A method and system for information reuse based on deep learning

Through the deep learning-based information reuse method, the MDP model is used to optimize the security measures selection in the HAZOP analysis report, which solves the problems of information sharing and reuse, improves the pertinence and efficiency of security measures, and improves the response speed and knowledge inheritance.

CN118710067BActive Publication Date: 2024-11-08CHINA SPECIAL EQUIP INSPECTION & RES INST +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411200743.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-29
Publication Date
2024-11-08
Estimated Expiration
2044-08-29

AI Technical Summary

Technical Problem

The existing HAZOP analysis methods have problems with data redundancy and repeated query in terms of information sharing and reuse, which leads to wasting valuable time for experts, and it is difficult for different projects and teams to effectively share and reuse information in design, construction and other stages.

Method used

Using deep learning-based information reuse method, we can collect and extract risk type tags and security measures in historical HAZOP analysis reports, establish an MDP model, optimize the security measure selection process, realize structured storage of information and fast query, and provide real-time Safety measures recommendations.

Benefits of technology

It significantly improves the pertinence and effectiveness of security measures, optimizes the decision-making process, improves the response speed, enhances knowledge inheritance and accumulation, and improves resource utilization efficiency and transparency and traceability of decision-making.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118710067B_ABST
    Figure CN118710067B_ABST
Patent Text Reader

Abstract

The present application relates to a method and system for information reuse based on deep learning, and to the technical field of information reuse, wherein the method comprises the steps of first acquisition, first extraction, first modeling, setting, second acquisition, first input, query and first output; the system comprises an information acquisition module, an information extraction module, a model building module, and an input and output module. The present application introduces an MDP model to predict safety measures in a HAZOP analysis report, which can significantly improve the pertinence and effectiveness of safety measures and optimize the decision-making process.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of information reuse, and in particular to an information reuse method and system based on deep learning. Background Art

[0002] Hazard and Operability Analysis (HAZOP) is a widely used methodology that aims to deeply identify and accurately assess potential operational risks, and then propose effective safety measures to significantly reduce the probability of accidents. This process relies on the cooperation of a cross-disciplinary team that integrates senior experts in engineering, operations and safety. They skillfully use a series of preset guide words (such as "increase", "decrease", "missing", "reverse", etc.) to conduct a comprehensive and systematic review of potential deviations in process flows and operations, and then analyze the root causes of deviations, predict their possible consequences, and plan corresponding mitigation strategies.

[0003] The HAZOP analysis report is the core result of this process. It not only records in detail all the potential risk points revealed during the analysis process, but also proposes targeted safety measures and estimates the risk mitigation effects that may be achieved after the implementation of these safety measures. The existing HAZOP analysis professional software can realize resource sharing to a certain extent and play a role in auxiliary analysis. However, the expert knowledge base formed by this method generally only stores HAZOP information in a "stacked" manner in the database, without establishing the connection between knowledge. There is a lot of data redundancy, and the target knowledge is screened from a large number of repeated HAZOP knowledge query results, which wastes the precious analysis time of HAZOP experts and is not conducive to information sharing and reuse among different projects and teams at various stages such as design and construction. Summary of the invention

[0004] In order to improve the reusability of information in HAZOP analysis reports, the present application provides an information reuse method and system based on deep learning.

[0005] In the first aspect, the present application provides an information reuse method based on deep learning, which adopts the following technical solution:

[0006] A deep learning-based information reuse method comprises the following steps:

[0007] First collection: Collect historical HAZOP analysis reports and record them as the first report;

[0008] First extraction: extract and store the risk type label and safety measures of the first report;

[0009] First modeling: establish the MDP model;

[0010] Setting: including setting the state set, setting the action set, setting the transition probability and setting the reward step;

[0011] Set state set: The set of all risk type labels is used as the state set of the MDP model;

[0012] Set action set: The set of security measures taken by all risk type labels is used as the action set of the MDP model;

[0013] Set the transition probability: The calculation model of the transition probability is as follows:

[0014] ;

[0015] Among them, p is the transition probability; is the frequency of occurrence of the mth state; is the severity of the mth state;

[0016] Set rewards: Set the rewards of the MDP model based on the risk type label. The calculation model is as follows:

[0017] ;

[0018] in, yes The reward of the MDP model at this moment, is the weight of the i-th risk type label among all risk type labels; is the proportion of the jth security measure among all security measures included in the current risk type label; is the reward compensation amount; a is the number of risk type labels in the state set; b is the number of safety measures in the current risk type label; is the implementation of the jth safety measure in the i-th risk label, and the calculation model is as follows:

[0019] ;

[0020] Second collection: real-time collection of HAZOP analysis reports, recorded as the second report;

[0021] First input: take the second report as the current state of the MDP model and obtain the conditional probability of the future state of the MDP model;

[0022] Query: Based on the risk type label and safety measures of the stored first report, query the first report corresponding to the maximum condition probability, record it as the third report, and use the safety measures corresponding to the third report as the safety measures of the second report;

[0023] First output: output the second report and the safety measures for the second report.

