Nuclear power valve design document information recognition and collection method and system

Abstract

The application provides a nuclear power valve design file information recognition and collection method and system, the collection method comprising receiving a nuclear power valve design file, and pre-processing the nuclear power valve file based on a CNN valve design file image recognition model to meet the format and geometric requirements, to obtain a target valve design file; identifying the information of the to-be-identified area of the target valve design file based on a CRNN valve drawing information recognition model; creating a corpus analysis library using the recognition result of step 3, and carrying out part-of-speech tagging and Chinese word segmentation processing on the corpus analysis library; carrying out text analysis on the processed corpus result of the nuclear power valve of step 4, extracting high-frequency parameters and key parameters as the processing basis for subsequent text keyword extraction, and carrying out text vectorization processing; filtering invalid information according to the text vectorization similarity, judging the correctness of the key information that has completed information integration, and then operating the correct information into the valve drawing information database.

Description

Technical Field

[0001] This invention relates to the field of nuclear power valve design technology, specifically to a method and system for identifying and collecting information from nuclear power valve design documents. Background Technology

[0002] Nuclear power valves are crucial equipment in the reactor loop of nuclear power plants, used in large quantities and with wide applications. They also encompass various types, such as gate valves, globe valves, ball valves, butterfly valves, check valves, and safety valves. The design information varies across different valve design document types, resulting in numerous design documents for different nuclear power valves. These documents cover a wide range of design requirements and information, including detailed specifications for applicable scope, standards and regulations, functional requirements, installation requirements, structural requirements, testing / qualification requirements, manufacturing requirements, quality assurance requirements, packaging / transportation / storage requirements, and documentation requirements.

[0003] Therefore, if the requirements in the nuclear power valve design documents are extracted and processed into two columns—a major category of design requirements and a specific design requirements column—and a design document information database is formed, the feasibility of the nuclear power valve design document requirements will be effectively improved. Furthermore, it will enable information management, data screening, and information correction operations for the nuclear power valve design requirements.

[0004] Furthermore, as an important component of nuclear power pipeline systems, valves may serve as critical equipment when the system performs its functions. Therefore, the optimization of the equipment design of nuclear power valves restricts the optimization of the design of nuclear power system functions. Improving the executability of the design document information will greatly help to clarify specific design requirements, distinguish the design needs of different types of valves, and ensure that the design requirements are correct and reliable.

[0005] The information contained in nuclear power plant valve design documents can be broadly categorized into: design requirements, detailed design information, tabular parameters, image parameters, and other information. Design requirements and detailed design information are text-based parameters, tabular parameters are either textual or numerical parameters, and the rest are classified as other parameters. Text-based parameters are crucial design information specified in nuclear power plant valve design documents.

[0006] Existing design document information acquisition systems or methods can identify parameter information in design documents, but they still have the following problems or shortcomings:

[0007] First, existing design solutions are mainly designed for industrial valve design documents and cannot meet the special requirements of nuclear power valve design documents, such as the requirement for nuclear power valves to undergo qualification tests compared to general industrial valves. Therefore, existing design methods are difficult to solve the engineering problems of nuclear power valve equipment in my country and lack transferability to nuclear power valve equipment design.

[0008] Second, the existing solution mainly focuses on the modeling process, but in reality, the model needs to be trained based on the drawings. Therefore, the drawings used in the model training will affect its recognition ability. The model cannot rely on new information to carry out self-learning and self-enhancement. It lacks iterative upgrades of the parameter information of the design documents. The actual capabilities of the model cannot be fundamentally updated. Therefore, the current solution lacks the ability to upgrade and iterate parameters, and its actual model capabilities cannot be updated and upgraded.

[0009] Third, the existing scheme focuses on the collection and identification of information in design documents, without carrying out the classification, storage, management and application of information. The parameter resources of nuclear power valve design documents are not used efficiently, resulting in a waste of parameter information resources. At the same time, there is a lack of parameter judgment on the design document information. Only the identification work is done, without information evaluation.

[0010] Fourth, the existing schemes use inefficient identification models and methods with limited processing capabilities and lack effectiveness in identifying nuclear power plant valve design documents. Furthermore, the current schemes have relatively simple information processing for design document segments, insufficient analysis, and no established workflow for text information processing. They also lack a comprehensive processing model that includes segmentation, word analysis, part-of-speech tagging, and information mining, thus limiting the capabilities of the model.

[0011] Based on this, the inventors of this application have designed a method and system for identifying and collecting information from nuclear power plant valve design documents to overcome at least one of the aforementioned technical problems. Summary of the Invention

[0012] The technical problem to be solved by the present invention is to overcome the shortcomings of low efficiency and poor effectiveness in the extraction and identification of nuclear power valve design document information in the prior art, and to provide a method and system for identifying and collecting nuclear power valve design document information.

[0013] The present invention solves the above-mentioned technical problems through the following technical solution:

[0014] The first aspect of this invention provides a method for identifying and collecting information from nuclear power plant valve design documents, characterized in that it includes:

[0015] Step 1: Receive the nuclear power valve design file and preprocess the nuclear power valve file according to the format and geometric requirements based on the CNN valve design file image recognition model to obtain the target valve design file;

[0016] Step 2: The valve drawing information recognition model based on CRNN identifies the information of the area to be identified in the target valve design document; wherein, the valve drawing information recognition model based on CRNN is trained using a valve drawing information database;

[0017] Step 3: Create a corpus analysis database using the recognition results of Step 3, and perform part-of-speech tagging and Chinese word segmentation on the corpus analysis database;

[0018] Step 4: Perform text analysis on the processed corpus results of the nuclear power valve in Step 4, extract high-frequency parameters and key parameters as the basis for subsequent text keyword extraction, and perform text vectorization processing.

[0019] Step 5: Filter invalid information according to text vectorization similarity, judge the correctness of key information that has been integrated, and then input the correct information into the valve drawing information database.

