warehouse monitoring methods lack a quantitative description of the dynamic development process of risk events, making it difficult to achieve early warning. Furthermore, they cannot effectively analyze the nonlinear coupling relationship between multiple factors, resulting in high false alarm and false alarm rates and insufficient warning accuracy.
A video image analysis-based approach is adopted to construct a risk event status identification model by acquiring historical and real-time video data of hazardous waste warehouses. Fuzzy clustering and factor decomposition are performed, multi-task learning is carried out using a Transformer encoder, and hierarchical early warning is achieved by combining Bayesian update and DS evidence theory.
It enables accurate perception and early identification of the risk status of hazardous waste warehouses, improves the real-time and intelligent level of monitoring, reduces false alarm and missed alarm rates, and improves the accuracy of early warning and decision support.
Smart Images
Figure CN122157078A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data monitoring technology, specifically to a method for monitoring and identifying hazardous waste warehouses based on video image analysis. Background Technology
[0002] Traditional methods for monitoring and identifying hazardous waste warehouses mainly rely on manual inspections and single sensor threshold alarms, which have significant drawbacks. On the one hand, they lack a quantitative description of the dynamic development process of risk events, making it difficult to achieve early warnings. They often only trigger responses after the risks have manifested, resulting in a significant lag. On the other hand, existing technologies are mostly based on isolated risk factors or simple linear superposition for judgment, which cannot effectively analyze the nonlinear coupling relationship and evidence conflicts between multiple factors. This leads to high false alarm and false alarm rates and insufficient warning accuracy, making it difficult to meet the urgent need for real-time, accurate, and intelligent risk management in complex hazardous waste warehouse environments. Summary of the Invention
[0003] To address the aforementioned technical issues, this paper provides a method for monitoring and identifying hazardous waste warehouses based on video image analysis. This technical solution solves the urgent need for real-time, accurate, and intelligent risk management in complex hazardous waste warehouse environments.
[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for monitoring and identifying hazardous waste warehouses based on video image analysis, comprising: S1. Obtain historical risk event trigger image data of each hazardous waste warehouse, mark the risk events in the historical risk event trigger image data to support multi-factor sets, construct the status identification model of each risk event of the hazardous waste warehouse, and generate the development vector of each risk event of the hazardous waste warehouse. S2. Obtain real-time image data of the hazardous waste warehouse, perform fuzzy clustering according to the development interval vector of each risk event in the hazardous waste warehouse, determine the category to which the real-time image data of the hazardous waste warehouse belongs, substitute it into the status recognition model of each risk event in the hazardous waste warehouse, and generate the real-time development vector of each risk event in the hazardous waste warehouse. S3. Based on the real-time development vectors of various risk events in the hazardous waste warehouse, decompose the real-time supporting factors for each risk event, verify the risk score of each real-time supporting factor, construct a risk coupling model, verify the trigger probability of each risk event under the superposition of real-time supporting factors, determine the preset graded early warning threshold range, and trigger the corresponding level alarm.
[0005] Preferably, step S1 specifically includes: Based on the records of the enterprise's EHS management system, access the historical monitoring videos of the hazardous waste warehouse, obtain the video stream data of the hazardous waste warehouse, perform data preprocessing, extract the recorded time points of historical risk events, mark each historical risk event with a specific timestamp, extract the video stream data of the hazardous waste warehouse t minutes before the occurrence of each historical risk event, and generate a set of timestamp-historical risk event video segments. The risk events mentioned include: leakage, fire, and improper storage; Based on a set of timestamp-historical risk event video clips, a frame-by-frame sampling and analysis method is designed to perform analysis at 1 frame per second. Using a target detection algorithm, the location of different types of hazardous waste containers (barrels, bags, and cans), designated stacking area boundaries, safety passages, workers, and handling equipment is identified. Statistical analysis and relational encoding are performed to obtain a target list with category and location for each frame, generating spatial features of each risk event in the timestamp-historical risk event video clip set.
[0006] Preferably, step S1 further includes: The timestamp-historical risk event video clip set is cropped into fixed-length video clips, input into the I3D network, and high-dimensional feature vectors representing dynamic changes within the video clips are obtained. The temporal features of each risk event in the timestamp-historical risk event video clip set are then extracted. Based on a set of video clips of timestamp-historical risk events, the images are converted from RGB color space to HSV color space. Using a Gaussian mixture model, an image model of the static background of the hazardous waste warehouse floor and walls is built. The current frame is compared with the image model of the static background to identify areas with abnormal color changes. A threshold h is set. If the color change exceeds the threshold h, it is marked as a potential leakage area. Based on a set of timestamp-historical risk event video clips, motion vectors are calculated between each frame in the video clip set to identify smoke motion features. The LBP method is used to extract texture features from each frame in the video clip set to detect subtle texture changes in the image and identify smoke signs. Using infrared cameras, temperature readings from timestamp-historical risk event video clips are extracted to obtain their temperature characteristics. Combined with potential leak areas and smoke signs, environmental characteristics of each risk event in the timestamp-historical risk event video clip set are generated.