[0024] By adopting the above technical solution, the present application extracts risk type labels and safety measures from the historical HAZOP analysis report (first report), and establishes an MDP model based on these data, takes the second report as the current state of the MDP model, obtains the conditional probability of the future state of the MDP model, and then obtains the first report that is most similar to the second report, takes the safety measures of the first report as the safety measures of the second report, and outputs the second report and the safety measures of the second report for reference by the staff, especially when different projects and different teams need to perform HAZOP analysis at various stages such as design, construction, and maintenance, the corresponding safety measures can be quickly obtained through this method. The real-time collected HAZOP analysis report (second report) can be immediately processed by the MDP model, and the recommended safety measures can be quickly output. This instant feedback mechanism greatly improves the company's response speed to potential risks, helps to take preventive measures before risks occur, and reduces the probability and loss of accidents. By storing the experience and knowledge in the historical HAZOP analysis report in the MDP model in a structured manner and using it to guide future decision-making processes, and assigning the safety measures of the first report to the second report for reference by the staff, the reuse of information in the HAZOP analysis report is achieved. This application optimizes the safety measure selection process in the HAZOP analysis report by introducing the MDP model, which can significantly improve the pertinence and effectiveness of safety measures, optimize the decision-making process, increase response speed, enhance knowledge inheritance and accumulation, improve resource utilization efficiency, and enhance the transparency and traceability of decision-making.

[0025] Optionally, after performing the first acquisition step and before performing the first extraction step, the method further includes:

[0026] First judgment: judging whether the first report has a risk type label, if so, executing the first extraction step; if not, executing the data processing step;

[0027] Data processing: including the second modeling, second extraction, third extraction, second judgment, fusion and dimensionality reduction steps;

[0028] Second modeling: Building a BERT model based on deep learning algorithms;

[0029] Second extraction: Based on the BERT model, the deep semantic features of the first report are extracted to obtain a deep semantic feature vector;

[0030] Third extraction: using the TF-IDF algorithm to extract keyword features in the first report and obtain a keyword feature vector;

[0031] Second judgment: judge whether the deep semantic feature vector and the keyword feature vector can be spliced. If yes, perform the fusion step; if not, perform the dimension reduction step;

[0032] Fusion: concatenate the deep semantic feature vector and the keyword feature vector to obtain the first fusion vector;

[0033] Dimensionality reduction: reduce the dimension of the deep semantic feature vector or keyword feature vector, unify the dimensions of the deep semantic feature vector and the keyword feature vector, and perform the fusion step;

[0034] Second output: Use the Softmax activation function to convert the first fusion vector into a probability distribution, and output the predicted probability distribution of each risk type in the first report;

[0035] Get label: The risk type with predicted probability distribution greater than the preset threshold is used as the risk type label for the first report.

[0036] By adopting the above technical solution, the present application introduces a pre-judgment step (first judgment), which can intelligently choose whether to directly use the existing first report to perform the first extraction step, and in the data processing step, the deep learning algorithm (BERT model) and TF-IDF algorithm are used to extract the deep semantic features and keyword features of the first report respectively, and through the judgment, fusion or dimensionality reduction steps, these features are effectively integrated, and finally the fused features are converted into the predicted probability distribution of the risk type through the Softmax activation function, and the risk type label of the report is accurately obtained based on the preset threshold. Through a series of operations, the present application realizes the label extraction of the first report without risk type label, so that the first report in the first extraction step has a risk type label. On the one hand, the number of key elements in the MDP model is increased, and on the other hand, some risk type labels of HAZOP analysis reports are supplemented.

[0037] Optionally, after executing the vector extraction step and before executing the second determination step, the method further includes:

[0038] Fourth extraction: extract the dimension of the deep semantic feature vector, recorded as the first dimension;

[0039] Fifth extraction: extract the dimension of the keyword feature vector, recorded as the second dimension.

[0040] By adopting the above technical solution, the present application extracts the dimension of the deep semantic feature vector (first dimension) and the dimension of the keyword feature vector (second dimension), quantifies the degree of dimensional mismatch between the two feature vectors, and uses the first dimension and the second dimension as a direct basis for determining whether a dimensionality reduction operation is required and how much dimensionality reduction to perform in the subsequent judgment step. If the first dimension and the second dimension are different, it indicates that the two feature vectors do not match in dimension and cannot be fused. At this time, a dimensionality reduction step needs to be performed to unify the dimensions. On the contrary, if the first dimension and the second dimension are the same, there is no need to perform a dimensionality reduction operation and fusion can be performed directly.

[0041] Optionally, in the dimension reduction step, the deep semantic feature vector or the keyword feature vector is reduced in dimension according to the first dimension and the second dimension.

[0042] By adopting the above technical solution, based on the first dimension and the second dimension, the dimension of a feature vector with a high dimension in the deep semantic feature vector or the keyword feature vector is reduced, so that the deep semantic feature vector and the keyword feature vector are consistent in dimension, which provides convenience for the subsequent fusion steps. The unified dimension ensures that different feature vectors can match each other during the fusion process, thereby improving the fusion effect.

[0043] Optionally, after executing the second modeling step and before executing the second extraction step, the method further includes:

[0044] Pre-training: Based on the Wikipedia corpus, the BERT model is trained using the masked language model task and the next sentence prediction task;

[0045] Fine-tuning: Use the trained BERT model as a starting point for a specific NLP task and adjust the parameters of the trained BERT model.

[0046] By adopting the above technical solution, the present application can learn rich language knowledge, including semantic relations and grammatical structures of vocabulary, by pre-training on a large-scale, general corpus (such as Wikipedia). This extensive learning enables the BERT model to have the ability to deeply understand and analyze text, laying a solid foundation for subsequent performance on specific tasks. The masked language model task enables the BERT model to infer reasonable candidates for masked vocabulary based on contextual information, thereby enhancing the BERT model's ability to understand text semantics. The next sentence prediction task helps the BERT model learn the logical relationship and semantic coherence between sentences, which is of great significance for understanding the overall structure and context of the text. On the basis of pre-training, the BERT model is applied to specific NLP tasks through the fine-tuning process. This step usually involves further training the model using labeled task-related data, adjusting the parameters of the BERT model to better fit the task requirements, and fine-tuning can make full use of the knowledge learned by the pre-trained BERT model, while optimizing for specific tasks, thereby significantly improving the performance of the BERT model on the task. After the pre-training and fine-tuning steps, the BERT model can have better prediction capabilities.