[0020] According to an embodiment of the present invention, step 1 includes:

[0021] Step 11: Use OpenCV's imread() function to read the nuclear power plant valve design file;

[0022] Step 12: Check and judge the format of the nuclear power valve design file, and convert the non-JPG format of the nuclear power valve design file to JPG format;

[0023] Step 13: Perform file angle checks and file pixel height transformations based on pixel coordinates on the nuclear power valve design file after format conversion in Step 12 using pixel operations;

[0024] Step 14: Use the CNN valve design document image recognition model to recognize the nuclear power valve design document that has been checked and transformed in Step 13, and remove the image information from the target valve design document;

[0025] Step 15: Use OpenCV's merge() method to merge the nuclear power valve design file after the image information recognition and removal in Step 14 into three channels;

[0026] Step 16: Perform bilateral filtering on the nuclear power valve design document obtained after channel merging in Step 15;

[0027] Step 17: Perform binarization thresholding on the target valve design file that has been filtered in step 16 to obtain the target valve design file.

[0028] According to an embodiment of the present invention, the valve drawing information recognition model of the CRNN in step 2 includes a CRNN convolutional layer, a CRNN recurrent layer and a CRNN transcription layer;

[0029] The CRNN convolutional layer is used to extract text features of the information to be recognized; the CRNN recurrent layer is used to extract sequence features and generate a probability feature sequence; the CRNN transcription layer is used to predict the probability feature sequence and convert it into a character sequence.

[0030] According to one embodiment of the present invention, the CRNN convolutional layer is used to extract the features of the information to be identified in the input design document, and the convolutional feature map is obtained after convolution, pooling and normalization processing, and is used as the input of the CRNN recurrent layer;

[0031] The CRNN recurrent layer uses gated recurrent units to calculate the probability feature sequence, which is then used as the input to the CRNN transcription layer.

[0032] The CRNN transcription layer uses a connection-time classification loss function to convert the predicted probability feature sequence input to the CRNN recurrent layer into a character sequence.

[0033] According to an embodiment of the present invention, step 3 includes:

[0034] Step 31: Receive the recognition result information from Step 2 and save it as a txt text file for use as the input file for the NLP corpus; wherein, the corpus is created and the recognition information result file in the corpus is obtained through the NLTK fileids() method;

[0035] Step 32: Perform simplified / traditional Chinese conversion on the text in the corpus;

[0036] Step 33: Remove stop words from the text segments in the corpus;

[0037] Step 34: Deconstruct the text information of the corpus segment and separate the words in the segment;

[0038] Step 35: Perform part-of-speech tagging on the words in the corpus text.

[0039] According to an embodiment of the present invention, step 4 includes:

[0040] Step 41: Receive the processed expected result information from Step 3 and extract high-frequency words from the text segment;

[0041] Step 42: Extract keywords from the text using NLP keyword extraction methods and form a keyword sequence Igj;

[0042] Step 43: Extract key numerical parameters from the passage;

[0043] Step 44: Train text paragraph vectors using the Python gensim library based on the Doc2Vec model.

[0044] According to an embodiment of the present invention, in step 42, the NLP employs the TF-IDF keyword extraction algorithm.

[0045] According to an embodiment of the present invention, step 5 includes:

[0046] Step 51: Receive the analyzed and processed information from step 4, and filter similar information in the key information sequence based on text vectorization similarity;

[0047] Step 52: Based on the existing key information sequence, the high-frequency sequence Igp and the keyword sequence Igj are selected as text keywords for the current text segment according to certain principles, and the text keyword set is output to represent the main meaning of the current text segment.

[0048] Step 53: Use a key information matching mechanism to evaluate and compare key information;

[0049] Step 54: Integrate the information data evaluated in Step 53 and use it to guide the design of the CRNN document information recognition model for further training.

[0050] A computer program product according to a second aspect of this application includes a computer program that, when executed by a processor, implements the steps performed by a computer as described in the first aspect.

[0051] According to a third aspect of this application, a computer-readable storage medium has a computer program that, when executed by a processor, implements the steps performed by a computer as described in the first aspect.

[0052] A data acquisition system according to a fourth aspect of this application includes: a memory capable of storing instructions executable by a processor; and a processor capable of executing the instructions to perform the steps performed by a computer as described in the first aspect of the method.

[0053] The positive and progressive effects of this invention are as follows:

[0054] This invention relates to a method for identifying and collecting information from nuclear power plant valve design documents. By sequentially preprocessing, identifying, analyzing, sorting, and storing the nuclear power plant valve design documents, the method efficiently completes the identification and collection of information from nuclear power plant valve design documents. At the same time, it performs model enhancement to ensure the adaptability and applicability of the model's self-learning and self-enhancement. This method can provide technical support for nuclear power plant valve selection, document review, equipment design, system optimization, data verification, and other related work. Attached Figure Description

[0055] The above and other features, properties and advantages of the present invention will become more apparent from the following description taken in conjunction with the accompanying drawings and embodiments, wherein:

[0056] Figure 1 This is a flowchart of the nuclear power plant valve design document information identification and acquisition method of the present invention;

[0057] Figure 2 A flowchart for document information processing was designed for this invention;

[0058] Figure 3 This is a schematic diagram of geometric rotation inspection of the design documents for this invention;

[0059] Figure 4 A flowchart for document information recognition was designed for this invention;

[0060] Figure 5 A flowchart for processing document text corpus was designed for this invention;

[0061] Figure 6 A flowchart for document text corpus analysis was designed for this invention;

[0062] Figure 7 A flowchart for evaluating document text information is designed for this invention;

[0063] Figure 8 This is a schematic diagram illustrating the integration and storage of text information from design documents into a database for this invention.

[0064] Figure 9 This is a schematic diagram of the structure of the electronic device of the present invention. Detailed Implementation

[0065] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0066] Embodiments of the invention will now be described in detail with reference to the accompanying drawings. Preferred embodiments of the invention will now be described in detail, examples of which are illustrated in the drawings. Wherever possible, the same reference numerals will be used in all the drawings to denote the same or similar parts. Furthermore, although the terminology used herein is selected from commonly known and used terminology, some terms mentioned in this specification may have been chosen by the applicant at his or her discretion, and their detailed meanings are explained in the relevant sections of the description herein. Moreover, the invention should be understood not only by the actual terms used, but also by the meaning implied by each term.

