Production safety risk grading and hidden danger troubleshooting double-control system and control method thereof

Abstract

The invention discloses a production safety risk grading and hidden danger troubleshooting double-control system and a control method thereof, which realize dynamic grading of risk points and continuous updating of a knowledge graph by utilizing a risk dynamic grading method based on a small sample knowledge graph and combining a natural language processing technology. The system comprises a risk management and control subsystem and a hidden danger investigation subsystem, the risk management and control subsystem is responsible for risk information extraction, importing, grading and management and control measure making, and the hidden danger investigation subsystem is responsible for executing inspection and investigation of management and control measures, including hidden danger monitoring, investigation plan making, hidden danger registration, hidden danger rectification and hidden danger statistical analysis, so that the influence of human subjective factors is reduced; according to the method, objective risk assessment is provided through modeling based on a large amount of data and a knowledge graph, and the risk level is dynamically updated in combination with a hidden danger checking result, so that risk management and closed-loop control with higher timeliness are realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of production safety, and in particular relates to a dual control system and method for production safety risk classification and hidden danger investigation. Background Art

[0002] Double control prevention refers to the dual prevention of production safety risk control and hidden danger investigation and treatment. Given the importance of production safety to the livelihood of thousands of households, building a scientific and systematic dual control prevention system is of great significance for enterprises to comprehensively and effectively control risks at the source and nip potential accidents in the bud. At present, the safety management measures of many enterprises still rely on traditional manual operations and expert experience, which leads to a series of problems: First, the existing risk assessment methods are often based on expert opinions and manual evaluation, which makes the assessment results affected by personal experience and preferences, lacking objectivity and consistency. Manual risk assessment is not only time-consuming and laborious, but also difficult to ensure the accuracy and timeliness of the assessment, especially in the face of large-scale or complex and changing production environments. Secondly, a large amount of production safety data is stored in semi-structured or unstructured form, and the potential value of this data has not been fully explored and utilized. Traditional data processing methods often cannot effectively extract key information from these data, thereby limiting the quality and efficiency of risk management decisions. Due to the lack of real-time monitoring and dynamic adjustment mechanisms, the existing risk assessment system is difficult to respond quickly to emergencies or changing environmental conditions. Summary of the invention

[0003] The purpose of the embodiment of the present invention is to provide a dual control system for production safety risk classification and hidden danger investigation, so as to solve the problems of objectivity, inconsistency and inefficiency caused by reliance on manual work and experience, and to achieve more objective, accurate, effective and real-time risk classification management and control;

[0004] The second purpose of an embodiment of the present invention is to provide a production safety risk classification and hidden danger detection dual control system control method.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is a dual control system for production safety risk classification and hidden danger investigation, including a risk control subsystem and a hidden danger investigation subsystem;

[0006] The risk management subsystem includes:

[0007] Risk information extraction and import module, used to extract risk-related data and import it into the risk database;

[0008] The risk classification knowledge graph module is used to extract data fields from the risk database to construct a risk classification knowledge graph, and then maintain and update the risk classification knowledge graph; the risk classification knowledge graph is used to identify risks and classify risks; including risk locations, risk causes, risk status, risk costs, number of hidden danger occurrences, and number of hidden danger inspections;

[0009] The risk control measures module is used to create corresponding control measures according to different risk levels;

[0010] Risk statistics module, used for automatic statistical analysis of risks;

[0011] The hidden danger investigation subsystem is used to inspect the implementation of risk control measures, including:

[0012] Hidden danger monitoring module, used to monitor the classified risk locations in the risk classification knowledge graph in real time and record them in the risk database;

[0013] The hidden danger investigation plan module is used to formulate an investigation plan for each risk point based on the residual risk after the implementation of control measures;

[0014] The hidden danger investigation and registration module is used to issue investigation tasks according to the formulated hidden danger investigation plan and receive hidden danger registration information;

[0015] Hidden danger rectification module, used to formulate rectification plans for registered hidden dangers;

[0016] The hidden danger statistics module is used to conduct statistical analysis of hidden dangers according to different risk points and departments.

[0017] Furthermore, it includes a risk control subsystem control method and a hidden danger investigation subsystem control method, and the risk control subsystem control method is specifically as follows:

[0018] S1. The risk information extraction and import module extracts risk-related data from production safety documents and organizes them into structured data sets, and then imports the structured data sets into the risk database; risk-related data includes structured data and unstructured data, where the unstructured data is extracted from the risk information extraction subsystem after small sample training, and then mapped with the risk knowledge base fields and stored in the risk database;

[0019] S2. The risk classification knowledge graph module constructs a risk classification knowledge graph using the risk location, risk cause, risk status, risk cost, number of hidden danger occurrences, number of hidden danger inspections, and number of hidden danger rectifications in the risk database, and completes and updates the risk classification knowledge graph; the number of hidden danger registrations, number of hidden danger inspections, and number of hidden danger rectifications are generated by the hidden danger inspection subsystem;

[0020] S3. Use the risk classification knowledge graph and automatically classify risks based on the risk database to create corresponding risk classifications;

[0021] S4. After risk classification, create corresponding control measures in the risk control measures module according to different risk levels, and establish a control measures library;

[0022] S5. Finally, the risk statistics module performs risk statistics; specifically, the number of risks in different regions and departments is counted.

[0023] Furthermore, the risk database includes risk analysis objects and risk points; the risk classification divides the risk points in the risk database into low risk, general risk, greater risk and major risk levels according to the severity of the risk.

