Construction quality inspection specification intelligent analysis method for automatic extraction of inspection information

Abstract

The invention belongs to the technical field of building specification named entity recognition, and relates to a construction quality inspection specification intelligent analysis method oriented to inspection information automatic extraction, in particular to a construction quality inspection specification intelligent analysis method based on deep learning and natural language processing. According to the method, a normative structured extraction method is constructed through a regular expression, a tag system for named entity recognition is designed, a corresponding data set is established, and a RoBERTa-wwm-BiLSTM-CRF model is used for automatically extracting quality information, so that the data blank in the field is filled up, and automation and intelligence of construction quality inspection work are promoted.

Description

Technical Field

[0001] The present invention belongs to the technical field of named entity recognition of building specifications, and relates to an intelligent parsing method for construction quality inspection specifications for automatic extraction of inspection information. The present invention is an intelligent parsing method for construction quality inspection specifications based on deep learning (DL) and natural language processing (NLP). Background Art

[0002] Construction quality is crucial to engineering projects. Its impact is not limited to the applicability of the project itself, but also has a profound impact on social and economic development, government credibility, personnel and property safety, and many other aspects. In construction projects, any quality problems will have a significant impact on the project schedule and consume a lot of manpower, material and financial resources. Quality inspection during the construction process is an important link to ensure project quality, safety and compliance. It runs through all stages of construction, from the initial foundation construction to the final completion acceptance. The quality of quality inspection work directly affects the overall performance and service life of the building.

[0003] An engineering construction project usually involves multiple professional fields, each of which has its corresponding construction quality acceptance specifications, and the construction quality inspection specifications cover all levels from national standards to local standards and then to group standards, which are highly complex and diverse. Due to the different regions and types of projects, the requirements for construction quality inspection during the construction of engineering construction projects may vary significantly, which makes it very difficult for practitioners, especially those with insufficient experience, to consult and retrieve standards and specifications. Standard specifications are complex and contain a lot of content. Manual consultation is not only time-consuming and labor-intensive, but also prone to omissions and misjudgments. Once the inspection is not timely, subsequent construction tasks may cover up quality problems in previous construction, leading to more serious consequences.

[0004] In recent years, with the continuous development of deep learning and NLP, powerful tools have been provided for the analysis of complex text documents, greatly improving the ability to extract valuable information from large amounts of text data. The development of pre-trained language models such as BERT has greatly improved the accuracy and efficiency of text understanding and processing, and with the maturity of NLP technologies such as named entity recognition (NER), we can effectively extract key information from unstructured text, which makes it possible for us to intelligently analyze construction quality inspection specifications. In the construction quality inspection specifications, the specifications from national to local levels cover a wide range and are complex. The extraction and analysis of the checkpoints involved are still in the exploratory stage, and no systematic solution has yet been formed. Therefore, the automated analysis and extraction of checkpoints in the construction quality inspection specifications is still a problem that needs to be solved. Summary of the invention

[0005] This paper proposes an intelligent parsing method for construction quality inspection specifications based on natural language processing technology for automatic extraction of construction quality inspection information, and realizes the parsing and automatic extraction of construction quality inspection point information based on the RoBERTa-wwm-BiLSTM-CRF model. This method can adapt to the parsing and quality inspection point extraction of different specifications, and provides strong technical support and methodological guidance for the intelligentization of construction quality management.

[0006] The technical solution of the present invention is as follows:

[0007] The intelligent parsing method of construction quality inspection specifications for automatic extraction of inspection information includes the following steps:

[0008] (1) Structured extraction of construction quality inspection specifications

[0009] First, the quality inspection specification document needs to be preprocessed to remove the preface, general principles, terms, basic regulations and appendix parts of the construction quality inspection specification, leaving only the main text of the document, and removing the corresponding watermarks, headers and footers. After the preprocessing process is completed, according to the text structure of the current construction quality inspection specification, a specification structured extraction algorithm is designed using regular expressions as a tool. The specification text structured extraction algorithm is as follows:

[0010] Input the pre-processed construction quality inspection specification and traverse the text content line by line. First, match the first-level title, which refers to the name of a large chapter in the construction quality inspection specification, that is, a specific sub-project; when the first-level title is matched, continue to traverse downward line by line to find the next adjacent first-level title, and complete the interception of the specification content between the two first-level titles. If the first-level title is the last section, intercept the remaining content of the specification, and output the content of each first-level title after completion.

