Engineering safety design document visualization review model training method

Abstract

The invention belongs to the technical field of artificial intelligence and data mining, and particularly discloses an engineering safety design document visual review model training method, which comprises the following steps of: marking entities and relationships in various engineering safety design documents, and constructing a scheme global template consisting of triples of a first entity, a relationship and a second entity; training a self-encoding language model by utilizing the scheme global template to obtain a span semantic vector; and training the recurrent neural network model by using the span semantic vector and the scheme global template to complete the training of the engineering safety design document visualization review model. According to the method, not only can the surface content of the document be examined, but also potential safety risks and problems in the document can be found, and the examination comprehensiveness, accuracy and efficiency are greatly improved.

Description

Technical Field

[0001] The invention discloses a method for training a visual review model for engineering safety design documents, and belongs to the technical field of artificial intelligence and data mining. Background Art

[0002] Safety design is crucial in the whole life cycle management of hydropower engineering design, construction and operation. The accuracy and completeness of safety design documents are one of the important indicators to ensure the safe operation of the project. Project safety design document review is a knowledge-intensive task. The current document review method has problems such as lack of unified standards, reliance on manual experience, time-consuming, error-prone, and low efficiency. Especially for projects of different types, the focus and methods of document review are also different. Therefore, it is extremely important to build a more scientific and efficient safety document review method.

[0003] With the rapid development of digital technology, artificial intelligence (AI) technology plays an increasingly important role in document review. In particular, technologies such as natural language processing and machine learning have brought new ideas and methods to document review to improve the efficiency and accuracy of review. Among them, natural language processing technology realizes intelligent review of document content by analyzing and understanding the text in the document. Natural language processing technology first converts the text in the document into a digital form that can be processed by the computer, and then extracts the key information and semantic relationships in the text through steps such as lexical analysis, syntactic analysis and semantic analysis. For example, in the review of engineering safety special documents, natural language processing technology can identify key information such as monitoring methods and monitoring instruments in the document, and analyze the logical relationship between these information. In addition, natural language processing technology can also classify and evaluate the content of the document through methods such as text classification and sentiment analysis, providing a more comprehensive and accurate basis for document review. Machine learning technology can automatically extract key features and patterns in the document by autonomously updating and learning a large number of design documents, thereby improving the accuracy and efficiency of document review. For example, in the review of engineering safety documents, machine learning technology can automatically identify format errors, abnormal monitoring methods, logical contradictions and other problems in documents by learning from a large number of engineering safety design reports, and give corresponding review opinions. In addition, machine learning technology can also improve its own review capabilities and accuracy through continuous learning and optimization, providing more reliable protection for engineering safety design document review. According to statistics, the use of machine learning technology for document review can increase review efficiency by more than 50%, while reducing the review error rate to less than 5%.

[0004] At present, intelligent review technology is still in its infancy. It is mostly used in contracts, specifications and other documents with clear clauses. The clauses of this type of document are clearly marked and the difficulty of machine learning is relatively low. Existing machine learning models perform well on specific data sets, but their performance may decline in new environments or under different conditions. Therefore, building a high-precision machine learning model for engineering safety design document data sets is one of the current technical difficulties.

[0005] At present, intelligent review technology is rarely used in engineering safety design documents and is in its early stages. Engineering safety design documents usually include hydrology, geology, monitoring, structure, computer and other multi-professional contents. Due to the diversity of its contents and the complexity of its data, it increases the difficulty of analysis for data processing and natural language understanding, which in turn affects the accuracy and efficiency of the review technology. In addition, the review of engineering safety design documents involves multiple fields, such as structural safety, equipment performance, instrument selection, etc. Different fields require different review rules and standards, and the subjectivity and flexibility of the review standards will lead to inconsistent review results. Therefore, exploring and establishing review rules and standards applicable to engineering safety design documents is another technical difficulty that currently exists.

