Safety education and training monitoring and evaluation method and system based on operation mechanism optimization
By building a security education and training knowledge graph and adaptive clustering model, dynamically adjusting evaluation indicators, and using blockchain to achieve automation, the problems of insufficient knowledge representation and insufficient data utilization in traditional methods are solved, and the monitoring and evaluation of security education and training are improved Intelligent and refined.
Patent Information
- Application Number
- CN202411204864.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-30
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-08-30
AI Technical Summary
Traditional safety education and training monitoring and evaluation methods lack comprehensive and systematic knowledge representation, and it is difficult to accurately characterize the semantic information of the training process, the use of multi-source heterogeneous data is insufficient, the selection of monitoring and evaluation indicators is unreasonable, and the dynamic changes in the training process cannot be adapted to the scientific Insufficient sex and effectiveness.
Using the method based on operation mechanism optimization, a safety education and training knowledge graph is constructed for semantic representation. The semantic features of multi-source heterogeneous data are extracted through the multi-grained feature mapping model, and an adaptive clustering model and causal graph model that integrates group intelligence are constructed. Dynamic Adjust the weight of evaluation indicators and use blockchain to automate and real-time updates of monitoring and evaluation processes.
The intelligence and refinement of safety education and training monitoring and evaluation have been improved, the efficiency and quality of data fusion utilization have been enhanced, the evaluation indicators have been dynamically adjusted, more scientific and efficient evaluation results have been achieved, manual operations have been reduced, and evaluation efficiency has been improved.
Smart Images

Figure CN118734254B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to education and training monitoring technology, and in particular to a safety education and training monitoring and evaluation method and system based on operation mechanism optimization. Background Art
[0002] With the rapid development of society, various safety accidents occur frequently, and safety education and training work faces many challenges. The traditional safety education and training monitoring and evaluation methods have the following shortcomings: First, there is a lack of comprehensive and systematic knowledge representation, which makes it difficult to accurately describe the semantic information of the training process; second, there is a lack of effective integration and utilization of multi-source heterogeneous data, which affects the quality and efficiency of monitoring data; third, the selection of monitoring and evaluation indicators is unreasonable, and the causal dependency between indicators is not fully considered; fourth, the evaluation strategy is static and solidified, and cannot adapt to the dynamic changes of the training process. The above problems restrict the scientificity and effectiveness of safety education and training monitoring and evaluation work, and a more intelligent and refined solution is urgently needed.
[0003] In response to the above problems, the patent of this invention proposes a safety education and training monitoring and evaluation method based on operation mechanism optimization. Starting from multiple levels such as knowledge representation, data fusion, intelligent modeling, and strategy optimization, the system improves the intelligence level and refinement of safety education and training monitoring and evaluation. Summary of the invention
[0004] The embodiment of the present invention provides a safety education and training monitoring and evaluation method and system based on operation mechanism optimization, which can solve the problems in the prior art.
[0005] According to a first aspect of the embodiments of the present invention,
[0006] Provide a safety education and training monitoring and evaluation method based on operation mechanism optimization, including:
[0007] Construct a safety education and training knowledge graph, perform semantic representation of the safety education and training process based on the safety education and training knowledge graph, obtain the semantic representation result of the safety education and training process, establish a multi-granularity feature mapping model based on the semantic representation result, and simultaneously obtain multi-source heterogeneous monitoring data of the safety education and training process, use the multi-granularity feature mapping model to perform multi-granularity semantic feature extraction on the multi-source heterogeneous monitoring data, and obtain a multi-granularity semantic feature set;
[0008] Construct an adaptive clustering model that integrates swarm intelligence optimization, takes a multi-granularity semantic feature set as input, outputs cluster labels, and constructs a causal graph model for safety education and training evaluation indicators. The causal graph model contains evaluation indicator nodes and causal relationship edges. The potential causal relationship between the evaluation indicator nodes is mined through a causal inference algorithm, and a causal effect estimation model is constructed. The dynamic weight vector of the evaluation indicator is obtained using the causal effect estimation model.
[0009] The clustering labels and dynamic weight vectors of evaluation indicators are taken as input, and iterative optimization is performed through the pre-trained strategy optimization model to obtain the adaptive monitoring and evaluation strategy. A blockchain network is built, and the obtained multi-source heterogeneous monitoring data, clustering labels, dynamic weight vectors of evaluation indicators and adaptive monitoring and evaluation strategies are stored on the chain. The monitoring and evaluation process smart contract is designed to realize the automatic triggering and scheduling of the monitoring and evaluation process, and the adaptive monitoring and evaluation strategy is updated in real time.
[0010] In an optional embodiment,
[0011] A safety education and training knowledge graph is constructed, and a semantic representation of the safety education and training process is performed based on the safety education and training knowledge graph. The semantic representation results of the safety education and training process include:
[0012] Construct a safety education and training knowledge graph, which includes entities and relationships;
[0013] The safety education and training knowledge graph is embedded using a pre-set knowledge graph embedding model. Triples containing head entities, relations, and tail entities are defined. By minimizing the energy function of correct triples and maximizing the energy function of incorrect triples, entities and relations are mapped into a low-dimensional continuous vector space to obtain embedded representations of entities and relations.
[0014] Extract the multi-hop relationship path in the safety education and training knowledge graph, use the path sorting model to encode the multi-hop relationship path, determine the path feature vector and the answer sorting function, sort the candidate answers according to the path feature vector, and obtain the candidate answer sorting result; at the same time, use the attention-based path encoder to learn the weight of the multi-hop relationship path, and obtain the aggregate vector representation of the multi-hop relationship path through attention weighting;
[0015] Constructing a graph neural network model, which includes a graph convolutional network and a graph attention network. Based on the structural information of the safety education and training knowledge graph, the neighbor information of the current entity is aggregated through the graph convolutional network, the weights of different neighbors are calculated through the graph attention network, the feature vector representation of the current entity is updated, and the feature vector representation of the target entity is obtained;
[0016] The embedded representations of entities and relations, the ranking results of candidate answers, the aggregate vector representation of multi-hop relationship paths, and the feature vector representation of target entities are integrated, and the attention mechanism is introduced for weighted fusion to obtain a multi-perspective fused knowledge graph representation of safety education and training;
[0017] Based on the fused safety education and training knowledge graph representation, the safety education and training process is semantically represented to obtain the semantic representation result of the safety education and training process.
[0018] In an optional embodiment,
[0019] An adaptive clustering model integrating swarm intelligence optimization is constructed, which takes multi-granularity semantic feature sets as input and outputs cluster labels including:
[0020] Based on the pre-obtained multi-granularity semantic feature set, an initial clustering model is constructed, and the number of cluster centers and cluster sample attribution of the initial clustering model are set according to the characteristics of safety education and training data and domain knowledge;
[0021] The particle swarm optimization algorithm is introduced into the initial clustering model. The particle position and velocity are updated through iterative optimization strategy to search for the optimal cluster center. The particle position represents the cluster center coordinates, the particle velocity represents the cluster center update direction and step size, and the particle fitness function is used to evaluate the clustering quality.
[0022] Based on the optimal cluster center obtained by the search, the ant colony optimization algorithm is introduced to dynamically divide the cluster samples. Each ant corresponds to a cluster sample. The ants move between different cluster centers according to the cluster center attraction and pheromone concentration, and dynamically adjust the sample attribution through the pheromone accumulation and attenuation mechanism to search for the optimal cluster division information.
[0023] The optimal cluster center and optimal cluster division information obtained by fusion search are adaptively adjusted to adjust the initial cluster model parameters, including the number of cluster centers and sample attribution threshold, and repeated iterations are performed until the maximum number of iterations is reached to obtain the optimal cluster model;
[0024] The multi-granularity semantic feature set is input into the optimal clustering model, and the cluster label corresponding to each sample is output.
[0025] In an optional embodiment,
[0026] The potential causal relationship between the evaluation indicator nodes is mined through a causal inference algorithm, a causal effect estimation model is constructed, and the dynamic weight vector of the evaluation indicator is obtained by using the causal effect estimation model, including:
[0027] Inputting the original features of the evaluation index into a preset causal invariant feature extractor, mapping the original features to a low-dimensional latent space, and obtaining a latent representation of the evaluation index;
[0028] Construct a multi-task learning framework including causal effect estimation tasks and representation learning tasks, and construct positive and negative sample pairs based on the original features of the evaluation indicators. For the same evaluation indicator, sample pairs sampled from different data distributions are constructed as positive sample pairs. For different evaluation indicators, sample pairs randomly sampled therefrom are constructed as negative sample pairs. Based on the constructed positive and negative sample pairs, a contrastive learning loss function is constructed.
