Intelligent potential safety hazard inspection system based on image acquisition and recognition

Abstract

The invention discloses a potential safety hazard intelligent inspection system based on image acquisition and identification, and relates to the technical field of potential safety hazard inspection, and the system comprises a multi-modal image acquisition module which generates a standardized image feature vector; the hidden danger recognition module performs hidden danger recognition and risk assessment on the standardized image feature vector based on a deep learning model to generate a hidden danger risk index; the disposal decision module is used for establishing an intelligent disposal decision model and generating a disposal work order based on the output of the disposal decision model; and the multi-dimensional association analysis and active prevention module is used for integrating the hidden danger identification data, the environmental parameters and historical disposal records to perform association rule mining, generating a hidden danger development trend graph and pushing an active prevention strategy based on the graph. According to the invention, through combination of multi-modal image acquisition, deep learning and reinforcement learning technologies, intelligent and dynamically adjusted potential safety hazard detection, assessment, disposal and prevention are realized, and the accuracy and efficiency of risk response are improved.

Description

Technical Field

[0001] This invention relates to the field of safety hazard inspection technology, specifically to an intelligent safety hazard inspection system based on image acquisition and recognition. Background Technology

[0002] Currently, safety hazard inspections rely heavily on manual inspections, which suffer from problems such as low efficiency, high rates of missed or false detections, and delayed response. Manual inspections are limited by manpower costs and time and energy, making it difficult to achieve all-weather, full-coverage monitoring of the monitoring area. Especially in complex industrial scenarios or dangerous areas, the safety of inspection personnel cannot be fully guaranteed.

[0003] Existing image recognition-based inspection systems mostly employ single-modal image acquisition and rely solely on visible light images for analysis. Their recognition accuracy drops significantly in complex environments such as insufficient lighting or obstruction. Furthermore, they lack correlation analysis with historical data and environmental parameters after hazard identification, making proactive risk prevention difficult. They often only passively respond to existing hazards, failing to prevent their spread and secondary disasters. In addition, the formulation of response strategies lacks intelligent decision support, and resource allocation is unreasonable, leading to low efficiency in hazard handling. Therefore, there is an urgent need for an intelligent safety hazard inspection system that integrates multimodal perception, intelligent recognition, precise decision-making, and proactive prevention to improve the level of intelligent safety management. Summary of the Invention

[0004] The purpose of this invention is to solve the problems mentioned above by proposing an intelligent security hazard inspection system based on image acquisition and recognition.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A smart safety hazard inspection system based on image acquisition and recognition includes:

[0007] Multimodal image acquisition module, hazard identification module, intelligent response decision-making module, and multi-dimensional correlation analysis and proactive prevention module;

[0008] The multimodal image acquisition module is used to acquire image data in real time based on a camera array deployed in the monitoring area, preprocess the acquired image data to generate standardized image feature vectors, and transmit the generated standardized image feature vectors to the hazard identification module.

[0009] The hazard identification module is used to pre-store a safety hazard feature library and a historical case database. Based on a deep learning model, it identifies and assesses risks from standardized image feature vectors and generates a hazard risk index.

[0010] The disposal decision module is used to establish an intelligent disposal decision model based on the hidden danger risk index and a preset disposal strategy library. Based on the output of the disposal decision model, a disposal work order is generated, which includes hidden danger location, disposal suggestions and resource allocation information.

[0011] The multi-dimensional correlation analysis and proactive prevention module is used to integrate hazard identification data, environmental parameters and historical handling records to mine correlation rules, generate hazard development trend maps, push proactive prevention strategies based on the maps, and realize data interaction and instruction distribution within the system.

[0012] Preferably, the multimodal image acquisition module specifically includes:

[0013] Real-time image acquisition unit: Based on visible light cameras, infrared thermal imagers, and panoramic cameras, it acquires image data of the monitoring area in real time, including equipment status images, environmental scene images, personnel behavior images, and scene images captured on site by mobile phones and panoramic cameras. The image data is appended with timestamps and location tags, and the acquired real-time image data is synchronized in time.

[0014] Noise filtering unit: Uses Gaussian filtering algorithm to filter noise in the acquired image and normalizes the size of image data of different resolutions and formats;

[0015] The Gaussian filtering algorithm formula is as follows:

[0016]

[0017] In the formula, the function value of the Gaussian filter kernel at coordinates (x,y) is used to represent the filter weight at that position, σ is the Gaussian standard deviation, which controls the smoothness of the filter, and π is the constant of pi.