[0024] Furthermore, the risk information extraction subsystem consists of three components: small sample relationship extraction, relationship-based small sample entity recognition, and small sample global entity pairing.

[0025] Furthermore, the risk information extraction subsystem implements small sample training through the natural language pre-training model. The specific process is as follows:

[0026] S1. Randomly extract a part of the manually collected and annotated historical risk data as a data set, which is divided into training data set D train , test data set D test , and then divide the two datasets into support dataset and query dataset respectively:

[0027]

[0028] in, D train Supporting datasets, D train The query dataset D test Supporting datasets, D test The query dataset;

[0029] S2. In the small sample relation extraction task, relation label knowledge is introduced to optimize the relation prototype representation:

[0030]

[0031] is the relation head entity of the k-th instance sentence of relation i, is the relation tail entity of the k-th instance sentence of relation i, For the kth instance sentence of the relation, the f_span function represents obtaining the entity representation from the sentence;

[0032]

[0033] Among them, entity types include equipment, operators, and security events;

[0034] represents the overall encoding representation of the sentence, Relation representation that contains information about the sentence, head entity, and tail entity;

[0035]

[0036] represents the prototype representation of the relation, K represents the total number of sentences contained in each relation;

[0037] p r,i =Similarity(h i q ,C r )

[0038] p r,i is the similarity between the query instance and the relation prototype, is an instance of query set relation i, C r is the relation prototype of the support set relation i;

[0039]

[0040] L rel is the cross entropy loss function for small sample relationship extraction, N is the total number of relationship types in the training data set, and y r,i is the relation label in relation i, p r,i is the similarity between the query instance and the relation prototype;

[0041] On the relation-based few-shot entity recognition task, a span-based entity-level prototype was constructed:

[0042]

[0043] E is the span representation of the sentence, l is the number of words in the instance, e is the embedding prototype span of each word, h is the tag encoding of the word embedding, and h n represents the tag encoding of the nth word embedding, h n-span_size is the tag encoding of the n-span_size-th word embedding, span_size is the maximum length of a single span, and f is the maximum pooling function that combines the embeddings of the spans;

[0044] S3. Introduce hierarchical contrastive learning to perform hierarchical contrastive learning on different span combinations and jointly refine the metric space for relation extraction and entity recognition.

[0045] Furthermore, the hierarchical contrastive learning includes contrastive learning between relational prototypes, contrastive learning between entities, and contrastive learning between entities and relations:

[0046]

[0047] L conRel is the relation prototype comparison constraint loss, m1 is the minimum distance margin of the same type of data, m2 is the maximum distance margin of heterogeneous data, N is the total number of relation types in the training data set, represents the prototype of relation i, represents the prototype of relation j;

[0048]

[0049] L conEntity is the entity prototype contrast constraint loss, E is the total number of entities in the training dataset, represents the prototype of the i-th entity under relation k, Represents the prototype of the j-th entity under relation k;

[0050]

[0051] Furthermore, the risk classification knowledge graph will be supplemented and updated as the frequency of hidden dangers, the execution of risk control measures and other factors change. The supplementation and updating of the risk classification knowledge graph is implemented by training the supplementation and updating model of the risk classification knowledge graph using a bionic prototype memory attention network. When the risk location, risk cause, risk status, risk cost, number of hidden danger occurrences and number of hidden danger inspections contained in the risk classification knowledge graph change, the supplementation and updating model of the risk classification knowledge graph is trained to automatically update and supplement the risk classification knowledge graph. The model training method is specifically as follows:

[0052] S1. Establish a memory bank M consisting of short-term memory, working memory and long-term memory. (i) ,M (w) ,M (l)}; M is the memory bank, M (i) is short-term memory, M (w) is the working memory, M (l) It is long-term memory;

[0053] S2. Introducing the bionic memory update mechanism:

[0054]

[0055] Among them, g(·) is the similarity function, It is M (i)The nth candidate prototype feature in; exp is the natural exponential function, μ n is the nth candidate average prototype feature, is the characteristic variance, M n (i) are n prototype features stored in short-term memory;

[0056] S3, using memory-enhanced prototype smooth triplet loss to constrain the similarity between new and old memory knowledge:

[0057] L=L m +λL P

[0058] Among them, L is the total loss of the model, L m is the margin-based ranking loss, L P is the prototype smooth triplet loss, λ is a hyperparameter, where the margin-based ranking loss L m :

[0059]

[0060] Among them, Q r is the positive query dataset for model construction, Q r- is the negative query dataset for model construction, S r is the supporting dataset of the model, φ(q r ,S r ) represent the positive and negative query datasets and the support dataset S respectively. r The similarity score between them, γ is the marginal hyperparameter, is a query unit in the negative query dataset, q r is a query unit in the query data set, and r is the number of relations;

[0061] Where L P Is the prototype smooth triple

[0062]

[0063] represents the farthest embedding of similar data of prototype z, Represents the nearest embedding of heterogeneous data for prototype z.

[0064] Furthermore, the control method of the hidden danger detection subsystem is specifically as follows:

[0065] S1. The hidden danger monitoring module monitors and controls hidden dangers in real time at risk locations or equipment according to risk classification levels;

[0066] S2. When the hidden danger monitoring module finds hidden dangers, the hidden danger investigation subsystem will promptly issue an alarm to the safety administrator and automatically register the hidden danger information to the risk database through the hidden danger investigation registration module;

[0067] S3. The hidden danger inspection plan module provides hidden danger inspection functions. By formulating hidden danger inspection plans, hidden danger inspections are carried out on risk points.

