Building material data classification method and system based on data enhancement and feature extraction

Abstract

The present disclosure relates to the technical field of data processing, and proposes a building material data classification method and system based on data enhancement and feature extraction, which comprises: obtaining building material text to be classified, preprocessing and expanding the text content of the building material name; extracting keywords from the expanded building material text content, and using a multi-level data enhancement method to gradually enhance the semantic information of the extracted building material text; converting the text obtained after data enhancement into a word vector, extracting full-text features through a convolutional neural network, and extracting features between contexts through a parallel convolutional neural network to capture the context semantic information and contextual meaning in the building material text description, and obtaining the building material text features after feature extraction; matching the building material text features after feature extraction with the corresponding building material type. The building material name and building material type can be accurately matched, so that the building material can be matched to the correct carbon emission factor, providing a basis for carbon emission management in the building field.

Description

Technical Field

[0001] This disclosure relates to the technical field of data processing technology, specifically to a method and system for classifying building material data based on data augmentation and feature extraction. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] Carbon emission management is crucial in the construction industry. Only with accurate calculation and control of carbon emissions can the achievement of dual carbon targets be effectively supported. The carbon emission factor method is widely used in calculating carbon emissions during the building materials production stage. It multiplies the carbon emission factor of each building material by its quantity to obtain the carbon emission amount. However, in practice, building materials in carbon emission factor libraries are categorized by broad categories, primarily based on material type, without detailed descriptions or specific carbon emission factors. Furthermore, the descriptive language of building materials in actual project bills of materials is often inconsistent. This makes it difficult to match the materials in the bill of materials with the correct carbon emission factors, thus affecting carbon emission calculations.

[0004] Patent CN113255805A provides an imbalanced data augmentation classification algorithm based on distance matching metrics, but it is only applicable to machine learning classification scenarios where there is a certain similarity metric between labeled and unlabeled data. Patent CN115099241A utilizes a combination of algorithms such as LDA and Simhash to extract textual information and obtain an optimal semantic dataset. The inventors discovered that existing technologies do not consider classification methods for short, multi-topic texts, resulting in low accuracy in classifying short text data. Summary of the Invention

[0005] To address the aforementioned issues, this disclosure proposes a building material data classification method and system based on data augmentation and feature extraction. This method can accurately match building material names with building material types, thereby matching building materials to the correct carbon emission factors and providing a foundation for carbon emission management in the construction sector.

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

[0007] One or more embodiments provide a building material data classification method based on data augmentation and feature extraction, including the following steps:

[0008] Obtain the text of building materials to be classified, and preprocess it to expand the text content of the building material names;

[0009] Keyword extraction was performed on the expanded building materials text, and a multi-level data augmentation method was used to progressively enhance the semantic information of the extracted building materials text.

[0010] The text obtained after data augmentation is converted into word vectors. Full-text features are extracted through convolutional neural networks, and features between contexts are extracted through parallel convolutional neural networks. This captures the contextual semantic information and contextual meaning in the description of building materials, resulting in the extracted features of the building materials text.

[0011] The extracted building material text features are matched with the corresponding building material types.

[0012] One or more embodiments provide a building material data classification system based on data augmentation and feature extraction, including:

[0013] Data augmentation module: Configured to obtain the text of building materials to be classified and preprocess it to augment the text content of the building material names;

[0014] Multi-level data augmentation module: It is configured to extract keywords from the expanded building materials text content and use a multi-level data augmentation method to enhance the semantic information of the extracted building materials text step by step;

[0015] Feature extraction module: It is configured to convert the text obtained after data augmentation into word vectors, extract full-text features through convolutional neural networks, and extract features between contexts through parallel convolutional neural networks, capturing the contextual semantic information and contextual meaning in the description of building materials, and obtaining the features of building materials text after feature extraction;

[0016] Matching module: Configured to match the extracted building material text features with the corresponding building material type.

[0017] An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the steps in the above-described building material data classification method based on data augmentation and feature extraction.

[0018] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the steps in the above-described building material data classification method based on data augmentation and feature extraction.

