A Method and System for Identifying Process Types Based on Neural Network-Introduced Domain Terminology

Abstract

This invention discloses a method and system for identifying process types based on neural networks and the introduction of domain terminology. It relates to the fields of construction engineering and natural language processing and machine learning, and includes: S1. Data acquisition, S2. Data processing, S3. Data quantification, S4. Model building, and S5. Process identification. This invention can accurately and quickly break down, reorganize, and classify scattered quality records, and achieve matching with corresponding specifications. This simplifies the steps for managers to find specifications during quality inspection, saves time, and thus assists in engineering quality management.

Description

Technical Field

[0001] This invention relates to the field of construction engineering, and also to the fields of natural language processing and machine learning, and in particular to a method and system for identifying process types based on neural networks that introduce domain terminology. Background Technology

[0002] Project management documents serve as reliable records of the entire project, summarizing past construction experience and promptly reflecting the project's quality status. They provide guidance for future work and improve efficiency. With the continuous development of construction technology, large-scale construction projects are increasing, leading to a surge in management documents generated during the construction process. Currently, quality management in the construction industry largely relies on on-site sampling and testing by supervisors. These tests contain numerous data records, and preliminary judgments on whether the results meet specifications largely depend on the supervisors' years of experience. Through long-term development of engineering quality systems, my country's engineering quality management regulations have formed a relatively complete categorization system, facilitating managers' access to information as needed. However, most construction quality management documents originate from managers' daily records, resulting in broad content, flexible formats, and significant fragmentation. Manually extracting quality information from various processes and finding corresponding specifications is cumbersome and prone to errors. Therefore, quickly and accurately extracting textual information is a powerful aid for implementing efficient construction quality management and a crucial prerequisite for accurately conducting construction quality compliance testing based on specifications.

[0003] To intelligently extract information from engineering documents, researchers have conducted numerous explorations, confirming the promising potential of combining natural language processing with deep learning techniques in the field of engineering construction. However, most scholars' research focuses on contract texts and risk records with relatively fixed formats, paying less attention to quality texts and failing to fully leverage the potential value of quality records in engineering quality inspection. Further research is needed on the in-depth analysis and intelligent processing of engineering quality texts.

[0004] Therefore, proposing a process type identification method and system based on neural networks to introduce domain terminology to solve the difficulties of existing technologies is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a method and system for identifying process types based on the introduction of domain terminology using neural networks. It comprehensively considers the professional meaning of the special expressions of professional terms and the semantics of commonly used words, improves the input port of the convolutional neural network, and classifies and integrates massive construction quality text records according to the prescribed process operations. This makes the process records in the scattered and fragmented construction quality texts orderly, and connects the corresponding quality specifications with the process as the link, so that quality management personnel can make a faster judgment on the compliance of construction quality.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A process type identification method based on neural networks and domain terminology is used for process identification in construction quality texts and to assist in construction quality control. The method includes the following steps:

[0008] S1. Data Acquisition: Obtain quality text data of on-site construction records and select the work process according to the nature of the project;

[0009] S2. Data Processing: Construct a process-specification index form based on the process type, introduce domain terminology to perform data cleaning and preprocessing on the quality text data in S1, and obtain a text set;

[0010] S3. Data Quantization: Transform the text set in S2 into a dense vector set using text quantization methods;

[0011] S4. Model Building: Input the vector set from S3 and train the multi-input port neural network model to obtain the trained neural network model;

[0012] S5. Process Identification: Input new quality text data, repeat S2 and S3 to process the new quality text, and then input the processed data into the trained neural network model to automatically identify the process type.

[0013] Optionally, in the above method, the domain terms in S2 are terms from a domain terminology lexicon.

[0014] Optionally, the data cleaning in S2 of the above method includes: splitting paragraphs into sentences and removing sentences that are irrelevant to quality and sentences that are too short to express semantics.