[0047] Optionally, after executing the step of setting the transition probability and before executing the step of setting the reward, the step further includes:

[0048] Setting the state transition probability matrix: According to the transition probability, a state transition probability matrix is ​​established. The calculation model of the state transition probability matrix is ​​as follows:

[0049] ;

[0050] Among them, P is the state transition probability matrix, The probability of the current state transforming into the future state when taking different actions.

[0051] By adopting the above technical solution, the transition probability is organized into a matrix form (i.e., the state transition probability matrix P), which can represent the transition relationship between different states in a structured manner to facilitate mathematical and computational processing. Each element in the state transition probability matrix P represents the probability corresponding to taking different actions to transform into different future states in the current state in the second report, which can represent the dynamic changes of the system with specific values, help to more accurately evaluate the performance and risks of the system, and provide a scientific basis for decision-making.

[0052] Optionally, in the step of setting the reward, an evaluation index is also used to set the reward. The evaluation index is used to evaluate the severity of the future state. The calculation model of the evaluation index is as follows:

[0053] ;

[0054] in, is the value of the evaluation index of the i-th risk type label, and the calculation model is as follows:

[0055] ;

[0056] in, is the value in the state transition probability matrix when the current state of the i-th risk type label is transformed into the future state; is the weight of the kth evaluation indicator; k is the number of evaluation indicators.

[0057] By adopting the above technical solution, the evaluation indicators quantify the severity of the future state into specific values ​​through mathematical models, so that the system can more intuitively understand the actual impact of the future state under each risk type label or each evaluation indicator, and thus make more accurate judgments. When calculating the evaluation indicators, the weights of the evaluation indicators are introduced to reflect the importance of different evaluation indicators in the overall evaluation. By reasonably allocating weights, the system can more flexibly adjust the focus on different future states, making the reward setting more in line with actual needs.

[0058] Optionally, in the step of setting rewards, rewards are also set in the following manner:

[0059] ;

[0060] in, It is the magnification factor, which is used to control the magnification of each reward component.

[0061] By adopting the above technical solution, the present application can flexibly control the distribution ratio and magnitude of the reward by adjusting the amplification factor. The present application can achieve refined control of the reward by setting different amplification factors for different reward components.

[0062] Optionally, after executing the setting step and before executing the second acquisition step, the method further includes:

[0063] Second input: The first report is taken as the past state of the MDP model.

[0064] By adopting the above technical solution, the first report is used as the past state of the MDP model, which provides an initial decision-making environment or background for the MDP model. The MDP model can evaluate the possible results of different actions based on this initial state when making subsequent decisions, so as to make more reasonable decisions. In the MDP model, the transition between states is continuous, and each state evolves based on the previous state and the actions taken. Taking the first report as the past state helps to maintain the coherence of the MDP model during the state transition process. In this way, when processing subsequent data or reports, the MDP model can better understand the correlation and evolution trend between them, so as to make more accurate predictions and decisions.

[0065] In the second aspect, the present application provides an information reuse system based on deep learning, which adopts the following technical solutions:

[0066] An information reuse system based on deep learning, comprising:

[0067] Information collection module, used to collect historical HAZOP analysis reports and real-time HAZOP analysis reports;

[0068] An information extraction module, comprising a first information extraction unit and a second information extraction unit;

[0069] The first information extraction unit is in communication connection with the information acquisition module and is used to extract risk type labels and safety measures from the historical HAZOP analysis report;

[0070] The second information extraction unit is in communication connection with the first information extraction unit and the output unit, and is used to match the risk type label and the safety measures extracted by the first information extraction unit and the future state of the MDP model output by the output unit, query the historical HAZOP analysis report corresponding to the maximum conditional probability of the future state of the MDP model, and use the safety measures corresponding to the historical HAZOP analysis report as the safety measures of the real-time HAZOP analysis report;

[0071] Model building module: used to build the MDP model and set the state set, action set, transition probability and reward of the MDP model;

[0072] The input-output module includes an input unit and an output unit. The input unit is connected to the model building module for inputting the real-time HAZOP analysis report into the MDP model. The output unit is connected to the model building module for outputting the future state and conditional probability of the MDP model. The output unit is also used to output a second report and safety measures.

[0073] By adopting the above technical solutions, the time cost of manual reading and analysis is greatly reduced by automatically collecting and extracting key information (such as risk type labels and safety measures) from historical HAZOP analysis reports. The MDP (Markov Decision Process) model is used to predict future states, further accelerating the decision-making process. This system uses the MDP model to predict the conditional probability of future states, and combines safety measures in similar situations in historical reports to provide scientific and reasonable safety recommendations for real-time HAZOP analysis reports. This data-driven and model-based prediction method helps to improve the scientificity and accuracy of decision-making. By systematically storing and retrieving experience and knowledge from historical HAZOP analysis reports, the HAZOP analysis of new projects can make full use of past successes and lessons, avoid repeating mistakes, and promote the reuse of information in HAZOP analysis reports.

[0074] In summary, the present application includes at least one of the following beneficial technical effects:

[0075] 1. This application stores the experience and knowledge in the historical HAZOP analysis report in a structured manner in the MDP model and uses it to guide future decision-making processes, and assigns the safety measures of the first report to the second report for reference by staff, thereby achieving the reuse of information in the HAZOP analysis report. This application optimizes the safety measures selection process in the HAZOP analysis report by introducing the MDP model, which can significantly improve the pertinence and effectiveness of safety measures, optimize the decision-making process, increase response speed, enhance knowledge inheritance and accumulation, improve resource utilization efficiency, and enhance the transparency and traceability of decision-making.