[0006] Existing natural language understanding technology has great limitations for different documents. For example, for some complex sentences and paragraphs, the system may not be able to accurately understand their meaning and contextual relationships, which affects the accuracy of the review. Security design documents are usually highly logical throughout the text, and the monitoring content, monitoring methods, and monitoring frequencies are progressive and closely related. Depending on the content and depth of the document, it is necessary to pay attention to the logical relationship and integrity between the various parts during the review to ensure the feasibility and effectiveness of the security design solution. Therefore, how to enable machines to "understand" documents is the third technical difficulty that currently exists. Summary of the invention

[0007] The purpose of the present invention is to provide a method for training a visual review model for engineering safety design documents, so as to solve the technical problems that the existing engineering safety design document review methods are mainly based on manual review, and have poor review accuracy and low efficiency.

[0008] The present invention provides a method for training a visual review model for engineering safety design documents, comprising:

[0009] Step 1: Annotate the entities and relationships in various engineering safety design documents, and construct a global solution template consisting of a triple of a first entity, a relationship, and a second entity;

[0010] Step 2: Use the global template of the scheme to train the autoencoder language model to obtain a span semantic vector, specifically including:

[0011] The autoencoder language model is trained using the global template of the scheme to obtain a word vector PAR and a sentence vector SEN;

[0012] After performing maximum pooling on the word vector PAR, the word vector is concatenated with the sentence vector SEN to obtain a span semantic vector;

[0013] Then step 2 is specifically:

[0014]

[0015] In the formula, The first entity position index Go to the second entity position index The position index of the span in the entire span sequence, is a constant function, For a sequence length of , the window length is The number of all spans of is the maximum pooling function, For word vector The vector after maximum pooling, For the indivual , is the concatenation function, For the general With sentence vector The concatenated vector, is the hyperbolic tangent function, is a linear mapping function, is the generated span semantic vector;

[0016] Step 3: Use the span semantic vector and the solution global template to train a recurrent neural network model to complete the training of the engineering safety design document visual review model.

[0017] Preferably, after completing the training of the engineering safety design document visual review model, the method further includes:

[0018] Build a global content template;

[0019] When it is determined that the safety design document of the project to be reviewed does not match the global template of the content, a review warning instruction is issued.

[0020] Preferably, after completing the training of the engineering safety design document visual review model, the method further includes:

[0021] When it is determined that the safety design document of the project to be reviewed does not match the global template of the solution, a review warning instruction is issued.

[0022] Preferably, the entities and relationships in various engineering safety design documents are annotated, including:

[0023] Mark word sequences in various engineering safety design documents;

[0024] The annotated word sequence is reorganized, arranged and flattened to obtain a span sequence containing entities and relations after annotation.

[0025] Preferably, the recurrent neural network model is a long short-term memory network model based on a multi-head self-attention mechanism.

[0026] Preferably, the loss function of the long short-term memory network model based on the multi-head self-attention mechanism includes a relationship classification loss function and an entity recognition loss function.

[0027] Preferably, the first entity and the second entity are respectively one of a monitoring item, a monitoring location, a monitoring method and a monitoring instrument.

[0028] Preferably, the monitoring items include at least one of environmental quantities, deformation, seepage, stress-strain, strong earthquakes in the hub area and patrol inspections.

[0029] Preferably, the relationship includes one of a spatial relationship, an associative relationship and an attribute relationship.

[0030] Compared with the prior art, the engineering safety design document visualization review model training method of the present invention has the following beneficial effects:

[0031] 1. Improve review efficiency: The present invention can complete the review of a large number of engineering safety design documents in a short period of time. Compared with manual page-by-page and line-by-line reading and analysis, AI can quickly scan the content of documents, quickly extract key information and make judgments, which greatly saves review time, especially for the massive documents generated in large engineering projects. By timely reviewing newly generated documents, it is ensured that safety issues can be discovered and handled in a timely manner.

[0032] 2. Enhanced review accuracy: The present invention can conduct in-depth mining and comprehensive analysis of documents. It can not only review the surface content of documents, but also discover potential security risks and problems in documents through semantic understanding of texts and correlation analysis of data, reduce errors caused by human factors, and improve the reliability of review results. For example, it can identify logical contradictions between different parts of texts, abnormal data fluctuations, etc., so as to more comprehensively evaluate the safety status of the project.

