Industrial scene safety assessment method and system based on generative adversarial network

Abstract

The invention relates to the technical field of safety assessment methods, in particular to an industrial scene safety assessment method and system based on a generative adversarial network, and the method comprises the steps: obtaining the historical data of an industrial scene; obtaining current industrial environment operation data to obtain a multi-source data set; generating an adversarial network based on historical data and training conditions to obtain a trained generator and discriminator; generating abnormal data of the industrial scene by using a generator based on the multi-source data set; judging the trueness of the abnormal data by using a discriminator, and adjusting the input data of the generator according to the trueness; real and effective abnormal data are obtained; key frames are extracted from real and effective abnormal data, and each frame of data is converted into nodes; generating an industrial safety risk map based on the nodes; converting the industrial safety risk map into a safety risk map by using a dynamic multi-source map reasoning module; and the protection strategy generator is used for extracting protection measures and outputting a protection scheme, so that the recognition capability of the system on unknown risks is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of security assessment methods, and more specifically, to an industrial scenario security assessment method and system based on a generative adversarial network. Background Art

[0002] As the level of intelligence and networking of industrial systems continues to increase, while improving production efficiency, it also brings unprecedented security challenges. Traditional industrial scenario security assessment methods mainly rely on expert experience and static rules, which are difficult to cope with the increasingly complex and dynamic industrial environment. In recent years, although some machine learning-based methods have been introduced into the field of industrial security assessment, there are still many shortcomings.

[0003] Existing industrial scenario safety assessment methods mainly face the following problems: First, data acquisition is difficult. Real industrial safety accident data is often scarce, and it is difficult to fully assess potential risks by relying solely on normal operation data. Second, the assessment model lacks generalization capabilities. Many methods can only deal with known types of safety risks and have limited ability to identify new or unknown risks. Furthermore, safety assessment results lack interpretability. Although some black box models can give risk scores, it is difficult to provide specific risk causes and protection recommendations. Finally, existing methods often regard safety risk assessment as a static problem, ignoring the dynamic characteristics of industrial systems and the complex interactions between various elements. Summary of the invention

[0004] In response to these problems, this paper proposes an industrial scenario safety assessment method and system based on generative adversarial networks. This method innovatively applies generative adversarial networks to the field of industrial safety, effectively solving a series of technical problems such as data scarcity, model generalization, interpretability, and dynamic evaluation.

[0005] The present invention provides an industrial scenario safety assessment method based on a generative adversarial network, comprising:

[0006] The acquisition steps include:

[0007] Acquire historical data of industrial scenarios, wherein the historical data includes micro sensor data, macro data, and safety risk assessment results;

[0008] Obtain current industrial environment operation data and obtain multi-source data sets;

[0009] Processing steps include:

[0010] Based on the historical data, training a conditional generative adversarial network to obtain a trained generator and discriminator;

[0011] Based on the multi-source data set, using the generator to generate abnormal data of the industrial scene;

[0012] Using the discriminator to judge the authenticity of the abnormal data, and adjusting the input data of the generator according to the authenticity;

[0013] Perform iterative training based on the difference between the real data corresponding to the abnormal data and the abnormal data to obtain real and effective abnormal data;

[0014] Extract key frames from the real and effective abnormal data, and convert each frame of data into a node;

[0015] Generate an industrial safety risk map based on the nodes;

[0016] Converting the industrial safety risk map into a safety risk map using a dynamic multi-source graph reasoning module;

[0017] Output steps include:

[0018] According to the security risk graph, protection measures are extracted through a protection strategy generator, and a protection plan is output.

[0019] Preferably, the training step of the conditional generative adversarial network specifically includes:

[0020] Constructing a two-layer generation network structure, including a first generation network G1 and a second generation network G2;

[0021] Generate macro data of industrial scenes using the first generation network G1;

[0022] Generate micro data of industrial scenes using the second generation network G2;

[0023] Construct a discriminant network D to judge the authenticity of the generated data.

[0024] Preferably, the step of determining the authenticity of abnormal data specifically includes:

[0025] Extracting an image from the abnormal data, and extracting features of the image;

[0026] quantizing the features into node vectors;

[0027] Comparing the node vector with the corresponding frame data to extract features in the node vector;

[0028] Convert the features into corresponding confidence levels;

[0029] The authenticity of the abnormal data is determined according to the confidence level.