[0018] Image feature extraction unit: Using the gray-level co-occurrence matrix calculation formula, semantic features, texture features, and spatial features are extracted from the preprocessed image. The extracted features are standardized to obtain a standardized image feature vector. The standardized image feature vector is then transmitted to the hazard identification module through the multi-dimensional correlation analysis and proactive prevention module for further hazard identification and risk assessment.

[0019] The formula for calculating the gray-level co-occurrence matrix is ​​as follows:

[0020]

[0021] In the formula, GLCM(i,j,d,θ) is the element value of the gray-level co-occurrence matrix when the gray values ​​are i and j, the distance is d, and the angle is θ, which represents the statistical characteristics of the gray value relationship in the image. i and j are the gray values ​​of two pixels in the image, d is the distance between the two pixels, θ is the angle between the line connecting the two pixels and the horizontal direction, x and y are the horizontal and vertical coordinates in the image pixel coordinate system, which represent the position of the reference pixel, and I(i,j,d,θ,x,y) is the indicator function.

[0022] Preferably, the hazard identification module specifically includes:

[0023] Hazard detection unit: Based on a trained YOLOv7 deep learning model, it performs target detection on standardized image feature vectors to identify safety hazard targets in the image;

[0024] Risk assessment unit: Combining historical data and current status information of potential hazards, the analytic hierarchy process (AHP) is used to solve for the weight parameters of the standardized image feature vector. The weight parameters of the standardized image feature vector are used as input, and the risk level prediction data is used as output to establish a risk assessment model. The risk level of the potential hazard is estimated based on classification analysis.

[0025] Hazard risk index generation unit: Based on hazard detection results and risk assessment data, the hazard risk index is calculated comprehensively using a weighted average method;

[0026] Early warning and alarm unit: Based on the comparison between the hidden danger risk index and the threshold, it determines whether there are potential safety hazards. If the hidden danger risk index is higher than the set threshold, the system will trigger an alarm mechanism to notify safety management personnel to take action.

[0027] Preferably, the risk assessment unit includes:

[0028] The analytic hierarchy process (AHP) is used to construct a judgment matrix, obtain the relative importance weights of each potential risk factor, construct a risk assessment model, optimize the model by adjusting the weights, perform a consistency test on the risk assessment model, and calculate the consistency ratio.

[0029] The consistency test formula is as follows:

[0030]

[0031] In the formula, CI is an indicator that measures the degree of consistency of a matrix; the smaller the value, the better the matrix consistency. λ max The largest eigenvalue of the judgment matrix is ​​obtained through eigenvalue decomposition. n is the matrix order. CR is used to check whether the consistency of the judgment matrix is ​​acceptable. It is the ratio of the consistency index to the average random consistency index. RI is the average random consistency index, which is the average consistency index calculated through a large number of random judgment matrices.

[0032] The consistency ratio is compared with a set threshold. If the consistency ratio is less than the threshold, the judgment matrix has satisfactory consistency; otherwise, the judgment matrix is ​​readjusted.

[0033] Based on the derived weight values ​​and standardized image feature vectors, a risk assessment model is established using a risk assessment formula to estimate the risk level of potential hazards. The risk assessment formula is as follows:

[0034]

[0035] In the formula, R is the risk level assessment value of the hidden danger calculated by the risk assessment model, and w i The relative importance weight of the i-th hidden danger factor is obtained from the judgment matrix using the analytic hierarchy process (AHP). i is the standardized image feature vector value corresponding to the i-th hidden danger factor, which is the feature data after preprocessing and standardization, and n is the number of hidden danger factors involved in the assessment.

[0036] Preferably, the hidden danger risk index generation unit specifically includes:

[0037] Obtain the hazard category label and corresponding confidence level output by the hazard detection unit, and retrieve the basic risk coefficient of the hazard category from the safety hazard feature library;

[0038] Extract the predicted risk levels generated by the risk assessment unit, and calculate the hazard risk index using a weighted fusion formula:

[0039]

[0040] In the formula, I represents the final calculated hazard risk index, which comprehensively reflects the overall risk level of safety hazards within the monitoring area; n is the number of hazards detected; i represents the hazard's serial number; and C represents the weighting ratio of the basic risk coefficient and the predicted risk level value. i B represents the detection confidence level of the i-th hidden danger. i R represents the basic risk coefficient of the i-th hidden danger. i Let represent the predicted risk level of the i-th hazard, where 0.6 and 0.4 are the basic risk coefficients B, respectively. i Risk level prediction value R i Weighting percentage;

[0041] The calculated risk index is compared with a preset threshold to generate a risk level label.