[0029] The causal effect estimation objective function and the contrastive learning loss function are jointly optimized to train the causal invariant feature extractor, so that the causal invariant feature extractor gradually learns the causal invariant representation of the evaluation index, and finally obtains the trained causal invariant feature extractor;
[0030] The original features of the new evaluation indicator samples are obtained, and the original features of the new evaluation indicator samples are input into the trained causal invariant feature extractor to obtain the causal invariant representation of the new evaluation indicator samples, and the causal invariant representation of the new evaluation indicator samples is input into the causal effect estimation model. Through the causal effect estimation model, based on the obtained causal invariant representation of the new evaluation indicator samples, the corresponding causal effect estimation results are obtained. According to the causal effect estimation results, the dynamic weight learning algorithm of the evaluation indicators is introduced, and the reward function is designed. The causal effect strength, indicator importance and model prediction error are comprehensively considered, and the optimization direction of the weight vector is guided by reinforcement learning, so as to finally obtain the dynamic weight vector of the evaluation indicator.
[0031] In an optional embodiment,
[0032] Jointly optimizing the causal effect estimation objective function and the contrastive learning loss function, training the causal invariant feature extractor includes:
[0033] Based on the causal effect estimation objective function, the representation learning regularization term and the contrastive learning loss function, a joint optimization objective function is constructed to train the causal invariant feature extractor. The calculation formula of the joint optimization objective function is as follows:
[0034] ;
[0035] Among them, L represents the joint optimization objective function, L c represents the causal effect estimation objective function, λ1 represents the balance factor of the learning regularization term, n represents the number of samples of the evaluation indicator, f(·) represents the causal invariant feature extraction function, x i represents the original features of the i-th evaluation indicator, c i represents interference factors, |·| F represents the Frobenius norm, λ2 represents the balance factor of the learning loss function, and z i represents the potential representation of the i-th evaluation indicator in the positive sample pair, z j represents the potential representation of the jth evaluation indicator in the positive sample pair, sim(·) represents the similarity metric function, τ represents the temperature parameter, K represents the number of negative samples, and z k represents the potential representation of the kth evaluation metric in the negative sample pair.
[0036] In an optional embodiment,
[0037] The dynamic weight vector of cluster labels and evaluation indicators is used as input, and iterative optimization is performed through the pre-trained strategy optimization model to obtain the adaptive monitoring evaluation strategy including:
[0038] Construct an initial strategy optimization model, which includes a strategy generator and a strategy evaluator. The strategy generator generates candidate monitoring and evaluation strategies based on the current environmental status and historical decision sequences. The strategy evaluator estimates the expected utility of different monitoring and evaluation strategies, ranks the candidate monitoring and evaluation strategies, and selects the optimal strategy as output.
[0039] Collect historical safety education and training monitoring data and construct a monitoring trajectory sequence dataset for trajectory prediction pre-training, where each monitoring trajectory sequence contains multiple time steps, and each time step corresponds to a monitoring state vector;
[0040] The monitoring trajectory sequence is divided into multiple subsequences to form a prediction data set, and a random sampling method is used to select some subsequences as the prediction model input to obtain the state prediction distribution. By minimizing the difference loss between the predicted trajectory and the true trajectory, the prediction model parameters are updated to obtain a trained prediction model, and the trained prediction model is integrated into the initial policy optimization model to obtain the target policy optimization model.
[0041] The clustering labels and evaluation index weight vectors are concatenated into a state feature vector, which is input into the target strategy optimization model. The model parameters of the target strategy optimization model are updated by a gradient descent algorithm, and the iteration is repeated until the model converges, thereby finally obtaining an adaptive monitoring evaluation strategy.
[0042] In an optional embodiment,
[0043] Build a blockchain network, store the obtained multi-source heterogeneous monitoring data, cluster labels, dynamic weight vectors of evaluation indicators, and adaptive monitoring and evaluation strategies on the chain, design a monitoring and evaluation process smart contract, realize the automatic triggering and scheduling of the monitoring and evaluation process, and update the adaptive monitoring and evaluation strategy in real time, including:
[0044] Constructing a blockchain network, designing a data structure and a transaction format in the blockchain network, using the data structure and the transaction format to encapsulate multi-source heterogeneous monitoring data, cluster labels, dynamic weight vectors of evaluation indicators, and adaptive monitoring and evaluation strategy parameters, and using an encryption algorithm to encrypt the encapsulated multi-source heterogeneous monitoring data, cluster labels, dynamic weight vectors of evaluation indicators, and adaptive monitoring and evaluation strategy parameters;
[0045] Based on the encrypted multi-source heterogeneous monitoring data, cluster labels, dynamic weight vectors of evaluation indicators and adaptive monitoring and evaluation strategy parameters, a smart contract is deployed in the blockchain network, and the triggering conditions and execution logic of the monitoring and evaluation process are set in the smart contract. The monitoring and evaluation process is automatically triggered and the execution logic is executed according to the triggering conditions;
[0046] In the process of executing the monitoring and evaluation process, the adaptive monitoring and evaluation strategy parameters are abstracted as the state variables of the smart contract, and the strategy parameter reading function, the strategy parameter writing function and the strategy update auxiliary function are designed in the smart contract. The strategy parameter reading function is used for external calls to read the current adaptive monitoring and evaluation strategy parameter value. When the strategy parameter writing function receives a write request that meets the authority verification, it writes the new adaptive monitoring and evaluation strategy parameters into the state variables;
[0047] After the strategy optimization model generates new adaptive monitoring and evaluation strategy parameters, the strategy parameter writing function of the smart contract is called to pass in the new adaptive monitoring and evaluation strategy parameters and write them into the state variables to achieve real-time update of the adaptive monitoring and evaluation strategy.
[0048] According to a second aspect of the embodiments of the present invention,
[0049] Provide a safety education and training monitoring and evaluation system based on operation mechanism optimization, including:
[0050] The first unit is used to construct a safety education and training knowledge graph, perform semantic representation of the safety education and training process based on the safety education and training knowledge graph, obtain the semantic representation result of the safety education and training process, establish a multi-granularity feature mapping model according to the semantic representation result, and simultaneously obtain multi-source heterogeneous monitoring data of the safety education and training process, and use the multi-granularity feature mapping model to perform multi-granularity semantic feature extraction on the multi-source heterogeneous monitoring data to obtain a multi-granularity semantic feature set;
[0051] The second unit is used to build an adaptive clustering model that integrates swarm intelligence optimization, takes a multi-granularity semantic feature set as input, outputs cluster labels, and builds a causal graph model for safety education and training evaluation indicators. The causal graph model contains evaluation indicator nodes and causal relationship edges. The potential causal relationship between the evaluation indicator nodes is mined through a causal inference algorithm, and a causal effect estimation model is built. The dynamic weight vector of the evaluation indicator is obtained using the causal effect estimation model;
[0052] The third unit is used to take the dynamic weight vectors of cluster labels and evaluation indicators as input, perform iterative optimization through a pre-trained strategy optimization model, obtain an adaptive monitoring and evaluation strategy, build a blockchain network, store the obtained multi-source heterogeneous monitoring data, cluster labels, dynamic weight vectors of evaluation indicators, and adaptive monitoring and evaluation strategies on the chain, design a monitoring and evaluation process smart contract, realize the automatic triggering and scheduling of the monitoring and evaluation process, and update the adaptive monitoring and evaluation strategy in real time.
[0053] According to a third aspect of the embodiments of the present invention,
[0054] An electronic device is provided, comprising:
[0055] processor;
[0056] a memory for storing processor-executable instructions;
[0057] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.
[0058] A fourth aspect of the embodiments of the present invention is:
[0059] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.
[0060] In this embodiment, by using the knowledge graph to systematically and structure the semantic representation of the safety education and training process, a solid knowledge foundation is laid for subsequent intelligent analysis. By adopting multi-granularity feature mapping, the semantic information of multi-source heterogeneous monitoring data can be mined from different dimensions such as time, space, and semantics, which improves the efficiency and quality of data fusion utilization. The adaptive clustering model that integrates group intelligence is introduced to dynamically group the monitored objects according to the similarity of semantic features at different granularities, overcoming the limitations of traditional fixed grouping. The causal dependency relationship between evaluation indicators is characterized by the causal effect estimation model, and the indicator weights are dynamically adjusted, making the selection of evaluation indicators more scientific and reasonable, and the evaluation results more interpretable. The strategy optimization model is used to jointly optimize and solve various monitoring indicators and evaluation indicators. The resulting adaptive evaluation strategy can be dynamically adjusted according to the real-time feedback of the training process, which improves the accuracy and timeliness of the evaluation. Blockchain is used to build a secure and reliable monitoring data storage environment, and smart contracts are designed to realize the automatic triggering and scheduling of the monitoring and evaluation process, reducing manual operation links and improving evaluation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0061] Figure 1 A schematic diagram of a process flow of a safety education and training monitoring and evaluation method based on operation mechanism optimization according to an embodiment of the present invention;
[0062] Figure 2 This is a schematic diagram of the structure of a safety education and training monitoring and evaluation system based on operation mechanism optimization according to an embodiment of the present invention. DETAILED DESCRIPTION
[0063] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0064] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.