[0076] 2. This application introduces a pre-judgment step (first judgment), which can intelligently choose whether to directly use the existing first report to perform the first extraction step, and in the data processing step, the deep learning algorithm (BERT model) and TF-IDF algorithm are used to extract the deep semantic features and keyword features of the first report respectively, and through the judgment, fusion or dimensionality reduction steps, these features are effectively integrated, and finally the fused features are converted into the predicted probability distribution of the risk type through the Softmax activation function, and the risk type label of the report is accurately obtained based on the preset threshold. This application realizes the label extraction of the first report without risk type labels through a series of operations, so that in the first extraction step, all first reports have risk type labels. On the one hand, the number of key elements in the MDP model is increased, and on the other hand, some risk type labels of HAZOP analysis reports are supplemented.

[0077] 3. This application uses the first report as the past state of the MDP model, providing an initial decision-making environment or background for the MDP model. The MDP model can evaluate the possible results of different actions based on this initial state when making subsequent decisions, thereby making more reasonable decisions. In the MDP model, the transition between states is continuous, and each state evolves based on the previous state and the actions taken. Using the first report as the past state helps maintain the coherence of the MDP model during the state transition process. In this way, when processing subsequent data or reports, the MDP model can better understand the correlation and evolution trend between them, thereby making more accurate predictions and decisions.

[0078] 4. By automatically collecting and extracting key information from historical HAZOP analysis reports (such as risk type labels and safety measures), the time cost of manual reading and analysis is greatly reduced. The MDP (Markov Decision Process) model is used to predict future states, further accelerating the decision-making process. This system uses the MDP model to predict the conditional probability of future states, and combines safety measures in similar situations in historical reports to provide scientific and reasonable safety recommendations for real-time HAZOP analysis reports. This data-driven and model-based prediction method helps to improve the scientificity and accuracy of decision-making. By systematically storing and retrieving experience and knowledge from historical HAZOP analysis reports, the real-time collected HAZOP analysis can make full use of past successes and lessons, avoid repeating mistakes, and promote the reuse of information in HAZOP analysis reports. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 is a flow chart of Example 1 of the present application;

[0080] Figure 2 This is a flowchart of adding a label in step S21 of Example 1 of the present application;

[0081] Figure 3 This is a flowchart of information reuse in step S5 of Example 1 of the present application;

[0082] Figure 4 It is a structural diagram of Example 2 of the present application. DETAILED DESCRIPTION

[0083] The following combination Figures 1 to 4 This application is described in further detail.

[0084] Embodiment 1: This embodiment discloses an information reuse method based on deep learning, referring to Figure 1 The method includes: step S1: first collection, step S2: extracting labels, step S3: modeling, step S4: second input, step S5: information reuse, firstly collect historical HAZOP analysis reports, then extract risk type labels and safety measures of historical HAZOP analysis reports, then establish an MDP model and set key elements of the model, then use the historical HAZOP analysis reports as the past state of the MDP model, then use the real-time HAZOP analysis report as the present state of the MDP model to obtain the future state of the MDP model, obtain the first report that is most similar to the second report, and then output the safety measures taken in the first report to the staff. The specific steps of this embodiment are as follows:

[0085] Step S1: First collection, collecting historical HAZOP analysis reports, and recording the historical HAZOP analysis reports as the first reports.

[0086] In step S1, the method of collecting the historical HAZOP analysis report includes:

[0087] Internal data sources: Sort out historical HAZOP analysis reports stored by various departments within the enterprise (such as safety, production, engineering, etc.).

[0088] External data source: Obtain publicly available HAZOP analysis reports from outside the enterprise.

[0089] After collecting the historical HAZOP analysis reports, the collected HAZOP analysis reports are converted into a unified electronic format, and the text content in the HAZOP analysis reports is extracted in the electronic format. The text content is then converted into a text sequence format. When converting to a text sequence, the text content is preprocessed by word segmentation, stop word removal, stem extraction or word form, which improves the efficiency and accuracy of subsequent processing.

[0090] The text sequences are stored in a structured manner in a database and indexed and metadata are created for them to quickly retrieve and locate relevant information for subsequent query, analysis, and application.

[0091] Step S2: extracting tags, including: step S21: adding tags and step S22: first extraction.

[0092] like Figure 2 As shown, step S21: adding a label includes: step S211: first judgment, step S212: data processing, step S213: second output and step S214: obtaining a label.

[0093] Step S211: first judgment, judging whether the first report has a risk type label, if so, executing step S2123: extracting vectors; if not, executing step S212: data processing.

[0094] Step S212: data processing, including: step S2121: second modeling, step S2122: training, step S2123: vector extraction, step S2124: dimension extraction, step S2125: second judgment, step S2126: fusion and step S2127: dimensionality reduction.

[0095] Step S2121: Second modeling, establishing a BERT (Bidirectional Encoder Representations from Transformers) model based on a deep learning algorithm. This embodiment uses a BERT-base-chinese model optimized for Chinese context (referred to as a BERT model in subsequent steps). This model has been trained with extensive Chinese corpus and can deeply understand the semantic subtlety and contextual changes of Chinese.

[0096] Step S2122: training, including: step S21221: pre-training and step S21222: fine-tuning.

[0097] Step S21221: Pre-training, based on the Wikipedia corpus, using the masked language model task and the next sentence prediction task to train the BERT model.

