Commodity material classification method, system, medium and equipment based on deep learning
By using deep learning algorithms to specialize material data and employing Bi-LSTM and one-dimensional convolutional layers to capture material features, the problem of poor material classification performance in existing technologies has been solved, and efficient classification of industrial materials has been achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 欧冶工业品股份有限公司
- Filing Date
- 2023-03-06
- Publication Date
- 2026-05-12
AI Technical Summary
Existing classification methods fail to effectively utilize relevant information about industrial materials, resulting in poor material classification performance.
A deep learning-based approach is adopted. By preprocessing the material description, technical attributes, specifications and names, a deep learning algorithm model is designed and constructed. Bi-LSTM and one-dimensional convolutional layers are used to capture long-range and local features of the text, and fully connected layers are combined for classification.
It has improved the accuracy and effectiveness of material classification, with particularly significant improvements in its application in the field of industrial materials.
Smart Images
Figure CN116244630B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of material detection, and more specifically, to a method, system, medium, and device for classifying commodity materials based on deep learning. Background Technology
[0002] There are numerous classification methods and their applications, but few are geared towards industrial applications. Existing classification methods do not utilize relevant material information and fail to capture material-related characteristics in the industrial field, resulting in poor material classification performance.
[0003] Patent document CN115410131A (application number: CN202211121406.6) discloses a method for intelligent classification of short videos, including the following steps: a) loading the original video; b) data preprocessing; c) video data feature extraction; d) BertModel semantic tag feature fusion training; e) automatic intelligent video classification. This method for intelligent classification of short videos is tailored to different target subjects. Furthermore, the method does not perform specific processing for industrial materials, and therefore cannot guarantee good results when applied to the material processing field.
[0004] Patent document CN114011750A (application number: CN202111253062.X) discloses a material visual inspection and classification device and method. The material visual inspection and classification device includes a residual material and dust removal component, a finished product classification box, a discharge air nozzle component, a front dimension detection camera, a side dimension detection camera, a bottom dimension detection camera, an appearance inspection camera one, an appearance inspection camera two, a feeding guide component, and a rotary feeding component. However, this invention does not provide specific treatment for industrial materials, and therefore cannot guarantee that the method will be effective in the field of material processing. Summary of the Invention
[0005] In view of the deficiencies in the existing technology, the purpose of this invention is to provide a commodity material classification method, system, medium and device based on deep learning.
[0006] A deep learning-based commodity material classification method provided by the present invention includes:
[0007] Step S1: Obtain material data and perform data preprocessing;
[0008] Step S2: Design and build a deep learning algorithm model, and train it using the processed data;
[0009] Step S3: Optimize the deep learning algorithm model architecture and data features, and conduct comparative experiments to prove the effectiveness of the deep learning algorithm model;
[0010] Step S4: Use the optimized deep learning algorithm model to classify the given materials.
[0011] Preferably, in step S1:
[0012] Acquire material data and perform data preprocessing on the material description, material technical attributes, material type and specification, and material name in the material data;
[0013] Step S1.1: Extract text from the material description field, clean and remove irrelevant symbols, obtain the corresponding code for each word in the text by querying a dictionary, and initially convert each line of text into an index vector as a feature item. A , used as input for Bi-LSTM;
[0014] Step S1.2: Based on the material's technical attributes, clean and segment the data, construct a dictionary, and obtain the corresponding code for each word in the processed text by querying the dictionary. This initially transforms each line of text into an index vector Item. B , used as input for Bi-LSTM;
[0015] Step S1.3: Process the text corresponding to the material specification field, remove symbols that cannot represent the text, retain letters and numbers according to the definition based on the writing specifications of the material specifications, obtain the corresponding code for each word in the processed text by querying a dictionary, and initially convert each line of text into a vector as a feature item. C , used as input for Bi-LSTM;
[0016] Step S1.4: The material name text contains existing material specification data. Chinese characters are retained, while redundant letters, numbers, punctuation, and stop words are removed. Each word in the processed text is encoded using a dictionary lookup, and each line of text is initially converted into an index vector Item. D , used as input for Bi-LSTM.
[0017] Preferably, in step S2:
[0018] Step S2.1: Design a word embedding layer, taking the vectorized short text as input and outputting word vectors;
[0019] Design a word embedding layer, with Item A Item B Item C and Item D As input, it is encoded and transformed into word vectors:
[0020] I A =Embedding(Item) A (1)
[0021] I B =Embedding(Item) B (2)
[0022] I C =Embedding(Item) C (3)
[0023] I D =Embedding(Item) D (4)
[0024] in, Where w ai ∈R L For this data item A The embedded representation of the i-th character of the field, T a For this data item A The length of the text embedded in the field; Where w bi ∈R L For this data item B The embedded representation of the i-th character of the field, T b For this data item B The length of the text embedded in the field; Where w ci ∈R L For this data item C The embedded representation of the i-th character of the field, T C For this data item C The length of the text embedded in the field; Where w di ∈R L For this data item D The embedded representation of the i-th character of the field, T d For this data item D The length of the text embedded in the field;
[0025] Step S2.2: Design four parallel bidirectional long short-term memory networks Bi-LSTM, using word vectors from the embedding layer as input, to capture long-distance feature dependencies of the text;
[0026] Step S2.3: Design a one-dimensional convolutional layer, using the output of the Bi-LSTM layer as input, to further capture local feature relationships;
[0027] Step S2.4: Design a fully connected layer, output the word vectors trained by Bi-LSTM and one-dimensional convolutional layer to the fully connected layer for classification, and output the classification result.
