HSE system auditing method based on artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence, and discloses an artificial intelligence-based HSE system auditing method, which comprises the following steps of: 1, acquiring visual data, environment data and equipment state data by using a camera, an unmanned aerial vehicle and an Internet of Things sensor, and extracting text data in a regulation library, a historical auditing record and an accident report, timestamp synchronization processing is carried out on the data after collection; 2, performing denoising, image enhancement and size normalization processing on the visual data; space-time correlation optimization processing is adopted, a dynamic time warping algorithm is utilized to calculate the similarity among visual data, environment data and equipment state data, and time alignment and modal alignment are performed, so that efficient fusion of cross-modal data is realized, the consistency and integrity of data input are improved, information deviation caused by data asynchronization is reduced, and the accuracy of data fusion is improved. The HSE auditing system can keep high accuracy under the condition of multi-source data input, and the reliability of an auditing result is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and specifically provides an HSE system audit method based on artificial intelligence. Background Art

[0002] The HSE (Health, Safety, Environment) management system is an important management framework for modern industrial enterprises to ensure production safety, environmental protection and occupational health. Traditional HSE audits mainly rely on manual inspections, paper document records and empirical judgment methods, with low audit efficiency, being easily affected by human factors, and unable to cover complex production environments. With the development of industrial Internet and artificial intelligence technologies, using data-driven methods to improve the automation and accuracy of HSE system audits has become a research hotspot.

[0003] Existing HSE audit methods usually rely on a single data source, such as analyzing potential safety hazards based on video surveillance and detecting abnormal situations through environmental sensors, but they cannot achieve the fusion of multi-modal data, resulting in limitations in the accuracy and comprehensiveness of audit results. In addition, traditional risk prediction models mostly adopt linear regression and statistical analysis methods, unable to capture complex relationships between potential safety hazards, leading to a decrease in the accuracy of risk assessment. At the same time, HSE risk early warning usually adopts the method of setting fixed thresholds and cannot be dynamically adjusted according to the characteristics of different production environments, resulting in a high false alarm rate and affecting the management decisions of enterprises.

[0004] Therefore, those skilled in the art provide an HSE system audit method based on artificial intelligence to solve the above-mentioned problems. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention provides an HSE system audit method based on artificial intelligence to solve the problems raised in the above background art.

[0006] To achieve the above objectives, the present invention is realized through the following technical solutions: An HSE system audit method based on artificial intelligence, including:

[0007] Step 1: Use cameras, drones, and Internet of Things sensors to obtain visual data, environmental data, and equipment status data, and extract text data from the regulation library, historical audit records, and accident reports. After the data is collected, timestamp synchronization processing is performed;

[0008] Step 2: Perform denoising, image enhancement, and size normalization processing on the visual data, perform missing value filling, outlier removal, and normalization processing on the environmental data and equipment status data, and perform word segmentation, stop word removal, and syntactic structure parsing on the text data;

[0009] Step 3: Use a convolutional neural network to extract the spatial features of visual data, and use a temporal neural network to model the time dependence of environmental data and device status data to obtain time-varying patterns. Use a pre-trained language model to perform embedding transformation on text data to generate a vectorized representation;

[0010] Step 4: Perform feature alignment processing on the feature vectors from different sources, use a cross-modal deep learning model to establish a unified feature space representation, and allocate the contribution degrees of different modal data through an attention mechanism to form a comprehensive feature vector. The comprehensive feature vector is dynamically adjusted according to the contribution of different data sources during the training stage;

[0011] Step 5: Based on the fused comprehensive feature vector, use a time series modeling method to learn the risk evolution law, and combine a supervised learning method to construct a prediction model. The prediction model calculates the risk score after inputting the feature vector of the current audit environment and is dynamically corrected based on historical audit data;

[0012] Step 6: According to the HSE risk prediction score, combine the safety threshold set by industry standards and the statistical distribution of historical accident data to calculate the risk level, generate corresponding audit suggestions and alarm information. The alarm information includes the detection results of violation behaviors, accident categories, and the scope of influence, and is sorted by priority according to the risk level;

[0013] Step 7: Explain the decision-making process of the HSE risk prediction model, use a feature contribution degree calculation method to quantify the influence degree of different data modalities on the audit results, and combine a visualization tool to generate an audit report. The audit report includes the basis for risk assessment, the contribution degree of key features, the prediction credibility, and the relevant regulation matching situation;

[0014] Step 8: Receive the feedback information from the auditors, record the false alarm and missed alarm situations, and use an adaptive optimization algorithm to dynamically adjust the parameters of the HSE risk prediction model. Perform incremental learning based on the newly added audit data to enable the model to optimize its decision-making ability, and update the industry regulation database to adapt to new audit standards and risk management requirements.