[0068] S4. After formulating the hidden danger investigation plan, the hidden danger investigation registration module will conduct hidden danger investigation in the hidden danger investigation area, and manually classify the hidden dangers into general hidden dangers and major hidden dangers according to the safety standards specified by the factory, and register the hidden dangers. The hidden danger rectification module will formulate a hidden danger rectification plan based on the registered hidden dangers, and then the hidden danger rectification personnel will rectify the hidden dangers according to the hidden danger rectification plan. After the rectification is completed, it will be submitted for hidden danger rectification acceptance;

[0069] S5. After the hidden dangers are checked and rectified, the hidden danger statistical analysis module conducts hidden danger statistical analysis to count the number of hidden danger registrations, hidden danger check times, and hidden danger rectification times in each area and department, and then regularly updates and completes the risk classification knowledge graph.

[0070] Furthermore, the hidden danger inspection plan includes: periodic inspection, scheduled inspection, and real-time inspection; periodic inspection includes periodic inspection on an annual, monthly, weekly, and daily basis; scheduled inspection is carried out according to the established inspection start time; real-time inspection starts inspection immediately according to the formulated inspection plan; the contents of the inspection plan include: hidden danger inspection start time, end time, hidden danger inspector, and hidden danger inspection area.

[0071] Compared with the prior art, the beneficial effects of the present invention include the following points:

[0072] 1. Existing risk assessment methods are often based on expert opinions and manual evaluation, which makes the assessment results affected by personal experience and preferences, lacking objectivity and consistency. The present invention analyzes a large amount of basic risk data and uses the knowledge graph of artificial intelligence to model the data, thereby obtaining a more objective assessment result.

[0073] 2. The existing risk assessment is not combined with the hidden danger investigation results. The risk assessment rarely changes after being determined, and it cannot reflect the risk level of the risk unit in real time. However, this discovery combines risk assessment with hidden danger investigation. After the risk assessment, the hidden danger investigation results collected by the system are analyzed, and the risk classification knowledge graph is updated according to a certain update strategy using knowledge graph completion technology, so that the entire system completes the closed loop and the risk assessment is more timely.

[0074] 3. The present invention establishes an algorithm model based on a bionic prototype memory attention network to achieve the completion and update of the risk grading knowledge graph. This model can automatically adjust and update the content of the knowledge graph more accurately when it detects changes in the data related to the risk grading knowledge graph in the system, ensuring effective dynamic management of various risks. In addition, the present invention also utilizes a small sample information extraction algorithm model, which only needs to use a relatively small data set to train knowledge triples (subject-relationship-object) with sufficiently high accuracy, greatly reducing the need for large-scale manually labeled data, thereby effectively saving labor costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0076] Figure 1 It is the overall workflow diagram of the dynamic intelligent adjustment of the safe production dual-defense control system in the present invention;

[0077] Figure 2 It is a step diagram of the risk dynamic adjustment method based on knowledge graph in the present invention; DETAILED DESCRIPTION

[0078] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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.

[0079] like Figure 1 This embodiment provides a dual control system for production safety risk grading and hidden danger investigation, which uses the knowledge graph technology in the field of natural language processing to grade the risk points in the dual control system, and dynamically updates the risk grading knowledge graph through the results of hidden danger investigation and rectification, so as to realize intelligent and dynamic management of risk grading.

[0080] In some specific implementations, the production safety risk classification and hidden danger detection dual control system includes a risk management subsystem and a hidden danger detection subsystem.

[0081] In some possible implementations, the risk management subsystem includes: a risk information extraction and import module, a risk classification knowledge graph module, and a risk management measures module; wherein the risk information extraction and import module obtains a safety entity information data set by extracting or importing information related to factory production risk data. Production risk data includes: structured data, which exists in the form of structured tables such as excel and csv. This type of data is directly mapped to the risk knowledge base field through the risk information import function and stored in the risk knowledge base; unstructured data, which exists in the form of unstructured text such as word and txt. This type of data needs to extract risk information through the risk information extraction subsystem, and then map it with the risk knowledge base field and store it in the risk database. Risk data includes: including risk information data including risk analysis objects, risk analysis units, risk points and the causes of risks, and the consequences of risks.

[0082] Furthermore, the risk grading knowledge graph module includes the construction and maintenance of the risk grading knowledge graph. The construction of the risk grading knowledge graph is to extract relevant data fields from the risk database to construct a risk grading knowledge graph. The risk grading knowledge graph is composed of risk location, risk cause, risk status, risk cost, number of hidden danger occurrences, and number of hidden danger inspections. The risk grading knowledge graph is not only used for risk identification, but also provides a basis for risk grading. It is divided into low risk, general risk, greater risk and major risk levels according to the severity of the risk, and is updated according to the statistical analysis of hidden danger inspections, so as to dynamically update the risk grading to monitor dynamically changing risks. The maintenance of the risk grading knowledge graph is to use the small sample knowledge graph completion technology based on the bionic prototype memory attention network to complete and update the knowledge graph after the risk grading knowledge graph is constructed and new dynamic data is generated.