[0067] Reference Figure 1 This invention proposes a method for identifying and collecting information from nuclear power plant valve design documents, including:

[0068] Step 1: Receive the nuclear power valve design file and preprocess the nuclear power valve file according to the format and geometric requirements based on the CNN valve design file image recognition model to obtain the target valve design file.

[0069] It should be noted that step 1 corresponds to the text information processing module of the present invention. The text information processing module is used to receive nuclear power valve design documents, then perform pre-processing work on the documents, and convert the nuclear power valve design documents into a document format that meets the required format and geometry.

[0070] Please refer to the detailed procedures. Figure 2 Step 1 includes:

[0071] Step 11: Use OpenCV's imread() function to read the nuclear power valve design file.

[0072] Step 12: Check and judge the format of the nuclear power valve design file, and convert the non-JPG format of the nuclear power valve design file to JPG format.

[0073] The corresponding design file format checking submodule is used to check and judge the file format. For example, for valve design files in "*.pdf" format, it performs file format conversion and converts them into image files (JPG format). That is, Figure 2 In the file format check node, if the format check is qualified, the file geometry check will proceed; if it is not qualified, the valve drawing file format will be converted, and the file geometry check will be performed after the conversion is completed.

[0074] Step 13: Perform file angle checks and file pixel height transformations based on pixel coordinates on the nuclear power valve design file after format conversion in Step 12 using pixel operations.

[0075] The corresponding design file geometry check submodule performs file angle checks and pixel height transformations based on pixel coordinates.

[0076] Specifically, the geometry check submodule of this design file uses Python Numpy to perform pixel operations on the design file and obtain the vertical coordinates of the first character on the left side of the design file. For example... Figure 3As shown, if all coordinates yield a pixel value of 0 (i.e., the three-channel pixel values ​​B=G=R=0), it indicates a pure black pixel, meaning all characters in the current design file are in a vertical position, requiring no rotation geometry operation. Otherwise, the characters in the current design file are skewed, which is not conducive to information recognition and extraction. It is necessary to solve the deflection angle of the character at this time and perform rotation geometry operation to correct the deflection. Using the previous column of pixel coordinates as the reference value, obtain the pixel coordinate position with a pixel value of 0 in the right column. At this time, the two pixel coordinates form an angle. Calculate the angle between the pixel coordinate position and the vertical direction. At the same time, use the OpenCV getRotationMatrix2D() method to rotate the design file with the lower left corner of the design file as the rotation coordinate center point and the obtained deflection angle as the rotation angle.

[0077] The `size()` method of OpenCV is used to obtain the pixel values ​​of the width and height of the design file. Since the CRNN model that recognizes the information in the design file requires the geometric pixel height of the design file input to be fixed at 32, the design file geometry check submodule uses the `dsize()` method of OpenCV to perform geometric transformation on the input file that does not meet the pixel height requirement in order to fix the file pixel information.

[0078] Step 14: Use the CNN valve design document image recognition model to recognize the nuclear power valve design document that has been checked and transformed in Step 13, and remove the image information from the target valve design document.

[0079] The corresponding design file image information recognition submodule removes image information from the design file.

[0080] Since this invention targets the character parameter information in the design documents of nuclear power valves for information recognition and extraction, while the image information in the design documents is mainly envelope image parameters or images inserted for contextual understanding, the image information in the design documents does not belong to the information to be recognized and extracted, does not need to be recognized and processed, and will interfere with the recognition and acquisition of character parameter information, or even cause information recognition failure.

[0081] The design document image information recognition submodule adopts a valve design document image recognition model based on CNN (Convolutional Neural Network). It uses existing valve image data for model training and then uses it for image recognition in this module. After recognition, it selects and locates the image position in the design document, and uses NumPy pixel operations to set the image pixel value within the location box to 255, that is, to turn the area at that position into pure white, so as to remove the corresponding image information. After this processing, the relevant information will not interfere with the recognition and acquisition of character parameter information.

[0082] Step 15: Use OpenCV's merge() method to merge the nuclear power valve design file after the image information recognition and removal in Step 14 into three channels.

[0083] The corresponding design file pixel channel merging submodule, after the design file image information is identified and removed, does not need to use three-way RGB channels. This submodule uses OpenCV's merge() method to merge the three-way channels of the design file in order to carry out subsequent design file image filtering and pixel thresholding processing.

[0084] Step 16: Perform bilateral filtering on the nuclear power valve design document that has been merged in Step 15.

[0085] The corresponding design file filtering submodule performs bilateral filtering on the design file after pixel channel merging. This removes noise and reduces detail layer information while effectively preserving text boundary information, thus smoothing the image. This submodule uses OpenCV's `bilaterFilter()` method, with the pixel centroid located at the design file's centroid pixel coordinates, and the color calculation range set to 255.

[0086] Step 17: Perform binarization thresholding on the target valve design file that has been filtered in step 16 to obtain the target valve design file.

[0087] The corresponding design file pixel thresholding submodule can make the pixel information in the design file stand out after the design file has been filtered, forming a strong pixel contrast and distinction effect. Highlighting the pixel information in the design file helps to enhance the recognition effect during information recognition.

[0088] This submodule uses OpenCV's `threshold()` method, with a pixel value of 200 as the dividing line and 255 as the maximum value. Pixels with a value less than or equal to 200 are set to 0 and displayed as pure black, while pixels with a value greater than 200 are set to 255 and displayed as pure white.

[0089]

[0090] After pixel binarization thresholding of the design document, the pixel values ​​of the area to be identified in the design document show obvious differentiation, which will effectively improve the recognition efficiency of the design document information.

[0091] Step 2: The valve drawing information recognition model based on CRNN identifies the information of the area to be identified in the target valve design document; wherein, the valve drawing information recognition model based on CRNN is trained using a valve drawing information database.

[0092] Reference Figure 4 Step 2 corresponds to the text information recognition module, which is used to recognize the information in the area to be recognized in the design document. This module adopts a valve design document information recognition model based on CRNN to recognize the information in the processed design document recognition area.