[0015] Preferably, the visual data, environmental data, and device status data obtained in Step 1 are processed through spatio-temporal correlation optimization to improve the consistency and integrity of the data. Among them, spatio-temporal correlation optimization uses a dynamic time warping algorithm to calculate the similarity of data sequences, which is defined as follows:

[0016] D m,n = min(D m-1,n-1 , D m-1,n , D m,n-1 ) + d(x m , y n ),

[0017] Among them, D m,n is the alignment distance between the visual data at the m-th time step and the environmental data at the n-th time step; x m is the eigenvalue of the visual data at time step m;

[0018] D m-1,n-1 is the alignment distance between the visual data at the (m - 1)-th time step and the environmental data at the (n - 1)-th time step; y n is the eigenvalue of the environmental data at time step n;

[0019] D m-1,n is the alignment distance between the visual data at the (m - 1)-th time step and the environmental data at the n-th time step; D m,n-1 is the alignment distance between the visual data at the m-th time step and the environmental data at the (n - 1)-th time step; d(x m , y n ) is the Euclidean distance calculation formula.

[0020] Preferably, in the preprocessing of the visual data in step 2, a method based on Fourier transform is used to remove noise signals and improve the image quality. The Fourier transform formula is as follows:

[0021]

[0022] Among them, F(u, v) is the image data in the frequency domain, f(x, y) is the original image data in the spatial domain, M is the width of the image, N is the height of the image, (u, v) is the frequency coordinate, j is the imaginary unit, is the basis function of the Fourier transform, and (x, y) is the pixel coordinate in the spatial domain image.

[0023] Preferably, when performing embedding transformation on the text data in step 3, an improved self-attention mechanism is used to optimize the text feature representation. The calculation method is as follows:

[0024]

[0025]

[0026] Among them, A i,j is the attention weight between the i-th word and the j-th word in the text data,

[0027] Q i is the query vector, K j is the key vector, T is the transpose symbol, d k is the dimension of the key vector K j , and n is the total number of words in the text data; exp(·) is the exponential function;

[0028] Z iis the final feature representation of the i-th word in the text data, V j is the value vector.

[0029] Preferably, the multi-modal data fusion in step 4 adopts an adaptive weighting strategy, dynamically adjusts the weights according to the importance of different modal data, and the calculation method is as follows:

[0030]

[0031] where, W p is the weighting coefficient of the p-th data modality, γ p is the data credibility score of the p-th data modality, F p is the feature vector of the p-th data modality, N is the total number of data modalities,

[0032] exp(·) is the exponential function, γ q is the data credibility score of the q-th data modality, and Z is the integrated feature vector after fusion.

[0033] Preferably, the HSE risk prediction model in step 5 combines a graph convolutional neural network and time series modeling to improve the accuracy of risk prediction by constructing a safety hazard association graph. The graph convolutional calculation method is as follows:

[0034] H (l+1) =σ(D -1 / 2 AD -1 / 2 H (l) W l ),

[0035] where, H (l) is the feature matrix of the l-th layer, H (l+1) is the feature matrix of the l+1-th layer,

[0036] A is the adjacency matrix, D is the degree matrix, D -1 / 2 is the graph normalization matrix,

[0037] W l is the training parameter matrix of the l-th layer, and σ(·) is the non-linear activation function.

[0038] Preferably, the early warning decision generation in step 6 adopts Bayesian optimization based on Gaussian process to dynamically adjust the safety threshold, and the optimization objective is as follows:

[0039]

[0040] where, T * is the optimal safety threshold, L(T) is the false alarm and missed alarm loss function of risk warning, and E[·] is the expected value calculation.

[0041] Preferably, the interpretability analysis of the audit results in step 7 uses the SHAP method to calculate the feature contribution, and the specific calculation is as follows:

[0042]

[0043] Among them, φ j Represents the importance contribution of feature j in the risk prediction model decision.

[0044] S is a feature subset, N is a set of features, |S| is the number of features in subset S, |N| is the size of the full set of features, v(S) is the model prediction score of feature subset S,

[0045] v(S∪{j}) is the change in risk prediction score after feature j is added to subset S.