[0083] Furthermore, the risk control measures module is used to create corresponding control measures according to different risk levels and establish a control measures library. The system can also evaluate the residual risk after the implementation of control measures to help users better control the overall risk of the enterprise;

[0084] The hidden danger investigation subsystem is used to inspect the implementation of risk control measures, including hidden danger monitoring module, hidden danger investigation plan module, hidden danger investigation registration module, hidden danger rectification module, hidden danger statistical analysis module. The system can automatically formulate investigation plans and rectification plans, and send SMS alarms when they are overdue. Real-time monitoring of major risk points is carried out through monitoring equipment, and hidden dangers are discovered and alarmed in time. The hidden danger statistical analysis module can analyze the hidden danger discovery rate and rectification rate of each department, and intuitively display the management effect;

[0085] Furthermore, the hidden danger inspection plan module is used to formulate an inspection plan for each risk point according to the residual risk after the implementation of control measures. The hidden danger inspection plan includes: periodic inspection, scheduled inspection, and real-time inspection. Periodic inspection includes periodic inspection by year, month, week, and day; scheduled inspection is carried out according to the established inspection start time; real-time inspection starts the inspection immediately according to the formulated inspection plan. The content of the inspection plan includes: hidden danger inspection start time, end time, hidden danger inspector, hidden danger inspection area and other information;

[0086] Furthermore, the hidden danger monitoring module uses gas detectors, cameras and other terminal devices to monitor hidden dangers in real time at locations that are classified as major risk points in the risk classification. If hidden dangers are found, they are automatically registered in a timely manner.

[0087] Furthermore, the hidden danger inspection and registration module distributes the hidden danger inspection tasks to the inspection personnel of each department according to the formulated hidden danger inspection plan. The inspection personnel conduct hidden danger inspections in the areas they are responsible for, and register the hidden dangers found according to general hidden dangers and major hidden dangers. The hidden danger registration information includes: the time when the hidden danger occurred, the place where the hidden danger occurred, photos of the hidden danger, the person who discovered the hidden danger and other information.

[0088] Furthermore, the hidden danger rectification module is used to formulate a rectification plan for the registered hidden dangers, including rectification deadlines, rectification measures, etc. If the rectification is overdue, the system will automatically send a text message to warn. After the rectification personnel confirm that the rectification is completed, they will report the rectification acceptance, and the acceptance personnel will accept the implementation of the hidden danger rectification. If the rectification is unsatisfactory, the system will require rectification again until the acceptance criteria are met.

[0089] Furthermore, the hidden danger statistics module is used to conduct statistical analysis of hidden dangers according to different risk points and departments. The number of hidden danger registrations, hidden danger inspections, and hidden danger rectifications in each area and department are counted. Through the statistical results, users can understand the hidden dangers of each risk point and department, provide a reference basis for hidden danger management, and dynamically modify the risk classification knowledge graph. According to the risk classification knowledge graph update time set by the system, the risk classification knowledge graph is updated and completed regularly.

[0090] In some possible implementations, the risk information extraction subsystem extracts knowledge triples from the enterprise's unstructured safety risk basic historical data, which include accident records, safety event descriptions, safety equipment information, safety operating procedures, etc. The extracted knowledge triples include risk analysis objects, risk units, and control measures.

[0091] The risk analysis objects include: areas, equipment, and process preparation processes that may have risks in the production process; risk units exist in the risk analysis objects, including: a component on the risk equipment, a sub-area in the risk area, and a link in the process preparation process. These data are mainly in the form of word documents, which are unstructured documents. The information extraction method can better extract the required structured risk information from these unstructured documents. Control measures are measures to manage and control risks that may exist in risk units. For example, in order to control the risk of machine injuries, the corresponding control measure is that the operator should maintain a certain safe distance from the machine.

[0092] The risk information extraction subsystem uses the small sample entity relationship extraction method in natural language processing, which includes three components: small sample relationship extraction, relationship-based small sample entity recognition, and small sample global entity pairing. Small sample relationship extraction: Through training with a small number of relationship samples, the model can identify the relationship between the head and tail entities in an instance. Relationship-based small sample entity recognition: Through training with a small number of entity samples, the model can identify each entity in an instance and derive entity attributes. Small sample global entity pairing: After identifying the entities in the instance through the small sample entity recognition module, the head entity and tail entity of the knowledge triple are matched.

[0093] In some specific implementations, the risk information import module imports existing structured data into the system risk information database, and these existing structured data are uploaded to the risk management subsystem through data, and then these structured data fields are mapped with the corresponding data table fields of the risk knowledge database, and finally stored in the risk knowledge base.

[0094] In some specific implementations, the risk classification knowledge graph module extracts risk information from historical data and constructs a risk classification knowledge graph with imported data. The risk classification knowledge graph includes risk location, risk cause, risk status, risk cost, number of hidden danger occurrences, and number of hidden danger inspections. The risk classification knowledge graph is used to automatically classify risks in the security field knowledge base. This implementation method divides risks into low risk, general risk, greater risk, and major risk levels according to the severity of the risk.

[0095] After the risk classification knowledge graph is constructed, if there is no corresponding risk entry in the risk knowledge base, the user can manually create it and add it to the knowledge base. The risk classification knowledge graph is updated dynamically based on the hidden danger investigation statistics.

[0096] In some possible implementations, the specific process of the risk management subsystem is as follows:

[0097] S1. Extract information based on the risk-related data in the safety documents in the factory workshops set up by each enterprise to obtain a safety entity information data set. These safety documents record various safety incidents and related information in the industrial production process, covering multiple entity types such as equipment, operators, and safety incidents. This embodiment extracts a large amount of risk-related information from the safety documents through the risk information extraction and import module, and stores it in the risk database to support risk classification management and hidden danger investigation. Risk-related data include: structured data, which exists in the form of structured tables such as excel and csv. This type of data is directly mapped to the risk knowledge base field through the risk information import function and stored in the risk knowledge base; unstructured data exists in the form of unstructured text such as word and txt. This type of data needs to extract risk information through a small sample risk information extraction method after training, and then mapped with the risk knowledge base field and stored in the risk database.