[0011] Then, the intercepted part is traversed line by line to match the secondary title. The secondary title represents the further subdivided content in the sub-project, or the main control project or general project that should be distinguished during the quality inspection. After matching the secondary title, traverse line by line to find the next secondary title, and complete the interception of the standard content between the two secondary titles. If the secondary title is the last section of the part, the remaining content is intercepted, and the content of each secondary title is output, filtering out the general provisions in the secondary title that do not involve specific quality inspection points.

[0012] Finally, in the content intercepted by the second-level title, we traverse line by line to match the third-level title and its corresponding inspection quantity, inspection method and other standard contents. The third-level title covers the detailed contents of quality inspection during the construction process. After completion, the matched contents are output in turn, and the above process is repeated continuously until the structured processing of the entire specification is completed, and the corresponding text information is stored in the Excel file to facilitate the subsequent research work.

[0013] (2) Constructing a labeling system for quality inspection information

[0014] There is no corpus that can be directly used for the extraction of checkpoints in the construction quality inspection specification. Therefore, the present invention constructs a labeling system for the named entity recognition (NER) task suitable for construction quality inspection, and determines the label types of eight named entities, namely: type, material, inspection subject, inspection sub-item, activity, attribute, location, and tense, and limits and refines the construction quality checkpoint information through the combination of these eight labels. The information of the quality checkpoint is further divided into four categories: checkpoint limiting conditions, checkpoint core information, checkpoint refinement information and other related information, among which "checkpoint limiting conditions" includes two labels of type and material, "checkpoint core information" includes the inspection subject label, "checkpoint refinement information" includes three labels of inspection sub-item, activity and attribute, and "other related information" includes two labels of location and tense.

[0015] "Inspection subject": The inspection subject is the information core of the construction quality inspection, and is a specific item for inspection and acceptance of specific building components or systems based on relevant specifications and standards. During the construction process, the inspection subject covers both building components such as load-bearing walls, beams, columns, and floor slabs, as well as functional systems such as waterproofing systems and ventilation systems. It is a key information node to ensure that the construction quality meets the requirements.

[0016] "Type": During the construction process, different building components or systems have their own quality requirements. The type label is used to describe the function, purpose or appearance of the inspection subject, so as to distinguish inspection subjects of different forms or purposes. For example, in "Equipment and materials used in monitoring and controlling energy-saving projects should be inspected upon arrival", "Equipment and materials used in monitoring and controlling energy-saving projects" describes the purpose of the inspection subject "equipment and materials", so it can be classified as a "type" entity.

[0017] "Material": Material labels are used to identify the material composition of the inspection subject or its components, and involve specific materials used in building components or systems. The materials used in the inspection point components directly affect the standards and methods of construction quality inspection. Different materials have different construction requirements and inspection standards. Clarifying the inspection point materials will help the entire construction quality inspection process.

[0018] "Inspection sub-item": Inspection sub-item label is a sub-inspection item under the inspection subject, which is used to describe the specific quality requirements or detailed content of the subject. Inspection sub-item is a further decomposition of the inspection subject, which is used to refine the inspection content of the subject in the specification.

[0019] "Attributes": Attribute tags describe certain physical or functional characteristics that the subject of the inspection should have, which are used to measure whether the quality level of the subject in a certain aspect meets the standard. For example, the "yield strength" and "bending performance" of steel bars. Attribute tags are used to help define and quantify the specific requirements of the inspection object to ensure that the functionality and safety of building components or systems meet the corresponding standards.

[0020] "Activity": The activity label indicates the specific construction actions or operation procedures related to the inspection subject during the construction process, and is used to identify construction behaviors related to quality inspection. Different construction activities will trigger different inspection requirements, which can help inspectors identify the construction behaviors that need to be inspected. For example, the installation of steel bars, the pouring of concrete, etc.

[0021] "Tense": The tense label is used to describe the time nodes that need to be inspected during the construction process, such as "after the steel bars are delivered to the site" and "before the concrete is poured", as well as the information on the corresponding judgment conditions that the inspection subject needs to meet, such as "when the bottom of the window well is above the highest groundwater level..." Here, "above the highest groundwater level" meets the information requirements of the tense label.

[0022] "Location": The location label is used to describe the physical spatial location of the inspection subject or the specific location relative to other components, and to identify the specific spatial information of the inspection object. For example, "the formwork and support at the post-casting strip should be set independently", and the location information here is "at the post-casting strip".