[0046] Preferably, the feedback and system optimization in step 8 adopts an incremental learning mechanism to dynamically update the HSE risk prediction model, and the optimization objectives are as follows:

[0047]

[0048] Among them, θ * is the optimized model parameter, y t is the actual risk level, f θ (X t ) is the current model prediction value, l(·) is the loss function, θ is the current model parameter, T is the time window length, X t is the input feature at time step t.

[0049] Preferably, the HSE system audit method integrates augmented reality technology to visualize the warning information, and combines Grad-CAM technology to generate a risk area heat map, and the calculation method is as follows:

[0050]

[0051] in, is the Grad-CAM heat map of category c, k is the channel index of the convolutional layer feature map, is the gradient weight coefficient of category c on the kth channel of the feature map, A k is the kth channel of the feature map.

[0052] The present invention provides an HSE system audit method based on artificial intelligence. It has the following beneficial effects:

[0053] 1. Through the adoption of spatio-temporal correlation optimization processing, the present invention uses the dynamic time warping algorithm to calculate the similarity between visual data, environmental data, and device status data, and performs time alignment and modality alignment to achieve efficient fusion of cross-modal data, improve the consistency and integrity of data input, reduce information deviation caused by data asynchronization, enable the HSE audit system to maintain high accuracy under the condition of multi-source data input, and improve the reliability of audit results.

[0054] 2. By combining the graph convolutional neural network and time series modeling, the present invention constructs a safety hazard correlation graph to capture the complex correlation relationships between HSE risk factors, and uses the graph structure learning method to optimize the risk assessment process to achieve HSE risk prediction, avoid the limitation that traditional linear prediction models cannot handle complex risk factor interactions, improve the accuracy of risk prediction, enhance the interpretability of audit results, and enable managers to clearly understand the source of risks.

[0055] 3. By dynamically adjusting the safety threshold based on Bayesian optimization of Gaussian processes, combining industry standards and the distribution of historical accident data, the present invention adaptively optimizes the alarm trigger mechanism of the HSE audit system, realizes dynamic adjustment of the risk warning sensitivity in different production environments, solves the problem of poor adaptability of traditional fixed-threshold warning methods, reduces the occurrence of false alarms and missed alarms, enables the HSE warning system to match the safety standards of different enterprises and production lines, and improves the reliability and practicality of warning decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is a flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0057] To enable those skilled in the art to understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Based on the embodiments of the present invention, other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0058] The present invention will be described in detail below with reference to the accompanying drawings:

[0059] Embodiment:

[0060] Please refer to the attached Figure 1 , the embodiment of the present invention provides an HSE system audit method based on artificial intelligence, including:

[0061] Step 1: Obtain visual data, environmental data, and equipment status data using cameras, drones, and Internet of Things sensors, and extract text data from the regulation library, historical audit records, and accident reports. After data collection, perform timestamp synchronization processing;

[0062] Step 2: Perform denoising, image enhancement, and size normalization on visual data, perform missing value filling, outlier removal, and normalization on environmental data and equipment status data, and perform word segmentation, stop word removal, and syntactic structure parsing on text data;

[0063] Step 3: Use a convolutional neural network to extract the spatial features of visual data, and use a temporal neural network to model the time dependence of environmental data and equipment status data to obtain time-varying patterns. Use a pre-trained language model to perform embedding transformation on text data to generate vectorized representations;

[0064] Step 4: Perform feature alignment processing on feature vectors from different sources, use a cross-modal deep learning model to establish a unified feature space representation, and assign the contribution degrees of different modal data through an attention mechanism to form a comprehensive feature vector. The comprehensive feature vector is dynamically adjusted according to the contribution of different data sources during the training stage;

[0065] Step 5: Based on the fused comprehensive feature vector, use a time series modeling method to learn the risk evolution law, and combine a supervised learning method to construct a prediction model. The prediction model calculates the risk score after inputting the feature vector of the current audit environment, and performs dynamic correction based on historical audit data;

[0066] Step 6: According to the HSE risk prediction score, combine the safety threshold set by industry standards and the statistical distribution of historical accident data to calculate the risk level, generate corresponding audit suggestions and alarm information. The alarm information includes the detection results of violations, accident categories, and the scope of influence, and is sorted by priority according to the risk level;