[0098] The masked language model task randomly masks (i.e. replaces with a special token [MASK]) some words in the input text, and then requires the BERT model to predict the original values ​​of these masked words based on the context. In this way, the BERT model can learn the dependencies and semantic information between words.

[0099] The BERT model also uses the next sentence prediction (NSP) task to enhance the BERT model's ability to understand the relationship between sentences. The BERT model is trained to predict whether two sentences actually appear consecutively in the original text. This task helps the BERT model learn sentence-level representations.

[0100] Step S21222: fine-tuning, using the trained BERT model as the starting point for a specific NLP task and adjusting the parameters of the trained BERT model.

[0101] In the fine-tuning phase, the pre-trained BERT model is used as a starting point to customize the model according to the needs of the specific task. During the fine-tuning process, the BERT model is trained using labeled data related to the specific task. Through the back-propagation algorithm, the parameters in the BERT model (including the parameters learned in the pre-training phase and the parameters in the newly added task-specific layers) are updated to minimize the loss function on the training set. The BERT model can gradually learn how to make accurate predictions or generation for specific tasks.

[0102] Step S2123: extracting vectors, including: step S21231: second extraction and step S21232: third extraction.

[0103] Step S21231: the second extraction, using the BERT model fine-tuned in step S21222: fine-tuning, extracts the deep semantic features of the first report and obtains a deep semantic feature vector.

[0104] The text sequence in step S1: first collection is segmented and converted into multiple .

[0105] Multiple are input into the BERT model respectively. The BERT model performs Generate an n-dimensional feature vector and finally output the deep semantic feature vector V, where .

[0106] ;

[0107] ;

[0108] Indicates that The feature vector converted by the BERT model, Indicates the i-th , each feature vector has a dimension of n, then the deep semantic feature vector V is a dimensional vector, where Is the text sequence The number of

[0109] Step S21232: The third extraction uses the TF-IDF algorithm to extract keyword features in the first report and obtain a keyword feature vector.

[0110] Based on the text sequence in step S1: the first collection, the TF-IDF value of each keyword in the text sequence is calculated, and the calculation model is as follows:

[0111] ;

[0112] Among them, E is the number of times the keyword appears in the text sequence, G is the total number of words in the document, I is the total number of documents in the text sequence, and K is the number of documents containing the keyword.

[0113] After calculating the TF-IDF value of each keyword, select the top n keywords with the highest TF-IDF values, and use the TF-IDF value of each keyword as a feature to generate multiple feature vectors. These feature vectors are input into a neural network composed of multiple fully connected layers to further mine and optimize the feature expression to obtain a keyword feature vector. The dimension of the keyword feature vector is n, and each dimension corresponds to the TF-IDF value of a keyword.

[0114] In this embodiment, the neural network composed of multiple fully connected layers is specifically:

[0115] Initially, a fully connected layer containing 64 neurons takes over to process multiple feature vectors , using the ReLU activation function to introduce nonlinearity, thereby enhancing the neural network model's ability to capture features. The calculation model of the first fully connected layer is as follows:

[0116] ;

[0117] in, and They are the weight and bias of the first fully connected layer. The weight of the first fully connected layer is initialized by Xavier initialization, and the bias is initialized to 0 or a value close to 0. The output of this layer is is the activation function.

[0118] Next, As input, it moves to a fully connected layer with 128 neurons to further refine the information. The calculation model of the second fully connected layer is as follows:

[0119] ;

[0120] in, and are the weights and biases of the second fully connected layer, which are the same as the first fully connected layer. The weights of the second fully connected layer are initialized by Xavier initialization, and the bias of the second fully connected layer is initialized to 0 or a value close to 0. is the output of this layer.

[0121] Finally, As input, it passes through a 256-neuron fully connected layer to achieve the purpose of highly abstracting the key information of the text. The calculation model of the third fully connected layer is as follows:

[0122] ;

[0123] in, and They are the weight and bias of the third fully connected layer. Similarly, the initialization weight of the third fully connected layer is initialized by Xavier initialization, and the bias of the third fully connected layer is initialized to 0 or a value close to 0. is the output of this layer, representing the keyword feature vector.

[0124] Step S2124: extracting dimensions, including: step S21241: fourth extraction and step S21242: fifth extraction.

[0125] Step S21241: Fourth extraction, extracting the dimension of the deep semantic feature vector , recorded as the first dimension.

[0126] Step S21242: The fifth extraction is to extract the dimension n of the keyword feature vector, which is recorded as the second dimension.

[0127] Step S2125: second judgment, based on step S2124: extracting the first dimension and the second dimension of the dimension, judge whether the deep semantic feature vector and the keyword feature vector can be spliced. If the first dimension is the same as the second dimension, execute step S2126: fusion; if the first dimension and the second dimension are different, execute step S2127: dimensionality reduction.

[0128] Step S2126: Fusion: concatenate the deep semantic feature vector and the keyword feature vector at the beginning and the end to obtain a first fused vector.

[0129] Step S2127: Dimensionality reduction: according to the first dimension and the second dimension, the deep semantic feature vector or the keyword feature vector with a higher dimension is reduced in dimension, the dimensions of the deep semantic feature vector and the keyword feature vector are unified, and a fusion step is performed.

[0130] For example, if the deep semantic feature vector has a higher dimension, the deep semantic feature vector is pooled to reduce its dimension so that the deep semantic feature vector is the same as the keyword feature vector and then concatenated with the keyword feature vector.

[0131] definition is the deep semantic feature vector after pooling, whose dimension is n, then the fused feature vector It can be expressed as:

[0132] ;

[0133] in, Represents a concatenation operation, which concatenates two vectors end-to-end to obtain a first fused vector.