[0065] Figure 1 Schematic diagram of the process of the safety education and training monitoring and evaluation method based on the operation mechanism optimization of the embodiment of the present invention, such as Figure 1 As shown, the method includes:
[0066] S101. Construct a safety education and training knowledge graph, perform semantic representation of the safety education and training process based on the safety education and training knowledge graph, obtain the semantic representation result of the safety education and training process, establish a multi-granularity feature mapping model according to the semantic representation result, and simultaneously obtain multi-source heterogeneous monitoring data of the safety education and training process, use the multi-granularity feature mapping model to extract multi-granularity semantic features from the multi-source heterogeneous monitoring data, and obtain a multi-granularity semantic feature set.
[0067] Among them, the construction of the safety education and training knowledge graph is composed of entities and the relationships between entities. Entities represent important concepts in the field of safety education and training, such as safety hazards, accident cases, and preventive measures. Relationships represent the semantic connection between entities, such as "hazards lead to accidents" and "measures prevent hazards". By constructing a knowledge graph, the knowledge in the field of safety education and training can be structured.
[0068] Among them, multi-granular features refer to the characterization of the training process from different levels and angles, including both summary features at the macro level and fine-grained features at the micro level. The multi-granular feature mapping model defines the mapping relationship between the semantic representation results and the multi-granular features. Through feature extraction and semantic association, the semantic information of the training process is mapped to feature spaces of different granularities. In the feature mapping process, ontology reasoning, natural language processing, pattern recognition and other technologies are comprehensively utilized to realize semantic-driven feature extraction. The multi-granular feature set can comprehensively characterize multiple aspects of the training process and provide rich feature representations for intelligent monitoring and evaluation.
[0069] The safety education and training process often involves multiple types of data, such as video, audio, text, sensor data, etc. These data come from different sources and formats, forming a heterogeneous data environment. In the feature extraction process, corresponding feature extraction algorithms are used for different types of data, such as convolutional neural networks for feature learning of video data, Mel-frequency cepstral coefficients for audio data to extract acoustic features, and word embedding models for text data to obtain semantic features. Through multi-granular feature mapping, heterogeneous monitoring data is converted into a unified multi-granular semantic feature set, laying a data foundation for subsequent clustering analysis and anomaly detection.
[0070] In an optional embodiment,
[0071] A safety education and training knowledge graph is constructed, and a semantic representation of the safety education and training process is performed based on the safety education and training knowledge graph. The semantic representation results of the safety education and training process include:
[0072] Construct a safety education and training knowledge graph, which includes entities and relationships;
[0073] The safety education and training knowledge graph is embedded using a pre-set knowledge graph embedding model. Triples containing head entities, relations, and tail entities are defined. By minimizing the energy function of correct triples and maximizing the energy function of incorrect triples, entities and relations are mapped into a low-dimensional continuous vector space to obtain embedded representations of entities and relations.
[0074] Extract the multi-hop relationship path in the safety education and training knowledge graph, use the path sorting model to encode the multi-hop relationship path, determine the path feature vector and the answer sorting function, sort the candidate answers according to the path feature vector, and obtain the candidate answer sorting result; at the same time, use the attention-based path encoder to learn the weight of the multi-hop relationship path, and obtain the aggregate vector representation of the multi-hop relationship path through attention weighting;
[0075] Constructing a graph neural network model, which includes a graph convolutional network and a graph attention network. Based on the structural information of the safety education and training knowledge graph, the neighbor information of the current entity is aggregated through the graph convolutional network, the weights of different neighbors are calculated through the graph attention network, the feature vector representation of the current entity is updated, and the feature vector representation of the target entity is obtained;
[0076] The embedded representations of entities and relations, the ranking results of candidate answers, the aggregate vector representation of multi-hop relationship paths, and the feature vector representation of target entities are integrated, and the attention mechanism is introduced for weighted fusion to obtain a multi-perspective fused knowledge graph representation of safety education and training;
[0077] Based on the fused safety education and training knowledge graph representation, the safety education and training process is semantically represented to obtain the semantic representation result of the safety education and training process.
[0078] Exemplarily, first, based on the construction of the knowledge graph, the knowledge graph embedding model is used to embed the safety education and training knowledge graph. The knowledge graph embedding model is a method that maps entities and relationships in the knowledge graph to a low-dimensional continuous vector space. Optionally, the classic TransE model is used as the knowledge graph embedding model. The TransE model defines a triple containing a head entity, a relationship, and a tail entity, and learns a low-dimensional vector representation of entities and relationships by minimizing the energy function of the correct triple and maximizing the energy function of the wrong triple. Through embedding processing, the symbolic representation in the knowledge graph can be converted into a continuous vector representation, which is convenient for subsequent calculations and modeling.
[0079] After obtaining the embedded representation of entities and relations, the multi-hop relation paths in the safety education and training knowledge graph are further extracted. The multi-hop relation path is a sequence of relations connecting two entities, reflecting the complex semantic association between entities. The path sorting model is used to encode the multi-hop relation path. The path sorting model learns the path feature vector and the answer sorting function, sorts the candidate answers according to the path feature vector, and obtains the candidate answer sorting result. The candidate answer is a set of entities related to the query entity in the knowledge graph. At the same time, the attention-based path encoder is used to learn the weights of the multi-hop relation paths. The attention mechanism can automatically learn the importance of different paths, and obtain the aggregate vector representation of the multi-hop relation path through attention weighting. This representation integrates the information of different paths and reflects the comprehensive semantic association between entities.
[0080] In addition to using path information, a graph neural network model is also constructed to capture the structural information of the knowledge graph. Graph neural network is a type of deep learning model based on graph structure data. Graph convolutional network and graph attention network are used as graph neural network models. Graph convolutional network aggregates the neighbor information of the current entity through graph convolution operation and updates the feature representation of the current entity. Graph attention network introduces the attention mechanism based on graph convolutional network, and assigns different importance to different neighbors by calculating the weights of different neighbors. Through the graph neural network model, the structural information of the knowledge graph can be fully utilized to learn high-quality feature representation of entities.
[0081] After obtaining the embedded representations of entities and relations, the ranking results of candidate answers, the aggregate vector representation of multi-hop relationship paths, and the feature vector representation of the target entity, the multi-perspective representations are fused. The fusion methods include feature concatenation and attention weighting. Feature concatenation directly concatenates different representations into a higher-dimensional vector representation. Attention weighting learns the weights of different representations through the attention mechanism and performs weighted summation. By fusing representations from multiple perspectives, a more comprehensive and accurate knowledge graph representation of safety education and training can be obtained.
[0082] Finally, based on the fused safety education and training knowledge graph representation, the safety education and training process is semantically represented. By using the knowledge graph to represent the semantic representation of learned entities and relationships, the key information in the safety education and training process can be represented and understood. For example, by calculating the semantic similarity between entities, relevant safety hazards and preventive measures can be discovered; through relational reasoning, potential accident risks can be predicted and corresponding response strategies can be given. Through semantic representation, the pertinence and effectiveness of safety education and training can be enhanced, and the training effect can be improved.
[0083] In this embodiment, the structured representation and semantic representation of safety education and training knowledge are achieved by building a knowledge graph, using a knowledge graph embedding model, extracting multi-hop relationship paths, building a graph neural network model, and fusing multi-perspective representations. This method makes full use of the structural information and semantic information of the knowledge graph, can effectively support the knowledge management and intelligent application of safety education and training, and improve the quality and efficiency of safety education and training.
[0084] S102. Construct an adaptive clustering model that integrates swarm intelligence optimization, take a multi-granularity semantic feature set as input, output clustering labels, and construct a causal graph model for safety education and training evaluation indicators. The causal graph model contains evaluation indicator nodes and causal relationship edges. The potential causal relationship between the evaluation indicator nodes is mined through a causal inference algorithm, and a causal effect estimation model is constructed. The dynamic weight vector of the evaluation indicator is obtained using the causal effect estimation model.