[0019] Compared with the prior art, the beneficial effects of this disclosure are as follows:

[0020] This disclosure employs multi-level data augmentation to improve the accuracy of text recognition; and uses CNN to extract global semantic information from building material text descriptions. By extracting semantic features between the context of building material text through a parallel CNN network, accurate extraction of data semantics is achieved, which improves the accuracy of text data classification and enables accurate matching of building material types to building material texts.

[0021] The advantages of this disclosure, as well as its additional advantages, will be described in detail in the following specific embodiments. Attached Figure Description

[0022] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute a limitation thereof.

[0023] Figure 1 This is an overall flowchart of the building material data classification method of Embodiment 1 of this disclosure;

[0024] Figure 2 This is a flowchart of the multi-level data augmentation process according to Embodiment 1 of this disclosure;

[0025] Figure 3 This is an example of the multi-level data augmentation result of Embodiment 1 of this disclosure;

[0026] Figure 4 This is a flowchart of extracting and classifying data features according to Embodiment 1 of this disclosure;

[0027] Figure 5 This is a schematic diagram illustrating the conversion of building material text into a word vector array in Embodiment 1 of this disclosure;

[0028] Figure 6 This is a schematic diagram of the aggregation feature in step 322 of Embodiment 1 of this disclosure;

[0029] Figure 7 This is a schematic diagram of the feature aggregation after feature extraction based on semantics before and after the CNN network in step 33 of embodiment 1 of this disclosure;

[0030] Figure 8 This is a flowchart illustrating the process of classifying building materials based on fully connected layers in Embodiment 1 of this disclosure. Detailed Implementation

[0031] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0032] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of this disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0033] It should be noted that the terminology used herein is for descriptive purposes only and is not intended to limit the exemplary embodiments according to this disclosure. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. It should be noted that, without conflict, the various embodiments and features within those embodiments can be combined with each other. The embodiments will now be described in detail with reference to the accompanying drawings.

[0034] Example 1

[0035] In one or more of the technical solutions disclosed in the embodiments, such as Figures 1 to 8 As shown, a building material data classification method based on data augmentation and feature extraction includes the following steps:

[0036] Step 1: Obtain the text of building materials to be classified and preprocess it to expand the text content of the building material names;

[0037] Step 2: Extract keywords from the expanded building materials text and use a multi-level data augmentation method to enhance the semantic information of the extracted building materials text step by step.

[0038] Step 3: Convert the text obtained after data augmentation into word vectors, extract full-text features through convolutional neural networks, and extract features between contexts through parallel convolutional neural networks to capture the contextual semantic information and contextual meaning in the building materials text description, and obtain the building materials text features after feature extraction.

[0039] Step 4: Match the extracted building material text features with the corresponding building material types.

[0040] In this embodiment, multi-level data augmentation is employed to improve the accuracy of text recognition; and a CNN is used to extract global semantic information from the building material text description. By extracting semantic features between the context of the building material text through a parallel CNN network, accurate extraction of data semantics is achieved, which can improve the accuracy of text data classification and accurately match building material text with building material types.

[0041] In this embodiment, the building material text includes the material name and also includes multi-level identifier text that describes its material type name.

[0042] The building material names obtained in step 1 are short texts, as shown in the first column of Table 1. They are the names of building materials in an actual construction project. The building material names are characterized by being short, having non-standard descriptions (including abbreviations or simplified characters), and containing little semantic information, which is not conducive to subsequent classification tasks.

[0043] Table 1

[0044]

[0045] Furthermore, in this embodiment, the preprocessing is as follows: the translation of building material names is obtained using web crawler technology to expand the text. Specifically, the nouns in the building material text are standardized, and the nouns and their explanations are obtained by crawling the Internet (such as Baidu Encyclopedia) using web crawler technology.

[0046] Translation, as a text enhancement technique, can increase the content and richness of labeled text by translating short technical terms, introducing more contextual information and semantic expression, thereby improving the performance of the model in classification tasks.

[0047] The preprocessed text and label examples are shown in Table 2. The data consists of seven columns. The first column is the building material name (also called the material name), which includes the text of the building material name and its explanation; the second column is the material type name (level 1); the third column is the type identifier (level 1); the fourth column is the material type name (level 2); the fifth column is the type identifier (level 2); the sixth column is the material type name (level 3); and the seventh column is the type identifier (level 3).