[0067] Step 7: Explain the decision-making process of the HSE risk prediction model, use a feature contribution calculation method to quantify the influence degree of different data modalities on the audit results, and combine a visualization tool to generate an audit report. The audit report includes the basis for risk assessment, the contribution degree of key features, the prediction credibility, and the relevant regulation matching situation;

[0068] Step 8: Receive the feedback information from the auditors, record the false alarm and missed alarm situations, and use an adaptive optimization algorithm to dynamically adjust the parameters of the HSE risk prediction model. Perform incremental learning based on the newly added audit data to enable the model to optimize its decision-making ability, and update the industry regulation database to adapt to new audit standards and risk management requirements.

[0069] In the first step, visual data, environmental data, and equipment status data are obtained by using cameras, drones, and Internet of Things sensors. Text data in the regulation library, historical audit records, and accident reports is extracted to achieve comprehensive collection of multi-source data. Timestamp synchronization processing ensures the temporal consistency of the data, enabling subsequent analysis to accurately correlate data of different modalities, reducing information deviation, and improving the reliability of HSE audits.

[0070] In the second step, denoising, image enhancement, and size normalization are performed on the visual data. Missing value filling, outlier removal, and normalization are carried out on the environmental data and equipment status data. Word segmentation, stop word removal, and syntactic structure parsing are performed on the text data, significantly improving the data quality. Clean and standardized data provides a solid foundation for subsequent feature extraction, reducing noise interference in model training and improving the input accuracy of the audit system.

[0071] In the third step, a convolutional neural network is used to extract spatial features of the visual data. A temporal neural network is used to model the temporal dependence of the environmental data and equipment status data. A pre-trained language model is used to perform embedding transformation on the text data to generate vectorized representations. Through deep learning methods, complex patterns in each data modality are captured, enhancing the feature expression ability and providing a rich information basis for subsequent risk assessment.

[0072] In the fourth step, by performing feature alignment on the feature vectors from different sources, a cross-modal deep learning model is used to establish a unified feature space representation. The contribution degrees of different modality data are allocated through an attention mechanism to form a comprehensive feature vector, and the weights are dynamically adjusted during the training phase to optimize the integration effect of multi-source data. The dynamic adjustment mechanism improves the flexibility and robustness of the fusion, meeting the requirements of different audit scenarios.

[0073] In the fifth step, based on the fused comprehensive feature vector, a time series modeling method is used to learn the risk evolution law. A prediction model is constructed by combining supervised learning methods and dynamically corrected based on historical audit data to enhance the dynamic adaptability of risk prediction. Time series modeling captures the time-varying characteristics of risks, supervised learning ensures the accuracy of prediction, and the dynamic correction mechanism enables the model to continuously optimize and adapt to changes in complex industrial environments.

[0074] In the sixth step, according to the HSE risk prediction score, combined with the safety threshold set by industry standards and the statistical distribution of historical accident data, the risk level is calculated. Audit suggestions and alarm information including violation detection results, accident categories, and impact ranges are generated, prioritized according to the risk level, and practical and prioritized early warning information is provided. Prioritization optimizes resource allocation, and detailed information supports rapid decision-making by managers, enhancing the response efficiency of HSE management.

[0075] In the seventh step, the decision-making process of the HSE risk prediction model is explained. The feature contribution calculation method is used to quantify the influence degree of different data modalities on the audit results. Combining with visualization tools, an audit report including risk assessment basis, key feature contribution, prediction credibility, and regulation matching situation is generated, significantly enhancing the transparency of the audit results. Interpretability analysis improves the credibility of the system, provides clear decision-making basis for auditors, and enhances the acceptance and practicality of HSE audits.

[0076] In the eighth step, feedback information from auditors is received, false alarms and missed alarms are recorded, and the parameters of the HSE risk prediction model are dynamically adjusted using an adaptive optimization algorithm. Incremental learning is carried out based on the newly added audit data, and the industry regulation database is updated to improve the long-term adaptability of the system. The incremental learning mechanism ensures that the model can continuously learn and improve, and the update of the regulation database maintains the compliance of the system, enhancing the sustainability and adaptability of HSE audits.