[0007] Preferably, step S1 further includes: By integrating and extracting the spatial, temporal, and environmental characteristics of each risk event from the historical risk event video clip set, a set of multiple factors supporting risk events in the triggering image data of each historical risk event is obtained. Using the set of multiple factors supporting risk events from historical risk event-triggered image data as input, a Transformer encoder is trained. The learning objectives are to predict the risk category caused by the development of the video segment set and the severity of the risk at the end of the video segment set. A loss function is designed, and the model parameters are continuously optimized through backpropagation. A risk event state recognition model for hazardous waste warehouses is constructed, and development vectors of each risk event in hazardous waste warehouses are generated.
[0008] Preferably, step S2 specifically includes: Based on the RTSP video streams from several cameras deployed in the hazardous waste warehouse, real-time image data of the hazardous waste warehouse is acquired, timestamps and frame numbers are aligned, and a dynamic background model of the hazardous waste warehouse is established using the ViBe algorithm. Real-time monitoring of foreground targets in the hazardous waste warehouse, including workers, handling equipment, and flowing liquids, is performed, and the real-time image data of the hazardous waste warehouse is preprocessed. Based on real-time image data of hazardous waste warehouses, and combined with the feature extraction technology of timestamp-historical risk event video clip sets, the spatial features, temporal features and environmental features of the image data of hazardous waste warehouses for each time period are extracted in real time, so as to obtain the multi-factor set of risk events supporting risk events in the real-time trigger image data of each risk event in hazardous waste warehouses.
[0009] Preferably, step S2 further includes: Based on the development vectors of each risk event in the hazardous waste warehouse, a risk event development vector library for each hazardous waste warehouse is established, which is a set of timestamp-historical risk event video clips. c points in this vector library are randomly selected as the initial cluster centers. For the risk events in the real-time image data of hazardous waste warehouses, which support multi-factor sets, the fuzzy C-means clustering algorithm is used to calculate the Euclidean distance from each real-time feature vector in the real-time image data of the hazardous waste warehouse to each initial cluster center. A fuzzy factor is set to calculate the membership degree of each risk event category and generate the membership degree matrix of the real-time image data of the hazardous waste warehouse.
[0010] Preferably, step S2 further includes: A threshold of 0.7 is set. If the membership degree of the real-time image data of the hazardous waste warehouse exceeds the threshold of 0.7, the real-time image data is determined to belong to the corresponding risk event category, thus determining the category to which the real-time image data of the hazardous waste warehouse belongs. Using real-time image data of hazardous waste warehouses and their respective risk event categories as input, the system associates and pairs risk event status recognition models for each hazardous waste warehouse, generates predictions of the risk categories caused by development in the image data and real-time development vectors of each risk event in the hazardous waste warehouse, calculates the similarity between the current risk status and historical events, and combines the predictions of the risk categories caused by development in the image data to obtain real-time development vectors of each risk event in the hazardous waste warehouse with confidence.
[0011] Preferably, step S3 specifically includes: Based on the real-time development vectors of various risk events in the hazardous waste warehouse, the SHAP analysis algorithm is used to calculate the contribution of each real-time supporting factor to the real-time development vector of the risk event, obtain the risk score of each real-time supporting factor, and identify the top k real-time supporting factors. Based on the identification of the top k real-time supporting factors, and combined with the set of supporting factors for risk events in the real-time triggered image data of each risk event in the hazardous waste warehouse, these factors are decomposed into three categories: spatial factors, temporal factors, and environmental factors. The contribution of the decomposed real-time supporting factors and the corresponding real-time development vectors of the risk events is summarized to generate a detailed report on the supporting factors.