[0033] 3. Reduce review costs: This invention reduces the need for a large number of professional reviewers and reduces labor costs. Especially for some long-term and large-scale engineering projects, AI-driven review technology can significantly reduce human resource investment while ensuring the quality of review. By improving the accuracy of the review, it reduces the additional costs of engineering rework and repair caused by safety issues not being discovered in time, as well as the huge economic losses and social impacts caused by safety accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is an overall framework diagram of the engineering safety design document visual review model training method in an embodiment of the present invention.

[0035] Figure 2 A schematic diagram of a span marking method and mapping strategy constructed for an embodiment of the present invention.

[0036] Figure 3 This is a diagram showing the effect of applying the engineering safety design document visual review model training method in an embodiment of the present invention. DETAILED DESCRIPTION

[0037] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present invention. However, it should be clear to those skilled in the art that the present invention may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present invention.

[0038] The embodiment of the present invention discloses a method for training a visual review model for engineering safety design documents, and its overall architecture is as follows: Figure 1 shown.

[0039] At the beginning of model construction, the following review rules are formulated:

[0040] The engineering safety design document review of the present invention is conducted from two dimensions, namely D1={content integrity} and D2={solution targeting}. D1 uses keywords to formulate a global content template. , , through the regular expression operation ('re' module) in Python, the rules are matched to traverse the integrity information of the engineering safety design document to be reviewed. If a certain dictionary information is missing during the traversal process, the D1 warning count .

[0041] The D2 review is conducted using natural language processing technology, combining the characteristics of engineering safety standards and specifications and safety design reports to establish a review rule knowledge base, namely the global template for the plan.

[0042] First, semantic modeling is performed on various types of engineering safety design documents. Document semantic modeling includes two tasks, namely entity recognition and relationship extraction. The embodiment of the present invention standardizes the terms, concepts and relationships in the document into triples of {first entity, relationship, second entity}. The extracted entities are classified, and the entity types of the first entity and the second entity are defined as: monitoring items (XM_S), monitoring locations (WZ_S), monitoring methods (FF_S), and monitoring instruments (YQ_S). Relationships are defined as: spatial relationships (KJ_G), association relationships (GL_G), and attribute relationships (SX_G).

[0043] The above monitoring items are at least one of environmental quantities, deformation, seepage, stress and strain, strong earthquakes in hub areas and patrol inspections; the monitoring instruments include pressure tubes, displacement gauges, piezometers, sedimentation meters and other equipment; the monitoring locations represent the names of buildings in various parts of the project, which are determined according to the type of project. Taking hydropower projects as an example, the monitoring locations are underground powerhouses, caverns, slopes, retaining dams, sand-trapping dams and other types; the monitoring methods represent the type and number of instruments deployed, such as the deployment of 27 sets of 4-point multi-point displacement gauges.

[0044] After the above review rules are prepared, the engineering safety design document visual review model training method of the embodiment of the present invention specifically includes:

[0045] Step 1: Annotate the entities and relationships in various engineering safety design documents, and construct a global solution template consisting of a triple of a first entity, a relationship, and a second entity.

[0046] The embodiment of the present invention takes into account the complex situations such as nesting and overlapping that may be encountered in the process of identifying entities and relationships in engineering safety design documents, so a span-based annotation method is constructed to annotate entities and relationships in engineering safety design documents, specifically: annotate word sequences in various types of engineering safety design documents; recombine and arrange the annotated word sequences to obtain span sequences containing entities and relationships after annotation. The span sequence of the embodiment of the present invention is a span sequence in binary form.