[0030] Preferably, the node vector is expressed as:

[0031] V n ={Sn ,T b ,M n ,P,G,F n ,T n},

[0032] Among them, V n is the node vector in the corresponding environment, S n is the device status of the nth frame, T b is the type of device, M n is the operating mode of the nth device, P is the production status of the nth frame, G is the working mode, and F n is the fault node feature, T n It is the time node feature graph.

[0033] Preferably, the step of comparing the node vector with the corresponding data to be predicted to obtain the confidence level specifically includes:

[0034] Extract the vector values ​​of the node vectors at the corresponding time and each dimension of the image in the abnormal data, and compare the vector values ​​with the data values ​​in the data to be predicted to obtain the difference ratio Δ i ;

[0035] The difference ratio is converted into confidence, expressed as: C ij =f(Δ i );

[0036] According to the confidence level, the authenticity of the data to be predicted under abnormal data is calculated, which is expressed as: R e =g(C ij ).

[0037] Preferably, according to the industrial safety risk map, the steps of analyzing safety risks at multiple scales from micro to macro specifically include:

[0038] In the security risk map, a circular area is obtained with each node as the center and the node vector as the radius;

[0039] Extract the overlapping areas between the circular areas, determine whether each node matches the node vector in the overlapping area, and obtain the influence range of the node;

[0040] According to the size of the overlapping area, the influence weight of the node is obtained;

[0041] At the micro level, each node is matched with the node vector to obtain the node's environmental information, operation data and time information;

[0042] At the macro level, the environmental information and operation data of the same node are aggregated, and the time information is not involved in the aggregation to determine the security status at the macro level;

[0043] Combine the information at the micro level with the macro level to determine the current security status and obtain a security risk map.

[0044] Preferably, the step of extracting protection measures through the protection strategy generator specifically includes:

[0045] Analyze the safety risk map and the safety risk graph separately, and store the analysis results in the expert database, literature data and historical risk data respectively;

[0046] Map the security status, macro security risks and protective measures to obtain the characteristics of the current security risks;

[0047] Compare the features with the expert database, literature data and historical risk data respectively to extract corresponding features;

[0048] Output protection solutions based on corresponding features.

[0049] Preferably, the step of extracting key frames from real and effective abnormal data specifically includes:

[0050] During the training process, set the abnormal data threshold;

[0051] Use cross-comparison method to detect abnormal data;

[0052] The abnormal degree of the frame data is determined by the abnormal data threshold, and the frame data with the maximum confidence is obtained;

[0053] The frame data is defined as a key frame, and the corresponding node vector is the key abnormal data.

[0054] Preferably, the method further includes the step of constructing an interpretability module:

[0055] Construct a reasoning module to interpret and reason about the quantitative evaluation results;

[0056] Construct an explanation generation module to generate risk explanation pictures based on the explanations obtained by the reasoning module;

[0057] Through reasoning, we can obtain the threat field, vulnerability field, and impact field corresponding to the quantitative assessment results;

[0058] The threat field, vulnerability field, and impact field corresponding to the interpreted reasoning picture and quantitative assessment results will be used to generate a risk assessment report.

[0059] An industrial scenario safety assessment system based on a generative adversarial network for executing the method comprises:

[0060] The training module is used to train the conditional generative adversarial network based on historical data to obtain the trained generator and discriminator;

[0061] The acquisition module is used to collect the current industrial environment operation data and obtain a multi-source data set;

[0062] A generator training module, used to generate abnormal data of industrial scenarios using the generator;

[0063] A discriminator training module, used to use the discriminator to judge the authenticity of the abnormal data and adjust the input data of the generator according to the authenticity;

[0064] A risk assessment module, used to extract key frames from real and effective abnormal data and generate an industrial safety risk map based on the key frames;

[0065] A safety detection module, used for converting the industrial safety risk map into a safety risk map using a dynamic multi-source map reasoning module;

[0066] The strategy generation module is used to extract protection measures through the protection strategy generator according to the security risk map and output a protection plan.

[0067] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0068] The method of the present invention has many significant advantages and beneficial effects. First, through the training of the conditional generative adversarial network, the method can generate a large amount of real and effective abnormal data, which greatly enriches the data basis of security assessment and effectively overcomes the problem of scarcity of real abnormal data. This not only improves the training effect of the model, but also enhances the system's ability to identify unknown risks.