[0023] (3) Training RoBERTa-wwm-BiLSTM-CRF recognition model

[0024] A data set is established based on the annotation system of the quality inspection information in step (2), and the RoBERTa-wwm-BiLSTM-CRF model is trained. The model is used to recognize and predict the standard text after structured extraction in step (1) and parse the corresponding quality inspection information.

[0025] The RoBERTa-wwm-BiLSTM-CRF model consists of the RoBERTa-wwm module, the BiLSTM module, and the CRF module, which are connected in sequence. In the RoBERTa-wwm module, RoBERTa, as an improved version of BERT, cancels the next sentence prediction (NSP) task, expands the training data set, and uses a dynamic masking mechanism to improve the stability and efficiency of the model and enhance the training effect; using the WWM method, all characters of a word are masked at the same time, rather than a certain character, so that the model can better understand the overall meaning of the word when processing text. The BiLSTM module consists of two LSTM networks, which process the forward and reverse information of the input sequence respectively. Through BiLSTM, the text features can be further enhanced to provide a more complete and accurate feature representation for downstream tasks. The CRF module learns and infers the transition probability and dependency of each label in the sequence by modeling the conditional dependency between labels, ensuring that the generated label sequence is logical, avoiding the problem of BiLSTM local optimal solution, and obtaining a more accurate and stable labeling effect.

[0026] The standard text sentence is used as input, and the text information is first embedded through the RoBERTa-wwm module, and then converted into a vector sequence for output to the BiLSTM module. BiLSTM uses its bidirectional structure to capture information from two directions (forward and reverse), further extract contextual information and semantic features in the text, and output a sequence containing more comprehensive sentence information. The output of BiLSTM is used as the input of the CRF layer to calculate the conditional probability of subsequent labels given the current label condition to consider the dependency between labels, so as to more accurately predict the target label sequence. After processing by the CRF layer, the model outputs the final annotation sequence to complete the parsing.

[0027] Beneficial effects of the present invention: The present invention constructs a standardized structured extraction method through regular expressions, designs a label system for named entity recognition, and establishes a corresponding data set. The RoBERTa-wwm-BiLSTM-CRF model is used to automatically extract quality information to fill the data gap in this field, thereby promoting the automation and intelligence of construction quality inspection work. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 A flow chart for the structured extraction of construction quality inspection specifications.

[0029] Figure 2 Schematic diagram of the labeling system for quality inspection information.

[0030] Figure 3 Schematic diagram of Bi-LSTM structure.

[0031] Figure 4 Schematic diagram of BIO labeling strategy.

[0032] Figure 5 It is the indicator for identifying various types of labels of the RoBERTa-wwm-BiLSTM-CRF model. DETAILED DESCRIPTION

[0033] The embodiments of the present invention are described in detail below with reference to the accompanying drawings and examples.

[0034] The intelligent analysis method for construction quality inspection specifications for automatic extraction of inspection information of the present invention comprises the following specific steps:

[0035] Step 1: Use rule-based methods to extract the regulatory text in a structured manner and store the extracted text content. The specific steps are as follows:

[0036] Step 1.1 Remove the preface, general principles, terminology, basic regulations and appendix of the construction quality inspection specification, retain only the main text of the specification document, and remove the corresponding watermarks, headers and footers.

[0037] Step 1.2: Use OCR to convert the text content in the image into editable and searchable text information, extract the selected standard body part, store it in TXT format, and manually correct any typos that may exist during the extraction process.

[0038] Step 1.3 Follow Figure 1 The method shown traverses the specification clauses line by line and exports the intercepted parts.

[0039] Step 2: According to the labeling system established by the present invention (such as Figure 2 As shown in Figure 2, the BIO annotation strategy is used to annotate the construction quality inspection specifications to form a construction quality inspection named entity recognition dataset. The BIO annotation strategy is as follows: Figure 4 .

[0040] Step 3: Use the dataset to train the RoBERTa-wwm-BiLSTM-CRF model to obtain the named entity recognition model. The specific steps are as follows:

[0041] Step 3.1: Divide the data into training set, test set and validation set according to the proportion, which are used for training the model, adjusting parameters and evaluating the model respectively.

[0042] Step 3.2: Establish the ROBERTA-WWM-LSTM-CRF model and train it. The Bi-LSTM structure is as follows: Figure 3 shown.