[0093] The CRNN valve design document information recognition model consists of CRNN convolutional layers, CRNN recurrent layers, and CRNN transcription layers, which respectively extract textual features from the information to be recognized, extract sequence features and generate probabilistic feature sequences, predict probabilistic feature sequences, and convert them into character sequences. The design document information recognition module, based on the CRNN valve design document information recognition model, can be divided into three sub-modules.

[0094] Specifically, the processing flow of the text information recognition module is as follows:

[0095] The CRNN convolutional layer is used to extract the features of the information to be identified in the input design file. After operations such as convolution, pooling, and normalization, the convolutional feature map is obtained and used as the input of the CRNN recurrent layer.

[0096] The CRNN convolutional layer uses a convolutional neural network as the core tool for feature extraction. A typical CRNN convolutional layer consists of 6 convolution operations, 4 max pooling operations, and 2 batch normalization operations, for a total of 13 convolutional layer operations. The relevant operation parameters are shown in Table 1, where C is the number of channels, k is the kernel size, s is the kernel stride, p is the padding size, and w is the pooling kernel size.

[0097] Table 1 CRNN Convolutional Layer Operations

[0098] 1 convolution C:64,k:3×3,s:1,p:1 ReLU 2 Max pooling w:2×2,s:2 - 3 convolution C:128,k:3×3,s:1,p:1 ReLU 4 Max pooling w:2×2,s:2 - 5 convolution C:256,k:3×3,s:1,p:1 ReLU 6 convolution C:256,k:3×3,s:1,p:1 ReLU 7 Max pooling w:1×2,s:2 - 8 convolution C:512,k:3×3,s:1,p:1 ReLU 9 Batch normalization - - 10 convolution C:512,k:3×3,s:1,p:1 ReLU 11 Batch normalization - - 12 Max pooling w:1×2,s:2 - 13 convolution C:512,k:2×2,s:1,p:0 ReLU

[0099] The purpose of convolution is feature detection. The input matrix of the convolution operation represents the image region to be identified, and the matrix elements correspond to image pixels, representing their grayscale information. The convolution kernel slides across the image with a stride, and convolution actually performs matrix inner product calculations. After convolution calculation, a feature map is obtained. Figure 1 It is usually a matrix or a tensor.

[0100] Pooling, also known as convergence or downsampling, is an operation for feature selection. Each element of the input matrix in the pooling operation identifies a feature, representing the detected value of that feature. The pooling kernel slides across the image with a stride, taking the maximum value from the matrix parameters covered by the sliding kernel. Pooling outputs a reduced matrix.

[0101] Batch normalization normalizes the net input of each layer in a neural network for each batch of samples. This operation performs a feature scaling transformation, which can accelerate the learning and convergence speed of the neural network and, to some extent, prevent gradient vanishing and gradient explosion. Gradient vanishing and gradient explosion can occur in the gradient descent calculation of neural networks, leading to parameter cessation of updates and parameter overflow, which in turn makes model learning difficult and computational convergence failure. Normalizing the input values ​​to keep them within a certain range, such as between 0 and 1, can reduce the difference in the value range, accelerate learning convergence, and suppress gradient vanishing and gradient explosion.

[0102] Activation functions are a crucial component of neural networks, used to establish nonlinear relationships between inputs and outputs, thereby enhancing the learning and expressive capabilities of the neural network. Real-world problems are generally nonlinear, and activation functions effectively represent these nonlinear relationships. Commonly used activation functions include the sigmoid function, the hyperbolic tangent function, and the rectified linear function.

[0103] This method uses the rectified linear function ReLU, whose derivative is always 1 in the positive interval, which can avoid the gradient explosion problem. At the same time, this function is simpler to calculate than other activation functions and can maintain good sparse activation performance, thus improving the generalization ability of the model. The relationship of this function is shown in the following formula.

[0104] f(x) = relu(x) = max(0,x);

[0105] The information to be identified undergoes multiple convolution, pooling, and normalization operations to extract textual features from the design document, resulting in a feature map that serves as the input to the CRNN recurrent layer.

[0106] The CRNN recurrent layer employs a Gate Recurrent Unit (GRU). GRU is a simplified network model of the Long Short-Term Memory Neural Network (LSTM), but with improved computational efficiency. Similar to LSTM, GRU has only two gates: an update gate and a reset gate. At each location in the network, these two gates and the state constitute a processing unit. GRU effectively represents and learns long-range dependencies, using the state and parameters of previous computational locations as input to synthesize the computational output of the current computational unit. The update gate controls and determines the state information from previous computational locations to be incorporated into the current computational unit, controlling parameter memory updates and propagation. The reset gate controls and determines the computational weight of the state from previous computational locations in the current unit's computation. When the output is close to 0, more information from previous computational locations is forgotten; when it is close to 1, more information is retained and propagated.

[0107] Compared to LSTM, GRU networks introduce fewer parameters and less complexity. They effectively solve the problems of vanishing and exploding gradients through gating mechanisms, maintaining computational efficiency while achieving high computational performance. They also avoid computational overfitting and are suitable for sequence computation and processing tasks.

[0108] The CRNN recurrent layer uses GRU to calculate the probability feature sequence, which is then used as the input to the CRNN transcription layer.

[0109] The CRNN transcription layer uses the Connectionist Temporal Classification (CTC) loss function to convert the predicted probability feature sequence input to the CRNN recurrent layer into a character sequence. Essentially, it finds the character sequence with the highest probability combination in the probability feature matrix as the final information recognition result.

[0110] In text recognition, the predicted result sequence does not have a fixed length, and the number of characters corresponding to different features is inconsistent. Traditional mapping function methods cannot solve the problem of variable-length text recognition, and variable sequence length text recognition is also one of the difficulties in designing text recognition models. In addition, blank positions or repeated character information may be introduced in the recurrent layer processing, which will lead to the distortion of recognition results and reduce the recognition ability of the model.

[0111] This invention's CRNN employs the CTC loss function, enabling end-to-end training that eliminates intermediate steps and focuses solely on the input and output of data, effectively improving recognition efficiency. Simultaneously, the CTC loss function introduces a whitespace "-" mechanism, inserting a whitespace "-" between repeated characters to separate redundant characters. Therefore, when the CRNN transcription layer converts the probabilistic feature sequence into a character sequence, it does not merge identical and consecutively occurring characters in the existing character sequence; instead, it first removes duplicate characters and whitespace "-" positions, effectively solving the character repetition and redundancy problem in text information recognition. This process is also known as the decoding process.