[0069] Secondly, the multi-scale safety risk analysis method proposed in this paper combines detailed analysis at the micro level with overall assessment at the macro level, and can fully capture various risk factors and their interactions in industrial systems. This method significantly improves the comprehensiveness and accuracy of safety assessments, making the assessment results closer to the complexity of actual industrial scenarios.

[0070] Furthermore, the dynamic multi-source graph reasoning module introduced in the present invention can update and adjust the security risk map in real time, effectively responding to the dynamic changes in the industrial environment. This feature greatly enhances the real-time and adaptability of the system, allowing security assessments to be continuously optimized and adjusted over time.

[0071] In addition, the method of the present invention provides clear risk explanation and visualization results by constructing an interpretability module, which not only enhances the credibility of the assessment results, but also provides intuitive and powerful support for formulating targeted protection strategies.

[0072] Finally, the protection strategy generator of the present invention combines expert knowledge base, historical data and machine learning algorithms to generate highly customized and actionable protection plans. This greatly improves the practicality of safety assessment results and helps industrial enterprises implement safety protection measures more effectively.

[0073] In general, the method proposed in this invention has achieved a comprehensive improvement in industrial scenario safety assessment by innovatively integrating multiple advanced technologies. It not only solves key technical problems such as data scarcity, model generalization, and dynamic assessment, but also has made significant progress in the comprehensiveness, accuracy, real-time and operability of the assessment. This method provides industrial enterprises with a powerful and flexible safety management tool, which is expected to play an important role in improving the safety of industrial systems, reducing accident risks, and ensuring production stability. It has important practical significance for promoting the safe development of industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 A flow chart of training a conditional adversarial network is generated for the present invention.

[0075] Figure 2 This is a flow chart of determining the authenticity of abnormal data of the present invention.

[0076] Figure 3 A flow chart is generated for the security risk map of the present invention.

[0077] Figure 4 It is a system module interaction diagram of the present invention. DETAILED DESCRIPTION

[0078] Please refer to Figure 1-4 The present invention provides an industrial scene security assessment method and system based on a generative adversarial network. The method first obtains the historical data and current operation data of the industrial scene, then uses a conditional generative adversarial network to generate abnormal data and conduct an assessment, and finally outputs a protection plan. Specifically, the method of the present invention includes the following steps:

[0079] First, in the acquisition step, this method acquires historical data of industrial scenarios, including micro sensor data, macro data, and safety risk assessment results. At the same time, it acquires the operating data of the current industrial environment to obtain a multi-source data set. These data lay the foundation for subsequent analysis and evaluation.

[0080] Next, in the processing step, the method trains a conditional generative adversarial network based on the acquired historical data to obtain a trained generator and discriminator. Preferably, in one embodiment of the present invention, the conditional generative adversarial network adopts a two-layer structure, including a first generative network G1 and a second generative network G2. Among them, G1 is used to generate macro data of industrial scenes, and G2 is used to generate micro data. This two-layer structure can better simulate the multi-scale characteristics of industrial scenes.

[0081] Then, based on the multi-source data set, the method uses the trained generator to generate abnormal data of industrial scenarios. The input of the generator may include noise vectors and conditional information, such as equipment status, environmental parameters, etc. Preferably, the network structure of the generator can be composed of a fully connected layer and multiple upsampling layers. For example, 3-5 layers of fully connected layers can be used, the number of neurons in each layer can be gradually increased from 128 to 512, and then 2-3 layers of upsampling layers are connected, and finally the abnormal data with the same dimension as the real data is output.

[0082] Next, the method uses a discriminator to judge the authenticity of the generated abnormal data, and adjusts the input data of the generator according to the authenticity. The input of the discriminator is the abnormal data, and the output is a scalar between 0 and 1, indicating the authenticity of the input data. Preferably, the discriminator can adopt a convolutional neural network structure, including 3-5 convolutional layers and 2-3 fully connected layers.

[0083] In the process of judging the authenticity of abnormal data, this method first extracts images from the abnormal data and extracts the features of the images. The features are then quantified into node vectors.

[0084] Next, the node vector is compared with the corresponding frame data to extract the features in the node vector. The features are then converted into corresponding confidence levels, and finally the authenticity of the abnormal data is determined based on the confidence levels.

[0085] This method performs iterative training based on the difference between the real data corresponding to the abnormal data and the abnormal data, and finally obtains real and effective abnormal data. Preferably, the number of iterative training can be set to 100-500 times, or until the loss function of the generator and the discriminator reaches a preset threshold (such as 0.01).