[0077] The visual data, environmental data, and equipment status data obtained in Step 1 are processed through spatio-temporal correlation optimization to improve the consistency and integrity of the data. Among them, spatio-temporal correlation optimization uses the dynamic time warping algorithm to calculate the similarity of data sequences, which is defined as follows:

[0078] D m,n =min(D m-1,n-1 ,D m-1,n ,D m,n-1 )+d(x m ,y n ),

[0079] where D m,n is the alignment distance between the m-th time step of visual data and the n-th time step of environmental data; x m is the feature value of visual data at time step m;

[0080] D m-1,n-1 is the alignment distance between the (m - 1)-th time step of visual data and the (n - 1)-th time step of environmental data; y n is the feature value of environmental data at time step n;

[0081] D m-1,n is the alignment distance between the (m - 1)-th time step of visual data and the n-th time step of environmental data; D m,n-1 is the alignment distance between the m-th time step of visual data and the (n - 1)-th time step of environmental data; d(x m ,y n ) is the Euclidean distance calculation formula.

[0082] The dynamic time warping algorithm can calculate the alignment distance of various types of data between different time steps, enabling the precise synchronization and fusion of visual data, environmental data, and device status data, avoiding data inconsistency problems caused by time differences, ensuring the accurate expression of the relationships between different data modalities, and improving the credibility of the data.

[0083] Through spatio-temporal correlation optimization processing, the system can effectively fill in the blank parts in the data. Especially during the process of aligning time steps between different data sources, the dynamic time warping algorithm can minimize information loss and redundancy caused by time asynchrony, ensure that the data of each time step can participate in subsequent analysis processes, and improve the integrity of the data.

[0084] During the alignment process, the dynamic time warping algorithm can discover potential similarities and differences between different data modalities and can reveal abnormal situations in the data. By comparing the similarity between visual data and environmental data, the system can identify potential abnormal states, correct or mark them through an automated mechanism, and provide support for subsequent risk prediction and management.

[0085] Spatio-temporal correlation optimization provides consistent data input for the training and analysis of subsequent models. Through precisely aligned data, the training efficiency and accuracy of deep learning models can be significantly improved, and errors caused by data inconsistency and missing data can be reduced.

[0086] In the visual data preprocessing in Step 2, a method based on Fourier transform is used to remove noise signals and improve the image quality. The Fourier transform formula is as follows:

[0087]

[0088] where F(u,v) is the image data in the frequency domain, f(x,y) is the original image data in the spatial domain, M is the width of the image, N is the height of the image, (u,v) are the frequency coordinates, j is the imaginary unit, is the basis function of the Fourier transform, and (x,y) are the pixel coordinates in the spatial domain image.

[0089] The Fourier transform can convert the signals in the image to the frequency domain, separating the noise signals from the effective signals. By filtering out high-frequency noise components, the influence of noise on the image quality is greatly reduced, and the accuracy of subsequent analysis and processing is improved.

[0090] The removal of noise makes the image clear, enhancing the details and features of the image. A clear image helps to accurately extract visual features, especially in object detection and pattern recognition, improving the effect of image processing.

[0091] The image processed by Fourier transform denoising has a high signal-to-noise ratio and reduces unnecessary interference components. Subsequent image analysis can be performed on pure image data, improving the accuracy and reliability of the analysis.

[0092] Through frequency domain processing, Fourier transform can effectively preserve the important low-frequency features in the image while removing unimportant high-frequency noise. Therefore, it can retain the effective information of the image while denoising, avoiding the loss of important information due to over-filtering.

[0093] Fourier transform can convert complex operations in the spatial domain into simple operations in the frequency domain, reducing the computational amount and improving the efficiency of image processing. When processing large-scale image data, Fourier transform can provide a fast calculation speed.

[0094] When performing embedding transformation on the text data in step 3, an improved self-attention mechanism is used to optimize the text feature representation, and the calculation method is as follows:

[0095]

[0096] where A i,j is the attention weight between the i-th word and the j-th word in the text data,

[0097] Q i is the query vector, K j is the key vector, T is the transpose symbol, d k is the dimension of the key vector K j and n is the total number of words in the text data, exp(·) is the exponential function;

[0098] Z i is the final feature representation of the i-th word in the text data, and V j is the value vector.

[0099] Through the self-attention mechanism, the model can consider the relationships between each word in the text and other words, especially long-distance dependencies. This means that even distant words in the text can influence each other, improving the accuracy and richness of the text feature representation.

[0100] The self-attention mechanism captures the semantic information of the text at a deep level by focusing on words in different positions. Especially when dealing with complex long sentences and texts with multiple meanings, it can help the model understand and accurately extract relevant information, improving the model's expressive ability.