[0042] Preferably, the preset threshold is determined through historical data modeling, collecting the hidden risk index and actual handling results over the past 12 months, and using the percentile method to divide the threshold range:

[0043] Low risk threshold: I <P 30 , where P 30 30th percentile;

[0044] Medium risk threshold: P 30 ≤I≤P 60 , where P 60 60th percentile;

[0045] High-risk threshold: P 60 <I≤P 90 , where P 90 It is the 90th percentile;

[0046] Emergency risk threshold: I>P 90 ;

[0047] The system automatically matches the response mechanism according to different risk threshold ranges. Low risk is only recorded and archived, medium risk is pushed with prompt information, high risk generates a handling work order, and emergency risk triggers an audible and visual alarm and pushes it to the emergency command platform simultaneously.

[0048] Preferably, the intelligent decision-making module specifically includes:

[0049] Disposal strategy generation unit: Based on the hazard risk index and historical disposal data, combined with different hazard types and disposal scenarios, it generates targeted disposal strategies;

[0050] Optimize decision-making unit: Adjust and optimize the handling strategy based on reinforcement learning algorithm to select the most suitable handling plan for the current hidden danger status;

[0051] Task allocation unit for response personnel: Response tasks are allocated based on the type of hazard, the location of the monitoring area, and the skills of the response personnel.

[0052] Preferably, the step of adjusting and optimizing the handling strategy based on reinforcement learning algorithms to select the most suitable handling plan for the current hidden danger status specifically includes:

[0053] Collect risk index, hazard type, historical handling data, and handling effect data of potential hazards; select handling strategies based on hazard status; and generate an initial solution.

[0054] A reward function is constructed based on the risk index, hazard type, and disposal cost factors of the hazard, with the aim of maximizing the disposal effect and minimizing the disposal cost.

[0055] The reward function expression is as follows:

[0056] G = γ1E - γ2C + γ3(10 - R)

[0057] In the formula, G is the reward value in reinforcement learning, γ1, γ2, and γ3 are weight coefficients, which respectively represent the importance of the treatment effect, treatment cost, and risk index in the reward calculation, E is the quantitative value of the treatment effect, C is the quantitative value of the treatment cost, and R is the hidden danger risk index.

[0058] The reward function is solved, and the disposal strategy is optimized based on the deep Q-network algorithm. In each iteration, the strategy is adjusted according to the learning mechanism of the algorithm to maximize the reward function. The most suitable disposal solution is output based on the optimization result.

[0059] Preferably, the allocation of disposal tasks based on the type of hazard, the location of the monitoring area, and the skills of the disposal personnel specifically includes: assessing the urgency of the hazard based on its risk index, type, and impact range; calculating the priority of each disposal task; and scheduling the disposal tasks based on their priorities. The priority calculation formula is as follows:

[0060] U = δ1R + δ2T + δ3S

[0061] In the formula, δ1, δ2, and δ3 are weighting coefficients. The priority of each disposal task is calculated, and the disposal tasks are scheduled based on the priority of the disposal tasks.

[0062] Preferably, the multi-dimensional correlation analysis and proactive prevention module specifically includes:

[0063] Cross-dimensional data association unit: Integrates image recognition data, environmental sensor data, and equipment operation logs to establish a multi-source data association rule base, and uses data association rule mining technology to identify potential correlations between hidden dangers and the environment and equipment status;

[0064] Hazard Chain Prediction Unit: Based on the association rule base, it predicts the secondary hazard chains that the current hazard may cause through time series analysis, and outputs the simulation results of the hazard diffusion path;

[0065] Proactive prevention strategy generation unit: Combining the hazard chain prediction results with historical handling cases, it generates prevention strategies that include pre-intervention measures and pushes them to the handling decision module.

[0066] Compared with the prior art, the beneficial effects of the present invention are:

[0067] 1. This system monitors different types of safety hazards in real time through multimodal image acquisition technology, and uses deep learning models to perform efficient hazard identification and risk assessment, thereby improving the accuracy and intelligence of safety inspections.

[0068] 2. The system combines the hazard risk index, historical data, and reinforcement learning algorithms to generate a risk assessment for each hazard, formulate targeted handling strategies, and optimize the handling plan based on the real-time hazard status, thereby improving the efficiency and economy of emergency response.

[0069] 3. Through cross-dimensional data analysis and prediction of potential hazard development trends, the system can formulate and push proactive prevention strategies. At the same time, it can intelligently allocate tasks according to the type and priority of potential hazards, and with the help of an efficient early warning mechanism, ensure rapid response and minimize losses. Attached Figure Description

[0070] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0071] Figure 1 This is a system framework diagram of the present invention;

[0072] Figure 2 This is a diagram showing the internal system framework of the multimodal image acquisition module in this invention.