[0048] Table 2

[0049]

[0050] Text augmentation is performed on the preprocessed material names. First, primary text augmentation is performed, grouping texts with the same primary identifier into the same corpus. Keywords are extracted from each corpus using LDA, and a triple model is generated for each text using an n-gram model. Based on the relationship between the keywords in the corpus and the text's own triple model, the original building material name text is expanded using triples to obtain the primary-level augmented building material text. Similarly, secondary and tertiary text augmentation are performed based on the primary-level augmentation, finally resulting in the tertiary-level augmented text Axi, which is the final form of data augmentation. The data augmentation flowchart is shown below. Figure 4 As shown.

[0051] In step 2, the data augmentation in this patent is a step-by-step augmentation, with each level of data augmentation having a sequential relationship. The number of data augmentation levels can be set. After the first level of data augmentation, the subsequent data augmentation steps will perform data augmentation on the data augmented at the previous level.

[0052] Specifically, the level of data augmentation is set according to the level of the building material type name. In this embodiment, it is set to three levels of data augmentation, that is, two levels of data augmentation are implemented on the basis of the first level of data augmentation, and three levels of data augmentation are further performed on the basis of the second level of data augmentation, so that the text after data augmentation is finally converted into the form of word vectors.

[0053] Optionally, in this embodiment, the method for each level of data augmentation includes the following steps:

[0054] Step 21: For the acquired building materials text, classify the texts with the same building type identifier at the corresponding level into the same corpus;

[0055] Step 22: Extract keywords from each corpus using the LDA method;

[0056] LDA (Latent Dirichlet Allocation) is a generative probability model used to represent each document in a text dataset as a mixture of multiple probability distributions, which can automatically extract topic keyword information from a large number of documents.

[0057] In this embodiment, the number of topic words is set to 1, and the topic words Top of the corpus are determined. These can be the building type identifier names of the corresponding level. The LDA model is trained using the text of the corpus. During the training process, the probability distribution of words related to the topic words Top is learned through LDA.

[0058] After training the LDA model, the probability distribution of words related to the topic term "Top" is obtained, where each word has a probability value related to the topic term "Top". The top sR words with the highest probability values ​​are selected as keywords related to the topic term "Top", forming the topic keyword list (K) for each corpus. R 1, K R 2, K R 3, ..., K R sR ).

[0059] Step 23: Use the n-gram model to generate a ternary model for each text in each corpus;

[0060] The n-gram model is a text processing method that makes text data more structured. It first divides the text into n consecutive elements (usually words) so that the text data can be represented as a sequence of n consecutive elements. This sequence is called an n-gram sequence group.

[0061] In this embodiment, the n-gram model is set to 3, called a tri-gram. Each building material name text in the corpus is input into the tri-gram model, and the output is a triplet composed of three consecutive words from the building material name text. For example, ... Figure 2 When the text "insulation nail" is input into the tri-gram model, the model will output: "insulation nail, is, will", "is, will, insulation board", "will, insulation board, fix", "insulation board, fix, to", "fix, to, wall", "to, wall, on", "wall, on, of", "on, of, a kind", "of, a kind, special", "a kind, special, engineering plastic", "special, engineering plastic, expansion nail", ..., "fix, round piece, composition".

[0062] Step 24: Combine the triplets in the ternary model with the original building material name text to expand the original building material name text and obtain the classification-enhanced building material text;

[0063] Specifically, taking first-level data augmentation as an example, the process is as follows: texts with the same first-level identifier are divided into the same corpus. Keywords for each corpus are extracted using the LDA method. At the same time, a triple model for each text is generated using an n-gram model. Based on the relationship between the keywords in the corpus where the text is located and its own triple model, the original building material name text is expanded using triples to obtain the first-level classification augmented building material text.

[0064] In step 21, data with the same identifier are grouped into the same corpus, which is called data merging.

[0065] Based on the R-level (R=1, 2, 3) classification labels of building materials texts, building materials texts with the same R-level classification label data are grouped into the same corpus.

[0066] The segmented building materials texts are categorized into corpus X. R 1, X R 2, X R 3, ..., X R n , where n is the number of R-level categories.