[0015] Optionally, the data preprocessing in S2 includes: using jieba segmentation to split the sentence into words, deleting stop words, retaining terms with professional attributes, and forming a word-level text set; splitting the sentence according to the level of individual characters and forming a character-level text set.

[0016] Optionally, in the above method, the text set in data preprocessing in S2 includes word-level text set and character-level text set, and the text set is divided into training set and validation set according to a certain ratio.

[0017] The above method, optionally, involves text quantization in S3, which uses ContinuousBag-of-Word technology to transform the central word into a low-dimensional dense vector through the text input of the central word's context. This transforms the character-level text set and word-level text set established in S2 into character-level vector sets and word-level vector sets that can participate in mathematical calculations, respectively.

[0018] The above method, optionally, involves the following specific steps for model construction in S4:

[0019] S401. Based on the CNN model, a basic convolutional neural network is built according to the hierarchical structure of input layer, convolutional layer, pooling layer and output layer;

[0020] S402. Extend the CNN into a dual-entry convolutional neural network, and set a concatenation layer after the convolutional layer to merge the feature extraction results of the dual entry points and input them into the next layer for further computation;

[0021] S403. Take out the character vectors and word vectors corresponding to the same sentence from the vector set respectively, input them into one port of the dual-entry convolutional neural network, use the established model to extract dual features, use the dual features as the learning material of the neural network, and learn the same features of similar processes and different features of dissimilar processes;

[0022] S404. Identify outlier processes based on the features learned by the neural network, test the learning effect of the neural network model with validation set data, evaluate the learning status of the neural network model with loss value and recognition accuracy as rating indicators, adjust the parameters of the neural network model, and obtain a well-trained neural network model.

[0023] Optionally, S5. Process identification also includes: integrating the quality text data into different subsets using the identified process types, searching the process-specification type form according to the subset type, matching the corresponding specification, and repeating this operation to achieve specification matching for the entire quality text data.

[0024] The process type identification system based on neural network-introduced domain terminology, which applies any of the above-mentioned methods for process type identification based on neural network-introduced domain terminology, includes: a data acquisition module, a data processing module, a data quantization module, a model building module, and a process identification module;

[0025] The data acquisition module, connected to the input end of the data processing module, is used to acquire on-site construction record quality text data and select the work process according to the nature of the project.

[0026] The data processing module, connected to the input of the data quantification module, is used to construct a process-specification index form based on the process type, and to introduce domain terminology to perform data cleaning and preprocessing on the quality text data in S1 to obtain a text set.

[0027] The data quantization module, connected to the input of the model building module, is used to transform a text set into a dense vector set using text quantization methods.

[0028] The model building module is connected to the input end of the process identification module. It is used to input the vector set into the multi-input port neural network model to obtain the trained neural network model.

[0029] The process identification module is connected to the output of the model building module. It is used to input new quality text data, repeat steps S2 and S3 to process the new quality text, and then input the processed data into the trained neural network model to automatically identify the process type.

[0030] As can be seen from the above technical solution, compared with the prior art, the beneficial effects of this invention are as follows: It fully considers the textual characteristics of professional terms in the field of engineering construction, and trains a neural network to learn the semantics of engineering texts. It mines and integrates the process information contained in the quality text data records generated during the engineering construction process, divides the text statements into different subsets according to the process, and identifies the corresponding building codes. This application can accurately and quickly split, reorganize, and classify scattered quality records, and achieve matching with corresponding codes, simplifying the steps for managers to find codes during quality inspection, saving time, and thus assisting in engineering quality management. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0032] Figure 1 Flowchart of the process type identification method based on neural network-introduced domain terminology provided by the present invention;

[0033] Figure 2 The structure diagram of the dual-entry convolutional neural network provided by this invention;

[0034] Figure 3 The loss value and accuracy of the dual-entry neural network validation set provided by this invention;

[0035] Figure 4This is an accuracy diagram of the integrated result of the text quality recording process provided in an embodiment of the present invention;

[0036] Figure 5 The structural block diagram of the process type identification system based on neural network-introduced domain terminology provided by the present invention. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] This invention can be used in a wide variety of general-purpose or special-purpose computing environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor devices, distributed computing environments including any of the above devices, etc.