[0073] Figure 3 This is a diagram showing the internal system framework of the hazard identification module in this invention.

[0074] Figure 4 This is an internal system framework diagram of the decision-making module in this invention;

[0075] Figure 5 This is a diagram of the internal system framework of the multi-dimensional correlation analysis and proactive prevention module in this invention. Detailed Implementation

[0076] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0077] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0078] Please see Figure 1 As shown, a smart safety hazard inspection system based on image acquisition and recognition includes:

[0079] Multimodal image acquisition module, hazard identification module, intelligent response decision-making module, and multi-dimensional correlation analysis and proactive prevention module;

[0080] The multimodal image acquisition module is used to acquire image data in real time based on a camera array deployed in the monitoring area, preprocess the acquired image data to generate standardized image feature vectors, and transmit the generated standardized image feature vectors to the hazard identification module.

[0081] The hazard identification module is used to pre-store a safety hazard feature library and a historical case database. Based on a deep learning model, it identifies and assesses risks from standardized image feature vectors and generates a hazard risk index.

[0082] The disposal decision module is used to establish an intelligent disposal decision model based on the hidden danger risk index and a preset disposal strategy library. Based on the output of the disposal decision model, a disposal work order is generated, which includes hidden danger location, disposal suggestions and resource allocation information.

[0083] The multi-dimensional correlation analysis and proactive prevention module is used to integrate hazard identification data, environmental parameters and historical handling records to mine correlation rules, generate hazard development trend maps, push proactive prevention strategies based on the maps, and realize data interaction and instruction distribution within the system.

[0084] By integrating multimodal image acquisition, hazard identification, intelligent response decision-making, and proactive prevention analysis, the system can comprehensively, in real time, and intelligently identify safety hazards, provide targeted preventive measures, and intervene before risks occur. Through deep learning models and multi-dimensional data analysis, the system can dynamically adjust and predict risks, thereby providing comprehensive safety protection at multiple levels.

[0085] Please see Figure 2 As shown, the multimodal image acquisition module specifically includes:

[0086] Real-time image acquisition unit: Based on visible light cameras, infrared thermal imagers, and panoramic cameras, it acquires image data of the monitoring area in real time, including equipment status images, environmental scene images, personnel behavior images, and scene images captured on site by mobile phones and panoramic cameras. The image data is appended with timestamps and location tags, and the acquired real-time image data is synchronized in time.

[0087] Noise filtering unit: Uses Gaussian filtering algorithm to filter noise in the acquired image and normalizes the size of image data of different resolutions and formats;

[0088] The Gaussian filtering algorithm formula is as follows:

[0089]

[0090] In the formula, the function value of the Gaussian filter kernel at coordinates (x,y) is used to represent the filter weight at that position, σ is the Gaussian standard deviation, which controls the smoothness of the filter, and π is the constant of pi.

[0091] Image feature extraction unit: Using the gray-level co-occurrence matrix calculation formula, semantic features, texture features, and spatial features are extracted from the preprocessed image. The extracted features are standardized to obtain a standardized image feature vector. The standardized image feature vector is then transmitted to the hazard identification module through the multi-dimensional correlation analysis and proactive prevention module for further hazard identification and risk assessment.

[0092] The formula for calculating the gray-level co-occurrence matrix is ​​as follows:

[0093]

[0094] In the formula, GLCM(i,j,d,θ) is the element value of the gray-level co-occurrence matrix when the gray values ​​are i and j, the distance is d, and the angle is θ, which represents the statistical characteristics of the gray value relationship in the image. i and j are the gray values ​​of two pixels in the image, d is the distance between the two pixels, θ is the angle between the line connecting the two pixels and the horizontal direction, x and y are the horizontal and vertical coordinates in the image pixel coordinate system, which represent the position of the reference pixel, and I(i,j,d,θ,x,y) is the indicator function.

[0095] By using various types of camera equipment combined with time synchronization and noise filtering technologies, different information in the monitored area can be captured comprehensively and efficiently, providing more accurate and diverse input data for subsequent hazard identification and risk assessment.

[0096] Please see Figure 3 As shown, the hazard identification module specifically includes:

[0097] Hazard detection unit: Based on a trained YOLOv7 deep learning model, it performs target detection on standardized image feature vectors to identify safety hazard targets in the image;

[0098] Risk assessment unit: Combining historical data and current status information of potential hazards, the analytic hierarchy process (AHP) is used to solve for the weight parameters of the standardized image feature vector. The weight parameters of the standardized image feature vector are used as input, and the risk level prediction data is used as output to establish a risk assessment model. The risk level of the potential hazard is estimated based on classification analysis.