[0012] Preferably, step S3 further includes: Based on the identified top k real-time support factors, risk scores for the top k real-time support factors are obtained. Using the analytic hierarchy process (AHP), an expert group is formed to construct a judgment matrix. The analysis scores of each real-time support factor are compared pairwise to calculate the initial risk weight vector of each real-time support factor. Through consistency testing, the initial risk weight of each real-time support factor is obtained. Different risk transformation functions are designed for different types of factors, including linear functions, sigmoid functions, and step functions, to obtain the risk score of each real-time support factor. Based on the Bayesian update algorithm and according to the historical early warning records of hazardous waste warehouses, a correct early warning is defined as a true positive, a false alarm is defined as a false positive, a missed alarm is defined as a false negative, and a correct warning without an early warning is defined as a true negative. For each false alarm and false negative event, the contribution of supporting factors is analyzed, the factors that cause false alarms are weighted less and the factors that cause false negatives are weighted more, and the risk score of each real-time supporting factor is verified by summarizing and updating once a month. Based on the risk score of each real-time support factor, the risk of each real-time support factor is grouped according to physical correlation. Temperature, smoke and open flame are defined as fire source related group, density, location and stability are defined as stacking related group, and humidity, ventilation and electricity are defined as environment related group. The possible combinations of the risk of each real-time support factor are calculated, and their geometric mean is obtained to obtain the coupled risk of all real-time support factors occurring at the same time, and a risk coupling model is constructed.
[0013] Preferably, step S3 further includes: Based on the coupling risk of all real-time supporting factors occurring simultaneously, using DS evidence theory, each real-time supporting factor is regarded as an evidence source. A basic probability allocation function is defined, and orthogonal sums are calculated for multiple evidence sources to obtain the trigger probability of each risk event under the superposition of real-time supporting factors. Based on the trigger probabilities of various risk events under the superposition of real-time supporting factors, a multi-level early warning system is set, triggering corresponding alarms. If the trigger probability of each risk event is between 0.3 and 0.5, it is determined to be a Level 1 warning, recorded internally by the system, and the safety officer is notified. If the trigger probability is between 0.6 and 0.7, it is determined to be a Level 2 warning, immediately triggering an audible and visual alarm and pushing warning information to on-duty personnel. If the trigger probability is between 0.8 and 0.85, it is determined to be a Level 3 warning, activating emergency broadcasting and notifying the emergency response team to handle the situation. If the trigger probability is greater than 0.85, it is determined to be a Level 4 warning, triggering the plant-wide emergency plan and linking the fire protection and evacuation systems. Compared with the prior art, the beneficial effects of this invention are: This invention proposes a hazardous waste warehouse monitoring and identification scheme based on video image analysis. This scheme achieves accurate perception and early identification of the risk status of hazardous waste warehouses by constructing a dynamic risk event development vector and matching it with fuzzy clustering of real-time image data. Furthermore, it improves the accuracy and reliability of key risk support factor analysis by utilizing multi-factor decomposition and Bayesian adaptive calibration. By constructing a risk coupling model based on evidence theory, it scientifically quantifies the trigger probability under multi-factor superposition conditions, achieving a leap from single-threshold alarms to probability-based hierarchical early warning, comprehensively reducing false alarm and missed alarm rates, and improving the real-time performance, intelligence, and decision support level of monitoring and early warning. Attached Figure Description
[0014] Figure 1 This is a flowchart of a hazardous waste warehouse monitoring and identification method based on video image analysis. Detailed Implementation
[0015] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.
[0016] Reference Figure 1 As shown, a method for monitoring and identifying hazardous waste warehouses based on video image analysis includes: S1. Obtain historical risk event trigger image data of each hazardous waste warehouse, mark the risk events in the historical risk event trigger image data to support multi-factor sets, construct the status identification model of each risk event of the hazardous waste warehouse, and generate the development vector of each risk event of the hazardous waste warehouse. Step S1 specifically includes: Based on the records of the enterprise's EHS management system, access the historical monitoring videos of the hazardous waste warehouse, obtain the video stream data of the hazardous waste warehouse, perform data preprocessing, extract the recorded time points of historical risk events, mark each historical risk event with a specific timestamp, extract the video stream data of the hazardous waste warehouse t minutes before the occurrence of each historical risk event, and generate a set of timestamp-historical risk event video segments. The risk events mentioned include: leakage, fire, and improper storage; Based on a set of timestamp-historical risk event video clips, a frame-by-frame sampling and analysis method is designed to perform analysis at 1 frame per second. Using a target detection algorithm, the location of different types of hazardous waste containers (barrels, bags, and cans), designated stacking area boundaries, safety passages, workers, and handling equipment is identified. Statistical analysis and relational encoding are performed to obtain a target list with category and location for each frame, generating spatial features of each risk event in the timestamp-historical risk event video clip set.