[0098] Further risk database includes risk analysis objects, risk points (internal submodules or process components of risk analysis objects, i.e. the causes and occurrence points of risks), including different risk types and characteristics, and possible consequences. Improve the risk database and provide more comprehensive and accurate information support.

[0099] S2. Use some data fields in the risk database, including: risk location, risk cause, risk status, risk cost, number of hidden danger occurrences, number of hidden danger inspections, and number of hidden danger rectifications to build a risk classification knowledge graph. The number of hidden danger registrations, the number of hidden danger inspections, and the number of hidden danger rectifications are generated in the hidden danger inspection module.

[0100] S3. The risk classification knowledge graph will automatically classify risks based on the risk database, and classify them into low risk, general risk, major risk and serious risk levels according to the severity of the risk. If there is no corresponding risk entry in the risk database, the user can create it manually and then complete the risk database.

[0101] S4. After risk classification, corresponding control measures are created in the control measures function module according to different risk levels, and a control measures library is established. At the same time, after formulating risk control measures, the existing risk assessment methods are used to assess the residual risks after the implementation of control measures, helping users to better control the overall risks of factories and enterprises. Risk analysis methods include risk matrix analysis method (LS), operating condition hazard analysis method (LEC), and risk degree analysis method (MES).

[0102] S5. Finally, risk statistics are performed through the risk statistics module to count the number of risks in different regions and departments and display them. Users can understand the risk status of each region and department by analyzing the statistical results and provide a reference for decision-making.

[0103] In some possible implementations, the specific process of the hidden danger detection subsystem is as follows:

[0104] S1. The hidden danger monitoring module uses gas detectors, cameras and other terminal devices to conduct real-time monitoring of hidden dangers in locations or equipment listed as major risk points in risk classification.

[0105] Furthermore, the real-time detection and monitoring of hidden dangers is used for training in small sample field target detection by artificial intelligence, and the system can use monitoring equipment such as cameras to monitor major risk points in real time. The specific contents are as follows:

[0106] Video data is collected through camera equipment and used for small sample field target detection model training of artificial intelligence. After that, deep learning algorithms are used to identify and locate objects in images, including machinery and employees. The monitoring targets include the following:

[0107] 1. Whether employees in the risk area are properly dressed, including work clothes, safety helmets, dust-proof glasses, earplugs, etc.;

[0108] 2. Whether employees in the risk area are kept at a certain safe distance from risk equipment;

[0109] 3. Whether the number of employees in the risk area exceeds the prescribed number;

[0110] 4. Whether there are any dangerous goods that violate regulations in the risk area;

[0111] 5. Check whether there is any equipment failure in the risk area.

[0112] S2. Once a hidden danger is found, the system will issue an alarm in time to help users respond and deal with it quickly. The system will send an alarm message and call the hidden danger investigation personnel to formulate a hidden danger investigation plan through the mobile phone number in the user information and the WeChat applet.

[0113] S3. The hidden danger inspection plan module provides hidden danger inspection functions. By formulating hidden danger inspection plans, hidden danger inspections are carried out on risk points. Hidden danger inspection plans include: periodic inspections, scheduled inspections, and real-time inspections. Periodic inspections include periodic inspections by year, month, week, and day; scheduled inspections are carried out according to the set inspection start time; real-time inspections start inspections immediately according to the formulated inspection plan. The inspection plan includes: hidden danger inspection start time, end time, hidden danger inspector, hidden danger inspection area and other information.

[0114] S4. After formulating the hidden danger inspection plan, the hidden danger inspection registration module will conduct hidden danger inspection in the hidden danger inspection area, and classify the hidden dangers found out into general hidden dangers and major hidden dangers manually according to the safety standards formulated by the factory, and register the hidden dangers in the risk database. The hidden danger rectification module will formulate a hidden danger rectification plan based on the registered hidden dangers, and then the hidden danger rectification personnel will rectify the hidden dangers according to the hidden danger rectification plan. After the rectification is completed, it will be submitted for acceptance. The acceptance personnel will approve the rectification and re-rectify it according to the submitted rectification situation.

[0115] S5. After the hidden dangers are checked and rectified, the hidden danger statistical analysis module conducts hidden danger statistical analysis to count the number of hidden danger registrations, hidden danger check times, and hidden danger rectification times in each area and department. The risk grading knowledge graph is updated and completed regularly according to the risk grading knowledge graph update time set by the system.

[0116] The method steps of the risk dynamic adjustment method based on small sample knowledge graph in the present invention are shown in FIG. Figure 2 ; The specific method steps are described as follows:

[0117] This implementation uses a natural language pre-training model as an encoder to model the text more accurately, and supports the training of a small sample entity relationship extraction subsystem to obtain risk relationships and corresponding entities. In this implementation, the natural language selection pre-training model Bert can be selected as needed. Other natural language pre-training models:

[0118] Due to the small sample size of production safety data and the lack of sufficient labeled data, the information extraction model trained by the existing supervised learning model in the field of risk knowledge cannot play a good role. In order to solve the problem that supervised learning cannot learn entities and relationship types that have never been seen before and solve the problem of lack of sufficient labeled data, this implementation method adopts a small sample training method, which only requires a small number of samples for learning to achieve a better extraction effect. In the small sample training method, the data set is usually divided into a support set and a query set. The small sample training method only learns a new concept from a few instances in the support set, while maintaining good generalization in the query set. The specific contents are as follows:

[0119] S1. Randomly extract a part of the manually collected and annotated historical risk data as a data set, which is divided into training data set D train , test data set D test , and then divide the two data sets into support data sets and query data sets respectively The structure and form of the support dataset are the same as those of the query dataset, but they contain different categories. Here we take the support dataset in the training dataset as an example:

[0120]

[0121] h i represents the head entity of instance i in the dataset, t i represents the tail entity of instance i in the dataset, r j represents the relationship in the relationship set of instance i, S is the supporting instance set, s i To support instance i in the instance set, R is the relationship set, E is the entity set, NK are K support instances randomly selected from the N triplet categories.