[0086] After obtaining real and effective abnormal data, this method extracts key frames from it and converts each frame data into a node. The extraction of key frames can be based on the difference between frames or the feature importance score. For example, a difference threshold (such as 0.1) can be set, and when the difference between adjacent frames exceeds the threshold, the frame is considered a key frame.

[0087] Based on the extracted nodes, the method generates an industrial safety risk map. The generation process of the risk map involves modeling the relationship between nodes and risk propagation analysis. Preferably, a graph neural network (such as Graph Convolutional Network) can be used to model the relationship between nodes, and an attention mechanism can be used to capture important risk propagation paths.

[0088] Finally, the method uses a dynamic multi-source graph reasoning module to convert the industrial safety risk map into a safety risk map. This step involves the dynamic update of the graph structure and the fusion of multi-source information. Preferably, a recurrent neural network (such as LSTM) can be used to model the dynamic changes of the graph structure, and multimodal fusion techniques (such as attention mechanism) can be used to integrate information from different sources.

[0089] In the output step, the method extracts protection measures through the protection strategy generator according to the generated security risk graph and outputs the protection plan. The protection strategy generator can be based on a method that combines a rule base and a machine learning model. For example, a decision tree or random forest can be used to learn the pattern of historical protection measures and combine it with the rules defined by experts to generate the final protection plan.

[0090] Through the above steps, the method of the present invention can comprehensively evaluate the safety risks of industrial scenarios and provide targeted protective measures. The method combines advanced technologies such as deep learning, graph theory and multi-source information fusion, and has strong innovation and practical value. The method of the present invention further defines and represents the node vector in an innovative way. In a preferred embodiment of the present invention, the node vector is represented as:

[0091] V n ={S n ,T b ,M n ,P,G,F n ,T n},

[0092] Among them, V n is the node vector in the corresponding environment, S n is the device status of the nth frame, T b is the type of device, M n is the operating mode of the nth device, Pn is the production status of the nth frame, G is the working mode, F n is the fault node feature, T n It is the time node feature graph.

[0093] This multi-dimensional node vector representation method can comprehensively capture the various characteristics of equipment and environment in industrial scenarios, providing a rich information basis for subsequent risk assessment.

[0094] In practical applications, the device status S n It may include the operating parameters of the equipment, such as temperature, pressure, vibration, etc.; Equipment type T b Can be a predefined category, such as pump, valve, sensor, etc.; operation mode M n It can include normal, standby, fault and other states; the production state P can reflect the current production progress or output; the working mode G can be batch production, continuous production, etc. Fault node feature F n And the time node feature graph T n The potential fault information and timing characteristics are captured respectively.

[0095] The method of the present invention adopts an innovative confidence calculation method in the process of comparing the node vector with the data to be predicted. Specifically, the method first extracts the vector values ​​of the node vectors at the corresponding time and each dimension of the image in the abnormal data, and compares these vector values ​​with the data values ​​in the data to be predicted, and obtains the difference ratio Δ i Preferably, the difference ratio may be calculated using methods such as Euclidean distance or cosine similarity.

[0096] Next, the method converts the difference ratio into a confidence level. In a preferred embodiment of the present invention, the confidence level is calculated as:

[0097]

[0098] Among them, C ij is the confidence of the jth node in the i-th frame, k is the adjustment factor, and θ is the threshold. This sig moid function form of transformation can map the difference ratio to between 0 and 1, and has good nonlinear characteristics. Usually, k can take a value between 1 and 10, and θ can take a value between 0.5 and 0.8. The specific values ​​can be adjusted according to the actual application scenario.

[0099] Based on the calculated confidence level, the method further calculates the authenticity of the data to be predicted under abnormal data. In one embodiment of the present invention, the calculation formula of the authenticity is:

[0100]

[0101] Among them, R e represents the authenticity in environment e, and N is the number of frames with the highest confidence in the abnormal data. This calculation method takes into account the highest confidence in multiple frames of data and can better reflect the overall authenticity of the abnormal data.

[0102] The method of the present invention adopts a multi-scale analysis method from micro to macro when generating an industrial safety risk map. First, in the safety risk map, each node is taken as the center of the circle and the node vector is taken as the radius to obtain a circular area. This representation method intuitively reflects the influence range of each node.