[0101] Compared with traditional convolutional neural networks or recurrent neural networks based on fixed windows, the improved self-attention mechanism can comprehensively focus on the relationships between all words in the text, reducing information loss.

[0102] The multi-modal data fusion in Step 4 adopts an adaptive weighting strategy, dynamically adjusts the weights according to the importance of different modal data, and the calculation method is as follows:

[0103]

[0104] Among them, W p is the weighting coefficient of the p-th data modality, γ p is the data credibility score of the p-th data modality, F p is the feature vector of the p-th data modality, N is the total number of data modalities,

[0105] exp(·) is the exponential function, γ q is the data credibility score of the q-th data modality, and Z is the integrated feature vector after fusion.

[0106] By adopting the adaptive weighting strategy for multi-modal data fusion, the weights in the final decision can be dynamically adjusted according to the credibility of each data modality, significantly improving the accuracy and flexibility of data fusion. Enhancing the adaptability and robustness of the system, enabling the system to effectively integrate and analyze data in a multi-source data environment, and providing comprehensive and accurate support for safety management and risk assessment.

[0107] The HSE risk prediction model in Step 5 combines graph convolutional neural network and time series modeling, and improves the accuracy of risk prediction by constructing a safety hazard association graph. The graph convolutional calculation method is as follows:

[0108] H (l+1) =σ(D -1 / 2 AD -1 / 2 H (l) W l ),

[0109] Among them, H (l) is the feature matrix of the l-th layer, H (l+1) is the feature matrix of the l+1-th layer,

[0110] A is the adjacency matrix, D is the degree matrix, D -1 / 2 is the graph normalization matrix,

[0111] W l is the training parameter matrix of the l-th layer, and σ(·) is the non-linear activation function.

[0112] Combining the graph convolutional neural network with the time series modeling method, by constructing a safety hazard association graph, the accuracy of HSE risk prediction can be significantly improved. This method can effectively capture the complex relationship between safety hazards, and can handle spatiotemporal dynamic dependencies, improving the timeliness and accuracy of the prediction model. At the same time, the graph structure enhances the model's learning ability for multi-dimensional data, improves the comprehensiveness and interpretability of risk assessment, and provides reliable and accurate prediction support for safety management. Provide technical support for safety management and risk prediction of industrial enterprises.

[0113] The early warning decision generation in step 6 uses Bayesian optimization based on Gaussian process to dynamically adjust the safety threshold. The optimization objectives are as follows:

[0114]

[0115] Among them, T * is the optimal safety threshold, L(T) is the false positive and false negative loss function of risk warning, and E[·] is the expected value calculation.

[0116] The Bayesian optimization method based on Gaussian process can significantly improve the accuracy and robustness of HSE risk early warning system by dynamically adjusting the safety threshold. This method can reduce false positives and false negatives, optimize the safety threshold, and make the early warning system flexible and accurate in dealing with various safety risks. By improving the decision-making efficiency and accuracy of the early warning system, it can provide reliable safety management and risk prediction support for industrial enterprises. The method strengthens the intelligence and automation of the safety management system and provides a technical basis for the safety assurance of enterprises.

[0117] The interpretability analysis of the audit results in step 7 uses the SHAP method to calculate the feature contribution, which is calculated as follows:

[0118]

[0119] Among them, φ j Represents the importance contribution of feature j in the risk prediction model decision.

[0120] S is a feature subset, N is a set of features, |S| is the number of features in subset S, |N| is the size of the full set of features, v(S) is the model prediction score of feature subset S,

[0121] v(S∪{j}) is the change in risk prediction score after feature j is added to subset S.

[0122] The SHAP method is used to analyze the audit results for interpretability and quantify the contribution of each feature to risk prediction decisions. This enhances the transparency and credibility of the model, provides data support for risk assessment and safety decision-making, helps optimize the model structure and formulate personalized safety strategies, and ultimately improves the effectiveness of the safety management system.

[0123] The feedback and system optimization in Step 8 adopt an incremental learning mechanism to dynamically update the HSE risk prediction model, and the optimization objectives are as follows:

[0124]

[0125] where θ * is the optimized model parameter, y t is the true risk level, f θ (X t ) is the current model prediction value, l(·) is the loss function, θ is the current model parameter, T is the time window length, and X t is the input feature at time step t.