[0099] Hazard risk index generation unit: Based on hazard detection results and risk assessment data, the hazard risk index is calculated comprehensively using a weighted average method;

[0100] Early warning and alarm unit: Based on the comparison between the hidden danger risk index and the threshold, it determines whether there are potential safety hazards. If the hidden danger risk index is higher than the set threshold, the system will trigger an alarm mechanism to notify safety management personnel to take action.

[0101] By using the YOLOv7 deep learning model for target detection and combining it with the analytic hierarchy process (AHP) for risk assessment, this innovative method intelligently identifies potential safety hazards and generates a hazard risk index through a data-driven approach. This method improves the accuracy of identification and the quantitative analysis of hazards, and demonstrates greater adaptability and flexibility, especially in complex environments.

[0102] The risk assessment unit includes:

[0103] The analytic hierarchy process (AHP) is used to construct a judgment matrix, obtain the relative importance weights of each potential risk factor, construct a risk assessment model, optimize the model by adjusting the weights, perform a consistency test on the risk assessment model, and calculate the consistency ratio.

[0104] The consistency test formula is as follows:

[0105]

[0106] In the formula, CI is an indicator that measures the degree of consistency of a matrix; the smaller the value, the better the matrix consistency. λ max The largest eigenvalue of the judgment matrix is ​​obtained through eigenvalue decomposition. n is the matrix order. CR is used to check whether the consistency of the judgment matrix is ​​acceptable. It is the ratio of the consistency index to the average random consistency index. RI is the average random consistency index, which is the average consistency index calculated through a large number of random judgment matrices.

[0107] The consistency ratio is compared with a set threshold. If the consistency ratio is less than the threshold, the judgment matrix has satisfactory consistency; otherwise, the judgment matrix is ​​readjusted.

[0108] Based on the derived weight values ​​and standardized image feature vectors, a risk assessment model is established using a risk assessment formula to estimate the risk level of potential hazards. The risk assessment formula is as follows:

[0109]

[0110] In the formula, R is the risk level assessment value of the hidden danger calculated by the risk assessment model, and w i The relative importance weight of the i-th hidden danger factor is obtained from the judgment matrix using the analytic hierarchy process (AHP). i is the standardized image feature vector value corresponding to the i-th hidden danger factor, which is the feature data after preprocessing and standardization, and n is the number of hidden danger factors involved in the assessment;

[0111] By constructing a judgment matrix using the analytic hierarchy process (AHP) and optimizing the assessment model, and ensuring the reasonableness of the assessment results through consistency checks, this method is more scientific and accurate than traditional assessment techniques. It effectively reduces the interference of human factors and can generate more reliable risk level estimates.

[0112] The specific components of the hidden danger risk index generation unit include:

[0113] Obtain the hazard category label and corresponding confidence level output by the hazard detection unit, and retrieve the basic risk coefficient of the hazard category from the safety hazard feature library;

[0114] Extract the predicted risk levels generated by the risk assessment unit, and calculate the hazard risk index using a weighted fusion formula:

[0115]

[0116] In the formula, I represents the final calculated hazard risk index, which comprehensively reflects the overall risk level of safety hazards within the monitoring area; n is the number of hazards detected; i represents the hazard's serial number; and C represents the weighting ratio of the basic risk coefficient and the predicted risk level value. i B represents the detection confidence level of the i-th hidden danger. i R represents the basic risk coefficient of the i-th hidden danger. i Let represent the predicted risk level of the i-th hazard, where 0.6 and 0.4 are the basic risk coefficients B, respectively. i Risk level prediction value R i Weighting percentage;

[0117] The calculated hidden danger risk index is compared with the preset threshold to generate a risk level label;

[0118] By combining the results of hazard detection with risk assessment data using a weighted average method, a hazard risk index is calculated. This method makes hazard risk assessment more comprehensive, allowing for a detailed breakdown of each type of hazard and generating a corresponding risk index, which facilitates risk prioritization and emergency response.

[0119] The preset threshold is determined through historical data modeling, collecting the hazard risk index and actual handling results over the past 12 months, and using the percentile method to divide the threshold range:

[0120] Low risk threshold: I <P 30 , where P 30 30th percentile;

[0121] Medium risk threshold: P 30 ≤I≤P 60 , where P 60 60th percentile;

[0122] High-risk threshold: P 60 <I≤P 90 , where P 90 It is the 90th percentile;

[0123] Emergency risk threshold: I>P 90 ;

[0124] The system automatically matches the response mechanism according to different risk threshold ranges. Low risk is only recorded and archived, medium risk is pushed with prompt information, high risk generates a handling work order, and emergency risk triggers an audible and visual alarm and pushes it to the emergency command platform simultaneously.