[0134] Step S213: Second output, using the Softmax activation function to convert the first fusion vector into a probability distribution, and outputting the predicted probability distribution of each risk type in the first report.

[0135] The first fusion vector is used as the input of the Softmax function. After being processed by the Softmax function, each element of the first fusion vector is converted into the predicted probability of the corresponding risk type. The sum of all predicted probabilities is 1, and the predicted probability distribution is used as the probability distribution of the first reported risk type.

[0136] Step S214: Obtain a label, and use the risk type with a predicted probability distribution greater than a preset threshold as the risk type label of the first report.

[0137] Iterate over each element in the predicted probability distribution (i.e., the predicted probability of each risk type), and use the risk types corresponding to those elements that are greater than a preset threshold as the risk type labels for the first report.

[0138] Step S22: First extraction, based on the historical HAZOP analysis report collected in step S1: First collection, and the predicted risk type label in step S214: Acquisition label, extract the risk type label of the first report, and extract the safety measures taken in the first report, and store the risk type label and safety measures of the first report in the database accordingly.

[0139] Step S3: Modeling, including: Step S31: First Modeling and Step S32: Setting.

[0140] Step S31: First modeling, establishing an MDP model.

[0141] Step S32: Setting, including: Step S321: Setting the state set, Step S322: Setting the action set, Step S323: Setting the transition probability, Step S324: Setting the state transition probability matrix and Step S325: Setting the reward.

[0142] Step S321: setting a state set, taking the set consisting of all risk type labels in the first report as the state set of the MDP model, and each risk type label corresponds to a state.

[0143] In this embodiment, the risk type label includes: process variable class labels such as temperature, pressure, liquid level, material level, flow rate, pH, voltage, current, viscosity, frequency, speed, reaction, etc., guide word class labels such as no (no), more (large, high), less (small, low), excess (and), part (local), opposite (reverse) and other (abnormal), as well as cause event class, consequence event class, intermediate event class labels, etc., then the state set in the MDP model includes: process variable class states such as temperature, pressure, liquid level, material level, flow rate, pH, voltage, current, viscosity, frequency, speed, reaction, etc., guide word class states such as no (no), more (large, high), less (small, low), excess (and), part (local), opposite (reverse) and other (abnormal), as well as cause event class, consequence event class, intermediate event class labels, etc.

[0144] Part of the state space of the MDP model is shown in Table 1.

[0145] Table 1 Partial state space of MDP model

[0146]

[0147] In Table 1, the numerical indicators 0, 1, and -1 represent the status of the state, 0 indicates that the state exists, 1 indicates that the state does not exist, and -1 indicates that the state cannot be determined. The numerical indicator values ​​of other risk type labels are the same as the status shown in Table 1.

[0148] Step S322: Set an action set, and take the set of security measures taken by all risk type labels as the action set of the MDP model. Part of the action space of this embodiment is shown in Table 2.

[0149] Table 2 Partial action space of MDP model

[0150]

[0151] The actions taken in Table 2 are security measures taken for the state, and all combinations of security measures constitute the action set of the MDP model.

[0152] Step S323: setting the transition probability, the calculation model of the transition probability is as follows:

[0153] ;

[0154] Among them, p is the transition probability; is the frequency of occurrence of the mth state; is the severity of the m-th state, and the frequency and severity are obtained by statistically analyzing all the first reports to obtain their specific values.

[0155] For example, the current state is: liquid level / high / drainage pipe blocked / overflow. After executing a certain action, its state is transformed into the future state: liquid level / normal. The numerical indicator is transformed from 0 to 1 accordingly, and the transition probability calculation formula of this state is obtained using the above formula.

[0156] Step S324: Setting the state transition probability matrix: According to the transition probability, a state transition probability matrix is ​​established. The calculation model of the state transition probability matrix is ​​as follows:

[0157] ;

[0158] Among them, P is the state transition probability matrix, The probability that the current state will transform into different future states when different actions are taken.

[0159] For the same state, different actions taken have different transition probabilities. The different transition probabilities corresponding to the same state are combined into one row. When the number of transition probabilities between different states is different, resulting in inconsistent number of columns in the state transition probability matrix, 0 is used to fill the vacant positions. The meaning of 0 here is different from the numerical indicators in Table 1. 0 is used here to fill the gaps so as not to interfere with the calculation process in subsequent calculations.

[0160] For example, the first row represents the current state: temperature / high / cooling module failure / equipment damage, and the action is to shut down the equipment. The corresponding transition probability is ; The action taken is to debug the device, and the corresponding transition probability is ; The action taken is to replace the equipment, and the corresponding transition probability is If the number of columns in other rows is different from that in the first row, and the number of columns in other rows is greater than that in the first row, then let =0 until the row is filled.

[0161] Step S325: Set rewards. Rewards are used to express the quality of the MDP model decision and to inform the staff of the quality of the future state after taking the current action. The value of is 0, which means that the risk label is removed and the reward value is high; conversely, the risk label exists and the reward value is low.

[0162] The reward of the MDP model is set based on the risk type label of the first report, and the calculation model of the reward is as follows:

[0163] ;

[0164] in, yes The reward of the MDP model at this moment, is the weight of the ith risk type label among all risk type labels, which can be obtained by statistically analyzing the first report; is the proportion of the jth safety measure in all safety measures included in the current risk type label, obtained by statistical analysis; is the reward compensation amount, which takes a value of 1; a is the number of risk type labels in the state set, which can be obtained through statistical analysis; b is the number of safety measures in the current risk type label, which can be obtained through statistical analysis; is the implementation of the jth safety measure in the i-th risk label, and the calculation model is as follows:

[0165] .