[0017] Step S1 also includes: The timestamp-historical risk event video clip set is cropped into fixed-length video clips, input into the I3D network, and high-dimensional feature vectors representing dynamic changes within the video clips are obtained. The temporal features of each risk event in the timestamp-historical risk event video clip set are then extracted. Based on a set of video clips of timestamp-historical risk events, the images are converted from RGB color space to HSV color space. Using a Gaussian mixture model, an image model of the static background of the hazardous waste warehouse floor and walls is built. The current frame is compared with the image model of the static background to identify areas with abnormal color changes. A threshold h is set. If the color change exceeds the threshold h, it is marked as a potential leakage area. Based on a set of timestamp-historical risk event video clips, motion vectors are calculated between each frame in the video clip set to identify smoke motion features. The LBP method is used to extract texture features from each frame in the video clip set to detect subtle texture changes in the image and identify smoke signs. Using infrared cameras, temperature readings from timestamp-historical risk event video clips are extracted to obtain their temperature characteristics. Combined with potential leak areas and smoke signs, environmental characteristics of each risk event in the timestamp-historical risk event video clip set are generated.
[0018] Step S1 also includes: By integrating and extracting the spatial, temporal, and environmental characteristics of each risk event from the historical risk event video clip set, a set of multiple factors supporting risk events in the triggering image data of each historical risk event is obtained. Using the set of multiple factors supporting risk events from historical risk event-triggered image data as input, a Transformer encoder is trained. The learning objectives are to predict the risk category caused by the development of the video segment set and the severity of the risk at the end of the video segment set. A loss function is designed, and the model parameters are continuously optimized through backpropagation. A risk event state recognition model for hazardous waste warehouses is constructed, and development vectors of each risk event in hazardous waste warehouses are generated.
[0019] When using it, please refer to the steps outlined above: Existing technologies mainly rely on single sensors or manual inspections for risk monitoring of hazardous waste warehouses, which suffers from shortcomings such as insufficient real-time performance, lack of multi-factor correlation, and delayed early risk identification. Traditional visual methods are difficult to integrate spatiotemporal and environmental features for dynamic risk development modeling. This step integrates multimodal feature extraction from video streams and uses a Transformer encoder for multi-task learning, achieving early and accurate identification and vectorized representation of various risk events such as leaks, fires, and illegal storage. This improves the automation level of risk monitoring, the ability to analyze multi-factor correlation, and the predictive effectiveness of risk evolution trends.
[0020] S2. Obtain real-time image data of the hazardous waste warehouse, perform fuzzy clustering according to the development interval vector of each risk event in the hazardous waste warehouse, determine the category to which the real-time image data of the hazardous waste warehouse belongs, substitute it into the status recognition model of each risk event in the hazardous waste warehouse, and generate the real-time development vector of each risk event in the hazardous waste warehouse. Step S2 specifically includes: Based on the RTSP video streams from several cameras deployed in the hazardous waste warehouse, real-time image data of the hazardous waste warehouse is acquired, timestamps and frame numbers are aligned, and a dynamic background model of the hazardous waste warehouse is established using the ViBe algorithm. Real-time monitoring of foreground targets in the hazardous waste warehouse, including workers, handling equipment, and flowing liquids, is performed, and the real-time image data of the hazardous waste warehouse is preprocessed. Based on real-time image data of hazardous waste warehouses, and combined with the feature extraction technology of timestamp-historical risk event video clip sets, the spatial features, temporal features and environmental features of the image data of hazardous waste warehouses for each time period are extracted in real time, so as to obtain the multi-factor set of risk events supporting risk events in the real-time trigger image data of each risk event in hazardous waste warehouses.
[0021] Step S2 also includes: Based on the development vectors of each risk event in the hazardous waste warehouse, a risk event development vector library for each hazardous waste warehouse is established, which is a set of timestamp-historical risk event video clips. c points in this vector library are randomly selected as the initial cluster centers. For the risk events in the real-time image data of hazardous waste warehouses, which support multi-factor sets, the fuzzy C-means clustering algorithm is used to calculate the Euclidean distance from each real-time feature vector in the real-time image data of the hazardous waste warehouse to each initial cluster center. A fuzzy factor is set to calculate the membership degree of each risk event category and generate the membership degree matrix of the real-time image data of the hazardous waste warehouse.
[0022] Step S2 also includes: A threshold of 0.7 is set. If the membership degree of the real-time image data of the hazardous waste warehouse exceeds the threshold of 0.7, the real-time image data is determined to belong to the corresponding risk event category, thus determining the category to which the real-time image data of the hazardous waste warehouse belongs. Using real-time image data of hazardous waste warehouses and their respective risk event categories as input, the system associates and pairs risk event status recognition models for each hazardous waste warehouse, generates predictions of the risk categories caused by development in the image data and real-time development vectors of each risk event in the hazardous waste warehouse, calculates the similarity between the current risk status and historical events, and combines the predictions of the risk categories caused by development in the image data to obtain real-time development vectors of each risk event in the hazardous waste warehouse with confidence.