[0047] For example, the span-based annotation method of the present invention is used to annotate entities and relationships in engineering safety design documents. It is necessary to annotate word sequences and then reorganize and arrange them into span sequences. The embodiment of the present invention also determines a mapping strategy for converting annotated word sequences into span sequences. First, the word sequences in the engineering safety design document are annotated. Mark as , where the subscripts (0, 1, 2, ...) represent the sequence index marked in the engineering safety design document, Indicates the length of the annotation of the word sequence. In the embodiment of the present invention, each annotation entity and relationship is defined as a span , subscript Indicates Words, subscript Indicates span The length threshold is set. Entities are represented as [start index, end index, entity type] triples, and entity relationships are represented as [entity 1 start index, entity 1 end index, relationship type, entity 2 start index, entity 2 end index] quintuples. All annotated entities and entity relationships are stored in the database in Json format. The constructed span annotation method and mapping strategy are as follows Figure 2 shown.

[0048] The embodiment of the present invention combines the classification standard of the relationship in the D2 review to construct a relationship-based entity relationship set including four categories: , , , Combining the four types of relationship-based entity relationship sets constructed above, the relationship (Relation) is modeled as the first entity ( ) is mapped to the second entity ( ), .

[0049] For sentences with multiple triples, based on the characteristics of the above four types of entity relationship sets, in order to solve the relationship overlap problem, the embodiment of the present invention generates a separate labeling sequence. Specifically, in the label sequence of a certain relationship, only the corresponding head and tail entities are labeled, and the remaining words are assigned labels. Given a sentence, if the sentence contains entity relationship categories, then generate A label sequence.

[0050] Step 2: Use the global template of the solution to train the autoencoder language model to obtain the span semantic vector.

[0051] The embodiment of the present invention first converts entities in various engineering safety design documents into machine-recognizable vectors and captures the semantic relationship between entities. The auto-encoding language model pre-training is used to dynamically extract the context information of the current entity in the global template of the solution. The auto-encoding language model of the embodiment of the present invention can be BERT, ALBert‌, RoBERTa, etc., preferably BERT.

[0052] Exemplarily, the embodiment of the present invention extracts features through the core bidirectional Transformer encoder of BERT and outputs the character-level vectorized representation of the text. , , is the sequence length, is the hidden state dimension of BERT. The core structure of BERT uses Transformer blocks. Each Transformer block consists of two sub-layers: a multi-head self-attention mechanism and a fully connected feed-forward network. Each sub-layer adds residual connections and normalization, so the output of the sub-layer can be expressed as: .in, represents the sublayer function, represents the input variable, Represents the sub-layer normalization function.

[0053] The embodiment of the present invention uses the global template of the solution to train the autoencoder language model to obtain the span semantic vector, which specifically includes:

[0054] The autoencoder language model is trained using a global template with labeled information to obtain word vectors and sentence vectors. The word vectors are then max-pooled and concatenated with the sentence vectors to obtain a span semantic vector.

[0055] For example, the encoding layer in the autoencoder language model of the embodiment of the present invention includes word encoding and span encoding. The word encoding adopts the above-mentioned BERT pre-training model. The vector output by BERT is divided into three parts, namely, word vector PAR, position vector LOC and sentence vector SEN. Then, PAR is max-pooled to highlight the significant information in the semantics and then combined with Concatenate and construct span semantic vector. Assume , Indicates the window length Next, the position index is The entity 1 to position index is The vector matrix composed of all words in entity 2, yes The dimension of the model's hidden state, the span semantic vector calculation formula is as follows:

[0056] (1)

[0057] In the formula, The first entity position index Go to the second entity position index The position index of the span in the entire span sequence, is a constant function, For a sequence length of , the window length is The number of all spans of is the maximum pooling function, For word vector The vector after maximum pooling, For the indivual , is the concatenation function, For the general With sentence vector The concatenated vector, is the hyperbolic tangent function, is a linear mapping function, is the generated span semantic vector. The constructed initial span vector is looped through equation (1) and the span sequence is formed in order according to the mapping result. , , is the length of the span sequence. This achieves the feature representation of the labeled span sentence vector.

[0058] In order to improve the review efficiency, the present invention first identifies the relationship tags in the span sentence vector, and then identifies the entity according to the specific relationship if there is a relationship in the sentence. )、Association relationship( ) and attribute relations ( )), traverse all span sentence vectors to extract relations, and obtain all relation types contained in the triples in the current sentence. Relation recognition uses the Sigmoid activation function, and the recognition function is expressed as , , Represents the number of relationship categories, The relationship category is The relationship dimension of the function returns a value ranging from 0 to 1 as the basis for judging whether the relationship exists.