[0126] Dynamically updating the HSE risk prediction model through the incremental learning mechanism can significantly improve the adaptability and prediction accuracy of the model. Compared with traditional learning methods, incremental learning can save computing resources and ensure the continuous effectiveness of the model after long-term operation. Through real-time feedback and optimization, the system can always maintain high performance and reliability, providing strong support for safety management. The method greatly enhances the flexibility and intelligence level of the system, and improves the efficiency of HSE risk prediction and management.

[0127] The HSE system audit method integrates augmented reality technology to visually present early warning information, and combines Grad-CAM technology to generate a heat map of the risk area. The calculation method is as follows:

[0128]

[0129] where is the Grad-CAM heat map of category c, k is the channel index of the convolutional layer feature map, is the gradient weighting coefficient of category c on the k-th channel of the feature map, and A k is the k-th channel of the feature map.

[0130] Combining the heat map generated by augmented reality technology and Grad-CAM can effectively improve the visualization and interactivity of HSE risk early warning information. The method improves the efficiency and accuracy of decision-making, and enhances the risk identification ability and emergency response ability of staff. Through visualization technology, the safety management system becomes intuitive and user-friendly, enhancing the effectiveness and efficiency of overall safety management. Technological innovation provides intelligent safety management support for industrial enterprises, improving the risk prediction and response capabilities.

[0131] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An HSE system audit method based on artificial intelligence, characterized in that Including: Step 1: Obtain visual data, environmental data, and device status data using cameras, drones, and Internet of Things sensors, and extract text data from the regulation library, historical audit records, and accident reports. After data collection, timestamp synchronization processing is performed. Step 2: Denoise, enhance the image, and normalize the size of the visual data. Fill in missing values, remove outliers, and normalize the environmental data and device status data. Perform word segmentation, remove stop words, and parse the syntactic structure of the text data. Step 3: Use a convolutional neural network to extract the spatial features of the visual data, and use a temporal neural network to model the time dependence of the environmental data and device status data to obtain time-varying patterns. Use a pre-trained language model to perform embedding transformation on the text data to generate a vectorized representation. Step 4: Perform feature alignment processing on the feature vectors from different sources, use a cross-modal deep learning model to establish a unified feature space representation, and allocate the contribution degrees of different modal data through an attention mechanism to form a comprehensive feature vector. The comprehensive feature vector is dynamically adjusted according to the contribution of different data sources during the training phase. Step 5: Based on the fused comprehensive feature vector, use a time series modeling method to learn the risk evolution law, and combine a supervised learning method to construct a prediction model. The prediction model calculates the risk score after inputting the feature vector of the current audit environment, and performs dynamic correction based on historical audit data. Step 6: According to the HSE risk prediction score, combine the safety threshold set by industry standards and the statistical distribution of historical accident data to calculate the risk level, generate corresponding audit suggestions and alarm information. The alarm information includes the detection results of violation behaviors, accident categories, and the scope of influence, and is sorted by priority according to the risk level. Step 7: Explain the decision-making process of the HSE risk prediction model, use a feature contribution degree calculation method to quantify the influence degree of different data modalities on the audit results, and generate an audit report in combination with a visualization tool. The audit report includes the basis for risk assessment, the contribution degree of key features, the prediction credibility, and the relevant regulation matching situation. Step 8: Receive the feedback information from the auditors, record the false alarm and missed alarm situations, and use an adaptive optimization algorithm to dynamically adjust the parameters of the HSE risk prediction model. Perform incremental learning based on the newly added audit data to enable the model to optimize its decision-making ability, and update the industry regulation database to adapt to the new audit standards and risk management requirements.

2. The HSE system audit method based on artificial intelligence according to claim 1, wherein The visual data, environmental data, and device status data obtained in Step 1 are processed through spatio-temporal correlation optimization to improve the consistency and integrity of the data. Among them, spatio-temporal correlation optimization uses the dynamic time warping algorithm to calculate the similarity of data sequences, defined as follows: D m,n = min(D m-1,n-1 , D m-1,n , D m,n-1 ) + d(x m , y n ), Among them, D m,n is the alignment distance between the m-th time step of visual data and the n-th time step of environmental data; x m is the eigenvalue of visual data at time step m; D m-1,n-1 is the alignment distance between the visual data at the (m - 1)-th time step and the environmental data at the (n - 1)-th time step; y n is the eigenvalue of the environmental data at time step n; D m-1,n is the alignment distance between the visual data at the (m-1)-th time step and the environmental data at the n-th time step; D m,n-1 is the alignment distance between the visual data at the m-th time step and the environmental data at the (n-1)-th time step; d(x m , y n ) is the Euclidean distance calculation formula.