[0039] Reference Figure 1 As shown, this invention discloses a method for identifying process types based on domain terminology introduced by a neural network, used for process identification in construction quality texts and assisting in construction quality control. Specifically, it includes the following steps:

[0040] S1. Data Acquisition: Obtain quality text data of on-site construction records and select the work process according to the nature of the project;

[0041] S2. Data Processing: Construct a process-specification index form based on the process type, introduce domain terminology to perform data cleaning and preprocessing on the quality text data in S1, and obtain a text set;

[0042] S3. Data Quantization: Transform the text set in S2 into a dense vector set using text quantization methods;

[0043] S4. Model Building: Input the vector set from S3 and train the multi-input port neural network model to obtain the trained neural network model;

[0044] S5. Process Identification: Input new quality text data, repeat S2 and S3 to process the new quality text, and then input the processed data into the trained neural network model to automatically identify the process type.

[0045] Furthermore, the on-site construction records in S1 include, but are not limited to, construction supervision reports, engineering quality inspection reports, and construction quality technical documents.

[0046] Furthermore, the process in S1 is the main construction procedure that may be involved in the project, determined based on the specific engineering construction project and the engineer's experience.

[0047] Furthermore, the process-quality form in S2 is a form that maps the specifications to individual processes based on the different processes involved in the specifications.

[0048] Furthermore, the domain terms in S2 are terms from the domain terminology lexicon.

[0049] Furthermore, data cleaning in S2 specifically includes: splitting paragraphs into sentences, and removing sentences that are irrelevant to quality and sentences that are too short to express semantics.

[0050] Furthermore, the data preprocessing in S2 specifically includes: using jieba segmentation to split sentences into words, deleting stop words, retaining terms with professional attributes, and forming word-level text sets; and splitting sentences according to the level of individual characters to form character-level text sets.

[0051] Furthermore, in S2, the text set in the data preprocessing includes word-level text set and character-level text set, and the text set is divided into training set and validation set according to a certain ratio.

[0052] Furthermore, the S3 text quantization method refers to using ContinuousBag-of-Word technology to transform the central word into a low-dimensional dense vector through the text input of the central word's context, thereby transforming the character-level text set and word-level text set established in S2 into character-level vector sets and word-level vector sets that can participate in mathematical calculations, respectively.

[0053] Furthermore, in S3, dense vectors refer to low-dimensional numerical vectors, generally not exceeding three hundred dimensions.

[0054] Furthermore, the specific steps for building the S4 model are as follows:

[0055] S401. Based on the CNN model, a basic convolutional neural network is built according to the hierarchical structure of input layer, convolutional layer, pooling layer and output layer;

[0056] S402. Extend the CNN into a dual-entry convolutional neural network, and set a concatenation layer after the convolutional layer to merge the feature extraction results of the dual entry points and input them into the next layer for further computation;

[0057] S403. Take out the character vectors and word vectors corresponding to the same sentence from the vector set respectively, input them into one port of the dual-entry convolutional neural network, use the established model to extract dual features, use the dual features as the learning material of the neural network, and learn the same features of similar processes and different features of dissimilar processes;

[0058] S404. Identify outlier processes based on the features learned by the neural network, test the learning effect of the neural network model with validation set data, evaluate the learning status of the neural network model with loss value and recognition accuracy as rating indicators, adjust the parameters of the neural network model, and obtain a well-trained neural network model.

[0059] Furthermore, the CNN basic model described in S401 is the Convolutional Neural Network (CNN) basic model.

[0060] Furthermore, S5 process identification also includes: integrating quality text data into different subsets based on the identified process type, searching the process-specification type form according to the subset type, matching the corresponding specification, and repeating this operation to achieve specification matching for the entire quality text data.