[0067] Each corpus X R i It can be represented as X R i ={x1, x2, ..., x u}, where u is the number of building materials texts in the R-level classification corpus; each building materials text sentence is segmented into words, and each text x i (i = 1, 2, 3, ..., u) can be represented as x i = (w1, w2, ..., w l ), where l is xi The number of words in the corpus X is then used to calculate the word count for each corpus X. R 1, X R 2, X R 3, ..., X R n LDA was used to extract keywords.

[0068] In step 22, keyword extraction: keywords for each corpus are extracted using the LDA method;

[0069] Optionally, the keyword extraction method is as follows: In the corpus under the corresponding level R classification, determine the topic words, keeping the number of topic words at 1, and extract the keyword list (K) for each corpus using LDA. R w1 K R w2 K R w3 , ..., K R wn These keywords are sorted from highest to lowest according to their relevance score (degree of relevance) to the corpus's topic terms. The top sR keywords from the sorted list are used to form the topic keyword list (K) for each corpus. R 1, K R 2, K R 3, ..., K R sR ).

[0070] Steps 23 and 24 are the data augmentation steps. First, a triple model is generated, and then the text in the corpus is augmented based on the triples in the triple model. The specific explanation is as follows:

[0071] For each building materials text x in the corpus under the R-level classification i The corresponding ternary model is generated using n-grams, that is, each text x in each corpus is used to generate a ternary model. i It is represented as a set of triples x i =(tr R 1, tr R 2, ..., tr R q ), where q is x i The total number of triples in the text. When a certain triple in the text is tr R i The i = 1, 2, 3, ..., q corpus contains the set of topic keywords K of its respective corpus. R i If at least one keyword is present, the original text is expanded using this triple, achieving R-level text enhancement. The expanded building materials text is represented as Axi = (w1, w2, ..., wl, tr). R1, tr R 2, ..., tr R j ), where w1, w2, ..., w l It is x i The word in the original text, tr R 1, tr R 2, ..., tr R j It is x i The triples after conversion to a ternary model must contain at least one of the text x. i The subject keywords of the R-level corpus in which it is located.

[0072] The building materials text after three levels of enhancement is represented as Ax. i = (w1, w2, ..., w l tr 1 1, tr 1 2, ..., tr 1 j ,tr 2 1, tr 2 2, ..., tr 2 k tr 3 1, tr 3 2, ..., tr 3 z ), where w1, w2, ..., w l It is x i The word in the original text, tr 1 1, tr 1 2, ..., tr 1 j tr 2 1, tr 2 2, ....tr 2 k tr 3 1, tr 3 2, ..., tr 3 z It is x i The triple after conversion to a ternary model, tr 1 1, tr 1 2, ..., tr 1 j It must contain at least one of the text x i The topic keywords of the primary corpus, tr 2 1, tr 2 2, ..., tr 2 kIt must contain at least one topic keyword from the secondary corpus where the text xi is located, tr 2 1, tr 2 2, ..., tr 2 k It must contain at least one topic keyword from the third-level corpus where the text xi is located.

[0073] For example, the text for "hanging rod" after level 1 enhancement is A 1 x1 = ("hanging bar", ",", "is to", "act on", "concrete", "beam", ..., "also known as", "treasure bar", "steel bar, refers to, reinforced concrete", "angle steel, commonly known as, angle iron"... "low carbon steel, of, small round steel").

[0074] Building materials text A after level 1 text enhancement 1 x i Using the same method described above, after secondary and tertiary text enhancement, we obtained the building materials text dataset A. 3 x i For example, the text set of the building material name "hanging rod" after three levels of text enhancement is as follows: Figure 3 As shown.

[0075] In step 3, feature extraction and aggregation are performed. Specifically, the augmented text A is processed using a BERT model. 3 x i Transform into an array P consisting of word vectors token Using a CNN network to extract P token The full-text features are aggregated to obtain array P. agg Using composite CNN networks of different sizes to extract P agg The semantic features before and after are used to obtain an array Ps. Ps is then used to classify building material text through a fully connected layer. The feature aggregation flowchart is as follows. Figure 4 As shown, the detailed steps are explained below.

[0076] Step 31: Convert the data-augmented text into word vectors, which can be achieved using the bert-base-chinese model.