[0112] Meanwhile, the CTC loss function has low efficiency in recognizing similar-looking characters. Therefore, in the transcription layer, a function approach is adopted that mainly uses the CTC loss function and secondarily uses Center Loss to reduce recognition errors. When calculating Center Loss, the relative distance and class discreteness are considered to assist the CTC loss function in recognition, which plays an important supplementary role in the recognition of similar-looking characters by the CTC loss function.

[0113] The CRNN transcription layer removes duplicates from the probability distribution sequence of the recurrent layer and extracts the character with the highest probability to form the final recognition result.

[0114] Step 3: Create a corpus analysis database using the recognition results from Step 3, and perform part-of-speech tagging and Chinese word segmentation on the corpus analysis database.

[0115] The corresponding text corpus processing module is used to create a corpus analysis library from the identified results of the design documents, and to perform part-of-speech tagging, Chinese word segmentation, and other processing on it. This module will output the results information for text corpus analysis.

[0116] Reference Figure 5 The processing flow of the text corpus processing module is as follows:

[0117] Step 31: Receive the recognized result information input by the text information recognition module and save it as a txt text file, which will be used as the corpus input file for NLP (Natural Language Processing). This module creates a corpus and obtains the recognition information result file in the corpus through the fileids() method of NLTK (Natural Language Toolkit).

[0118] Step 32: Proceed to the text conversion submodule to perform text conversion operations on the corpus text. For text containing traditional characters in the passage, load the Converter library and use the Converter(“zh-hans”).convert() method to convert the traditional characters in the text.

[0119] Step 33: Proceed to the text segment stop word removal submodule to remove stop words from the text segment.

[0120] It should be noted that stop words are words that are unnecessary in a text passage, and their removal does not actually affect the understanding of the overall sentence or passage. Generally, Chinese text passages contain a large number of function words, pronouns, or spaces, verbs, and nouns without specific referential meaning. These should be removed during language processing and analysis to reduce processing difficulty and speed up computation. In this submodule, the stop word file stopword.txt is loaded to remove stop words from the text passage.

[0121] Step 34: Proceed to the Chinese word segmentation processing submodule to break down the text information and separate the words to facilitate text corpus analysis. The purpose of Chinese word segmentation is to divide the character sequence of a text into word sequences according to certain rules. A word is the smallest unit of language used in Chinese. In English, words are separated by spaces, so this space-separation method can be directly used for English word segmentation. However, this method cannot be used for Chinese texts because there are no explicit separators. Therefore, NLP-related Chinese word segmentation methods are needed to solve this problem. The jieba Chinese word segmentation method is a word segmentation method based on segmentation rules and statistical calculations. It supports multiple programming languages, including Python and C++. It mainly supports three working modes: precise mode, full mode, and search engine mode. Its word segmentation process mainly includes:

[0122] 1. Quickly scan the word graph based on the prefix dictionary (which can be created from an offline statistical dictionary), build a directed acyclic graph for possible word segmentation results, open up word segmentation paths, and establish node connections;

[0123] Second, the dynamic programming method is used to determine the path with the highest probability, and the highest probability is calculated in reverse to obtain the word segmentation path with the highest calculated probability, which is used as the result of word segmentation.

[0124] Third, HMM (Hidden Markov Model) is used to process out-of-vocabulary words (i.e., new words that have not appeared in the dictionary calculation and cannot be segmented by the word segmentation algorithm). The model state is used for calculation, and the Viterbi algorithm is used to calculate the optimal word segmentation path.

[0125] This submodule employs the jieba Chinese word segmentation method based on precise mode, using the jieba.cut() function to perform Chinese word segmentation on the text segment, with the cut_all parameter using the default setting (precise mode). After Chinese word segmentation, the text segment has a basic foundation for text analysis and processing.

[0126] Step 35: Proceed to the text segment part-of-speech tagging processing submodule to perform part-of-speech tagging processing on the words in the text segment.

[0127] Specifically, part-of-speech tagging is the process of clearly identifying the part of speech for each word in the word segmentation results, assigning a part-of-speech tag, and annotating it. In Chinese, the same word may have multiple parts of speech, and its meaning in a text can be completely different.

[0128] In the development of NLP, there are two types of part-of-speech tagging methods: rule-based and statistical. Rule-based tagging methods require the establishment of a rule base that expresses the context and relationship of a text, which is more cumbersome and laborious to implement; while statistical tagging methods automatically tag parts of speech through probability and statistical calculations, but are subject to semantic limitations of long-distance positions in the text context.

[0129] Therefore, this module adopts the jieba part-of-speech tagging method, which combines rule-based and statistical approaches. This method is efficient and highly adaptable. The main features of the jieba part-of-speech tagging method are:

[0130] 1. If the labeled word is a Chinese character, a directed acyclic graph is created based on the prefix dictionary, its maximum probability path is calculated, and its part of speech is searched in the prefix dictionary at the same time. If it is not found, it is labeled as "x" (part of speech unknown, such as punctuation marks). For words with unknown part of speech and out-of-vocabulary words, the HMM method is called to perform part of speech labeling.

[0131] 2. If the tagged word is not a Chinese character, regular expression will be used to determine it and assign the corresponding part of speech, where "x" represents an unknown part of speech, "m" represents a numeral, and "eng" represents an English word.

[0132] This submodule uses the NLP psg.cut() function to implement jieba part-of-speech tagging.

[0133] Step 4: Perform text analysis on the processed corpus results of nuclear power valves in Step 4, extract high-frequency parameters and key parameters as the basis for subsequent text keyword extraction, and perform text vectorization processing.

[0134] Please refer to Figure 6 Step 4 corresponds to the text corpus analysis module, which is used to perform text analysis on the processed corpus results of the nuclear power valve design documents, extract high-frequency parameters and key parameters as the basis for subsequent text keyword extraction, and perform text vectorization operation. This module will output the results information for information evaluation.