[0125] Based on historical data modeling, the system uses percentile methods to define risk thresholds and automatically matches response mechanisms to different risk levels. This innovation provides risk management with adaptive dynamic adjustment capabilities. The system can automatically adjust the response intensity according to the actual danger of potential hazards, thereby improving the system's reaction speed and handling efficiency.

[0126] Please see Figure 4 As shown, the intelligent decision-making module specifically includes:

[0127] Disposal strategy generation unit: Based on the hazard risk index and historical disposal data, combined with different hazard types and disposal scenarios, it generates targeted disposal strategies;

[0128] Optimize decision-making unit: Adjust and optimize the handling strategy based on reinforcement learning algorithm to select the most suitable handling plan for the current hidden danger status;

[0129] Task allocation unit for response personnel: Response tasks are assigned based on the type of hazard, the location of the monitoring area, and the skills of the response personnel.

[0130] By combining the hazard risk index, historical handling data, and different hazard types, targeted handling strategies are generated. By optimizing the decision-making unit and adjusting and optimizing the handling strategies based on reinforcement learning algorithms, the optimal handling plan is selected under different hazard conditions, thereby improving the accuracy and efficiency of handling.

[0131] Adjusting and optimizing the response strategy based on reinforcement learning algorithms to select the most suitable response plan for the current potential hazard status specifically includes:

[0132] Collect risk index, hazard type, historical handling data, and handling effect data of potential hazards; select handling strategies based on hazard status; and generate an initial solution.

[0133] A reward function is constructed based on the risk index, hazard type, and disposal cost factors of the hazard, with the aim of maximizing the disposal effect and minimizing the disposal cost.

[0134] The reward function expression is as follows:

[0135] G = γ1E - γ2C + γ3(10 - R)

[0136] In the formula, G is the reward value in reinforcement learning, γ1, γ2, and γ3 are weight coefficients, which respectively represent the importance of the treatment effect, treatment cost, and risk index in the reward calculation, E is the quantitative value of the treatment effect, C is the quantitative value of the treatment cost, and R is the hidden danger risk index.

[0137] The reward function is solved, and the disposal strategy is optimized based on the deep Q-network algorithm. In each iteration, the strategy is adjusted according to the learning mechanism of the algorithm to maximize the reward function. The most suitable disposal solution is output based on the optimization result.

[0138] By optimizing response strategies based on reinforcement learning algorithms and selecting the most suitable response plan for the current hazard status, the system can maximize the response effect while minimizing the response cost by comprehensively evaluating factors such as risk index, hazard type, and response cost, thus fully realizing intelligent decision-making.

[0139] The allocation of response tasks, based on the type of hazard, the location of the monitoring area, and the skills of the personnel, specifically includes: assessing the urgency of the hazard based on its risk index, type, and impact range; calculating the priority of each response task; and scheduling the response tasks based on their priorities. The priority calculation formula is as follows:

[0140] U = δ1R + δ2T + δ3S

[0141] In the formula, δ1, δ2, and δ3 are weighting coefficients. The priority of each disposal task is calculated, and the disposal tasks are scheduled based on the priority of the disposal tasks.

[0142] Based on the urgency and risk index of the hazard and the skills of the personnel handling it, the module intelligently allocates tasks. It can automatically assess the priority of hazards and rationally schedule tasks according to the assessment results, ensuring that limited handling resources are used most effectively.

[0143] Please see Figure 5 As shown, the multi-dimensional correlation analysis and proactive prevention module specifically includes:

[0144] Cross-dimensional data association unit: Integrates image recognition data, environmental sensor data, and equipment operation logs to establish a multi-source data association rule base, and uses data association rule mining technology to identify potential correlations between hidden dangers and the environment and equipment status;

[0145] Hazard Chain Prediction Unit: Based on the association rule base, it predicts the secondary hazard chains that the current hazard may cause through time series analysis, and outputs the simulation results of the hazard diffusion path;

[0146] Proactive prevention strategy generation unit: Combining the hazard chain prediction results with historical handling cases, it generates prevention strategies that include pre-intervention measures and pushes them to the handling decision module;

[0147] By integrating multi-source data such as image recognition data, environmental sensing data, and equipment operation logs, a data association rule base is established to identify the correlation between potential hazards and the environment and equipment status. This innovation enables the system not only to predict current hazards, but also to prevent possible secondary hazards through hazard chain analysis, thereby achieving proactive prevention and early risk management.