[0122] S2. In the task of small sample relationship extraction, the goal is to extract a single relationship type from a sentence, which allows the task to be formalized as a classification problem, in which each relationship type corresponds to a category, and the classifier needs to learn to correctly classify sentences into corresponding relationship types based on training data. Specifically, this method adopts a prototype network-based method to address the challenge of small sample relationship extraction. First, it is necessary to construct a prototype representation for each relationship. In this process, this implementation will support the organization of all sentences belonging to the same relationship, and ensure that each relationship category contains K example sentences, while annotating the entity information in these sentences. Unlike traditional methods that rely solely on sentence information, this implementation further integrates the information of the head entity and the tail entity to jointly construct the prototype representation of the relationship, because there is an intrinsic correlation between the type of entity and the type of relationship it involves. For example, given a relationship "born", the corresponding head entity type is usually "person" and the tail entity type is "location"; the specific process is:

[0123]

[0124] is the relation head entity of the k-th instance sentence of relation i, is the relation tail entity of the k-th instance sentence of relation i, For the kth instance sentence of the relation, the f_span function represents obtaining the entity representation from the sentence.

[0125]

[0126] represents the overall encoding representation of the sentence, The relation representation contains information about the sentence, the head entity, and the tail entity.

[0127]

[0128] represents the prototype representation of the relationship, K represents the total number of sentences contained in each relationship,

[0129] p r,i=Similarity(h i q ,C r )

[0130] p r,i is the similarity between the query instance and the relation prototype, is an instance of query set relation i, C r is the relation prototype of the support set relation i

[0131]

[0132] L rel is the cross entropy loss function for small sample relationship extraction, N is the total number of relationship types in the training data set, and y r,i is the relation label in relation i, p r,i is the similarity between the query instance and the relation prototype.

[0133] S3. In the small sample entity recognition task, the relationship prototype representation is optimized by introducing the relationship label knowledge. In the named entity recognition task, a span-based entity-level prototype is constructed to effectively alleviate the problems of label dependence and semantic incompleteness:

[0134]

[0135] E is the span representation of the sentence, l is the number of words in the instance, e is the embedding prototype span of each word, h is the tag encoding of the word embedding, and h n represents the tag encoding of the nth word embedding, h n-span_size is the tag encoding of the n-span_size word embedding, span_size is the maximum length of a single span, and f is the maximum pooling function that combines the embeddings of the spans.

[0136] S4. In order to better model the association between relations and entity prototypes, hierarchical contrastive learning is introduced to perform hierarchical contrastive learning on different span combinations to improve the accuracy of span recognition. By jointly refining the metric space used for relation extraction and entity recognition, the dispersed distribution of entity prototypes is enhanced, and the accuracy and effect of risk information extraction are improved. Hierarchical contrastive learning includes contrastive learning between relation prototypes, contrastive learning between entities, and contrastive learning between entities and relations:

[0137]

[0138] L conRel is the relation prototype comparison constraint loss, m1 is the minimum distance margin of the same type of data, m2 is the maximum distance margin of heterogeneous data, N is the total number of relation types in the training data set, represents the prototype of relation i, Represents the prototype of relation j.

[0139]

[0140] L conEntity is the entity prototype comparison constraint loss, m1 is the minimum distance margin of the same type of data, m2 is the maximum distance margin of heterogeneous data, N is the total number of relationship types in the training data set, E is the total number of entities in the training data set, represents the prototype of the i-th entity under relation k, Represents the prototype of the i-th entity under relation k.

[0141]

[0142] L conRelEntity is the entity relationship prototype comparison loss, N is the total number of relationship types in the training dataset, E is the total number of entities in the training dataset, represents the prototype of relation i, represents the prototype of entity j, R i Represents the number of entities in relation i.

[0143] In some possible implementations, the risk grading knowledge graph will be completed and updated accordingly based on dynamic factors such as the frequency of hidden dangers and the implementation of risk control measures. In order to achieve this goal, a bionic prototype memory attention network can be used to train a model specifically for completing and updating the risk grading knowledge graph. When the risk location, risk cause, risk status, risk cost, frequency of hidden dangers, and frequency of hidden danger inspections involved change, the completion and update model of the risk grading knowledge graph is trained to automatically update and complete the risk grading knowledge graph. The model training method is specifically as follows:

[0144] S1. Imitating human memory mechanism: The three components of human memory are working memory, long-term memory and short-term memory. Information is initially stored in short-term memory. After a period of accumulation, the short-term memory information is transferred to working memory. Then, after being processed in working memory, it is stored in long-term memory. This implementation method applies the human memory mechanism to enhance the learning ability of the small sample knowledge graph completion and update model:

[0145] Establish a memory bank M consisting of short-term memory, working memory and long-term memory = {M (i) ,M (w) ,M (l)}; M is the memory bank, M (i) is short-term memory, M (w) is the working memory, M (l) It is long-term memory.