[0135] Specifically, step 4 includes:

[0136] Step 41: Receive the processed expected result information from step 3 and extract high-frequency words from the text.

[0137] In other words, this submodule receives the processed corpus results from the text corpus processing module and extracts high-frequency words from the text segment. High-frequency words appear frequently in the text segment, similar to trending words in news articles, indicating that they effectively represent the main meaning and important features of the segment. Therefore, high-frequency words are the foundation for extracting text keywords. This submodule uses the sorted() function to extract high-frequency words from the word segmentation results, where the number of high-frequency words extracted, Ngp, is set to 8% (rounded up) of the text segment's word count. After extraction, the high-frequency words are stored in the high-frequency word sequence Igp.

[0138] Step 42: Use NLP keyword extraction methods to extract keywords from the text and form a keyword sequence Igj.

[0139] The corresponding keyword extraction submodule uses NLP keyword extraction methods to extract keywords from the text and form a keyword sequence Igj. Keywords are words that can reflect the main idea or main content of the text and can effectively summarize the central meaning of the entire text.

[0140] NLP includes various keyword extraction algorithms, among which TF-IDF (Term Frequency-Inverse Document Frequency) is a classic algorithm based on statistical calculation. The main idea of ​​this algorithm is that the importance of a word increases with its frequency of occurrence in a passage and decreases with its frequency of occurrence in the corpus. Therefore, the TF-IDF algorithm calculates importance from two parameters: term frequency and inverse document frequency.

[0141] I. Word frequency reflects the expressive power of words. It is the statistical result of how often a word appears in a document. The higher the frequency of a word in a document, the stronger its expressive power. Word frequency is defined as:

[0142]

[0143] Where, N i,j Indicates the frequency of a word in document jj. This indicates the total number of words in document jj.

[0144] II. Inverse document frequency (IVF) reflects the distinguishing ability of words. It is a statistical result of the frequency of a word appearing in documents within a document set. The less frequently a word appears in documents within a document set, the stronger its distinguishing ability. IVF is defined as:

[0145]

[0146] In the formula, |D| represents the total number of documents in the document set, and |{j:t i ∈d j}| represents document d in the document set. jThe number of documents containing the word is calculated by adding 1 to the denominator to avoid it being zero.

[0147] TF-IDF algorithm parameters f TF-IDF Defined as the product of term frequency and inverse document frequency, to comprehensively consider the importance of both to document analysis, i.e.:

[0148] f TF-IDF = tf * idf;

[0149] This submodule calculates term frequency and inverse document frequency during text processing and analysis, and solves for the TF-IDF algorithm parameter f. TF-IDF After calculation, the keyword sequence Igj is extracted from the calculation result, and the number of extracted keywords Ngp is set to 8% of the word count of the text (rounded up).

[0150] Step 43: Extract key numerical parameters from the passage.

[0151] The corresponding key data parameter extraction and processing submodule extracts key numerical parameters from the text segment.

[0152] For certain key numerical parameters in nuclear power plant valve design documents, such as the power supply parameters of the actuator, the test pressure required for testing, the test time, and the acceptance criteria, these are considered important and critical numerical parameters and should be carefully selected and retained. This submodule uses regular expressions to make judgments, employing the concordance() function based on context information to filter numerical terms from each input segment, determine the key numerical parameters, and extract the key numerical sequence Isz from the results.

[0153] Step 44: Train text paragraph vectors using the Python gensim library based on the Doc2Vec model.

[0154] The corresponding text vectorization processing submodule provides vectorized text information for subsequent text similarity evaluation and text information assessment.

[0155] Text vectorization is used to represent text as machine-readable vectors that express its meaning. It is an important form of text computer representation, mapping the text space to a vector space, enabling computational operations on the text, and providing crucial support for deep text analysis such as information retrieval, content filtering, parameter extraction, and text verification. Based on text granularity, text vectorization can be divided into vectorization processing methods at the character, word, and sentence levels. Text vectorization mainly includes steps such as extracting text keywords, vectorizing keywords using a model, and summing word vectors to obtain the vectorized representation of the text.

[0156] This submodule uses the Python gensim library to train text paragraph vectors based on the Doc2Vec model, laying the foundation for text information similarity evaluation.

[0157] Step 5: Filter invalid information according to text vectorization similarity, judge the correctness of key information that has been integrated, and then input the correct information into the valve drawing information database.

[0158] Step 5 corresponds to the text information evaluation module, which is used to filter and output text keywords from the high-frequency words and keywords sequences of the nuclear power valve design documents that have completed text analysis, filter invalid information according to text vectorization similarity, judge the correctness of the key information that has completed information integration, and process and store the correct and valid information into the database. This module will output the final valid and reliable design document information data, and use it to guide the training of the design document information recognition model, thereby improving the system's self-learning and self-reinforcing capabilities.

[0159] Specifically, please refer to Figure 7 Step 5 includes:

[0160] Step 51: Receive the analyzed and processed information from step 4, and filter similar information in the key information sequence based on text vectorization similarity.

[0161] In other words, the system receives the analyzed and processed information from the text corpus analysis module and filters out similar information in the key information sequence based on text vectorization similarity. In nuclear power valve design documents, some expressions such as "inlet—inlet," "seller—supplier," and "bolt—stud" have the same actual meaning but differ in wording. Such information can affect the effectiveness of the evaluation and should be filtered out based on text similarity before the evaluation.

[0162] This submodule uses the similarity() function to perform text similarity analysis on processed high-frequency word sequences, keyword sequences, and other text sequences. Different textual expressions with a similarity score greater than 70% are considered to have the same meaning, and only one of these expressions is retained. The text sequences processed by similarity deduplication will be able to maintain text uniqueness effectively.

[0163] Step 52: Based on the existing key information sequence, conduct text keyword screening on the high-frequency sequence Igp and the keyword sequence Igj according to certain principles, and use them as the text keywords of the current text segment, and output the text keyword set to represent the main meaning of the current text segment.