[0103] Next, the method extracts the overlapping areas between the circular areas, and determines whether each node matches the node vector in the overlapping area, thereby obtaining the influence range of the node.

[0104] Preferably, the match degree can be determined by vector similarity calculation, such as cosine similarity. When the similarity exceeds a preset threshold (such as 0.8), the nodes are considered to be matched.

[0105] According to the size of the overlapping area, the method further obtains the influence weight of the node. In a preferred embodiment of the present invention, the influence weight can be calculated by the following formula:

[0106]

[0107] Among them, W i is the influence weight of the i-th node, A i is the overlapping area of ​​the ith node, and N is the total number of nodes. This calculation method takes into account the mutual influence between nodes and can more accurately reflect the importance of each node in the entire system.

[0108] At the micro level, this method matches each node with the node vector to obtain the node's environmental information, operating data, and time information. This micro analysis can capture the detailed changes and local features of the system.

[0109] At the macro level, this method aggregates the environmental information and operation data of the same node, but does not include time information in the aggregation, so as to determine the security status at the macro level. This macro analysis can reflect the overall trend and global characteristics of the system.

[0110] Finally, this method combines the information at the micro and macro levels to determine the current safety status, thereby obtaining a complete safety risk map. This multi-scale analysis method can comprehensively assess the safety risks of industrial scenarios, taking into account local details while not ignoring overall trends.

[0111] In the process of generating protective measures, the method of the present invention adopts a comprehensive analysis method based on multi-source data. First, the safety risk map and the safety risk map are analyzed separately, and the analysis results are stored in the expert database, literature data and historical risk data respectively. This integration of multi-source data provides a rich knowledge basis for the subsequent generation of protective measures.

[0112] Next, the method maps the security status, macro security risks and protective measures to obtain the characteristics of the current security risks. This step realizes the transformation from risk assessment results to specific protective measures.

[0113] Then, this method compares the obtained features with the expert database, literature data and historical risk data to extract the corresponding features. This multi-source comparison method can make full use of existing knowledge and experience to improve the pertinence and effectiveness of protective measures.

[0114] Finally, according to the corresponding features, the method outputs a protection plan. Preferably, the generation of the protection plan can adopt a rule-based reasoning system or a machine learning model, such as a decision tree or a random forest. These methods can quickly generate appropriate protection measures according to the current risk characteristics.

[0115] Through the above steps, the method of the present invention realizes the automation of the whole process from risk assessment to the generation of protective measures, greatly improving the efficiency and accuracy of industrial scene safety assessment. The method of the present invention adopts an innovative cross-comparison method when extracting key frames from real and effective abnormal data. During the training process, the abnormal data threshold is first set. Preferably, the threshold can be determined according to the statistical characteristics of the historical data. For example, 3 times the standard deviation of the historical data can be taken as the initial threshold, and dynamically adjusted during the training process.

[0116] Next, the method uses a cross-comparison method to detect abnormal data. Specifically, for each frame of data, the method compares it with multiple adjacent frames of data to calculate its abnormality. In a preferred embodiment of the present invention, the abnormality can be calculated by the following formula:

[0117]

[0118] Among them, A i is the abnormality degree of the i-th frame, f i is the feature vector of the i-th frame, d(·,·) is the distance function (such as Euclidean distance), and n is the range of frames to be compared (usually 5-10). This cross-comparison method can effectively identify local abnormal patterns.

[0119] Then, the method determines the abnormal degree of the frame data by the abnormal data threshold, and obtains the frame data with the maximum confidence. Preferably, the following sigmoid function can be used to map the abnormal degree to a confidence between 0 and 1:

[0120]

[0121] Among them, C iis the confidence of the i-th frame, k is the adjustment factor (usually 1-5), and θ is the preset abnormal threshold.

[0122] Finally, this method defines the frame data with the highest confidence as the key frame, and the corresponding node vector is the key abnormal data. This method can effectively capture the most representative frames in the abnormal data and provide key information for subsequent risk assessment.

[0123] In order to improve the interpretability of the security assessment results, the method of the present invention also constructs an innovative interpretability module, which mainly includes a reasoning module and an explanation generation module.

[0124] The reasoning module is used to interpret and reason the quantitative evaluation results. In one embodiment of the present invention, the reasoning module can adopt a method that combines a rule-based reasoning system and a machine learning model. For example, a decision tree can be used to learn the pattern of historical evaluation results and generate explanations in combination with expert-defined rules. Preferably, the depth of the decision tree can be set to 3-5 layers to balance the detail and understandability of the explanation.