[0023] When using it, please refer to the steps outlined above: Existing technologies for risk monitoring in hazardous waste warehouses mainly rely on manual inspections and fixed sensors, which suffer from drawbacks such as poor real-time performance, delayed risk identification, and inability to perform dynamic clustering analysis on multi-factor coupled events. This step achieves real-time extraction of foreground targets by integrating RTSP video streams and ViBe dynamic background modeling. Furthermore, it utilizes fuzzy C-means clustering algorithm and multi-feature fusion technology to accurately segment real-time image data into risk event development intervals. By generating real-time development vectors with confidence scores through a state recognition model, the automation level, temporal correlation, and accuracy of multi-factor coupled risk assessment in risk identification are improved.
[0024] S3. Based on the real-time development vectors of various risk events in the hazardous waste warehouse, decompose the real-time supporting factors for each risk event, verify the risk score of each real-time supporting factor, construct a risk coupling model, verify the trigger probability of each risk event under the superposition of real-time supporting factors, determine the preset graded early warning threshold range, and trigger the corresponding level alarm.
[0025] Step S3 specifically includes: Based on the real-time development vectors of various risk events in the hazardous waste warehouse, the SHAP analysis algorithm is used to calculate the contribution of each real-time supporting factor to the real-time development vector of the risk event, obtain the risk score of each real-time supporting factor, and identify the top k real-time supporting factors. Based on the identification of the top k real-time supporting factors, and combined with the set of supporting factors for risk events in the real-time triggered image data of each risk event in the hazardous waste warehouse, these factors are decomposed into three categories: spatial factors, temporal factors, and environmental factors. The contribution of the decomposed real-time supporting factors and the corresponding real-time development vectors of the risk events is summarized to generate a detailed report on the supporting factors.
[0026] Step S3 also includes: Based on the identified top k real-time support factors, risk scores for the top k real-time support factors are obtained. Using the analytic hierarchy process (AHP), an expert group is formed to construct a judgment matrix. The analysis scores of each real-time support factor are compared pairwise to calculate the initial risk weight vector of each real-time support factor. Through consistency testing, the initial risk weight of each real-time support factor is obtained. Different risk transformation functions are designed for different types of factors, including linear functions, sigmoid functions, and step functions, to obtain the risk score of each real-time support factor. Based on the Bayesian update algorithm and according to the historical early warning records of hazardous waste warehouses, a correct early warning is defined as a true positive, a false alarm is defined as a false positive, a missed alarm is defined as a false negative, and a correct warning without an early warning is defined as a true negative. For each false alarm and false negative event, the contribution of supporting factors is analyzed, the factors that cause false alarms are weighted less and the factors that cause false negatives are weighted more, and the risk score of each real-time supporting factor is verified by summarizing and updating once a month. Based on the risk score of each real-time support factor, the risk of each real-time support factor is grouped according to physical correlation. Temperature, smoke and open flame are defined as fire source related group, density, location and stability are defined as stacking related group, and humidity, ventilation and electricity are defined as environment related group. The possible combinations of the risk of each real-time support factor are calculated, and their geometric mean is obtained to obtain the coupled risk of all real-time support factors occurring at the same time, and a risk coupling model is constructed.
[0027] Step S3 also includes: Based on the coupling risk of all real-time supporting factors occurring simultaneously, using DS evidence theory, each real-time supporting factor is regarded as an evidence source. A basic probability allocation function is defined, and orthogonal sums are calculated for multiple evidence sources to obtain the trigger probability of each risk event under the superposition of real-time supporting factors. Based on the trigger probability of each risk event under the superposition of real-time supporting factors, multi-level early warning is set, and corresponding alarms are triggered. If the trigger probability of each risk event is in the range of 0.3 to 0.5, it is determined to be a Level 1 warning, which is recorded internally by the system and the safety officer is notified. If the trigger probability of each risk event is in the range of 0.6 to 0.7, it is determined to be a Level 2 warning, which immediately triggers an audible and visual alarm and pushes the warning information to the on-duty personnel. If the trigger probability of each risk event is in the range of 0.8 to 0.85, it is determined to be a Level 3 warning, which activates the emergency broadcast and notifies the emergency team to handle the situation. If the trigger probability of each risk event is greater than 0.85, it is determined to be a Level 4 warning, which triggers the plant-wide emergency plan and links the fire protection and evacuation systems.
[0028] When using it, please refer to the steps outlined above: Existing technologies for risk early warning in hazardous waste warehouses often rely on a single static threshold or simple multi-factor weighting, lacking quantitative tracking of the dynamic development process of risk events and making it difficult to accurately identify key supporting factors. Traditional coupling analysis often ignores nonlinear interactions and evidence conflicts between factors, resulting in high false alarm and missed alarm rates and insufficient early warning timeliness. This step introduces SHAP analysis to dynamically analyze the contribution of factors, combines the analytic hierarchy process (AHP) and Bayesian updates to achieve adaptive calibration of risk scores, and utilizes DS evidence theory to fuse multi-source evidence, effectively improving the accuracy and real-time performance of risk coupling assessment, reducing the risk of false alarms and missed alarms, and achieving accurate triggering of tiered early warnings.