[0164] That is, the process moves to the output text keyword processing submodule, where high-frequency word sequence (Igp) and keyword sequence (Igj) are filtered according to certain principles based on the existing key information sequence. These are then used as the text keywords for the current text segment, and a set of text keywords is output to represent the main meaning of the current text segment. The main principles of sequence filtering are:

[0165] 1. The total number of keywords to be filtered in the text is set to 5% of the total number of words in the text (rounded up);

[0166] 2. When selecting keywords for text, the selection process should be based primarily on keyword sequences and secondarily on high-frequency word sequences.

[0167] Third, high-frequency words belonging to the keyword sequence are directly defined as text keywords and added to the text keyword set;

[0168] IV. For subsequent text keyword sequences that meet condition 3 but do not meet the quantity requirement of condition 1, fill the text keyword sequence with keywords from the keyword sequence in descending order of weight until the text keyword set meets the quantity requirement.

[0169] 5. If the number of overlapping high-frequency words belonging to the keyword sequence is too large, making it impossible to ultimately meet the quantity requirement of condition 4, then the corresponding text segments should be annotated, and the accuracy of Chinese word segmentation in the text corpus processing module should be improved to solve the problem.

[0170] Step 53: Use a key information matching mechanism to evaluate and compare key information.

[0171] That is, the process moves to the design document information evaluation and comparison processing submodule, where a key information matching mechanism is used to evaluate and compare key information.

[0172] In this submodule, a key feature matching mechanism is employed. Key information originates from valve parameter information required by the valve design document information database. The units of these parameters (e.g., temperature, pressure) are used as identifiable key features, as they are more prominent than other textual information, facilitating accurate parameter acquisition. These key features serve as screening indicators, comparing the matched information with currently identified data. Successfully matched information is extracted and stored; invalid information is discarded, and the process returns to the extraction stage until all key features have been used. Simultaneously, a correctness assessment is performed, and erroneous information is corrected. If an incorrect parameter information is identified, the error is reported back to the valve design document information database, and the information is rewritten to correct the error. If the information is correct, the process proceeds to the next submodule.

[0173] Step 54: Integrate the information data evaluated in Step 53 and use it to guide the design of the CRNN document information recognition model for further training.

[0174] That is, after transferring to the design document information integration and database submodule, and after the design document information evaluation and comparison submodule has completed the correct verification of the output information, the information data integration work is carried out, and it is used to guide the further training of the design document information recognition model, and the key parameter information of the design documents is sorted according to... Figure 8The information is saved in the valve design document database in the form of [database name].

[0175] The required information data is mainly divided into two columns: design requirement categories and design requirement summaries. The design requirement categories are the main categories of relevant design requirements in the nuclear power valve design documents, encompassing all primary requirements for nuclear power valve design. The design requirement summaries are generated by mapping the output text keyword set to the design requirement categories, thus abstracting the design documents into two columns of design requirement data.

[0176] Meanwhile, the information in the database is not only related to the current design documents, but also ensures the correctness and validity of the data. Therefore, the valve design document information database will be used for further training and strengthening of the CRNN valve design document information recognition model. The CRNN valve design document information recognition model trained based on the existing design document information will be more suitable for the information recognition and collection work of nuclear power project design documents, and its model work will be more targeted and reliable.

[0177] In summary, the present invention has at least the following advantages:

[0178] I. This invention addresses the information recognition of nuclear power plant valve design documents, meeting the requirements for information processing in nuclear power plant valve design documents, encompassing "text processing—information recognition—information collection—corpus analysis—data sorting—data storage—information error correction—model enhancement," and efficiently completing the information recognition and collection work. Furthermore, this invention employs natural language processing methods to abstract and extract the design document text into a refined dataset that meets the corresponding design requirements, enabling the computer to visualize and vectorize the textual information of the design documents.

[0179] II. This invention establishes an information-based data acquisition system, including the aforementioned text information processing module, text information recognition module, text corpus processing module, text corpus analysis module, and text information evaluation module. This system is used for managing design document information data and can guide the self-learning and self-reinforcement of the nuclear power valve design document information recognition model, thereby improving the model's capabilities. It should be noted that this system also includes sub-modules within each module, which will not be elaborated upon here.

[0180] Third, this invention establishes a design document information evaluation method to assess the correctness of the identification information and to correct and rewrite erroneous information to ensure the reliability of the information.

[0181] Fourth, this invention adopts an excellent nuclear power valve design document information recognition model, which has superior judgment and recognition capabilities and can realize information recognition and collection for text sequences of variable length.

[0182] Fifth, this invention can provide technical support for the selection of valves, document review, equipment design, system optimization, and data verification in nuclear power plants, reduce the cost of reviewing valve design documents and optimizing system design, and can be applied in other related fields.

[0183] refer to Figure 9 As shown, this application also provides a data acquisition system 1000, including a memory 1001 and a processor 1002. The memory is capable of storing instructions that can be executed by the processor 1002. The processor 1002 is capable of executing instructions to implement the steps performed by the computer in the nuclear power valve design document information identification and acquisition method described in the above embodiments.

[0184] It is understood that the aforementioned memory and processor are not limited to a specific memory or processor. Furthermore, in embodiments employing a distributed architecture, the specific execution terminal for each step can be adjusted according to actual circumstances, and the specific implementation scheme of each step on a particular terminal should not limit the scope of protection of this application.

[0185] According to another aspect of this application, a computer-readable medium is also provided.

[0186] The computer-readable medium provided in this application has computer instructions thereon. When executed by a processor, these computer instructions can implement the steps performed by the program in the methods described in the above embodiments.

[0187] According to another aspect of this application, a computer program product is also provided.

[0188] The computer-readable medium provided in this application includes a computer program that, when executed by a processor, can implement the steps performed by the program in the method described in the above embodiments.

[0189] The various illustrative logic modules and circuits described in conjunction with the embodiments disclosed herein may be implemented or performed using a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. The general-purpose processor may be a microprocessor, but in alternatives, it may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors cooperating with a DSP core, or any other such configuration.

[0190] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of both. The software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor such that the processor can read and write information to / from the storage medium. In an alternative, the storage medium may be integrated into the processor. The processor and storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In an alternative, the processor and storage medium may reside as discrete components in the user terminal.