[0148] In summary, the advantages of this invention are as follows:

[0149] The system employs multimodal image acquisition technology, enabling real-time monitoring of various types of potential hazards, such as equipment failures, abnormal personnel behavior, and environmental dangers. It also utilizes deep learning models for efficient hazard identification and risk assessment, comprehensively enhancing the accuracy and intelligence of safety inspections.

[0150] By combining the hazard risk index with historical data, the system can automatically generate risk assessment results for each hazard and formulate corresponding handling strategies based on the risk level, ensuring that the handling work is targeted and efficient.

[0151] With the help of reinforcement learning algorithms, the system can optimize the handling strategy according to the real-time status of potential hazards, maximize the handling effect and minimize the handling cost, thereby improving the efficiency and economy of emergency response.

[0152] Through cross-dimensional data correlation analysis, the system can identify the relationship between potential hazards and the status of the environment and equipment, and predict the development trend of hazards in advance. Based on this, the system can generate proactive prevention strategies and push them in real time to reduce the probability and consequences of hazards.

[0153] The system assigns tasks based on the type of hazard, the location of the monitoring area, and the skills of the personnel handling the situation, and schedules tasks according to their priority to ensure the timeliness and appropriateness of the response work.

[0154] By setting risk thresholds, the system can promptly trigger alarms and take corresponding emergency measures when the risk index of potential hazards reaches a preset high risk or emergency risk level. These measures include pushing information to the emergency command platform or triggering audible and visual alarms to ensure rapid response and minimize losses.

[0155] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A smart safety hazard inspection system based on image acquisition and recognition, characterized in that, include: Multimodal image acquisition module, hazard identification module, intelligent response decision-making module, and multi-dimensional correlation analysis and proactive prevention module; The multimodal image acquisition module is used to acquire image data in real time based on a camera array deployed in the monitoring area, preprocess the acquired image data to generate standardized image feature vectors, and transmit the generated standardized image feature vectors to the hazard identification module. The hazard identification module is used to pre-store a safety hazard feature library and a historical case database. Based on a deep learning model, it identifies and assesses risks from standardized image feature vectors and generates a hazard risk index. The disposal decision module is used to establish an intelligent disposal decision model based on the hidden danger risk index and a preset disposal strategy library. Based on the output of the disposal decision model, a disposal work order is generated, which includes hidden danger location, disposal suggestions and resource allocation information. The multi-dimensional correlation analysis and proactive prevention module is used to integrate hazard identification data, environmental parameters and historical handling records to mine correlation rules, generate hazard development trend maps, push proactive prevention strategies based on the maps, and realize data interaction and instruction distribution within the system.

2. The intelligent safety hazard inspection system based on image acquisition and recognition according to claim 1, characterized in that, The multimodal image acquisition module specifically includes: Real-time image acquisition unit: Based on visible light cameras, infrared thermal imagers, and panoramic cameras, it acquires image data of the monitoring area in real time, including equipment status images, environmental scene images, personnel behavior images, and scene images captured on site by mobile phones and panoramic cameras. The image data is appended with timestamps and location tags, and the acquired real-time image data is synchronized in time. Noise filtering unit: Uses Gaussian filtering algorithm to filter noise in the acquired image and normalizes the size of image data of different resolutions and formats; Image feature extraction unit: Using the gray-level co-occurrence matrix calculation formula, semantic features, texture features, and spatial features are extracted from the preprocessed image. The extracted features are standardized to obtain standardized image feature vectors. The standardized image feature vectors are then transmitted to the hazard identification module through the multi-dimensional correlation analysis and proactive prevention module for further hazard identification and risk assessment.

3. The intelligent safety hazard inspection system based on image acquisition and recognition according to claim 1, characterized in that, The hazard identification module specifically includes: Hazard detection unit: Based on a trained YOLOv7 deep learning model, it performs target detection on standardized image feature vectors to identify safety hazard targets in the image; Risk assessment unit: Combining historical data and current status information of potential hazards, the analytic hierarchy process (AHP) is used to solve for the weight parameters of the standardized image feature vector. The weight parameters of the standardized image feature vector are used as input, and the risk level prediction data is used as output to establish a risk assessment model. The risk level of the potential hazard is estimated based on classification analysis. Hazard risk index generation unit: Based on hazard detection results and risk assessment data, the hazard risk index is calculated comprehensively using a weighted average method; Early warning and alarm unit: Based on the comparison between the hidden danger risk index and the threshold, it determines whether there are potential safety hazards. If the hidden danger risk index is higher than the set threshold, the system will trigger an alarm mechanism to notify safety management personnel to take action.