[0166] When the number of first reports is large enough, the relevant knowledge of statistical analysis shows that the data obtained by statistical analysis can represent its actual value.

[0167] In another embodiment, in the step of setting the reward, an evaluation index is also used to set the reward. The evaluation index is used to evaluate the severity of the future state. The calculation model of the evaluation index is as follows:

[0168] ;

[0169] in, is the value of the evaluation index of the i-th risk type label, and the calculation model is as follows:

[0170] ;

[0171] in, is the value in the state transition probability matrix when the current state of the i-th risk type label is transformed into the future state, that is, ; is the weight of the kth evaluation indicator, which is a preset value; k is the number of evaluation indicators.

[0172] Table 3 Evaluation index weight value table

[0173]

[0174] In the first report, if the future state becomes worse after taking an action, the probability of its subsequent adoption will be smaller, and the weight of its evaluation indicator will be It is also very small, that is to say, will be very small, or even not take the action at all, i.e. When the future state is worse than the current state, The value of will be small, and the weight of this action in the first report will be small. The value of will be small, otherwise, the reward The calculated result is also too small.

[0175] In other embodiments, in the step of setting the reward, the reward is also set in the following manner:

[0176] ;

[0177] in, It is the magnification factor, which is used to control the magnification of each reward component. The staff sets it according to the actual situation. In this embodiment, the values ​​of the two are: 4, 1 respectively.

[0178] Step S4: The second input is to use the first report as the past state of the MDP model. With the "past state" as the starting point, the future states that may be caused by different actions can be evaluated based on the transition probability, reward function and other factors in the MDP model, so as to make the best decision.

[0179] like Figure 3 As shown, step S5: information multiplexing includes: step S51: second collection, step S52: first input, step S53: query and step S54: first output.

[0180] Step S51: Second collection, real-time collection of the HAZOP analysis report, recorded as the second report. Similar to step S1, the second report is converted into a text sequence.

[0181] Step S52: First input, take the text sequence of the second report as the current state of the MDP model, and obtain the conditional probability of the future state of the MDP model. Input the current state of the second report into the MDP model, and obtain the state similar to or the same as the current state of the second report and the conditional probability of the state in the state set of the MDP model.

[0182] Step S53: Query the risk type label and safety measures of the stored first report. The actions (ie, safety measures) of the first report similar to the second report (ie, the risk type label is the same or similar) can provide valuable reference for the response to the current risk (second report).

[0183] The first report corresponding to the maximum conditional probability is obtained and recorded as the third report, and the safety measures corresponding to the third report are used as the safety measures of the second report.

[0184] Step S54: First output, outputting the second report and the safety measures of the second report to the staff so that they can take corresponding actions based on the information to reduce risks or improve the safety of the system.

[0185] This application achieves the reuse of information in the HAZOP analysis report by storing the experience and knowledge in the historical HAZOP analysis report in the MDP model in a structured manner and using it to guide future decision-making processes. This application optimizes the safety measure selection process in the HAZOP analysis report by introducing the MDP model, which can significantly improve the pertinence and effectiveness of safety measures and optimize the decision-making process.

[0186] Embodiment 2: This embodiment discloses an information reuse system based on deep learning, referring to Figure 4 , the system comprising:

[0187] The information collection module is used to collect historical HAZOP analysis reports and real-time HAZOP analysis reports.

[0188] The information extraction module includes a first information extraction unit and a second information extraction unit.

[0189] The first information extraction unit is in communication with the information acquisition module and is used to extract risk type labels and safety measures from the historical HAZOP analysis report, and transmit the extracted risk type labels and safety measures to the second information extraction unit and store them for subsequent use.

[0190] The second information extraction unit is communicatively connected with the first information extraction unit and the output unit, and is used to match the risk type label and safety measures extracted by the first information extraction unit and the future state of the MDP model output by the output unit, query the historical HAZOP analysis report corresponding to the maximum conditional probability of the future state of the MDP model, and use the safety measures corresponding to the historical HAZOP analysis report as the safety measures of the real-time HAZOP analysis report.

[0191] Model building module: used to build the MDP model and set the state set, action set, transition probability and reward of the MDP model, the MDP model itself and its key elements, which are used to predict future states and calculate conditional probabilities under given states.

[0192] The input-output module includes an input unit and an output unit. The input unit is connected to the model building module for inputting the real-time HAZOP analysis report into the MDP model as the starting point of the MDP model analysis. The output unit is connected to the model building module for outputting the conditional probability of the future state of the MDP model, passing the conditional probability to the second information extraction unit for matching and recommendation, or directly displaying it to the user. The output unit is also used to output the second report and safety measures.

[0193] The information acquisition module of this system obtains historical HAZOP analysis reports and real-time HAZOP analysis reports. The first information extraction unit extracts risk type labels and safety measures from historical reports. The model building module uses historical data to build an MDP model. The input unit inputs real-time reports into the MDP model. The MDP model calculates the conditional probabilities of future states and transmits these probabilities through the output unit. The second information extraction unit matches historical reports based on conditional probabilities and recommends safety measures. The final safety measure recommendations are displayed to users through the output unit or used in other decision support systems.