[0077] Specifically, the data-enhanced building materials text Ax i = (w1, w2, ..., w l tr 1 1, tr 1 2, ..., tr 1 j tr 2 1, tr 2 2, ..., tr 2 k tr 31, tr 3 2, ..., tr 3 z Input the pre-trained bert-base-chinese model.

[0078] The "bert-base-chinese" model is a specific version of BERT, a natural language processing model specifically designed for Chinese language processing. It employs a Transformer architecture, enabling it to learn contextual information bidirectionally. The output of the BERT model is an array P consisting of n tokens. token In natural language processing, a "token" refers to a series of independent units into which text is segmented. Here, a token represents a word and serves as the basic unit of input to the model. For example... Figure 5 As shown. If the text length is insufficient, it will be automatically padded with pad characters. Each token has a dimension of f, which represents the f features of the token.

[0079] P token =(T1,T2,T3,......T) n (3.1)

[0080] in

[0081] T i =(F1,F2,F3,......F f (3.2)

[0082] Step 32: Perform convolution calculations on the obtained word vectors and aggregate the features obtained from the convolution. The specific steps are as follows:

[0083] Step 321: Perform convolution calculation operation, perform element-wise convolution calculation on the local features of each word vector to obtain the global features of each word vector;

[0084] Specifically, a CNN network with one convolutional layer is configured, including X filters of size 1*Fs, and the number of channels in the convolutional kernel is set to X. The number of filters corresponds to the number of channels in the convolutional layer, with each filter corresponding to one channel. The input data undergoes a convolution operation on each channel, and features are extracted from the input data through the filters of each channel, generating weighted output values. These output values ​​are then fed into an activation function for non-linear transformation to obtain richer feature representations.

[0085] Each filter combines f local features of the token and expresses them as a single global feature.

[0086] array P tokenAfter passing through a filter, the filter will perform element-wise convolution calculations on f local features of each token. The output array size after passing through each filter is C. row Okay, C' col This global feature vector represents the global features of each token. After extracting features through all filters, X tokens of size C are obtained. row Okay, C' col The feature array of the column.

[0087]

[0088]

[0089] Where Fh is the height of the filter, which is equal to 1; Fs is the width of the filter, which is equal to the dimension of the token; p is the padding size; and s is the stride.

[0090] Step 322: Aggregate features: Aggregate the obtained global feature vectors.

[0091] To more richly express the semantic information of words in building materials text, the global feature vector is aggregated through all filters. The output at this point is C. row Okay, C col Columns. The aggregated array Pagg is as follows: Figure 6 As shown, the vectors can more fully express the global features of building materials text.

[0092] C col =C′ col X (3.5)

[0093]

[0094] in:

[0095]

[0096] F i =f(w·T) i +b) (3.8)

[0097] Where f is a nonlinear activation (ReLU) function, W is the parameter matrix of the filter, and b is the bias.

[0098] Step 33: Extract the features between contexts from the aggregated features obtained in Step 32 using a parallel convolutional neural network (CNN network), and concatenate the extracted features to form an array to achieve feature extraction of semantics between the preceding and following contexts;

[0099] In some embodiments, this step is implemented by setting up a composite CNN network, which can be multiple parallel CNN networks. In this embodiment, setting up three parallel CNN networks is used for illustration.

[0100] In this composite CNN network, each CNN consists of sequentially connected convolutional layers and pooling layers. Each convolutional layer uses a kernel of different sizes, and these layers are used to extract the relationships and contextual meanings between words in different positions within the building materials text. The pooling layers are used to reduce the dimensionality and extract features from the outputs of the convolutional layers, summarizing and compressing the feature values ​​to better capture different combinations of word features, highlight the contextual meaning of the building materials text, and provide more useful feature information for subsequent text classification.

[0101] In this embodiment, by using convolutional kernels of different sizes, the convolutional layer can capture contextual information of different ranges, thereby revealing the contextual meaning of the building material text.