[0059] After that, the embodiment of the present invention performs entity recognition on the span sentence vector of the extracted relationship. Considering that a sentence may contain multiple entity relationships, and each entity relationship is associated with different entities, the constructed entity recognition method is decomposed according to different relationships. The relationship vector is represented as (0,1), and then the sentence and the related relationship vector representation are concatenated. During the recognition process, a relationship in the sentence is randomly extracted each time to generate a sentence vector representation based on the relationship. , The relationship is generated kind The initial relation embedding uses random initialization parameters, and then constructs Perform entity extraction.

[0060] Step 3: Use the span semantic vector and the global template of the solution to train the recurrent neural network model to complete the training of the engineering safety design document visual review model.

[0061] After using span semantic vectors to generate sentence vectors based on entity relationships, the embodiment of the present invention constructs a joint extraction model, specifically using sentence vectors based on entity relationships to train a recurrent neural network model. The recurrent neural network model in the embodiment of the present invention is a recursive neural network model, a long short-term memory network model, a bidirectional recurrent neural network model, etc. To improve the accuracy of the model, the embodiment of the present invention uses a long short-term memory network model based on a multi-head self-attention mechanism. The long short-term memory network model based on a multi-head self-attention mechanism uses a long short-term memory network ( ) and multi-head self-attention mechanism ( ) integrated into Methods for processing input , the output list is then passed through Activation function, for each Perform a multi-classification task and construct a two-dimensional matrix of "entity fragment-entity type" to achieve accurate entity extraction, as follows:

[0062] (2)

[0063] In the formula, For the The span semantic vector of characters, For the a relationship, for Characters in relationship Conditional sentence vector, is the concatenation function, and Respectively Characters in relationship The head entity and the tail entity under , Respectively Characters in relationship The probability of predicting the category of the head entity and the tail entity under the condition, represents the number of entity categories, The entity recognition model based on relation extraction can reduce the number of triples under each relation, thereby solving the problem of triple overlap.

[0064] According to the two subtasks of relation extraction and entity recognition, the model loss function is constructed, which includes two parts, namely the relation classification loss function and the entity recognition loss function. The entity recognition loss part includes the head entity loss and the tail entity loss. The model training loss function is as follows:

[0065] (3)

[0066] In the formula, The number of categories representing the relationship, Indicates the first Relation vector representation; and Respectively expressed in the relationship Under the conditions of The probability that a character is predicted as a head entity and a tail entity label; , , Represent the predicted values ​​of the relationship, head entity, and tail entity respectively.

[0067] After completing the training of the visual review model for the engineering safety design document, the embodiment of the present invention further includes: constructing a content global template; and issuing a review warning instruction when it is determined that the engineering safety design document to be reviewed does not match the content global template. Further, after completing the training of the visual review model for the engineering safety design document, the embodiment of the present invention further includes: issuing a review warning instruction when it is determined that the engineering safety design document to be reviewed does not match the solution global template.

[0068] For example, construct review warning rules and knowledge semantic network: construct hierarchical warning rules based on the review rules established in the previous text, as shown in Table 1, the warning levels are divided into three levels: general, important, and serious, and the symbols are represented by The early warning project is built for D1 and D2 in the review rules. D1 is used as the first-layer framework of the semantic web. The scale of the semantic web is expanded according to the extracted entities and relationships to form a complete information chain of the monitoring system. The early warning rules for the review indicators are as follows: the monitoring system information chain and the professional knowledge base are matched through the regular expression operation ('re' module) in Python to traverse the evaluation report D2={scheme targeting} indicator. If a certain information chain node information is missing during the traversal process, the D2 early warning count The review model constructed by the present invention completely uses artificial intelligence technology to conduct intelligent review throughout the entire process, without the need for manual intervention or the establishment of review templates, and is applicable to a wide range of text types.