[0146] S2, introduces a bionic memory update mechanism to achieve a smooth transition between short-term memory, working memory, and long-term memory. This mechanism not only stores information in a more compact form in working memory, but also transfers frequently accessed information from working memory to long-term memory to solve the problem of information being forgotten over time:

[0147]

[0148] Among them, g(·) is the similarity function, It is M (i) The nth candidate prototype feature in . exp is the natural exponential function, μ n is the nth candidate average prototype feature, is the characteristic variance, are n prototype features stored in short-term memory.

[0149] S3, using memory-enhanced prototype smooth triple loss and constraining the similarity between new and old memory knowledge to effectively improve the accuracy of graph completion updates:

[0150] L=L m +λL P

[0151] Among them, L is the total loss of the model, L m is the margin-based ranking loss, L P is the prototype smooth triplet loss, λ is a hyperparameter, where the margin-based ranking loss L m :

[0152]

[0153] Among them, Q r is the positive query dataset for model construction, Q r- is the negative query dataset for model construction, S r is the supporting dataset of the model, φ(q r ,S r ) represent the positive and negative query datasets and the support dataset S respectively. r The similarity score between them, γ is the marginal hyperparameter, is a query unit in the negative query dataset, q r is a query unit in the query data set;

[0154] Where L P Is the prototype smooth triple

[0155]

[0156] represents the farthest embedding of similar data of prototype z, The nearest embedding of heterogeneous data representing the prototype z;

[0157] In some possible implementations, risk grading is performed by using the knowledge of the risk grading knowledge graph and according to a risk analysis method, wherein the risk analysis method includes a risk matrix analysis method (LS), an operating condition hazard analysis method (LEC), and a risk degree analysis method (MES), and risk grading is formulated in combination with the opinions of risk experts.

[0158] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0159] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A dual control system for production safety risk classification and hidden danger investigation, characterized in that: Including risk control subsystem and hidden danger investigation subsystem; The risk management subsystem includes: Risk information extraction and import module, used to extract risk-related data and import it into the risk database; The risk classification knowledge graph module is used to extract data fields from the risk database to construct a risk classification knowledge graph, and then maintain and update the risk classification knowledge graph; the risk classification knowledge graph is used to identify risks and classify risks; including risk locations, risk causes, risk status, risk costs, number of hidden danger occurrences, and number of hidden danger inspections; The risk control measures module is used to create corresponding control measures according to different risk levels; Risk statistics module, used for automatic statistical analysis of risks; The hidden danger investigation subsystem is used to inspect the implementation of risk control measures, including: Hidden danger monitoring module, used to monitor the classified risk locations in the risk classification knowledge graph in real time and record them in the risk database; The hidden danger investigation plan module is used to formulate an investigation plan for each risk point based on the residual risk after the implementation of control measures; The hidden danger investigation and registration module is used to issue investigation tasks according to the formulated hidden danger investigation plan and receive hidden danger registration information; Hidden danger rectification module, used to formulate rectification plans for registered hidden dangers; The hidden danger statistics module is used to conduct statistical analysis of hidden dangers according to different risk points and departments.

2. A dual control system control method for production safety risk classification and hidden danger investigation as claimed in claim 1, characterized in that: It includes a risk control subsystem control method and a hidden danger investigation subsystem control method. The risk control subsystem control method is specifically as follows: S1. The risk information extraction and import module extracts risk-related data from production safety documents and organizes them into structured data sets, and then imports the structured data sets into the risk database; Risk-related data includes structured data and unstructured data. The unstructured data is extracted through the risk information extraction subsystem after small sample training, and then mapped with the risk knowledge base fields and stored in the risk database; S2. The risk classification knowledge graph module constructs a risk classification knowledge graph using the risk location, risk cause, risk status, risk cost, number of hidden danger occurrences, number of hidden danger inspections, and number of hidden danger rectifications in the risk database, and completes and updates the risk classification knowledge graph; the number of hidden danger registrations, number of hidden danger inspections, and number of hidden danger rectifications are generated by the hidden danger inspection subsystem; S3. Use the risk classification knowledge graph and automatically classify risks based on the risk database to create corresponding risk classifications; S4. After risk classification, create corresponding control measures in the risk control measures module according to different risk levels, and establish a control measures library; S5. Finally, the risk statistics module performs risk statistics; Specifically count the number of risks in different regions and departments.

3. A dual control system control method for production safety risk classification and hidden danger investigation according to claim 2, characterized in that: The risk database includes risk analysis objects and risk points; the risk classification divides the risk points in the risk database into low risk, general risk, greater risk and major risk levels according to the severity of the risk.

4. A dual control system control method for production safety risk classification and hidden danger investigation according to claim 2, characterized in that: The risk information extraction subsystem consists of three components: small sample relationship extraction, relationship-based small sample entity recognition, and small sample global entity pairing.