[0029] Based on the above, the specific implementation method is as follows: This embodiment takes the installation of 8 high-definition cameras, 4 infrared thermal imagers, and an environmental sensor network covering the entire area of a hazardous waste warehouse in a large petrochemical park as an example. Based on the warehouse's EHS records of three chemical leaks, two initial fire warnings, and multiple incidents of improper storage over the past two years, the system extracted a total of approximately 120 minutes of video footage preceding these historical events. Technicians used object detection algorithms to identify 13 types of key targets, including IBC containers, iron drums, leak-proof pallets, and fire lane lines. Combined with the I3D network, dynamic temporal features such as liquid diffusion and smoke drift were extracted. The system extracts the temperature drop features of leakage events and the local temperature rise features of fire events from infrared video. It uses a Transformer encoder to fuse and train the above spatial, temporal, and environmental multimodal features, and constructs a risk event status recognition model that includes three categories: leakage, fire, and illegal stacking. It also generates a corresponding baseline risk development vector library. One afternoon, the system's real-time video stream analysis module issued a notification. The system detected a dynamic foreground anomaly near shelf number 3 in area A using the ViBe algorithm. After extracting real-time features, the system showed the following: Spatial features: An iron drum labeled as waste solvent was detected to be tilted at 15 degrees; Environmental features: Infrared images showed that the local temperature of the drum was abnormally low, approximately 8°C below the ambient temperature, and HSV color analysis detected dark, irregular patches spreading on the ground beneath the drum; Temporal features: Motion vector analysis showed that the dark patches spread approximately 20 centimeters within 30 seconds. The system performs fuzzy C-means clustering analysis on the above-mentioned real-time support multi-factor set and the historical risk development vector library. The calculated membership degree of the risk to the leak event is as high as 0.86, which far exceeds the threshold of 0.7. It is accurately classified as a potential chemical leak risk event. The system inputs it into the leak risk status identification model to generate the real-time development vector of the current risk. Its vector similarity with a historical solvent leak event reaches 0.82, and the confidence level is evaluated as high. The system immediately conducted an in-depth analysis of this leakage risk event. Through the SHAP algorithm, it identified the barrel tilt angle, the area of ground color change, and the local temperature drop as the three most critical supporting factors for this risk, with risk scores of 0.79, 0.84, and 0.72, respectively. Combining the factor weights determined by the Analytic Hierarchy Process (AHP) and calibrated by the Bayesian module based on past warning accuracy, the risk coupling model further analyzes the interaction of various factors. Calculations using DS evidence theory show that the probability of a large-area leak event occurring under the superposition of multiple factors including container tipping, liquid leakage, and evaporation endothermic reaction is 0.78. This probability falls between the preset Level 2 warning (0.6-0.7) and Level 3 warning (0.8-0.85). Based on the rules, the system activates a higher-level warning, determining it to be a Level 2 warning enhancement state. The system immediately triggers the on-site audible and visual alarms and automatically pushes warning information, including the specific location, risk factors, on-site captured images, and an assessment report, to the mobile terminals of the warehouse shift supervisor and safety officer via the enterprise security platform. The on-duty personnel quickly arrive at the scene according to the instructions and confirm that a waste solvent drum has experienced a minor leak due to a broken pallet. They then handle the situation according to procedures. After confirming the anomaly has been eliminated, the system automatically deactivates the alarm and feeds the event data back to the model to optimize future identification accuracy.
[0030] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.
Claims
1. A method for monitoring and identifying hazardous waste warehouses based on video image analysis, characterized in that, include: S1. Obtain historical risk event trigger image data of each hazardous waste warehouse, mark the risk events in the historical risk event trigger image data to support multi-factor sets, construct the status identification model of each risk event of the hazardous waste warehouse, and generate the development vector of each risk event of the hazardous waste warehouse. S2. Obtain real-time image data of the hazardous waste warehouse, perform fuzzy clustering according to the development interval vector of each risk event in the hazardous waste warehouse, determine the category to which the real-time image data of the hazardous waste warehouse belongs, substitute it into the status recognition model of each risk event in the hazardous waste warehouse, and generate the real-time development vector of each risk event in the hazardous waste warehouse. S3. Based on the real-time development vectors of various risk events in the hazardous waste warehouse, decompose the real-time supporting factors for each risk event, verify the risk score of each real-time supporting factor, construct a risk coupling model, verify the trigger probability of each risk event under the superposition of real-time supporting factors, determine the preset graded early warning threshold range, and trigger the corresponding level alarm.