[0069] Table 1 Warning rules

[0070]

[0071] The present invention combines span annotation, BERT pre-training model, long short-term memory network (LSTM) and multi-head self-attention mechanism to construct a set of entity relationship joint extraction models (SPBERT-Multi-LSTM) for safety design documents. This method does not rely on annotated entities, but jointly identifies entities and relationships. The present invention simulates the mechanism and method of manual review, aiming to propose an efficient and accurate intelligent review method for engineering safety-related documents in combination with artificial intelligence technology, integrating natural language processing, machine learning and other technologies, and based on relevant specifications, legal provisions, relevant documents, experience and other review bases, the content integrity and solution pertinence of the design documents are reviewed in two professional dimensions, and finally the review results are evaluated and assisted in early warning, breaking through the limitations of traditional manual review and realizing refined intelligent review of engineering safety design documents.

[0072] Compared with the prior art, the present invention has the following beneficial effects:

[0073] 1. Improve review efficiency: The present invention can complete the review of a large number of engineering safety design documents in a short period of time. Compared with manual page-by-page and line-by-line reading and analysis, AI can quickly scan the content of documents, quickly extract key information and make judgments, which greatly saves review time, especially for the massive documents generated in large engineering projects. By timely reviewing newly generated documents, it is ensured that safety issues can be discovered and handled in a timely manner.

[0074] 2. Enhanced review accuracy: The present invention can conduct in-depth mining and comprehensive analysis of documents. It can not only review the surface content of documents, but also discover potential security risks and problems in documents through semantic understanding of texts and correlation analysis of data, reduce errors caused by human factors, and improve the reliability of review results. For example, it can identify logical contradictions between different parts of texts, abnormal data fluctuations, etc., so as to more comprehensively evaluate the safety status of the project.

[0075] 3. Reduce review costs: This invention reduces the need for a large number of professional reviewers and reduces labor costs. Especially for some long-term and large-scale engineering projects, AI-driven review technology can significantly reduce human resource investment while ensuring the quality of review. By improving the accuracy of the review, it reduces the additional costs of engineering rework and repair caused by safety issues not being discovered in time, as well as the huge economic losses and social impacts caused by safety accidents.

[0076] To determine the effectiveness of the solution of the present invention, the embodiment of the present invention also carried out verification of the following operation steps.

[0077] Dataset construction: This paper uses 12 typical engineering safety design documents in China as a large-scale corpus to construct a data set, with nearly 1 million characters. The report content contains all types of entities and relationships of all safety designs, a total of 230,000 triples, 110,000 Chinese sentences, and 5,000 predefined three types of relationship types, including 3,700 binary relationships and 1,300 multivariate relationships. The data annotation is generated by manual annotation. The data set is divided into training set, validation set and test set according to 8:1:1.

[0078] The experimental environment uses the central processing unit: Intel(R) Core(TM) i7-8700 CPU, the GPU is NVIDIA T400 graphics card, each with 8GB of video memory, the operating system is Windows 10, the running memory is 128G, the programming language is Python3.11.4, the deep learning framework is Pytorch 1.8.1 and the open source deep learning library Keras 2.3.1. The model parameters constructed during the experiment are shown in Table 2.

[0079] Table 2 Model parameter settings

[0080]

[0081] Evaluation index: In order to evaluate the effectiveness of the relationship extraction and entity recognition model proposed in this paper, the currently commonly used index accuracy ( )、Recall( )、 value( ) to measure the accuracy of model classification. First, for any category in the classification system Construct the confusion matrix ( ), as shown in Table 3.

[0082] Table 3 Confusion matrix

[0083]

[0084] In the table: express Class samples are correctly classified into kind; Indicates that samples of other classes are misclassified kind; express Class samples are incorrectly classified into other classes; Indicates that other class samples are correctly classified. Therefore, the evaluation indicators can be calculated as follows:

[0085] (4)

[0086] (5)

[0087] (6)

[0088] Where: It represents the ratio of the number of correct texts recognized by the model to the total number of recognized texts; Indicates the proportion of positive samples in the correctly divided samples; It is a comprehensive indicator, which represents the harmonic mean of precision and recall.