[0085] Among them, the adaptive clustering model realizes automatic optimization and parameter adaptation of the clustering process by introducing swarm intelligence algorithms such as particle swarm optimization and ant colony optimization. Cluster analysis is an unsupervised learning method that aims to divide similar samples into the same cluster, so that the sample similarity within the cluster is high and the sample similarity between clusters is low. In the adaptive clustering model, a multi-granularity semantic feature set is used as input, and the optimal cluster center and sample division are found through iterative optimization. The swarm intelligence algorithm simulates the intelligent mechanism of group behavior in nature, and achieves global optimization through collaboration and information interaction between individuals. In particle swarm optimization, each particle represents a potential clustering solution, and the optimal cluster center is continuously searched through the position update and speed update of the particle. In ant colony optimization, by simulating the behavior of ants looking for food, the pheromone mechanism is used to realize the dynamic division of cluster samples. By integrating multiple swarm intelligence algorithms, the adaptive clustering model can automatically find the optimal clustering solution in a complex data environment and output accurate clustering labels.
[0086] In addition to cluster analysis, a causal graph model was also constructed for the safety education and training evaluation indicators. Evaluation indicators are key factors in measuring the effectiveness of training. The causal graph model is a directed acyclic graph used to represent the causal dependency between variables. In the causal graph model, the evaluation indicators are represented as nodes, and the causal relationship between indicators is represented as directed edges. Through the causal inference algorithm, the potential causal relationship between the evaluation indicators is mined to reveal the causal dependency structure of the indicators. The causal inference algorithm identifies significant causal associations from the data through methods such as statistical tests and conditional independence judgment. On the basis of mining causal relationships, a causal effect estimation model is further constructed to quantify the causal influence intensity between the evaluation indicators. The causal effect reflects the degree of influence of a certain indicator change on other indicators, and can be used to guide the dynamic adjustment of the evaluation indicator weights. Through the causal graph model and causal effect estimation, the dynamic weight vector of the evaluation indicator is obtained, which reflects the dynamic change of the indicator importance and provides a basis for the dynamic optimization monitoring and evaluation strategy.
[0087] In an optional embodiment,
[0088] An adaptive clustering model integrating swarm intelligence optimization is constructed, which takes multi-granularity semantic feature sets as input and outputs cluster labels including:
[0089] Based on the pre-obtained multi-granularity semantic feature set, an initial clustering model is constructed, and the number of cluster centers and cluster sample attribution of the initial clustering model are set according to the characteristics of safety education and training data and domain knowledge;
[0090] The particle swarm optimization algorithm is introduced into the initial clustering model. The particle position and velocity are updated through iterative optimization strategy to search for the optimal cluster center. The particle position represents the cluster center coordinates, the particle velocity represents the cluster center update direction and step size, and the particle fitness function is used to evaluate the clustering quality.
[0091] Based on the optimal cluster center obtained by the search, the ant colony optimization algorithm is introduced to dynamically divide the cluster samples. Each ant corresponds to a cluster sample. The ants move between different cluster centers according to the cluster center attraction and pheromone concentration, and dynamically adjust the sample attribution through the pheromone accumulation and attenuation mechanism to search for the optimal cluster division information.
[0092] The optimal cluster center and optimal cluster division information obtained by fusion search are adaptively adjusted to adjust the initial cluster model parameters, including the number of cluster centers and sample attribution threshold, and repeated iterations are performed until the maximum number of iterations is reached to obtain the optimal cluster model;
[0093] The multi-granularity semantic feature set is input into the optimal clustering model, and the cluster label corresponding to each sample is output.
[0094] Exemplarily, an initial clustering model is first constructed through a multi-granularity semantic feature set of safety education and training data. The construction of the initial clustering model requires setting relevant parameters according to the characteristics of the safety education and training data and domain knowledge, such as the number of cluster centers and the initial attribution of cluster samples. The number of cluster centers determines the number of categories of the final clustering result, and needs to be reasonably set according to the actual distribution of the data and the clustering target. The initial attribution of cluster samples affects the starting point of clustering optimization, and can be initialized using random allocation or heuristic strategies.
[0095] Furthermore, in order to achieve adaptive optimization of the clustering model, the particle swarm optimization algorithm is introduced. The particle swarm optimization algorithm simulates the foraging behavior of bird flocks and uses information sharing and experience learning between particles to find the optimal solution in the search space. In the clustering problem, each particle represents a potential clustering solution. The position of the particle corresponds to the coordinates of the cluster center, and the speed of the particle represents the update direction and step size of the cluster center. Through the iterative optimization strategy, the particles continuously update their positions and speeds and move towards the direction of the optimal cluster center. In each iteration, the fitness function of the particle is used to evaluate the quality of the current clustering solution. The fitness function can comprehensively consider indicators such as intra-class compactness and inter-class separation. Through the iterative search of the particle swarm optimization algorithm, the optimal cluster center can be found.
[0096] After obtaining the optimal cluster center, the ant colony optimization algorithm was introduced to dynamically divide the cluster samples. The ant colony optimization algorithm simulates the behavior of ants in finding food and realizes information interaction and collaboration between ants through the pheromone mechanism. In the clustering problem, each ant corresponds to a cluster sample, and the ants move between different cluster centers according to the attractiveness of the cluster center and the pheromone concentration. The attractiveness of the cluster center indicates the similarity between the sample and the cluster center, and the pheromone concentration reflects the clustering decision of other ants. Through the accumulation and attenuation mechanism of pheromones, the ants dynamically adjust the attribution of samples and continuously optimize the cluster division. After multiple rounds of iterations, the ant colony optimization algorithm can search for the optimal cluster division information.
[0097] In order to further improve the adaptability of the clustering model, the optimal cluster center and optimal cluster division information obtained by the particle swarm optimization algorithm and the ant colony optimization algorithm are integrated to adaptively adjust the parameters of the initial clustering model. According to the number of optimal cluster centers, the number of cluster centers of the clustering model is dynamically adjusted to match the actual distribution of the data. At the same time, according to the optimal cluster division information, the sample attribution threshold of the clustering model is adjusted to better divide the boundary samples. Through the adaptive adjustment of parameters, the clustering model can better adapt to the characteristics of the data and the clustering objectives.
[0098] After completing the adaptive optimization of the clustering model, repeat the above steps until the maximum number of iterations is reached or other termination conditions are met. Through multiple rounds of iterative optimization, the clustering model is continuously improved and updated, and the optimal clustering result is finally obtained. By inputting the multi-granularity semantic feature set into the optimal clustering model, the corresponding clustering label can be output for each sample, realizing the effective clustering of safety education and training data.
[0099] In this embodiment, by integrating the particle swarm optimization algorithm and the ant colony optimization algorithm, it is possible to automatically search for the optimal clustering solution in a complex safety education and training data environment, and realize adaptive optimization and parameter adjustment of the clustering process. This method fully utilizes the global search capability and dynamic optimization characteristics of the swarm intelligence optimization algorithm, overcomes the limitations of traditional clustering methods, and improves the accuracy and efficiency of clustering.
[0100] In an optional embodiment,
[0101] The potential causal relationship between the evaluation indicator nodes is mined through a causal inference algorithm, a causal effect estimation model is constructed, and the dynamic weight vector of the evaluation indicator is obtained by using the causal effect estimation model, including:
[0102] Inputting the original features of the evaluation index into a preset causal invariant feature extractor, mapping the original features to a low-dimensional latent space, and obtaining a latent representation of the evaluation index;
[0103] Construct a multi-task learning framework including causal effect estimation tasks and representation learning tasks, and construct positive and negative sample pairs based on the original features of the evaluation indicators. For the same evaluation indicator, sample pairs sampled from different data distributions are constructed as positive sample pairs. For different evaluation indicators, sample pairs randomly sampled therefrom are constructed as negative sample pairs. Based on the constructed positive and negative sample pairs, a contrastive learning loss function is constructed.
[0104] The causal effect estimation objective function and the contrastive learning loss function are jointly optimized to train the causal invariant feature extractor, so that the causal invariant feature extractor gradually learns the causal invariant representation of the evaluation index, and finally obtains the trained causal invariant feature extractor;
[0105] The original features of the new evaluation indicator samples are obtained, and the original features of the new evaluation indicator samples are input into the trained causal invariant feature extractor to obtain the causal invariant representation of the new evaluation indicator samples, and the causal invariant representation of the new evaluation indicator samples is input into the causal effect estimation model. Through the causal effect estimation model, based on the obtained causal invariant representation of the new evaluation indicator samples, the corresponding causal effect estimation results are obtained. According to the causal effect estimation results, the dynamic weight learning algorithm of the evaluation indicators is introduced, and the reward function is designed. The causal effect strength, indicator importance and model prediction error are comprehensively considered, and the optimization direction of the weight vector is guided by reinforcement learning, so as to finally obtain the dynamic weight vector of the evaluation indicator.