[0061] In one specific embodiment, to test the engineering applicability of a deep neural network-based method for matching engineering text quality records with building codes, a construction project was conducted. The main project was concrete engineering, involving multiple sub-projects such as excavation, grouting, and metal structure installation. After analyzing the characteristics of the project, eight main construction procedures were extracted: dam grouting, bedrock excavation and support, concrete mixing and preparation, roller-compacted concrete construction, concrete pouring, concrete quality inspection, metal structure fabrication and installation, and operation of construction machinery and instruments. These procedures were numbered alphabetically as labels to train the proposed model to extract features of the procedures, thereby distinguishing dissimilar procedures. Simultaneously, a procedure-code form was defined based on these procedures for searching construction codes. A partial form in this embodiment is shown in Table 1.

[0062] Table 1. Illustrative Correspondence Form of the Processes and Specifications of the Invention

[0063]

[0064] The training corpus consisted of 7525 construction quality description statements collected during the project's construction process, divided into a training set (5655 statements) and a validation set (1870 statements). A multi-entry convolutional neural network model was built using the PyTorch deep learning framework. A schematic diagram of the model structure is shown below. Figure 2 As shown in the diagram, the Continuous Bag-of-Words technique is used to quantize each sentence into dense vectors at both the word and character levels. These vectors are then input into the established model to extract different features from each process operation. These different process operations are used as the basis for text subset partitioning, grouping high-quality text into more organized subsets. During training, accuracy gradually increases while the loss value continuously decreases until convergence. The changes in both are shown in the diagram. Figure 3As shown in the figure. After the loss value converged, the model performance was tested using 1866 unlabeled datasets. The final results showed that the weighted accuracy reached 91.47%, demonstrating high accuracy and proving the engineering applicability of the proposed method. The accuracy, recall, and F1 score for each step are shown in the figure. Figure 4 As shown in the diagram, the letters correspond to the eight processes given in the example. Figure 4 It can be seen that even with new text without labels, the model still has a high accuracy rate, successfully integrating messy text records into subsets of the same type. After the process subset is determined, the corresponding specification can be quickly found by looking up the process-specification correspondence table, and the compliance of the subset content can be judged.

[0065] and Figure 1 Corresponding to the method described above, this embodiment of the invention also provides a process type identification system based on neural networks to introduce domain terminology, used for... Figure 1 The specific implementation of the method, the process type recognition system based on neural network-introduced domain terminology provided in this embodiment of the invention, can be applied to computer terminals or various mobile devices, and its structural diagram is shown below. Figure 5 As shown, it specifically includes: a data acquisition module, a data processing module, a data quantification module, a model building module, and a process identification module;

[0066] The data acquisition module, connected to the input end of the data processing module, is used to acquire on-site construction record quality text data and select the work process according to the nature of the project.

[0067] The data processing module, connected to the input of the data quantification module, is used to construct a process-specification index form based on the process type, and to introduce domain terminology to perform data cleaning and preprocessing on the quality text data in S1 to obtain a text set.

[0068] The data quantization module, connected to the input of the model building module, is used to transform a text set into a dense vector set using text quantization methods.

[0069] The model building module is connected to the input end of the process identification module. It is used to input the vector set into the multi-input port neural network model to obtain the trained neural network model.

[0070] The process identification module is connected to the output of the model building module. It is used to input new quality text data, repeat steps S2 and S3 to process the new quality text, and then input the processed data into the trained neural network model to automatically identify the process type.