2. The method for monitoring and identifying hazardous waste warehouses based on video image analysis according to claim 1, characterized in that, Step S1 specifically includes: Based on the records of the enterprise's EHS management system, access the historical monitoring videos of the hazardous waste warehouse, obtain the video stream data of the hazardous waste warehouse, perform data preprocessing, extract the recorded time points of historical risk events, mark each historical risk event with a specific timestamp, extract the video stream data of the hazardous waste warehouse t minutes before the occurrence of each historical risk event, and generate a set of timestamp-historical risk event video segments. The risk events mentioned include: leakage, fire, and improper storage; Based on a set of timestamp-historical risk event video clips, a frame-by-frame sampling and analysis method is designed to perform analysis at 1 frame per second. Using a target detection algorithm, the location of different types of hazardous waste containers (barrels, bags, and cans), designated stacking area boundaries, safety passages, workers, and handling equipment is identified. Statistical analysis and relational encoding are performed to obtain a target list with category and location for each frame, generating spatial features of each risk event in the timestamp-historical risk event video clip set.
3. The method for monitoring and identifying hazardous waste warehouses based on video image analysis according to claim 2, characterized in that, Step S1 also includes: The timestamp-historical risk event video clip set is cropped into fixed-length video clips, input into the I3D network, and high-dimensional feature vectors representing dynamic changes within the video clips are obtained. The temporal features of each risk event in the timestamp-historical risk event video clip set are then extracted. Based on a set of video clips of timestamp-historical risk events, the images are converted from RGB color space to HSV color space. Using a Gaussian mixture model, an image model of the static background of the hazardous waste warehouse floor and walls is built. The current frame is compared with the image model of the static background to identify areas with abnormal color changes. A threshold h is set. If the color change exceeds the threshold h, it is marked as a potential leakage area. Based on a set of timestamp-historical risk event video clips, motion vectors are calculated between each frame in the video clip set to identify smoke motion features. The LBP method is used to extract texture features from each frame in the video clip set to detect subtle texture changes in the image and identify smoke signs. Using infrared cameras, temperature readings from timestamp-historical risk event video clips are extracted to obtain their temperature characteristics. Combined with potential leak areas and smoke signs, environmental characteristics of each risk event in the timestamp-historical risk event video clip set are generated.
4. The method for monitoring and identifying hazardous waste warehouses based on video image analysis according to claim 3, characterized in that, Step S1 also includes: By integrating and extracting the spatial, temporal, and environmental characteristics of each risk event from the historical risk event video clip set, a set of multiple factors supporting risk events in the triggering image data of each historical risk event is obtained. Using the set of multiple factors supporting risk events from historical risk event-triggered image data as input, a Transformer encoder is trained. The learning objectives are to predict the risk category caused by the development of the video segment set and the severity of the risk at the end of the video segment set. A loss function is designed, and the model parameters are continuously optimized through backpropagation. A risk event state recognition model for hazardous waste warehouses is constructed, and development vectors of each risk event in hazardous waste warehouses are generated.
5. The method for monitoring and identifying hazardous waste warehouses based on video image analysis according to claim 1, characterized in that, Step S2 specifically includes: Based on the RTSP video streams from several cameras deployed in the hazardous waste warehouse, real-time image data of the hazardous waste warehouse is acquired, timestamps and frame numbers are aligned, and a dynamic background model of the hazardous waste warehouse is established using the ViBe algorithm. Real-time monitoring of foreground targets in the hazardous waste warehouse, including workers, handling equipment, and flowing liquids, is performed, and the real-time image data of the hazardous waste warehouse is preprocessed. Based on real-time image data of hazardous waste warehouses, and combined with the feature extraction technology of timestamp-historical risk event video clip sets, the spatial features, temporal features and environmental features of the image data of hazardous waste warehouses for each time period are extracted in real time, so as to obtain the multi-factor set of risk events supporting risk events in the real-time trigger image data of each risk event in hazardous waste warehouses.
6. The method for monitoring and identifying hazardous waste warehouses based on video image analysis according to claim 5, characterized in that, Step S2 also includes: Based on the development vectors of each risk event in the hazardous waste warehouse, a risk event development vector library for each hazardous waste warehouse is established, which is a set of timestamp-historical risk event video clips. c points in this vector library are randomly selected as the initial cluster centers. For the risk events in the real-time image data of hazardous waste warehouses, which support multi-factor sets, the fuzzy C-means clustering algorithm is used to calculate the Euclidean distance from each real-time feature vector in the real-time image data of the hazardous waste warehouse to each initial cluster center. A fuzzy factor is set to calculate the membership degree of each risk event category and generate the membership degree matrix of the real-time image data of the hazardous waste warehouse.