[0106] Exemplarily, first, in order to extract causally invariant latent representations from the original features of the evaluation indicators, a causal invariant feature extractor is designed. The extractor adopts a deep neural network structure to map high-dimensional original features to a low-dimensional latent feature space through multi-layer nonlinear transformations. The feature representation in this latent space is constrained by a specific optimization objective, so that it can capture the intrinsic causal mechanism of the evaluation indicator without being affected by accidental disturbances and distribution drift during the data collection process. The training of the causal invariant feature extractor adopts a multi-task learning framework, considering two optimization objectives: the causal effect estimation task and the representation learning task.
[0107] In the causal effect estimation task, causal inference methods are used to estimate the strength of causal relationships between different evaluation indicators. Specifically, each evaluation indicator is regarded as a treatment variable, and its average causal effect on other indicators is calculated through counterfactual analysis. Here, a matching-based causal effect estimation method is adopted, that is, the samples are first divided into treatment groups and control groups according to the values of the treatment variables, and then the most similar sample pairs on the covariates are found in the two groups of samples for matching, and the average treatment effect is estimated by comparing the differences in the outcome variables between the matched sample pairs. The matching process is implemented by the nearest neighbor algorithm or by optimizing the matching distance objective function. The goal of the causal invariant feature extractor is to learn a feature transformation so that the distribution of the transformed latent features in the treatment group and the control group is as similar as possible, thereby reducing the covariate differences between the matched sample pairs and improving the accuracy of causal effect estimation. The causal effect estimation loss function measures the error between the estimated value and the true value, and the extractor parameters are optimized by gradient backpropagation.
[0108] In the representation learning task, the contrastive learning method is used to optimize the latent representation of the evaluation index, so that the representation similarity of the same index in different environments is higher than the similarity between different indexes. First, the positive and negative sample pairs of representation learning are constructed based on the original features. For each evaluation index, the sample pairs sampled from different data distributions (such as different time periods, regions, populations, etc.) are constructed as positive sample pairs, and the sample pairs randomly sampled between different evaluation indicators are constructed as negative sample pairs. Then, the similarity between the positive and negative sample pairs is calculated in the latent feature space, and the contrastive learning loss function is constructed. By minimizing the difference between the similarity of the positive sample pairs and the similarity of the negative sample pairs, the causal invariant feature extractor can learn the latent representation that can distinguish different indicators and adapt to different environments.
[0109] During the training process, the causal effect estimation loss and the representation learning loss are jointly optimized, and the parameters of the causal invariant feature extractor are iteratively updated through the stochastic gradient descent algorithm, ultimately obtaining a feature extractor that can both accurately estimate the causal effect and extract the causal invariant representation. This multi-task learning framework can simultaneously utilize the statistical information of a large number of unlabeled samples on limited labeled samples, thereby improving the generalization performance of causal effect estimation and potential representation learning tasks.
[0110] In the application phase, given a new evaluation indicator sample, first input its original features into the trained causal invariant feature extractor to obtain the causal invariant latent representation of the sample. Then input the latent representation into a pre-trained causal effect estimation model, which can be a regression model or a causal structure learning model, and output the causal effect strength between the evaluation indicator sample and other indicators. The causal effect strength here can be a scalar value, indicating the average impact of the indicator on other indicators, or a vector or matrix, which describes the causal relationship between the indicator and each other indicator. The causal effect estimation result can be used as an objective basis for the importance of the evaluation indicator and to guide the optimization of the subsequent evaluation indicator weights.
[0111] Finally, in order to dynamically adjust the weights of different evaluation indicators to adapt to changes in evaluation tasks and data environments, an evaluation indicator dynamic weight learning algorithm is introduced. This algorithm models the optimization problem of evaluation indicator weights as a sequential decision process and learns the optimal weight allocation strategy through reinforcement learning. First, a reward function is designed to evaluate the model prediction performance under the current weight vector. The reward function comprehensively considers three factors: causal effect strength, indicator importance, and model prediction error. The causal effect strength is given by the aforementioned causal effect estimation model, the indicator importance is pre-set by expert knowledge or domain experience, and the model prediction error is measured by the generalization performance of the evaluation model on new samples. The reward function integrates the three factors into a scalar value by weighted summation or nonlinear combination. Then, through reinforcement learning algorithms such as policy gradient or value function approximation, the weight vector is continuously explored and optimized to maximize the long-term cumulative reward. The weight learning algorithm updates the weight vector once at each decision step, gradually improves the reward function value, and finally converges to a dynamically optimal weight allocation scheme. This dynamic weight optimization mechanism can adjust the importance of each indicator according to the real-time feedback of the evaluation indicator data, and improve the adaptability and robustness of the evaluation model.
[0112] In an optional embodiment,
[0113] Jointly optimizing the causal effect estimation objective function and the contrastive learning loss function, training the causal invariant feature extractor includes:
[0114] Based on the causal effect estimation objective function, the representation learning regularization term and the contrastive learning loss function, a joint optimization objective function is constructed to train the causal invariant feature extractor. The calculation formula of the joint optimization objective function is as follows:
[0115] ;
[0116] Among them, L represents the joint optimization objective function, L crepresents the causal effect estimation objective function, λ1 represents the balance factor of the learning regularization term, n represents the number of samples of the evaluation indicator, f(·) represents the causal invariant feature extraction function, x i represents the original features of the i-th evaluation indicator, c i represents interference factors, |·| F represents the Frobenius norm, λ2 represents the balance factor of the learning loss function, and z i represents the potential representation of the i-th evaluation indicator in the positive sample pair, z j represents the potential representation of the jth evaluation indicator in the positive sample pair, sim(·) represents the similarity metric function, τ represents the temperature parameter, K represents the number of negative samples, and z k represents the potential representation of the kth evaluation metric in the negative sample pair.
[0117] In this embodiment, by introducing causal invariant representation learning, the problems of inconsistent data distribution and selection bias of evaluation indicators are well solved, and the reliability of cross-domain evaluation effect prediction is greatly improved. At the same time, the introduction of multi-task learning framework and dynamic weight optimization strategy further enhances the ability of this method to handle complex evaluation environments and dynamic evaluation needs, enabling it to quickly respond to changes in evaluation tasks, adaptively adjust the weights of evaluation indicators, and make full use of the feedback information of evaluation data to guide weight evolution, and continuously improve the performance of the evaluation model. In addition, the causal invariant representation of evaluation indicators and causal effect estimation results learned by this method also provide new ideas for the interpretability analysis of evaluation indicators and the screening of key indicators.
[0118] S103. Take the dynamic weight vectors of cluster labels and evaluation indicators as input, perform iterative optimization through the pre-trained strategy optimization model, obtain the adaptive monitoring and evaluation strategy, build a blockchain network, store the obtained multi-source heterogeneous monitoring data, cluster labels, dynamic weight vectors of evaluation indicators and adaptive monitoring and evaluation strategies on the chain, design the monitoring and evaluation process smart contract, realize the automatic triggering and scheduling of the monitoring and evaluation process, and update the adaptive monitoring and evaluation strategy in real time.
[0119] In an optional embodiment,
[0120] The dynamic weight vector of cluster labels and evaluation indicators is used as input, and iterative optimization is performed through the pre-trained strategy optimization model to obtain the adaptive monitoring evaluation strategy including:
[0121] Construct an initial strategy optimization model, which includes a strategy generator and a strategy evaluator. The strategy generator generates candidate monitoring and evaluation strategies based on the current environmental status and historical decision sequences. The strategy evaluator estimates the expected utility of different monitoring and evaluation strategies, ranks the candidate monitoring and evaluation strategies, and selects the optimal strategy as output.
[0122] Collect historical safety education and training monitoring data and construct a monitoring trajectory sequence dataset for trajectory prediction pre-training, where each monitoring trajectory sequence contains multiple time steps, and each time step corresponds to a monitoring state vector;
[0123] The monitoring trajectory sequence is divided into multiple subsequences to form a prediction data set, and a random sampling method is used to select some subsequences as the prediction model input to obtain the state prediction distribution. By minimizing the difference loss between the predicted trajectory and the true trajectory, the prediction model parameters are updated to obtain a trained prediction model, and the trained prediction model is integrated into the initial policy optimization model to obtain the target policy optimization model.
[0124] The clustering labels and evaluation index weight vectors are concatenated into a state feature vector, which is input into the target strategy optimization model. The model parameters of the target strategy optimization model are updated by a gradient descent algorithm, and the iteration is repeated until the model converges, thereby finally obtaining an adaptive monitoring evaluation strategy.