[0071] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on its differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0072] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A procedure type recognition method based on neural network introduction of domain terminology, characterized by, The process recognition and auxiliary construction quality control for construction quality text specifically comprises the following steps: S1. Data acquisition: acquire on-site construction record quality text data, and select a process according to engineering properties; S2. Data processing: construct a process-specification index table according to process types, introduce domain terms to perform data cleaning and data preprocessing on the quality text data in S1, and obtain a text set; S3. Data quantification: convert the text set in S2 into a dense vector set by using a text quantification method; S4. Model construction: input the vector set in S3 into a neural network model with multiple input ports, and obtain a trained neural network model; S5. Process recognition: input new quality text data, repeat S2 and S3 to process the input new quality text, and input the processed data into the trained neural network model to automatically recognize the process type; The specific steps of S4. Model construction are as follows: S401. Build a basic convolutional neural network according to the hierarchical structure of an input layer, a convolutional layer, a pooling layer, and an output layer; S402. Expand the CNN into a double-entry convolutional neural network, and set a splicing layer after the convolutional layer to combine the feature extraction results of the double entries and input them into the lower layer for continuous operation; S403. Take out the word vectors and the word vectors corresponding to the same sentence from the vector set respectively, input them into one port of the double-entry convolutional neural network respectively, extract double features by using the established model, and use the double features as learning materials of the neural network to learn the same features of the same process and the different features of different processes; S404. Identify different processes according to the features learned by the neural network, verify the learning effect of the neural network model by using the verification set data, use the loss value and the recognition accuracy as evaluation indexes to evaluate the learning status of the neural network model, adjust the parameters of the neural network model, and obtain a trained neural network model; S5. Process recognition further comprises: integrating the quality text data into different subsets by using the recognized process type, searching for a process-specification type table according to the subset type, matching the corresponding specification, repeating the operation, and realizing specification matching of the entire quality text data; The neural network is trained to learn the engineering field text semantics by fully considering the text features of the engineering construction field professional terms, mining and integrating the process information contained in the quality text data records generated in the engineering construction process, dividing the text sentences into different subsets according to the process, and finding out the corresponding building specifications. The application can accurately and quickly split and reorganize the scattered quality records, and realize matching with the corresponding specifications, thereby simplifying the steps of the management personnel in searching for the specifications in the quality detection process, saving time consumption, and thus assisting engineering quality management.

2. The process type recognition method based on a neural network and introducing domain terms according to claim 1, wherein the domain terms in S2 are terms in a domain term library dictionary.

3. The process type recognition method based on a neural network and introducing domain terms according to claim 1, wherein ​ The data cleaning in S2 specifically includes: splitting a paragraph into sentences, and removing sentences irrelevant to quality and sentences too short to express semantics; The data preprocessing in S2 specifically includes: splitting a sentence into words by using jieba, deleting stop words, retaining professional terms, and forming a word-level text set; and splitting a sentence into single characters to form a character-level text set.

4. The process type recognition method based on neural network and introducing field terms according to claim 3, characterized in that, The text set in the data preprocessing in S2 includes a word-level text set and a character-level text set, and the text set is divided into a training set and a validation set according to a certain proportion.

5. The process type recognition method based on neural network and introducing field terms according to claim 1, characterized in that, The text quantification method in S3 refers to converting a central word into a low-dimensional dense vector by using a Continuous Bag-of-Word technology and inputting a context of the central word, so as to convert the character-level text set and the word-level text set established in S2 into a character-level vector set and a word-level vector set for mathematical calculation.

6. A procedure type recognition system that introduces field terminology based on a neural network, characterized by, The process type recognition method based on neural network and introducing field terms according to any one of claims 1-5 comprises: a data acquisition module, a data processing module, a data quantification module, a model construction module, and a process recognition module; The data acquisition module is connected with the input end of the data processing module, and is used to acquire quality text data of field construction records and select a process according to engineering properties; The data processing module is connected with the input end of the data quantification module, and is used to construct a process-specification index table according to a process type, introduce field terms to perform data cleaning and data preprocessing on the quality text data in S1, and obtain a text set; The data quantification module is connected with the input end of the model construction module, and is used to convert the text set into a dense vector set by using a text quantification method; The model construction module is connected with the input end of the process recognition module, and is used to input the vector set into a neural network model with multiple input ports, and obtain a trained neural network model; The process recognition module is connected with the output end of the model construction module, and is used to input new quality text data, repeat S2 and S3 to process the input new quality text, and input the processed new quality text into the trained neural network model to automatically recognize a process type.

Images