3. The HSE system audit method based on artificial intelligence according to claim 2, wherein, The preprocessing of the visual data in Step 2 uses a method based on Fourier transform to remove noise signals and improve the image quality. The Fourier transform formula is as follows: Among them, F(u, v) is the image data in the frequency domain, f(x, y) is the original image data in the spatial domain, M is the width of the image, N is the height of the image, (u, v) is the frequency coordinate, and j is the imaginary unit. is the basis function of the Fourier transform, and (x, y) are the pixel coordinates in the spatial domain image.

4. The method for auditing the HSE system based on artificial intelligence according to claim 3, characterized in that, When performing embedding transformation on the text data in Step 3, an improved self-attention mechanism is used to optimize the text feature representation. The calculation method is as follows: Among them, A i,j is the attention weight between the i-th word and the j-th word in the text data, Q i is the query vector, K j is the key vector, T is the transpose symbol, d k is the dimension of the key vector K j is the total number of words in the text data, exp(·) is the exponential function; Z i is the final feature representation of the i-th word in the text data, and V j is the value vector.

5. The method for auditing the HSE system based on artificial intelligence according to claim 4, characterized in that The multimodal data fusion in step 4 adopts an adaptive weighting strategy to dynamically adjust the weights according to the importance of different modal data. The calculation method is as follows: Among them, W p is the weighting coefficient of the p-th data modality, γ p is the data credibility score of the p-th data modality, F p is the feature vector of the p-th data modality, and N is the total number of data modalities. exp(·) is the exponential function, and γ q is the data credibility score of the q-th data modality, and Z is the fused comprehensive feature vector.

6. The HSE system audit method based on artificial intelligence according to claim 5, characterized in that The HSE risk prediction model in step 5 combines graph convolutional neural network with time series modeling to improve the accuracy of risk prediction by constructing a safety hazard association graph. The graph convolution calculation method is as follows: H (l+1) = σ(D -1 / 2 AD -1 / 2 H (l) W l ), Among them, H (l) is the feature matrix of the l-th layer, and H (l+1) is the feature matrix of the (l + 1)-th layer. A is the adjacency matrix, D is the degree matrix, and D -1 / 2 is the graph normalization matrix, W l is the training parameter matrix of the l-th layer, and σ(·) is the non-linear activation function.

7. The method for auditing the HSE system based on artificial intelligence according to claim 6, wherein The early warning decision generation in step 6 adopts Bayesian optimization based on Gaussian process to dynamically adjust the safety threshold, and the optimization objectives are as follows: Among them, T * is the optimal safety threshold, L(T) is the false alarm and missed alarm loss function of risk warning, and E[·] is the calculation of the expected value.

8. The method for auditing the HSE system based on artificial intelligence according to claim 7, wherein The interpretability analysis of the audit results in step 7 uses the SHAP method to calculate the feature contribution, which is specifically calculated as follows: Among them, φ j represents the importance contribution of feature j in the decision-making of the risk prediction model. S is a feature subset, N is a set of features, |S| is the number of features in subset S, |N| is the size of the full set of features, v(S) is the model prediction score of feature subset S, v(S∪{j}) is the change in risk prediction score after feature j is added to subset S.

9. The method for auditing the HSE system based on artificial intelligence according to claim 8, wherein, The feedback and system optimization in step 8 adopts an incremental learning mechanism to dynamically update the HSE risk prediction model. The optimization objectives are as follows: Among them, θ * is the optimized model parameter, y t is the true risk level, f θ (X t ) is the current model prediction value, is the loss function, θ is the current model parameter, T is the time window length, X t is the input feature at time step t.

10. The method for auditing the HSE system based on artificial intelligence according to claim 9, characterized in that The HSE system audit method integrates augmented reality technology to visualize the warning information and generates a risk area heat map in combination with Grad-CAM technology. The calculation method is as follows: Among them, is the Grad-CAM heatmap of class c, k is the channel index of the convolutional layer feature map, is the gradient weighting coefficient of class c on the k-th channel of the feature map, and A k is the k-th channel of the feature map.

Images