[0043] Step 3.3 uses precision P (Precision), recall R (Recall) and F1 value (F-score) as evaluation indicators of the model effect experiment, where precision P refers to the ratio between the number of correctly identified entities and the total number of correctly identified entities, recall R refers to the ratio between the number of correctly identified entities and all entities, and F1 value combines precision P with recall R to comprehensively reflect the overall indicator.

[0044]

[0045]

[0046] Among them, TP represents the number of correctly recognized entities, FP represents the number of incorrectly recognized entities, and the number of incorrectly recognized entities is marked as FN.

[0047] Step 4: Use the trained model to predict the extracted specification clauses, automatically identify the construction quality inspection information, and store it in an EXCEL file.

[0048] Example

[0049] In order to verify the effectiveness of the method proposed in the invention, the "Concrete Structure Engineering Construction Quality Acceptance Code" (GB50204-2015) is used as an example to extract construction quality inspection information. This embodiment uses the deep learning framework Pytorch, uses Python 3.10 for development, and uses the GPU NVDIA GeForce RTX2060.

[0050] First, some current construction quality acceptance specifications were used as training data sets. The method described in step 1 was used to process these specifications. A total of 964 specification texts were obtained, and a total of 3550 label objects were marked. The data was divided into training set, test set and validation set in a ratio of 8:1:1. In order to verify the effectiveness of the RoBERTa-wwm-BiLSTM-CRF model, it was compared with the mainstream BERT-BiLSTM-CRF model, and the precision (P), recall (R) and F1 value were selected as evaluation indicators.

[0051] Table 1 Experimental results

[0052]

[0053] The RoBERTa-wwm-BiLSTM-CRF model achieved excellent performance in the test set, with a precision of 91.58, a recall of 92.87, and an F1 value of 92.22. All three evaluation indicators are higher than the BERT-BiLSTM-CRF model. Specifically, in terms of named entity labeling, the performance of RoBERTa-wwm-BiLSTM-CRF is also better than that of BERT-BiLSTM-CRF, which shows the effectiveness of the method model used. Figure 5 .

[0054] The "Concrete Structure Engineering Construction Quality Acceptance Code" (GB50204-2015) was selected. First, according to the method described in step 1, the specification text was preprocessed and then structured extracted using a Python script. Regular expressions were used to match the corresponding first-level titles, second-level titles, third-level titles, inspection quantities and inspection methods, as shown in Table 2, where "\u4e00-\u9a5" represents the range of Chinese characters that match the Unicode character encoding.

[0055] Table 2 Regular Expression

[0056]

[0057] Through the structured processing of the specification clauses, six inspection sub-projects with a total of 94 quality inspection information were obtained. The six sub-projects are formwork sub-project, steel bar sub-project, prestressed sub-project, concrete sub-project, cast-in-place structure sub-project, and prefabricated sub-project. The matched quality inspection information is stored in the "Quality Inspection" content column of EXCEL, followed by the "Inspection Quantity" and "Inspection Method" columns to store the corresponding specification information.

[0058] The inspection information is extracted using the trained RoBERTa-wwm-BiLSTM-CRF model. Through the Python script, the "Inspection Content" column in EXCEL is used as text information and input into the model for prediction. The prediction results output by the model are automatically written into the "Prediction Result" column corresponding to the "Inspection Content" column, completing the analysis of the inspection information.

[0059] Taking Article 4.2.3 under the category of "4.2 Formwork Installation" in the "Formwork Sub-project" of the specification as an example, its inspection content is "The formwork and bracket at the post-casting strip should be set independently". Through the structured extraction of the specification, the inspection quantity is "full inspection" and the inspection method is "observation" has been automatically stored in the corresponding position in EXCEL. Then the model is used to parse and predict the inspection content, and the quality inspection information is successfully obtained. Table 3 is the quality inspection information predicted by the model. Through the structured extraction method proposed in the present invention, using the designed information annotation system, using the RoBERTa-wwm-BiLSTM-CRF model, the inspection information extraction and analysis of all 94 specification clauses in the "Concrete Structure Engineering Construction Quality Acceptance Code" (GB50204-2015) were successfully completed, proving the effectiveness and feasibility of the method proposed in the present invention.

[0060] Table 3 Model prediction results

[0061]

[0062] The present invention successfully completes the extraction and analysis of construction quality inspection point information, can process the current quality inspection specifications, fills the gap in the industry, improves the efficiency of construction quality inspection, and lays a solid foundation for the automated management of construction quality.