[0125] The explanation generation module is used to generate a risk explanation picture based on the explanation inference obtained by the reasoning module. In a preferred embodiment of the present invention, the risk explanation picture can be in the form of a heat map to intuitively display the risk level of different areas or equipment. Specifically, different color shades can be used to represent the risk level, such as red for high risk, yellow for medium risk, and green for low risk.

[0126] In addition, this method also obtains the threat field, vulnerability field and impact field corresponding to the quantitative assessment results through reasoning. The definitions of these three fields are as follows:

[0127] 1. Threat field: refers to external factors that may cause harm to the system, such as cyber attacks, natural disasters, etc.

[0128] 2. Vulnerability field: represents the weaknesses or loopholes of the system itself, such as software vulnerabilities, configuration errors, etc.

[0129] 3. Impact field: Indicates the possible consequences of threats exploiting vulnerabilities, such as data leakage, equipment damage, etc.

[0130] Finally, this method will generate a comprehensive risk assessment report by interpreting the threat field, vulnerability field, and impact field corresponding to the reasoning picture and the quantitative assessment results. This multi-dimensional risk representation method can help users fully understand the current security situation and provide strong support for decision-making.

[0131] Based on the above method, the present invention also proposes a corresponding industrial scene safety assessment system. The system includes the following core modules:

[0132] The training module 1 is used to generate an adversarial network based on historical data training conditions to obtain a trained generator and discriminator. The training module 1 adopts the double-layer generation network structure proposed by the present invention, which can effectively generate abnormal data that conforms to the characteristics of actual industrial scenarios.

[0133] The acquisition module 2 is responsible for collecting the operating data of the current industrial environment and obtaining a multi-source data set. The acquisition module 2 can obtain real-time data from multiple data sources such as various sensors and control systems to provide a basis for subsequent analysis.

[0134] The generator training module 3 generates abnormal data of industrial scenarios using the trained generator. The generator training module 3 generates high-quality abnormal data samples by continuously adjusting input parameters.

[0135] The discriminator training module 4 uses the trained discriminator to judge the authenticity of the abnormal data and adjusts the input data of the generator according to the authenticity. The discriminator training module 4 continuously improves its recognition ability during the training process, thereby improving the quality of the generated data.

[0136] The risk assessment module 5 is responsible for extracting key frames from real and effective abnormal data and generating industrial safety risk maps based on these key frames. The risk assessment module 5 adopts the multi-scale analysis method proposed by the present invention, which can comprehensively assess the safety risks of the system.

[0137] The safety detection module 6 converts the industrial safety risk map into a safety risk map using the dynamic multi-source graph reasoning module. The safety detection module 6 can update the risk status in real time and detect potential safety threats in a timely manner.

[0138] The strategy generation module 7 extracts protection measures based on the security risk map through the protection strategy generator and outputs a protection plan. The strategy generation module 7 combines the expert knowledge base and the machine learning algorithm to generate highly targeted and operable protection measures.

[0139] Through the collaborative work of these modules, the system of the present invention can achieve comprehensive assessment and effective protection of safety risks in industrial scenarios, greatly improving the safety and reliability of industrial systems.

[0140] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modification, replacement, and improvement made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An industrial scenario safety assessment method based on generative adversarial networks, characterized in that: include: The acquisition steps include: Acquire historical data of industrial scenarios, wherein the historical data includes micro sensor data, macro data, and safety risk assessment results; Obtain current industrial environment operation data and obtain multi-source data sets; Processing steps include: Based on the historical data, training a conditional generative adversarial network to obtain a trained generator and discriminator; Based on the multi-source data set, using the generator to generate abnormal data of the industrial scene; Using the discriminator to judge the authenticity of the abnormal data, and adjusting the input data of the generator according to the authenticity; Perform iterative training based on the difference between the real data corresponding to the abnormal data and the abnormal data to obtain real and effective abnormal data; Extract key frames from the real and effective abnormal data, and convert each frame of data into a node; Generate an industrial safety risk map based on the nodes; Converting the industrial safety risk map into a safety risk map using a dynamic multi-source graph reasoning module; Output steps include: According to the security risk graph, protection measures are extracted through a protection strategy generator, and a protection plan is output.