[0089] Experimental results: The entity relationship joint extraction process constructed by the embodiment of the present invention and The method is used on the same dataset Model and The model is compared, and the training process still adopts the The training results of the labeling method are shown in Table 4. The bold font in the table represents the maximum value of the same indicator. It can be seen from Table 4 that the method proposed by the present invention is and The highest among the three models, relative Model and They increased by 7.72% and 2.77% respectively. Model , and The improvements were 17.49%, 14.69% and 16.03% respectively. The experimental results show that the entity relationship joint extraction method constructed in this paper has higher extraction performance, can effectively improve the accuracy of the review, and verify the effectiveness of the method proposed in this invention.

[0090] Table 4 Experimental results statistics

[0091]

[0092] The present invention constructs triples and a visualized semantic web based on the entities and relationships in the extracted security design document, performs a visual review of the document based on the information chain in the semantic web, and finally outputs the review result to a word file according to the early warning rules, thereby realizing early warning assistance and intelligent review. An example of the review result is shown in Table 5.

[0093] Table 5 Example of review results

[0094]

[0095] Device: The visual review device for engineering safety design documents constructed in the embodiment of the present invention is based on the Windows 10 operating system, and is developed using the Python 3.6 programming language, the Django development framework, and the Neo4j graph database. The MVT (Model-View-Template) development model is used to implement the retrieval and visual display of entity and relationship-related semantics. The operating mechanism of the device is to transmit the query request issued by the browser to the view layer via the HTTP transmission protocol, the Web server interface, and the routing. The view layer encapsulates the operation of the application page. The model layer encapsulates the query operation on the Neo4j graph database, and the template layer encapsulates the page display function written in HTML / CSS / JS and other files. After receiving the query results, the browser can display various information chains in the semantic web, and judge the integrity of the content in the information chain and the pertinence of the solution according to the review rules and early warning rules.

[0096] The device has three main functions: (1) Information recognition: It is mainly used to quickly extract the preset four-category entity relationship set based on the relationship from the input document: , , , , , , , . For example, in the safety design document of a hydropower station, information chains such as <thermometer, located at, dam foundation>, <thermometer, monitoring, dam body temperature>, <dam foundation, layout, 3 thermometers> are accurately identified, among which {located at} belongs to the spatial relationship category, {monitoring} belongs to the association relationship type, and {layout} belongs to the attribute relationship category. The output of this module is a triple of various relationships in the document. (2) Automatic review: The content integrity and targetedness of the document are comprehensively reviewed according to the preset review rules. The review method adopts the rules determined in the previous text of the present invention. Ensure that key information such as monitoring items, monitoring instruments, and monitoring methods in the document meets the requirements. (3) Intelligent early warning: According to the preset early warning method, different levels of early warning information in the document are output to assist designers in quickly modifying the design plan and ensuring the design quality. The effect diagram of the intelligent review device is as follows Figure 3 shown.

[0097] The engineering safety design document visualization review model of the present invention has the following beneficial effects:

[0098] 1. Improve review efficiency: Traditional engineering safety document reviews may rely mainly on manual reading, analysis and judgment line by line, which is relatively slow. Especially when faced with a large number of documents and complex data, the review work will consume a lot of time and manpower costs. For some simple and repetitive review tasks, although certain standards and processes can be formulated, the overall efficiency improvement is limited. The present invention can quickly process a large amount of document data and conduct a comprehensive review of a large number of engineering safety documents in a short period of time. Through rapid scanning and analysis of the text, as well as rapid calculation and comparison of the data, possible problems and risks in the document can be quickly identified, greatly improving the review efficiency.

[0099] 2. Improve the accuracy of review: Existing manual review is easily affected by factors such as the reviewer's experience, professional level, and fatigue, and may result in missed judgments, misjudgments, etc. Moreover, for some complex technical specifications and standards, there may be deviations in manual understanding and application, which affects the accuracy of the review results. The present invention is based on a large amount of data training and advanced algorithm models, and can more accurately understand and apply relevant standards and specifications for engineering safety. More accurate analysis and judgment can be performed on professional terms, data indicators, etc. in the document, reducing errors caused by human factors and improving the accuracy of the review.