Claims

1. An intelligent analysis method for construction quality inspection specifications for automatic extraction of inspection information, characterized in that: The following steps are involved: (1) Structured extraction of construction quality inspection specifications First, the quality inspection specification document needs to be preprocessed, and the preface, general principles, terms, basic regulations and appendix parts of the construction quality inspection specification need to be removed, and only the main text of the document needs to be retained, and the corresponding watermarks, headers and footers need to be removed; After the preprocessing process is completed, a specification structured extraction algorithm is designed using regular expressions according to the text structure of the current construction quality inspection specification; the specification text structured extraction algorithm is as follows: Input the pre-processed construction quality inspection specification and traverse the text content line by line; first match the first-level title, which refers to the name of a large chapter in the construction quality inspection specification, that is, a specific sub-project; when the first-level title is matched, continue to traverse downward line by line to find the next adjacent first-level title, and complete the interception of the specification content between the two first-level titles. If the first-level title is the last section, intercept the remaining content of the specification, and output the content of each first-level title after completion; Then, the intercepted part is traversed line by line to match the secondary title. The secondary title represents the further subdivided content in the sub-project, or the main control items or general items that should be distinguished during quality inspection. After matching the secondary title, traverse line by line to find the next secondary title, and complete the interception of the standard content between the two secondary titles. If the secondary title is the last section of the part, the remaining content is intercepted, and the content of each secondary title is output, and the general provisions in the secondary title that do not involve specific quality inspection points are screened out; Finally, in the content intercepted by the second-level title, traverse line by line to match the third-level title and its corresponding inspection quantity, inspection method and other standard contents. The third-level title covers the detailed content of quality inspection during the construction process. After completion, the matched content is output in turn, and the above process is repeated continuously until the structured processing of the entire specification is completed; (2) Constructing a labeling system for quality inspection information A labeling system for named entity recognition tasks suitable for construction quality inspection is constructed, and eight types of named entity labels are determined, namely: type, material, inspection subject, inspection sub-item, activity, attribute, location, and tense. The information of construction quality checkpoints is limited and refined through the combination of these eight labels; the information of quality checkpoints is further divided into four categories: checkpoint restriction conditions, checkpoint core information, checkpoint refinement information, and other related information, among which "checkpoint restriction conditions" include two labels: type and material, "checkpoint core information" includes inspection subject label, "checkpoint refinement information" includes three labels: inspection sub-item, activity, and attribute, and "other related information" includes two labels: location and tense; (3) Training RoBERTa-wwm-BiLSTM-CRF recognition model A data set is established based on the annotation system of the quality inspection information in step (2), and a RoBERTa-wwm-BiLSTM-CRF model is trained. The model is used to recognize and predict the standard text after structured extraction in step (1), and to parse the corresponding quality inspection information. The RoBERTa-wwm-BiLSTM-CRF model consists of a RoBERTa-wwm module, a BiLSTM module, and a CRF module, and the three modules are connected in sequence; The standard text sentence is taken as input, and the text information is first embedded through the RoBERTa-wwm module, and then converted into a vector sequence for output to the BiLSTM module; BiLSTM uses its bidirectional structure to capture information from both the forward and reverse directions, further extracts contextual information and semantic features in the text, and outputs a sequence containing more comprehensive sentence information; the output of BiLSTM is used as the input of the CRF layer to calculate the conditional probability of subsequent labels given the current label condition, to consider the dependency between labels, so as to more accurately predict the target label sequence. After processing by the CRF layer, the model outputs the final annotation sequence to complete the parsing.

2. The intelligent analysis method for construction quality inspection specifications for automatic extraction of inspection information according to claim 1 is characterized in that: In step (2), the eight tags are as follows: "Inspection subject" refers to the specific project of inspection and acceptance of specific building components or systems in accordance with relevant specifications and standards; "Type": used to describe the function, purpose or appearance of the inspection subject, so as to distinguish inspection subjects of different forms or purposes. "Material": used to identify the material composition of the inspection subject or its components, involving the specific materials used in building components or systems; "Inspection sub-item": used to describe the specific quality requirements or detailed content of the subject; "Attributes": describe certain physical or functional characteristics that the inspection subject should possess, and are used to measure whether the quality level of the subject in a certain aspect meets the standard; "Activity": refers to the specific construction actions or operation procedures related to the inspection subject during the construction process, and is used to identify construction behaviors related to quality inspection; "Tense": used to describe the time points during the construction process when inspections need to be carried out; "Position": used to describe the physical spatial location of the inspection subject or its specific position relative to other components, and to identify the specific spatial information of the inspection object.

Images