2. The method according to claim 1, characterized in that The training steps of the conditional generative adversarial network specifically include: Constructing a two-layer generation network structure, including a first generation network G1 and a second generation network G2; Generate macro data of industrial scenes using the first generation network G1; Generate micro data of industrial scenes using the second generation network G2; Construct a discriminant network D to judge the authenticity of the generated data.

3. The method according to claim 1, characterized in that The step of judging the authenticity of abnormal data specifically includes: Extracting an image from the abnormal data, and extracting features of the image; quantizing the features into node vectors; Comparing the node vector with the corresponding frame data to extract features in the node vector; Convert the features into corresponding confidence levels; The authenticity of the abnormal data is determined according to the confidence level.

4. The method according to claim 3, characterized in that The node vector is represented as: V n ={S n ,T b ,M n ,P,G,F n ,T n }, Among them, V n is the node vector in the corresponding environment, S n is the device status of the nth frame, T b is the type of device, M n is the operating mode of the nth device, P is the production status of the nth frame, G is the working mode, and F n is the fault node feature, T n It is the time node feature graph.

5. The method according to claim 4, characterized in that The step of comparing the node vector with the corresponding data to be predicted to obtain the confidence level specifically includes: Extract the vector values ​​of the node vectors at the corresponding time and each dimension of the image in the abnormal data, and compare the vector values ​​with the data values ​​in the data to be predicted to obtain the difference ratio Δ i ; The difference ratio is converted into confidence, expressed as: C ij =f(Δ i ); According to the confidence level, the authenticity of the data to be predicted under abnormal data is calculated, which is expressed as: R e =g(C ij ).

6. The method according to claim 1, characterized in that According to the industrial safety risk map, the steps for multi-scale analysis of safety risks from micro to macro include: In the security risk map, a circular area is obtained with each node as the center and the node vector as the radius; Extract the overlapping areas between the circular areas, determine whether each node matches the node vector in the overlapping area, and obtain the influence range of the node; According to the size of the overlapping area, the influence weight of the node is obtained; At the micro level, each node is matched with the node vector to obtain the node's environmental information, operation data and time information; At the macro level, the environmental information and operation data of the same node are aggregated, and the time information is not involved in the aggregation to determine the security status at the macro level; Combine the information at the micro level with the macro level to determine the current security status and obtain a security risk map.

7. The method according to claim 1, characterized in that The steps to extract protection measures through the protection strategy generator include: Analyze the safety risk map and the safety risk graph separately, and store the analysis results in the expert database, literature data and historical risk data respectively; Map the security status, macro security risks and protective measures to obtain the characteristics of the current security risks; Compare the features with the expert database, literature data and historical risk data respectively to extract corresponding features; Output protection solutions based on corresponding features.

8. The method according to claim 1, characterized in that The steps of extracting key frames from real and effective abnormal data specifically include: During the training process, set the abnormal data threshold; Use cross-comparison method to detect abnormal data; The abnormal degree of the frame data is determined by the abnormal data threshold, and the frame data with the maximum confidence is obtained; The frame data is defined as a key frame, and the corresponding node vector is the key abnormal data.

9. The method according to claim 1, characterized in that: Also included are steps to build the interpretability module: Construct a reasoning module to interpret and reason about the quantitative evaluation results; Construct an explanation generation module to generate risk explanation pictures based on the explanations obtained by the reasoning module; Through reasoning, we can obtain the threat field, vulnerability field, and impact field corresponding to the quantitative assessment results; The threat field, vulnerability field, and impact field corresponding to the interpreted reasoning picture and quantitative assessment results will be used to generate a risk assessment report.

10. An industrial scenario safety assessment system based on a generative adversarial network that executes the method according to any one of claims 1 to 9, characterized in that: include: The training module is used to train the conditional generative adversarial network based on historical data to obtain the trained generator and discriminator; The acquisition module is used to collect the current industrial environment operation data and obtain a multi-source data set; A generator training module, used to generate abnormal data of industrial scenarios using the generator; A discriminator training module, used to use the discriminator to judge the authenticity of the abnormal data and adjust the input data of the generator according to the authenticity; A risk assessment module, used to extract key frames from real and effective abnormal data and generate an industrial safety risk map based on the key frames; A safety detection module, used for converting the industrial safety risk map into a safety risk map using a dynamic multi-source map reasoning module; The strategy generation module is used to extract protection measures through the protection strategy generator according to the security risk map and output a protection plan.

Images