7. The method for monitoring and identifying hazardous waste warehouses based on video image analysis according to claim 6, characterized in that, Step S2 also includes: A threshold of 0.7 is set. If the membership degree of the real-time image data of the hazardous waste warehouse exceeds the threshold of 0.7, the real-time image data is determined to belong to the corresponding risk event category, thus determining the category to which the real-time image data of the hazardous waste warehouse belongs. Using real-time image data of hazardous waste warehouses and their respective risk event categories as input, the system associates and pairs risk event status recognition models for each hazardous waste warehouse, generates predictions of the risk categories caused by development in the image data and real-time development vectors of each risk event in the hazardous waste warehouse, calculates the similarity between the current risk status and historical events, and combines the predictions of the risk categories caused by development in the image data to obtain real-time development vectors of each risk event in the hazardous waste warehouse with confidence.
8. The method for monitoring and identifying hazardous waste warehouses based on video image analysis according to claim 7, characterized in that, Step S3 specifically includes: Based on the real-time development vectors of various risk events in the hazardous waste warehouse, the SHAP analysis algorithm is used to calculate the contribution of each real-time supporting factor to the real-time development vector of the risk event, obtain the risk score of each real-time supporting factor, and identify the top k real-time supporting factors. Based on the identification of the top k real-time supporting factors, and combined with the set of supporting factors for risk events in the real-time triggered image data of each risk event in the hazardous waste warehouse, these factors are decomposed into three categories: spatial factors, temporal factors, and environmental factors. The contribution of the decomposed real-time supporting factors and the corresponding real-time development vectors of the risk events is summarized to generate a detailed report on the supporting factors.
9. A method for monitoring and identifying hazardous waste warehouses based on video image analysis according to claim 8, characterized in that, Step S3 also includes: Based on the identified top k real-time support factors, risk scores for the top k real-time support factors are obtained. Using the analytic hierarchy process (AHP), an expert group is formed to construct a judgment matrix. The analysis scores of each real-time support factor are compared pairwise to calculate the initial risk weight vector of each real-time support factor. Through consistency testing, the initial risk weight of each real-time support factor is obtained. Different risk transformation functions are designed for different types of factors, including linear functions, sigmoid functions, and step functions, to obtain the risk score of each real-time support factor. Based on the Bayesian update algorithm and according to the historical early warning records of hazardous waste warehouses, a correct early warning is defined as a true positive, a false alarm is defined as a false positive, a missed alarm is defined as a false negative, and a correct warning without an early warning is defined as a true negative. For each false alarm and false negative event, the contribution of supporting factors is analyzed, the factors that cause false alarms are weighted less and the factors that cause false negatives are weighted more, and the risk score of each real-time supporting factor is verified by summarizing and updating once a month. Based on the risk score of each real-time support factor, the risk of each real-time support factor is grouped according to physical correlation. Temperature, smoke and open flame are defined as fire source related group, density, location and stability are defined as stacking related group, and humidity, ventilation and electricity are defined as environment related group. The possible combinations of the risk of each real-time support factor are calculated, and their geometric mean is obtained to obtain the coupled risk of all real-time support factors occurring at the same time, and a risk coupling model is constructed.
10. The method for monitoring and identifying hazardous waste warehouses based on video image analysis according to claim 9, characterized in that, Step S3 also includes: Based on the coupling risk of all real-time supporting factors occurring simultaneously, using DS evidence theory, each real-time supporting factor is regarded as an evidence source. A basic probability allocation function is defined, and orthogonal sums are calculated for multiple evidence sources to obtain the trigger probability of each risk event under the superposition of real-time supporting factors. Based on the trigger probability of each risk event under the superposition of real-time supporting factors, multi-level early warning is set, and corresponding alarms are triggered. If the trigger probability of each risk event is in the range of 0.3 to 0.5, it is determined to be a Level 1 warning, which is recorded internally by the system and the safety officer is notified. If the trigger probability of each risk event is in the range of 0.6 to 0.7, it is determined to be a Level 2 warning, which immediately triggers an audible and visual alarm and pushes the warning information to the on-duty personnel. If the trigger probability of each risk event is in the range of 0.8 to 0.85, it is determined to be a Level 3 warning, which activates the emergency broadcast and notifies the emergency team to handle the situation. If the trigger probability of each risk event is greater than 0.85, it is determined to be a Level 4 warning, which triggers the plant-wide emergency plan and links the fire protection and evacuation systems.