4. The intelligent safety hazard inspection system based on image acquisition and recognition according to claim 3, characterized in that, The risk assessment unit includes: The analytic hierarchy process (AHP) is used to construct a judgment matrix, obtain the relative importance weights of each potential risk factor, construct a risk assessment model, optimize the model by adjusting the weights, perform a consistency test on the risk assessment model, and calculate the consistency ratio. The consistency ratio is compared with a set threshold. If the consistency ratio is less than the threshold, the judgment matrix has satisfactory consistency; otherwise, the judgment matrix is ​​readjusted. Based on the derived weight values ​​and standardized image feature vectors, a risk assessment model is established using a risk assessment formula to estimate the risk level of potential hazards.

5. The intelligent safety hazard inspection system based on image acquisition and recognition according to claim 3, characterized in that, The hidden danger risk index generation unit specifically includes: Obtain the hazard category label and corresponding confidence level output by the hazard detection unit, and retrieve the basic risk coefficient of the hazard category from the safety hazard feature library; Extract the predicted risk level values ​​generated by the risk assessment unit, and calculate the hidden danger risk index using a weighted fusion formula; The calculated risk index is compared with a preset threshold to generate a risk level label.

6. The intelligent safety hazard inspection system based on image acquisition and recognition according to claim 5, characterized in that: The preset threshold is determined through historical data modeling, collecting the hazard risk index and actual handling results over the past 12 months, and using the percentile method to divide the threshold range: Low risk threshold: I <P 30 , where P 30 30th percentile; Medium risk threshold: P 30 ≤I≤P 60 , where P 60 60th percentile; High-risk threshold: P 60 <I≤P 90 , where P 90 90th percentile; Emergency risk threshold: I>P 90 ; The system automatically matches the response mechanism according to different risk threshold ranges. Low risk is only recorded and archived, medium risk is pushed with prompt information, high risk generates a handling work order, and emergency risk triggers an audible and visual alarm and pushes it to the emergency command platform simultaneously.

7. The intelligent safety hazard inspection system based on image acquisition and recognition according to claim 6, characterized in that, The intelligent decision-making module specifically includes: Disposal strategy generation unit: Based on the hazard risk index and historical disposal data, combined with different hazard types and disposal scenarios, it generates targeted disposal strategies; Optimize decision-making unit: Adjust and optimize the handling strategy based on reinforcement learning algorithm to select the most suitable handling plan for the current hidden danger status; Task allocation unit for response personnel: Response tasks are allocated based on the type of hazard, the location of the monitoring area, and the skills of the response personnel.

8. The intelligent safety hazard inspection system based on image acquisition and recognition according to claim 6, characterized in that, The process of adjusting and optimizing the handling strategy based on reinforcement learning algorithms to select the most suitable handling plan for the current potential hazard status specifically includes: Collect risk index, hazard type, historical handling data, and handling effect data of potential hazards; select handling strategies based on hazard status; and generate an initial solution. A reward function is constructed based on the risk index, hazard type, and disposal cost factors of the hazard, with the aim of maximizing the disposal effect and minimizing the disposal cost. The reward function is solved, and the disposal strategy is optimized based on the deep Q-network algorithm. In each iteration, the strategy is adjusted according to the learning mechanism of the algorithm to maximize the reward function. The most suitable disposal solution is output based on the optimization result.

9. A smart safety hazard inspection system based on image acquisition and recognition according to claim 6, characterized in that, The allocation of disposal tasks based on the type of hazard, the location of the monitoring area, and the skills of the disposal personnel specifically includes: assessing the urgency of the hazard based on its risk index, type, and scope of impact; calculating the priority of each disposal task; and scheduling the disposal tasks based on their priorities.

10. A smart safety hazard inspection system based on image acquisition and recognition according to claim 6, characterized in that, The multi-dimensional correlation analysis and proactive prevention module specifically includes: Cross-dimensional data association unit: Integrates image recognition data, environmental sensor data, and equipment operation logs to establish a multi-source data association rule base, and uses data association rule mining technology to identify potential correlations between hidden dangers and the environment and equipment status; Hazard Chain Prediction Unit: Based on the association rule base, it predicts the secondary hazard chains that the current hazard may cause through time series analysis, and outputs the simulation results of the hazard diffusion path; Proactive prevention strategy generation unit: Combining the hazard chain prediction results with historical handling cases, it generates prevention strategies that include pre-intervention measures and pushes them to the handling decision module.

Images