[0125] Exemplarily, an initial policy optimization model is first constructed, which consists of two parts: a policy generator and a policy evaluator. The policy generator is a deep neural network model that takes the current monitoring environment state and the historical monitoring decision sequence as input, and generates a series of candidate monitoring and evaluation strategies through forward propagation. These candidate strategies represent different monitoring schemes that can be selected under the current circumstances, such as adjusting the monitoring frequency, changing the monitoring content, optimizing the monitoring resource configuration, etc. The policy evaluator is also a deep neural network model, and its function is to estimate the expected utility of different candidate strategies under long-term execution, that is, the expected benefits that can be obtained by using the strategy for monitoring and evaluation under the current environmental state. The policy evaluator compares the expected utilities of different candidate strategies, ranks them by their merits, and selects the strategy with the highest expected utility as the optimal strategy output.
[0126] Then, in order to enable the policy optimization model to better understand and predict the evolution of the monitoring process, it is proposed to use unlabeled historical monitoring data for pre-training. Specifically, the monitoring log data recorded during the past safety education and training are collected to construct a monitoring trajectory sequence data set. Each sample in the data set corresponds to a complete monitoring activity and consists of monitoring states at multiple time steps. The monitoring state of each time step can be represented by a feature vector, which contains various monitoring indicator values at that time step, such as personnel attendance rate, training completion rate, and assessment pass rate. Each monitoring trajectory sequence is divided into several subsequences of the same length to form a state prediction data set. A part of the time steps are extracted from each subsequence by random sampling, and they are spliced into an incomplete state sequence as the input of the prediction model. The prediction model is a sequence-to-sequence neural network that takes an incomplete state sequence as input and generates the state prediction distribution of subsequent multiple time steps by autoregression. The parameters of the prediction model are optimized by minimizing the difference between the generated predicted state sequence and the true complete state sequence. This difference is measured by a loss function, such as mean square error loss or negative log-likelihood loss. The model parameters are iteratively updated through optimization algorithms such as stochastic gradient descent until the loss function converges to a lower level, and a trained monitoring trajectory prediction model is obtained. The trained prediction model is embedded into the initial policy optimization model to assist the policy generator and evaluator to more accurately estimate and optimize the long-term policy utility, thereby obtaining the target policy optimization model.
[0127] Finally, when the target strategy optimization model is actually applied, it is necessary to set the corresponding state characteristics and optimization goals according to the current specific monitoring tasks and environment. A two-stage method is adopted. First, different monitoring objects are divided into several categories through a clustering algorithm, and each category corresponds to a cluster label. At the same time, the importance weights of various monitoring and evaluation indicators are set according to expert experience to form an evaluation indicator weight vector. The cluster label of the current monitoring object and the evaluation indicator weight vector are spliced together to form a complete state feature vector, which is input into the target strategy optimization model. The strategy optimization model generates candidate strategies and corresponding expected utility values through forward calculation, and then optimizes the model parameters through gradient back propagation, so that the generated strategy can maximize the long-term expected utility. This optimization process is iteratively performed through the stochastic gradient descent algorithm, and the model parameters are continuously updated until the model converges and the performance indicators on the validation set are no longer significantly improved. In this way, an adaptive monitoring and evaluation strategy optimized for the current monitoring task and environment is obtained. This strategy can dynamically adjust the monitoring behavior according to real-time monitoring data and feedback, improve monitoring efficiency while meeting regulatory needs, and save monitoring costs.
[0128] In this embodiment, by combining unsupervised pre-training and supervised fine-tuning, full use is made of a large amount of unlabeled historical monitoring data and a small amount of labeled policy preference data, thereby greatly reducing the data labeling cost of monitoring policy optimization. At the same time, by incorporating clustering information and evaluation index preferences into state characteristics, the optimized strategy can adapt to different monitoring object characteristics and regulatory objectives, and has greater flexibility and applicability. In addition, this method adopts an end-to-end differentiable neural network structure, and realizes the joint optimization of policy generation and evaluation through gradient back propagation, avoiding the sample inefficiency problem caused by the alternating iteration of policy search and policy evaluation in traditional reinforcement learning algorithms, and accelerating the convergence speed of policy optimization.
[0129] In an optional embodiment,
[0130] Build a blockchain network, store the obtained multi-source heterogeneous monitoring data, cluster labels, dynamic weight vectors of evaluation indicators, and adaptive monitoring and evaluation strategies on the chain, design a monitoring and evaluation process smart contract, realize the automatic triggering and scheduling of the monitoring and evaluation process, and update the adaptive monitoring and evaluation strategy in real time, including:
[0131] Constructing a blockchain network, designing a data structure and a transaction format in the blockchain network, using the data structure and the transaction format to encapsulate multi-source heterogeneous monitoring data, cluster labels, dynamic weight vectors of evaluation indicators, and adaptive monitoring and evaluation strategy parameters, and using an encryption algorithm to encrypt the encapsulated multi-source heterogeneous monitoring data, cluster labels, dynamic weight vectors of evaluation indicators, and adaptive monitoring and evaluation strategy parameters;
[0132] Based on the encrypted multi-source heterogeneous monitoring data, cluster labels, dynamic weight vectors of evaluation indicators and adaptive monitoring and evaluation strategy parameters, a smart contract is deployed in the blockchain network, and the triggering conditions and execution logic of the monitoring and evaluation process are set in the smart contract. The monitoring and evaluation process is automatically triggered and the execution logic is executed according to the triggering conditions;
[0133] In the process of executing the monitoring and evaluation process, the adaptive monitoring and evaluation strategy parameters are abstracted as the state variables of the smart contract, and the strategy parameter reading function, the strategy parameter writing function and the strategy update auxiliary function are designed in the smart contract. The strategy parameter reading function is used for external calls to read the current adaptive monitoring and evaluation strategy parameter value. When the strategy parameter writing function receives a write request that meets the authority verification, it writes the new adaptive monitoring and evaluation strategy parameters into the state variables;
[0134] After the strategy optimization model generates new adaptive monitoring and evaluation strategy parameters, the strategy parameter writing function of the smart contract is called to pass in the new adaptive monitoring and evaluation strategy parameters and write them into the state variables to achieve real-time update of the adaptive monitoring and evaluation strategy.
[0135] For example, first, build a blockchain network that is suitable for monitoring and evaluation scenarios. According to the needs of monitoring and evaluation, select appropriate consensus mechanisms, network topologies, and node deployment solutions. The choice of consensus mechanism needs to take into account factors such as decentralization, scalability, and real-time performance, such as using algorithms that support fast consensus, such as PBFT and RAFT. The network topology can adopt a consortium chain or private chain model and be designed according to the participants and trust relationships of monitoring and evaluation. The node deployment plan needs to comprehensively consider factors such as geographical distribution and disaster recovery backup to ensure the availability and robustness of the network.
[0136] In the blockchain network, data structures and transaction formats that are suitable for monitoring and evaluation scenarios are designed. The data structure is used to encapsulate information such as multi-source heterogeneous monitoring data, cluster labels, dynamic weight vectors of evaluation indicators, and adaptive monitoring and evaluation strategy parameters. The transaction format defines the rules for data propagation and verification in the blockchain network. The design of data structures and transaction formats needs to take into account the integrity, consistency, and scalability of data, while considering the storage and transmission efficiency of data. General serialization formats such as Protocol Buffers and JSON can be used, or targeted custom formats can be designed.
[0137] In order to ensure the security and privacy of data, the encapsulated multi-source heterogeneous monitoring data, cluster labels, dynamic weight vectors of evaluation indicators, and adaptive monitoring and evaluation strategy parameters are encrypted. Symmetric encryption algorithms (such as AES) or asymmetric encryption algorithms (such as RSA) are used to encrypt data and generate encrypted ciphertext. The encryption process can be completed before the data is uploaded to the chain, or it can be implemented in a smart contract. The management of encryption keys can adopt centralized key management services, or use the distributed characteristics of blockchain to achieve decentralized key management.
[0138] Based on encrypted multi-source heterogeneous monitoring data, cluster labels, dynamic weight vectors of evaluation indicators, and adaptive monitoring and evaluation strategy parameters, monitoring and evaluation smart contracts are deployed in the blockchain network. Smart contracts are self-executing programs running on the blockchain that automate the monitoring and evaluation process through pre-defined logic and rules.
[0139] In the smart contract, set the trigger conditions and execution logic of the monitoring and evaluation process. The trigger condition defines the start time of the monitoring and evaluation process, which can be set based on factors such as time, events or external calls. The execution logic describes the specific steps and algorithms of the monitoring and evaluation process, including data preprocessing, feature extraction, cluster analysis, indicator calculation, strategy execution and other links. The execution logic can be encoded in smart contract programming languages such as Solidity, and implemented using mechanisms such as state variables, functions and events of the contract.