[0191] In one or more exemplary embodiments, the described functionality may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software as a computer program product, the functionality may be stored or transmitted as one or more instructions or code on or via a computer-readable medium. A computer-readable medium includes both computer storage media and communication media, encompassing any medium that facilitates the transfer of a computer program from one location to another. A storage medium may be any available medium accessible to a computer. By way of example and not limitation, such a computer-readable medium may include RAM, ROM, EEPROM, CD-ROM or other optical disc storage, disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and is accessible to a computer. Any connection is also legitimately referred to as a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of a medium. As used in this article, disk and disc include compact discs (CDs), laser discs, optical discs, digital multi-purpose discs (DVDs), floppy disks, and Blu-ray discs. Disks typically reproduce data magnetically, while discs reproduce data optically using lasers. Combinations of these should also be included within the scope of computer-readable media.

[0192] While this application discloses preferred embodiments as described above, it is not intended to limit the scope of this application. Any changes and modifications can be made by those skilled in the art without departing from the spirit and scope of this application. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of this application without departing from the content of the technical solution of this application shall fall within the protection scope defined by the claims of this application.

Claims

1. A method for identifying and collecting information from nuclear power plant valve design documents, characterized in that, include: Step 1: Receive the nuclear power valve design file and preprocess the nuclear power valve file according to the format and geometric requirements based on the CNN valve design file image recognition model to obtain the target valve design file; Step 1 includes: Step 11: Use OpenCV's imread() function to read the nuclear power plant valve design file; Step 12: Check and judge the format of the nuclear power valve design file, and convert the non-JPG format of the nuclear power valve design file to JPG format; Step 13: Perform file angle checks and file pixel height transformations based on pixel coordinates on the nuclear power valve design file after format conversion in Step 12 using pixel operations; Step 14: Use the CNN valve design document image recognition model to recognize the nuclear power valve design document that has been checked and transformed in Step 13, and remove the image information from the target valve design document; Step 15: Use OpenCV's merge() method to merge the nuclear power valve design file after the image information recognition and removal in Step 14 into three channels; Step 16: Perform bilateral filtering on the nuclear power valve design document obtained after channel merging in Step 15; Step 17: Perform binarization thresholding on the target valve design file that has been filtered in Step 16 to obtain the target valve design file; Step 2: The valve drawing information recognition model based on CRNN identifies the information of the area to be identified in the target valve design document; wherein, the valve drawing information recognition model based on CRNN is trained using a valve drawing information database; Step 3: Create a corpus analysis database using the recognition results from Step 2, and perform part-of-speech tagging and Chinese word segmentation on the corpus analysis database; Step 4: Perform text analysis on the processed corpus results of the nuclear power valve in Step 3, extract high-frequency words, keywords and key numerical parameters, and perform text vectorization processing; Step 5: Filter invalid information according to text vectorization similarity, judge the correctness of key information that has been integrated, and then input the correct information into the valve drawing information database.

2. The method for identifying and collecting information from nuclear power plant valve design documents according to claim 1, characterized in that, The valve drawing information recognition model of the CRNN in step 2 includes a CRNN convolutional layer, a CRNN recurrent layer, and a CRNN transcription layer; The CRNN convolutional layer is used to extract text features of the information to be recognized; the CRNN recurrent layer is used to extract sequence features and generate a probability feature sequence; the CRNN transcription layer is used to predict the probability feature sequence and convert it into a character sequence.

3. The method for identifying and collecting information from nuclear power plant valve design documents according to claim 2, characterized in that, The CRNN convolutional layer is used to extract the features of the information to be identified in the input design file. After convolution, pooling and normalization, the convolutional feature map is obtained and used as the input of the CRNN recurrent layer. The CRNN recurrent layer uses gated recurrent units to calculate the probability feature sequence, which is then used as the input to the CRNN transcription layer. The CRNN transcription layer uses a connection-time classification loss function to convert the predicted probability feature sequence input to the CRNN recurrent layer into a character sequence.

4. The method for identifying and collecting information from nuclear power plant valve design documents according to claim 1, characterized in that, Step 3 includes: Step 31: Receive the recognition result information from Step 2 and save it as a txt text file for use as the input file for the NLP corpus; wherein, the corpus is created and the recognition information result file in the corpus is obtained through the NLTK fileids() method; Step 32: Perform simplified / traditional Chinese conversion on the text in the corpus; Step 33: Remove stop words from the text segments in the corpus; Step 34: Deconstruct the text information of the corpus segment and separate the words in the segment; Step 35: Perform part-of-speech tagging on the words in the corpus text.

5. The method for identifying and collecting information from nuclear power plant valve design documents according to claim 1, characterized in that, Step 4 includes: Step 41: Receive the processed corpus results from Step 3 and extract high-frequency words from the text. Step 42: Extract keywords from the text using NLP keyword extraction methods and form a keyword sequence Igj; Step 43: Extract key numerical parameters from the passage; Step 44: Train text paragraph vectors using the Python gensim library based on the Doc2Vec model.

6. The method for identifying and collecting information from nuclear power plant valve design documents according to claim 5, characterized in that, In step 42, the NLP uses the TF-IDF keyword extraction algorithm.

7. The method for identifying and collecting information from nuclear power plant valve design documents according to claim 5, characterized in that, Step 5 includes: Step 51: Receive the analyzed and processed information from step 4, and filter similar information in the key information sequence based on text vectorization similarity; Step 52: Based on the existing key information sequence, the high-frequency word sequence Igp and the keyword sequence Igj are selected as text keywords for the current text segment according to certain principles, and the text keyword set is output to represent the main meaning of the current text segment. Step 53: Use a key information matching mechanism to evaluate and compare key information; Step 54: Integrate the information data evaluated in Step 53 and use it to guide the design of the CRNN document information recognition model for further training.

8. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the steps performed by a computer in the method as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The method comprises a computer program that, when executed by a processor, implements the steps performed by a computer in the method as described in any one of claims 1-7.

10. A data acquisition system, characterized in that, include: Memory is capable of storing instructions that can be executed by a processor; A processor capable of executing the instructions to implement the steps performed by a computer in the method as described in any one of claims 1-7.

Images