[0102] Specifically, this embodiment sets up three parallel CNN networks. The first CNN has a convolutional kernel size of Fh1*Fw, where Fh1=2 to highlight the semantic relationship between two consecutive words in the building materials text. The second CNN has a convolutional kernel size of Fh2*Fw, where Fh2=3 to extract the expressive features of every three words in the context. The third CNN has a convolutional kernel size of Fh3*Fw, where Fh3=4 to extract the best features from every four words that reveal the contextual meaning. The size of the three pooling layers is Fh. p *Fw, Fh p =1.

[0103] In this embodiment, the input array P agg The output array Pc is obtained after passing through three CNN networks respectively. i The sizes are m i Row, column c, i = 1, 2, 3. Then, each column passes through its pooling layer, and the pooling layer outputs an array Pp. i The sizes are n j Row, column 1, i is 1, 2 or 3.

[0104]

[0105]

[0106]

[0107] in:

[0108] T i =[F1,F2,F3,......,F c i = 1, 2, 3 (3.12)

[0109]

[0110]

[0111] Where Fw is the width of the convolutional and pooling layers, which is equal to the dimension of the input array, p is the padding size, and s is the stride.

[0112] Furthermore, concatenating the outputs of all pooling layers into a single array Ps integrates the building material text features extracted by the parallel CNNs, resulting in richer textual information and providing a more comprehensive and integrated feature representation for subsequent building material classification tasks. The feature extraction output array is shown below. Figure 7 As shown.

[0113]

[0114] In step 4, the extracted building material text features are matched with the corresponding building material types. This involves classifying the features of the building material text, which can be done using a fully connected layer. The specific steps include the following:

[0115] Step 41: Input the data augmentation and feature extraction processed building material text feature Ps array into the fully connected layer;

[0116] Step 42: Perform a nonlinear transformation in the fully connected layer using the ReLU function of the rectified linear unit. Activate the transformation result to obtain a probability distribution of the building material name in the building material text for all building material categories. The distribution with the highest score corresponds to the building material type.

[0117] The fully connected layer contains N neurons, where N is the number of categories in the three-level classification of all building materials. Each neuron sends its output to the Softmax function, which is the activation function for the multi-classification task. The output is a probability distribution of the input building material for all three-level classifications of building materials. The distribution with the highest probability is the three-level classification of the material type corresponding to the input building material.

[0118] This embodiment categorizes input building materials into three levels of material types, thereby correctly matching the carbon emission factor and calculating carbon emission information. The method in this embodiment can quickly and effectively classify building materials for different project descriptions, solving the problems of inaccurate carbon emission calculations caused by non-standard building material descriptions in project lists and simplified classification methods in the carbon emission factor library.

[0119] Example 2

[0120] Based on Example 1, this example provides a building material data classification system based on data augmentation and feature extraction, including:

[0121] Data augmentation module: Configured to obtain the text of building materials to be classified and preprocess it to augment the text content of the building material names;

[0122] Multi-level data augmentation module: It is configured to extract keywords from the expanded building materials text content and use a multi-level data augmentation method to enhance the semantic information of the extracted building materials text step by step;

[0123] Feature extraction module: It is configured to convert the text obtained after data augmentation into word vectors, extract full-text features through convolutional neural networks, and extract features between contexts through parallel convolutional neural networks, capturing the contextual semantic information and contextual meaning in the description of building materials, and obtaining the features of building materials text after feature extraction;

[0124] Matching module: Configured to match the extracted building material text features with the corresponding building material type.

[0125] It should be noted that each module in this embodiment corresponds one-to-one with each step in embodiment 1, and their specific implementation process is the same, so it will not be repeated here.

[0126] Example 3

[0127] This embodiment provides an electronic device, including a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, they complete the steps in the building material data classification method based on data augmentation and feature extraction in Embodiment 1.

[0128] Example 4

[0129] This embodiment provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps in the building material data classification method based on data augmentation and feature extraction in Embodiment 1.