[0140] According to the preset trigger conditions, the smart contract automatically triggers the monitoring and evaluation process and executes the corresponding execution logic. When the trigger conditions are met, the smart contract automatically reads the encrypted multi-source heterogeneous monitoring data, clustering labels, dynamic weight vectors of evaluation indicators, and adaptive monitoring and evaluation strategy parameters, decrypts them, and processes them according to the execution logic. The intermediate results and final results during the execution process can be written into the blockchain to generate an audit log that cannot be tampered with. The execution results can trigger subsequent actions, such as generating alarms, adjusting policy parameters, etc.
[0141] In the process of executing the monitoring and evaluation process, the adaptive monitoring and evaluation strategy parameters are abstracted into the state variables of the smart contract. The state variable is the persistent storage in the smart contract, which is used to save the key data of the contract. By mapping the strategy parameters to the state variables, the strategy parameters can be shared and synchronized in the blockchain network.
[0142] Design the policy parameter reading function, policy parameter writing function and policy update auxiliary function in the smart contract. The policy parameter reading function is used for external calls to read the current adaptive monitoring and evaluation policy parameter values. The policy parameter writing function receives external write requests, and after verifying the legitimacy and permissions of the request, writes the new adaptive monitoring and evaluation policy parameters to the state variables. The policy update auxiliary function encapsulates the common logic of policy updates, such as parameter validity checks, version control, etc.
[0143] When new adaptive monitoring and evaluation strategy parameters are generated, the new strategy parameters are passed in and written into the state variables by calling the strategy parameter writing function of the smart contract. The writing process is submitted to the blockchain network in the form of a transaction. After verification and confirmation by the consensus mechanism, the new strategy parameters are synchronized to all nodes to achieve real-time update of the strategy. The updated strategy parameters can take effect immediately when the monitoring and evaluation process is triggered next time, without manual intervention.
[0144] In this embodiment, the decentralized, tamper-proof and automatic execution characteristics of blockchain are utilized to achieve transparent, reliable, automated and real-time updating of the monitoring and evaluation process. The policy update process is protected by the blockchain consensus mechanism, which ensures the consistency and tamper-proofness of the policy. The design of the smart contract encapsulates the complex logic of the monitoring and evaluation process and improves the automation level of the evaluation process. At the same time, the encryption mechanism of the blockchain network ensures the security of monitoring data and policy parameters to prevent unauthorized access and tampering. By building a blockchain network for monitoring and evaluation, designing data structure and transaction format, and using smart contracts to achieve real-time updating and automatic execution of policies, the monitoring and evaluation process is decentralized, automated, secure and reliable.
[0145] Figure 2 Schematic diagram of the structure of the security education and training monitoring and evaluation system based on the optimized operation mechanism according to the embodiment of the present invention. Figure 2 As shown, the system comprises:
[0146] The first unit is used to construct a safety education and training knowledge graph, perform semantic representation of the safety education and training process based on the safety education and training knowledge graph, obtain the semantic representation result of the safety education and training process, establish a multi-granularity feature mapping model according to the semantic representation result, and simultaneously obtain multi-source heterogeneous monitoring data of the safety education and training process, and use the multi-granularity feature mapping model to perform multi-granularity semantic feature extraction on the multi-source heterogeneous monitoring data to obtain a multi-granularity semantic feature set;
[0147] The second unit is used to build an adaptive clustering model that integrates swarm intelligence optimization, takes a multi-granularity semantic feature set as input, outputs cluster labels, and builds a causal graph model for safety education and training evaluation indicators. The causal graph model contains evaluation indicator nodes and causal relationship edges. The potential causal relationship between the evaluation indicator nodes is mined through a causal inference algorithm, and a causal effect estimation model is built. The dynamic weight vector of the evaluation indicator is obtained using the causal effect estimation model;
[0148] The third unit is used to take the dynamic weight vectors of cluster labels and evaluation indicators as input, perform iterative optimization through a pre-trained strategy optimization model, obtain an adaptive monitoring and evaluation strategy, build a blockchain network, store the obtained multi-source heterogeneous monitoring data, cluster labels, dynamic weight vectors of evaluation indicators, and adaptive monitoring and evaluation strategies on the chain, design a monitoring and evaluation process smart contract, realize the automatic triggering and scheduling of the monitoring and evaluation process, and update the adaptive monitoring and evaluation strategy in real time.
[0149] According to a third aspect of the embodiments of the present invention,
[0150] An electronic device is provided, comprising:
[0151] processor;
[0152] a memory for storing processor-executable instructions;
[0153] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.
[0154] A fourth aspect of the embodiments of the present invention is:
[0155] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.
[0156] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.
[0157] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A safety education and training monitoring and evaluation method based on operation mechanism optimization, characterized in that: include: Construct a safety education and training knowledge graph, perform semantic representation of the safety education and training process based on the safety education and training knowledge graph, obtain the semantic representation result of the safety education and training process, establish a multi-granularity feature mapping model based on the semantic representation result, and simultaneously obtain multi-source heterogeneous monitoring data of the safety education and training process, use the multi-granularity feature mapping model to perform multi-granularity semantic feature extraction on the multi-source heterogeneous monitoring data, and obtain a multi-granularity semantic feature set; Construct an adaptive clustering model that integrates swarm intelligence optimization, takes a multi-granularity semantic feature set as input, outputs cluster labels, and constructs a causal graph model for safety education and training evaluation indicators. The causal graph model contains evaluation indicator nodes and causal relationship edges. The potential causal relationship between the evaluation indicator nodes is mined through a causal inference algorithm, and a causal effect estimation model is constructed. The dynamic weight vector of the evaluation indicator is obtained using the causal effect estimation model. The cluster labels and dynamic weight vectors of evaluation indicators are used as inputs, and the pre-trained strategy optimization model is used for iterative optimization to obtain the adaptive monitoring and evaluation strategy. A blockchain network is built to store the obtained multi-source heterogeneous monitoring data, cluster labels, dynamic weight vectors of evaluation indicators, and adaptive monitoring and evaluation strategies on the chain. A monitoring and evaluation process smart contract is designed to realize the automatic triggering and scheduling of the monitoring and evaluation process, and to update the adaptive monitoring and evaluation strategy in real time. The dynamic weight vector of cluster labels and evaluation indicators is used as input, and iterative optimization is performed through the pre-trained strategy optimization model to obtain the adaptive monitoring evaluation strategy including: Construct an initial strategy optimization model, which includes a strategy generator and a strategy evaluator. The strategy generator generates candidate monitoring and evaluation strategies based on the current environmental status and historical decision sequences. The strategy evaluator estimates the expected utility of different monitoring and evaluation strategies, ranks the candidate monitoring and evaluation strategies, and selects the optimal strategy as output. Collect historical safety education and training monitoring data and construct a monitoring trajectory sequence dataset for trajectory prediction pre-training, where each monitoring trajectory sequence contains multiple time steps, and each time step corresponds to a monitoring state vector; The monitoring trajectory sequence is divided into multiple subsequences to form a prediction data set, and a random sampling method is used to select some subsequences as the prediction model input to obtain the state prediction distribution. By minimizing the difference loss between the predicted trajectory and the true trajectory, the prediction model parameters are updated to obtain a trained prediction model, and the trained prediction model is integrated into the initial policy optimization model to obtain the target policy optimization model. The clustering labels and evaluation index weight vectors are concatenated into a state feature vector, which is input into the target strategy optimization model. The model parameters of the target strategy optimization model are updated by the gradient descent algorithm. The iteration is repeated until the model converges, and finally an adaptive monitoring and evaluation strategy is obtained. The candidate monitoring and evaluation strategies represent different monitoring schemes that can be selected under the current situation, including: adjusting the monitoring frequency, changing the monitoring content, and optimizing the monitoring resource configuration.
2. The method according to claim 1, characterized in that: A safety education and training knowledge graph is constructed, and a semantic representation of the safety education and training process is performed based on the safety education and training knowledge graph. The semantic representation results of the safety education and training process include: Construct a safety education and training knowledge graph, which includes entities and relationships; The safety education and training knowledge graph is embedded using a pre-set knowledge graph embedding model. Triples containing head entities, relations, and tail entities are defined. By minimizing the energy function of correct triples and maximizing the energy function of incorrect triples, entities and relations are mapped into a low-dimensional continuous vector space to obtain embedded representations of entities and relations. Extract multiple relationship paths from the safety education and training knowledge graph, encode the multiple relationship paths using a path sorting model, determine the path feature vector and the answer sorting function, sort the candidate answers according to the path feature vector, and obtain the candidate answer sorting result; at the same time, use an attention-based path encoder to learn the weights of multiple relationship paths, and obtain the aggregate vector representation of multiple relationship paths through attention weighting; Constructing a graph neural network model, which includes a graph convolutional network and a graph attention network. Based on the structural information of the safety education and training knowledge graph, the neighbor information of the current entity is aggregated through the graph convolutional network, the weights of different neighbors are calculated through the graph attention network, the feature vector representation of the current entity is updated, and the feature vector representation of the target entity is obtained; The embedded representations of entities and relations, the ranking results of candidate answers, the aggregated vector representations of multiple relationship paths, and the feature vector representations of target entities are integrated, and the attention mechanism is introduced for weighted fusion to obtain a multi-perspective fusion safety education and training knowledge graph representation; Based on the fused safety education and training knowledge graph representation, the safety education and training process is semantically represented to obtain the semantic representation result of the safety education and training process.