[0194] The above are all preferred embodiments of the present application, and the protection scope of the present application is not limited thereto. Therefore, any equivalent changes made according to the structure, shape, and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for information reuse based on deep learning, characterized in that: include: First collection: Collect historical HAZOP analysis reports and record them as the first report; First judgment: judging whether the first report has a risk type label, if so, executing the first extraction step; if not, executing the data processing step; Data processing: including the second modeling, second extraction, third extraction, second judgment, fusion and dimensionality reduction steps; Second modeling: Building a BERT model based on deep learning algorithms; Second extraction: Based on the BERT model, the deep semantic features of the first report are extracted to obtain a deep semantic feature vector; Third extraction: using the TF-IDF algorithm to extract keyword features in the first report and obtain a keyword feature vector; Second judgment: judge whether the deep semantic feature vector and the keyword feature vector can be spliced. If yes, perform the fusion step; if not, perform the dimension reduction step; Fusion: concatenate the deep semantic feature vector and the keyword feature vector to obtain the first fusion vector; Dimensionality reduction: reduce the dimension of the deep semantic feature vector or keyword feature vector, unify the dimensions of the deep semantic feature vector and the keyword feature vector, and perform the fusion step; Second output: Use the Softmax activation function to convert the first fusion vector into a probability distribution, and output the predicted probability distribution of each risk type in the first report; Obtaining labels: taking the risk type with a predicted probability distribution greater than a preset threshold as the risk type label for the first report; First extraction: extract and store the risk type label and safety measures of the first report; First modeling: establish the MDP model; Setting: including setting the state set, setting the action set, setting the transition probability and setting the reward step; Set state set: The set of all risk type labels is used as the state set of the MDP model; Set action set: The set of security measures taken by all risk type labels is used as the action set of the MDP model; Set the transition probability: The calculation model of the transition probability is as follows: ; Among them, p is the transition probability; is the frequency of occurrence of the mth state; is the severity of the mth state; Set rewards: Set the rewards of the MDP model based on the risk type label. The calculation model is as follows: ; in, yes The reward of the MDP model at this moment, is the weight of the i-th risk type label among all risk type labels; is the proportion of the jth security measure among all security measures included in the current risk type label; is the reward compensation amount; a is the number of risk type labels in the state set; b is the number of safety measures in the current risk type label; is the implementation of the jth safety measure in the i-th risk type label, and the calculation model is as follows: ; Second collection: real-time collection of HAZOP analysis reports, recorded as the second report; First input: take the second report as the current state of the MDP model and obtain the conditional probability of the future state of the MDP model; Query: Based on the risk type label and safety measures of the stored first report, query the first report corresponding to the maximum condition probability, record it as the third report, and use the safety measures corresponding to the third report as the safety measures of the second report; First output: output the second report and the safety measures for the second report.

2. The information reuse method based on deep learning according to claim 1, characterized in that: After executing the vector extraction step and before executing the second determination step, the method further includes: Fourth extraction: extract the dimension of the deep semantic feature vector, recorded as the first dimension; Fifth extraction: extract the dimension of the keyword feature vector, recorded as the second dimension.

3. The information reuse method based on deep learning according to claim 2, characterized in that: In the dimension reduction step, the deep semantic feature vector or the keyword feature vector is reduced in dimension according to the first dimension and the second dimension.

4. The information reuse method based on deep learning according to claim 3, characterized in that: After executing the second modeling step and before executing the second extraction step, the method further includes: Pre-training: Based on the Wikipedia corpus, the BERT model is trained using the masked language model task and the next sentence prediction task; Fine-tuning: Use the trained BERT model as a starting point for a specific NLP task and adjust the parameters of the trained BERT model.

5. The information reuse method based on deep learning according to any one of claims 1 to 4, characterized in that: After executing the step of setting the transition probability and before executing the step of setting the reward, it also includes: Setting the state transition probability matrix: According to the transition probability, a state transition probability matrix is ​​established. The calculation model of the state transition probability matrix is ​​as follows: ; Among them, P is the state transition probability matrix, The probability of the current state transforming into the future state when taking different actions.

6. The information reuse method based on deep learning according to claim 5, characterized in that: In the reward setting step, an evaluation index is also used to set the reward. The evaluation index is used to evaluate the severity of the future state. The calculation model of the reward is as follows: ; in, is the value of the evaluation index of the i-th risk type label, and the calculation model is as follows: ; in, is the value in the state transition probability matrix when the current state of the i-th risk type label is transformed into the future state; is the weight of the kth evaluation indicator; d is the number of evaluation indicators.

7. The information reuse method based on deep learning according to claim 6, characterized in that: In the reward setting step, the following method is also used: ; in, It is the magnification factor, which is used to control the magnification of each reward component.

8. The information reuse method based on deep learning according to any one of claims 1 to 4, characterized in that: After executing the setting step and before executing the second acquisition step, the method further includes: Second input: The first report is taken as the past state of the MDP model.

9. An information reuse system based on deep learning, the system being applicable to the method according to any one of claims 1 to 8, characterized in that: include: Information collection module, used to collect historical HAZOP analysis reports and real-time HAZOP analysis reports; An information extraction module comprises a first information extraction unit and a second information extraction unit; The first information extraction unit is in communication with the information acquisition module and is used to extract risk type labels and safety measures from the historical HAZOP analysis report; The second information extraction unit is in communication connection with the first information extraction unit and the output unit, and is used to match the risk type label extracted by the first information extraction unit with the safety measures and the future state of the MDP model output by the output unit, query the historical HAZOP analysis report corresponding to the maximum conditional probability of the future state of the MDP model, and use the safety measures corresponding to the historical HAZOP analysis report as the safety measures of the real-time HAZOP analysis report; Model building module: used to build the MDP model and set the state set, action set, transition probability and reward of the MDP model; The input-output module includes an input unit and an output unit. The input unit is connected to the model building module for inputting the real-time HAZOP analysis report into the MDP model. The output unit is connected to the model building module for outputting the future state and conditional probability of the MDP model. The output unit is also used to output a second report and safety measures.