[0130] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Various modifications and variations can be made to this disclosure by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

[0131] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. A building material data classification method based on data augmentation and feature extraction, characterized in that, Includes the following steps: Obtain the text of building materials to be classified, and preprocess it to expand the text content of the building material names; Keyword extraction is performed on the expanded building materials text content, and a multi-level data augmentation method is used to enhance the semantic information of the extracted building materials text level by level. The data augmentation levels are sequential, and the number of data augmentation levels is set. After the first level of data augmentation, the subsequent data augmentation steps will perform data augmentation on the data augmented by the previous level. The methods for each level of data augmentation include the following steps: For the acquired building materials text, texts with the same building type identifier at the corresponding level are grouped into the same corpus; Keywords for each corpus were extracted using the LDA method; The n-gram model is used to generate a trigram model for each text in each corpus; By combining the triplet in the ternary model with the original building material name text, the original building material name text is expanded, resulting in the enhanced building material text after classification. The text obtained after data augmentation is converted into word vectors. Full-text features are extracted through convolutional neural networks, and features between contexts are extracted through parallel convolutional neural networks. This captures the contextual semantic information and contextual meaning in the description of building materials, resulting in the extracted features of the building materials text. Features between contexts are extracted by using a parallel convolutional neural network, and the extracted features are concatenated into an array to achieve feature extraction of semantics between contexts. Each convolutional neural network in a parallel convolutional neural network consists of convolutional layers and pooling layers connected in sequence. Each convolutional layer uses a convolutional kernel of a different size. Different convolutional layers are used to extract the relationships and contextual meanings between words in different positions in building materials text. Pooling layers are used to reduce the dimensionality and extract features from the output of convolutional layers, summarizing and compressing the feature values ​​output by the convolutional layers; The extracted building material text features are matched with the corresponding building material types, and the extracted features are classified through a fully connected layer.

2. The building material data classification method based on data augmentation and feature extraction according to claim 1, characterized in that: The preprocessing involves using web crawling technology to obtain translations of building material names for text expansion. First, the nouns in the building material text are standardized, and then web crawling technology is used to crawl the nouns for translation, resulting in the nouns and their definitions in the building material text.

3. The building material data classification method based on data augmentation and feature extraction according to claim 1, characterized in that, The LDA method is used to extract keywords. The steps are as follows: In the corpus under the corresponding level classification, identify the subject words and keep the number of subject words at 1; The keyword list for each corpus was extracted using LDA; Calculate the relevance scores of the keyword domain corpus terms in the keyword list and sort them from high to low; Take the sorted keyword list and select a set number of highly relevant keywords to form the topic keyword list for each corpus.

4. A building materials data classification system based on data augmentation and feature extraction, characterized in that, include: Data augmentation module: Configured to obtain the text of building materials to be classified and preprocess it to augment the text content of the building material names; Multi-level data augmentation module: It is configured to extract keywords from the expanded building materials text content and use a multi-level data augmentation method to enhance the semantic information of the extracted building materials text level by level; the data augmentation levels are sequential, and the number of data augmentation levels is set. After the first level of data augmentation, the subsequent data augmentation steps will perform data augmentation on the data augmented by the previous level. The methods for each level of data augmentation include the following steps: For the acquired building materials text, texts with the same building type identifier at the corresponding level are grouped into the same corpus; Keywords for each corpus were extracted using the LDA method; The n-gram model is used to generate a trigram model for each text in each corpus; By combining the triplet in the ternary model with the original building material name text, the original building material name text is expanded, resulting in the enhanced building material text after classification. Feature extraction module: It is configured to convert the text obtained after data augmentation into word vectors, extract full-text features through convolutional neural networks, and extract features between contexts through parallel convolutional neural networks, capturing the contextual semantic information and contextual meaning in the building materials text description, and obtaining the feature-extracted building materials text features; it also extracts features between contexts through parallel convolutional neural networks, and concatenates the extracted features into an array to realize the feature extraction of semantics between contexts; Each convolutional neural network in a parallel convolutional neural network consists of convolutional layers and pooling layers connected in sequence. Each convolutional layer uses a convolutional kernel of a different size. Different convolutional layers are used to extract the relationships and contextual meanings between words in different positions in building materials text. Pooling layers are used to reduce the dimensionality and extract features from the output of convolutional layers, summarizing and compressing the feature values ​​output by the convolutional layers; Matching module: It is configured to match the extracted building material text features with the corresponding building material types, and classifies the extracted features through a fully connected layer.

5. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the steps in the building material data classification method based on data augmentation and feature extraction as described in any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, complete the steps in the building material data classification method based on data augmentation and feature extraction as described in any one of claims 1-3.

Images