3. The method according to claim 1, characterized in that An adaptive clustering model integrating swarm intelligence optimization is constructed, which takes multi-granularity semantic feature sets as input and outputs cluster labels including: Based on the pre-obtained multi-granularity semantic feature set, an initial clustering model is constructed, and the number of cluster centers and cluster sample attribution of the initial clustering model are set according to the characteristics of safety education and training data and domain knowledge; The particle swarm optimization algorithm is introduced into the initial clustering model. The particle position and velocity are updated through iterative optimization strategy to search for the optimal cluster center. The particle position represents the cluster center coordinates, the particle velocity represents the cluster center update direction and step size, and the particle fitness function is used to evaluate the clustering quality. Based on the optimal cluster center obtained by the search, the ant colony optimization algorithm is introduced to dynamically divide the cluster samples. Each ant corresponds to a cluster sample. The ants move between different cluster centers according to the cluster center attraction and pheromone concentration, and dynamically adjust the sample attribution through the pheromone accumulation and attenuation mechanism to search for the optimal cluster division information. The optimal cluster center and optimal cluster division information obtained by fusion search are adaptively adjusted to adjust the initial cluster model parameters, including the number of cluster centers and sample attribution threshold, and repeated iterations are performed until the maximum number of iterations is reached to obtain the optimal cluster model; The multi-granularity semantic feature set is input into the optimal clustering model, and the cluster label corresponding to each sample is output.
4. The method according to claim 1, characterized in that: The potential causal relationship between the evaluation indicator nodes is mined through a causal inference algorithm, a causal effect estimation model is constructed, and the dynamic weight vector of the evaluation indicator is obtained by using the causal effect estimation model, including: Inputting the original features of the evaluation index into a preset causal invariant feature extractor, mapping the original features to a low-dimensional latent space, and obtaining a latent representation of the evaluation index; Construct a multi-task learning framework including causal effect estimation tasks and representation learning tasks, and construct positive and negative sample pairs based on the original features of the evaluation indicators. For the same evaluation indicator, sample pairs sampled from different data distributions are constructed as positive sample pairs. For different evaluation indicators, sample pairs randomly sampled therefrom are constructed as negative sample pairs. Based on the constructed positive and negative sample pairs, a contrastive learning loss function is constructed. The causal effect estimation objective function and the contrastive learning loss function are jointly optimized to train the causal invariant feature extractor, so that the causal invariant feature extractor gradually learns the causal invariant representation of the evaluation index, and finally obtains the trained causal invariant feature extractor; The original features of the new evaluation indicator samples are obtained, and the original features of the new evaluation indicator samples are input into the trained causal invariant feature extractor to obtain the causal invariant representation of the new evaluation indicator samples, and the causal invariant representation of the new evaluation indicator samples is input into the causal effect estimation model. Through the causal effect estimation model, based on the obtained causal invariant representation of the new evaluation indicator samples, the corresponding causal effect estimation results are obtained. According to the causal effect estimation results, the dynamic weight learning algorithm of the evaluation indicators is introduced, and the reward function is designed. The causal effect strength, indicator importance and model prediction error are comprehensively considered, and the optimization direction of the weight vector is guided by reinforcement learning, so as to finally obtain the dynamic weight vector of the evaluation indicator.
5. The method according to claim 4, characterized in that Jointly optimizing the causal effect estimation objective function and the contrastive learning loss function, training the causal invariant feature extractor includes: Based on the causal effect estimation objective function, the representation learning regularization term and the contrastive learning loss function, a joint optimization objective function is constructed to train the causal invariant feature extractor. The calculation formula of the joint optimization objective function is as follows: ; Among them, L represents the joint optimization objective function, L c represents the causal effect estimation objective function, λ1 represents the balance factor of the learning regularization term, n represents the number of samples of the evaluation indicator, f(·) represents the causal invariant feature extraction function, x i represents the original features of the i-th evaluation indicator, c i represents interference factors, |·| F represents the Frobenius norm, λ2 represents the balance factor of the learning loss function, and z i represents the potential representation of the i-th evaluation indicator in the positive sample pair, z j represents the potential representation of the jth evaluation indicator in the positive sample pair, sim(·) represents the similarity metric function, τ represents the temperature parameter, K represents the number of negative samples, and z k represents the potential representation of the kth evaluation metric in the negative sample pair.
6. The method according to claim 1, characterized in that Build a blockchain network, store the obtained multi-source heterogeneous monitoring data, cluster labels, dynamic weight vectors of evaluation indicators, and adaptive monitoring and evaluation strategies on the chain, design a monitoring and evaluation process smart contract, realize the automatic triggering and scheduling of the monitoring and evaluation process, and update the adaptive monitoring and evaluation strategy in real time, including: Constructing a blockchain network, designing a data structure and a transaction format in the blockchain network, using the data structure and the transaction format to encapsulate multi-source heterogeneous monitoring data, cluster labels, dynamic weight vectors of evaluation indicators, and adaptive monitoring and evaluation strategy parameters, and using an encryption algorithm to encrypt the encapsulated multi-source heterogeneous monitoring data, cluster labels, dynamic weight vectors of evaluation indicators, and adaptive monitoring and evaluation strategy parameters; Based on the encrypted multi-source heterogeneous monitoring data, cluster labels, dynamic weight vectors of evaluation indicators and adaptive monitoring and evaluation strategy parameters, a smart contract is deployed in the blockchain network, and the triggering conditions and execution logic of the monitoring and evaluation process are set in the smart contract. The monitoring and evaluation process is automatically triggered and the execution logic is executed according to the triggering conditions; In the process of executing the monitoring and evaluation process, the adaptive monitoring and evaluation strategy parameters are abstracted as the state variables of the smart contract, and the strategy parameter reading function, the strategy parameter writing function and the strategy update auxiliary function are designed in the smart contract. The strategy parameter reading function is used for external calls to read the current adaptive monitoring and evaluation strategy parameter value. When the strategy parameter writing function receives a write request that meets the authority verification, it writes the new adaptive monitoring and evaluation strategy parameters into the state variables; After the strategy optimization model generates new adaptive monitoring and evaluation strategy parameters, the strategy parameter writing function of the smart contract is called to pass in the new adaptive monitoring and evaluation strategy parameters and write them into the state variables to achieve real-time update of the adaptive monitoring and evaluation strategy.
7. A safety education and training monitoring and evaluation system based on operation mechanism optimization, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to construct a safety education and training knowledge graph, perform semantic representation of the safety education and training process based on the safety education and training knowledge graph, obtain the semantic representation result of the safety education and training process, establish a multi-granularity feature mapping model according to the semantic representation result, and simultaneously obtain multi-source heterogeneous monitoring data of the safety education and training process, and use the multi-granularity feature mapping model to perform multi-granularity semantic feature extraction on the multi-source heterogeneous monitoring data to obtain a multi-granularity semantic feature set; The second unit is used to build an adaptive clustering model that integrates swarm intelligence optimization, takes a multi-granularity semantic feature set as input, outputs cluster labels, and builds a causal graph model for safety education and training evaluation indicators. The causal graph model contains evaluation indicator nodes and causal relationship edges. The potential causal relationship between the evaluation indicator nodes is mined through a causal inference algorithm, and a causal effect estimation model is built. The dynamic weight vector of the evaluation indicator is obtained using the causal effect estimation model; The third unit is used to take the dynamic weight vectors of cluster labels and evaluation indicators as input, perform iterative optimization through a pre-trained strategy optimization model, obtain an adaptive monitoring and evaluation strategy, build a blockchain network, store the obtained multi-source heterogeneous monitoring data, cluster labels, dynamic weight vectors of evaluation indicators, and adaptive monitoring and evaluation strategies on the chain, design a monitoring and evaluation process smart contract, realize the automatic triggering and scheduling of the monitoring and evaluation process, and update the adaptive monitoring and evaluation strategy in real time.
8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.
9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.