5. A dual control system control method for production safety risk classification and hidden danger investigation according to any one of claims 2 and 4, characterized in that: The risk information extraction subsystem implements small sample training through the natural language pre-training model. The specific process is as follows: S1. Randomly extract a part of the manually collected and annotated historical risk data as a data set, which is divided into training data set D train , test data set D test , and then divide the two datasets into support dataset and query dataset respectively: in, D train Supporting datasets, D train The query dataset D test Supporting datasets, D test The query dataset; S2. In the small sample relation extraction task, relation label knowledge is introduced to optimize the relation prototype representation: is the relation head entity of the k-th instance sentence of relation i, is the relation tail entity of the k-th instance sentence of relation i, For the kth instance sentence of the relation, the f_span function represents obtaining the entity representation from the sentence; Among them, entity types include equipment, operators, and security events; represents the overall encoding representation of the sentence, Relation representation that contains information about the sentence, head entity, and tail entity; represents the prototype representation of the relation, K represents the total number of sentences contained in each relation; p r,i =Similarity(h i q ,C r ) p r,i is the similarity between the query instance and the relation prototype, is an instance of query set relation i, C r is the relation prototype of the support set relation i; L rel is the cross entropy loss function for small sample relationship extraction, N is the total number of relationship types in the training data set, and y r,i is the relation label in relation i, p r,i is the similarity between the query instance and the relation prototype; On the relation-based few-shot entity recognition task, a span-based entity-level prototype was constructed: E is the span representation of the sentence, l is the number of words in the instance, e is the embedding prototype span of each word, h is the tag encoding of the word embedding, and h n represents the tag encoding of the nth word embedding, h n-span_size is the tag encoding of the n-span_size-th word embedding, span_size is the maximum length of a single span, and f is the maximum pooling function that combines the embeddings of the spans; S3. Introduce hierarchical contrastive learning to perform hierarchical contrastive learning on different span combinations and jointly refine the metric space for relation extraction and entity recognition.

6. A dual control system control method for production safety risk management and hidden danger investigation according to claim 5, characterized in that: The hierarchical contrastive learning includes contrastive learning between relational prototypes, contrastive learning between entities, and contrastive learning between entities and relations: L conRel is the relation prototype comparison constraint loss, m1 is the minimum distance margin of the same type of data, m2 is the maximum distance margin of heterogeneous data, N is the total number of relation types in the training data set, represents the prototype of relation i, represents the prototype of relation j; L conEntity is the entity prototype contrast constraint loss, E is the total number of entities in the training dataset, represents the prototype of the i-th entity under relation k, Represents the prototype of the j-th entity under relation k; 7. A dual control system control method for production safety risk classification and hidden danger investigation according to claim 2, characterized in that: The risk classification knowledge graph will be supplemented and updated as the frequency of hidden dangers, the execution of risk control measures and other factors change. The completion and update of the risk classification knowledge graph is realized by training the completion and update model of the risk classification knowledge graph using a bionic prototype memory attention network. When the risk location, risk cause, risk status, risk cost, number of hidden dangers, and number of hidden danger inspections contained in the risk classification knowledge graph change, the completion and update model of the risk classification knowledge graph is trained to automatically update and supplement the risk classification knowledge graph. The model training method is specifically as follows: S1. Establish a memory bank M consisting of short-term memory, working memory and long-term memory. (i) ,M (w) ,M (l) }; M is the memory bank, M (i) is short-term memory, M (w) is the working memory, M (l) It is long-term memory; S2. Introducing the bionic memory update mechanism: Among them, g(·) is the similarity function, P qrn It is M (i) The nth candidate prototype feature in; exp is the natural exponential function, μ n is the nth candidate average prototype feature, is the characteristic variance, are n prototype features stored in short-term memory; S3, using memory-enhanced prototype smooth triplet loss to constrain the similarity between new and old memory knowledge: L=L m +λL P Among them, L is the total loss of the model, L m is the margin-based ranking loss, L P is the prototype smooth triplet loss, λ is a hyperparameter, where the margin-based ranking loss L m : Among them, Q r is the positive query dataset for model construction, Q r- is the negative query dataset for model construction, S r is the supporting dataset of the model, Represent the positive and negative query datasets and the support dataset S respectively r The similarity score between them, γ is the marginal hyperparameter, is a query unit in the negative query dataset, q r is a query unit in the query data set, and r is the number of relations; Where L P Is the prototype smooth triple represents the farthest embedding of similar data of prototype z, Represents the nearest embedding of heterogeneous data for prototype z.

8. A dual control system control method for production safety risk management and hidden danger investigation according to claim 2, characterized in that: The control method of the hidden danger detection subsystem is specifically as follows: S1. The hidden danger monitoring module monitors and controls hidden dangers in real time at risk locations or equipment according to risk classification levels; S2. When the hidden danger monitoring module finds hidden dangers, the hidden danger investigation subsystem will promptly issue an alarm to the safety administrator and automatically register the hidden danger information to the risk database through the hidden danger investigation registration module; S3. The hidden danger inspection plan module provides hidden danger inspection functions. By formulating hidden danger inspection plans, hidden danger inspections are carried out on risk points. S4. After formulating the hidden danger investigation plan, the hidden danger investigation registration module will conduct hidden danger investigation in the hidden danger investigation area, and manually classify the hidden dangers into general hidden dangers and major hidden dangers according to the safety standards specified by the factory, and register the hidden dangers. The hidden danger rectification module will formulate a hidden danger rectification plan based on the registered hidden dangers, and then the hidden danger rectification personnel will rectify the hidden dangers according to the hidden danger rectification plan. After the rectification is completed, it will be submitted for hidden danger rectification acceptance; S5. After the hidden dangers are checked and rectified, the hidden danger statistical analysis module conducts hidden danger statistical analysis to count the number of hidden danger registrations, hidden danger check times, and hidden danger rectification times in each area and department, and then regularly updates and completes the risk classification knowledge graph.

9. A dual control system control method for production safety risk classification and hidden danger investigation according to claim 8, characterized in that: The hidden danger inspection plan includes: periodic inspection, scheduled inspection, and real-time inspection; periodic inspection includes periodic inspection on an annual, monthly, weekly, and daily basis; scheduled inspection is carried out according to the established inspection start time; real-time inspection starts immediately according to the formulated inspection plan; the contents of the inspection plan include: hidden danger inspection start time, end time, hidden danger inspector, and hidden danger inspection area.

Images