[0100] 3. Comprehensive review and evaluation: Manual review may be limited by the reviewer’s knowledge and experience, and it may be difficult to discover some deeply hidden security issues or complex logical relationships. Moreover, manual review can usually only focus on the main content of the document, and the review of some details or related information is not comprehensive enough. The present invention uses AI technology to conduct in-depth mining and comprehensive analysis of documents, which can not only review the surface content of the document, but also discover potential security risks and problems in the document through semantic understanding of the text, correlation analysis of the data, etc., and combine historical data and related cases to conduct a more comprehensive assessment of the current engineering safety design documents.

[0101] The above are only several embodiments of the present invention and are not intended to limit the present invention in any form. Although the present invention is disclosed as above in the form of a preferred embodiment, it is not intended to limit the present invention. Any technician familiar with the profession, without departing from the scope of the technical solution of the present invention, using the above disclosed technical content to make slight changes or modifications are equivalent to equivalent implementation cases and fall within the scope of the technical solution.

Claims

1. A method for training a visual review model for engineering safety design documents, characterized in that: include: Step 1: Annotate the entities and relationships in various engineering safety design documents, and construct a global solution template consisting of a triple of a first entity, a relationship, and a second entity; Step 2: Use the global template of the scheme to train the autoencoder language model to obtain a span semantic vector, specifically including: The autoencoder language model is trained using the global template of the scheme to obtain a word vector PAR and a sentence vector SEN; Performing maximum pooling on the word vector PAR and concatenating it with the sentence vector SEN to obtain a span semantic vector; Then step 2 is specifically: ， In the formula, The first entity position index Go to the second entity position index The position index of the span in the entire span sequence, is a constant function, For a sequence length of , the window length is The number of all spans of is the maximum pooling function, For word vector The vector after maximum pooling, For the indivual , is the concatenation function, For the general With sentence vector The concatenated vector, is the hyperbolic tangent function, is a linear mapping function, is the generated span semantic vector; Step 3: Use the span semantic vector and the solution global template to train a recurrent neural network model to complete the training of the engineering safety design document visual review model.

2. The method for training a visual review model for engineering safety design documents according to claim 1 is characterized in that: After completing the training of the visual review model for engineering safety design documents, it also includes: Build a global content template; When it is determined that the safety design document of the project to be reviewed does not match the global template of the content, a review warning instruction is issued.

3. The method for training a visual review model for engineering safety design documents according to claim 1 or 2, characterized in that: After completing the training of the visual review model for engineering safety design documents, it also includes: When it is determined that the safety design document of the project to be reviewed does not match the global template of the solution, a review warning instruction is issued.

4. The method for training a visual review model for engineering safety design documents according to claim 1, characterized in that: Label entities and relationships in various engineering safety design documents, including: Mark word sequences in various engineering safety design documents; The annotated word sequence is reorganized, arranged and flattened to obtain a span sequence containing entities and relations after annotation.

5. The method for training a visual review model for engineering safety design documents according to claim 1, characterized in that: The recurrent neural network model is a long short-term memory network model based on a multi-head self-attention mechanism.

6. The method for training a visual review model for engineering safety design documents according to claim 5 is characterized in that: The loss function of the long short-term memory network model based on the multi-head self-attention mechanism includes a relationship classification loss function and an entity recognition loss function.

7. The method for training a visual review model for engineering safety design documents according to claim 1, characterized in that: The first entity and the second entity are respectively one of a monitoring item, a monitoring location, a monitoring method and a monitoring instrument.

8. The method for training a visual review model for engineering safety design documents according to claim 7 is characterized in that: The monitoring items include at least one of environmental quantities, deformation, seepage, stress-strain, strong earthquakes in the hub area and patrol inspections.

9. The method for training a visual review model for engineering safety design documents according to claim 1, characterized in that: The relationship includes one of a spatial relationship, an association relationship and an attribute relationship.

Images