[0028] Preferably, in step S3:
[0029] Step S3.1: For Item B Clean up special characters and calculate the word frequency of all entries. Manually extract the pre-set features. By using whether each material has a certain sub-attribute, use binary classification variables to represent the corresponding text information and convert the material technical attribute description into a matrix composed of 0 and 1 values.
[0030] Step S3.2: Convolve the output of the Bi-LSTM layer in the original model with four types of one-dimensional convolutional layers, and finally input the proposed handcrafted feature Gaussian kernel support vector machine to obtain the classification result;
[0031] Step S3.3: Experiments were conducted using machine learning models, including random forest and support vector machine, for comparison. The models before and after optimization were compared, and the models were evaluated using four evaluation metrics: accuracy, precision, recall, and F1 score, to demonstrate the effectiveness of the models.
[0032] A deep learning-based commodity material classification system according to the present invention includes:
[0033] Module M1: Acquires material data and performs data preprocessing;
[0034] Module M2: Design and build deep learning algorithm models, and train them using processed data;
[0035] Module M3: Optimizes the architecture and data features of deep learning algorithm models, and conducts comparative experiments to prove the effectiveness of deep learning algorithm models;
[0036] Module M4: Uses an optimized deep learning algorithm model to classify a given material.
[0037] Preferably, in module M1:
[0038] Acquire material data and perform data preprocessing on the material description, material technical attributes, material type and specification, and material name in the material data;
[0039] Module M1.1: Extracts text from the material description field, cleans and removes irrelevant symbols, obtains the corresponding code for each word in the text by querying a dictionary, and initially transforms each line of text into an index vector as a feature item. A , used as input for Bi-LSTM;
[0040] Module M1.2: Based on the material's technical attributes, the data is cleaned, segmented, and a dictionary is built. Each word in the processed text is coded by querying the dictionary, and each line of text is initially converted into an index vector Item. B, used as input for Bi-LSTM;
[0041] Module M1.3: Processes the text corresponding to the material specification field, removes symbols that cannot represent the text, retains letters and numbers according to the definition based on the writing specifications of the material, obtains the corresponding code for each word in the processed text by querying a dictionary, and initially transforms each line of text into a vector as a feature item. C , used as input for Bi-LSTM;
[0042] Module M1.4: The material name text contains existing material specification data, retaining Chinese characters while removing redundant letters, numbers, punctuation, and stop words. Each word in the processed text is encoded using a dictionary lookup, and each line of text is initially converted into an index vector Item. D , used as input for Bi-LSTM.
[0043] Preferably, in module M2:
[0044] Module M2.1: Designs a word embedding layer that takes vectorized short text as input and outputs word vectors;
[0045] Design a word embedding layer, with Item A Item B Item C and Item D As input, it is encoded and transformed into word vectors:
[0046] I A =Embedding(Item) A (1)
[0047] I B =Embedding(Item) B (2)
[0048] I C =Embedding(Item) C (3)
[0049] I D =Embedding(Item) D (4)
[0050] in, Where w ai ∈R L For this data item A The embedded representation of the i-th character of the field, T a For this data item A The length of the text embedded in the field; Where w bi ∈R L For this data item B The embedded representation of the i-th character of the field, T b For this data item B The length of the text embedded in the field; Where w ci ∈R L For this data item C The embedded representation of the i-th character of the field, T C For this data item C The length of the text embedded in the field; Where w di ∈R L For this data item D The embedded representation of the i-th character of the field, T d For this data item D The length of the text embedded in the field;
[0051] Module M2.2: Design four parallel bidirectional long short-term memory networks Bi-LSTM, which take word vectors from the embedding layer as input to capture long-distance feature dependencies of the text;
[0052] Module M2.3: Design a one-dimensional convolutional layer, using the output of the Bi-LSTM layer as input, to further capture local feature relationships;
[0053] Module M2.4: Design a fully connected layer to output the word vectors trained by Bi-LSTM and one-dimensional convolutional layer to the fully connected layer for classification and output the classification results.
[0054] Preferably, in module M3:
[0055] Module M3.1: For Item B Clean up special characters and calculate the word frequency of all entries. Manually extract the pre-set features. By using whether each material has a certain sub-attribute, use binary classification variables to represent the corresponding text information and convert the material technical attribute description into a matrix composed of 0 and 1 values.
[0056] Module M3.2: Convolve the output of the Bi-LSTM layer in the original model with four types of one-dimensional convolutional layers, and finally input the proposed handcrafted feature into a Gaussian kernel support vector machine to obtain the classification result;
[0057] Module M3.3: Experiments were conducted using machine learning models, including random forest and support vector machine, for comparison. The models before and after optimization were compared, and the models were evaluated using four evaluation metrics: accuracy, precision, recall, and F1 score, to demonstrate the effectiveness of the models.
[0058] According to the present invention, a computer-readable storage medium storing a computer program is provided, wherein when the computer program is executed by a processor, it implements the steps of any of the deep learning-based commodity material classification methods described herein.
[0059] A commodity material classification device based on deep learning according to the present invention includes: a controller;
[0060] The controller includes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the deep learning-based commodity material classification methods; or, the controller includes any of the deep learning-based commodity material classification systems.
[0061] Compared with the prior art, the present invention has the following beneficial effects:
[0062] 1. This invention avoids the situation where the vector length of the text corresponding to each field is too short and the semantic information is fragmented by constructing feature vectors independently for the text of each field and performing feature learning.
[0063] 2. This invention uses a parallel quadruple Bi-LSTM network and Conv1D concatenation to generate feature vectors, thereby capturing long-range feature dependencies in text and compensating for the shortcomings of Bi-LSTM in capturing local features, thus obtaining important features of text information.
[0064] 3. This invention performs specialized processing on industrial material data, ensuring that the method has good results when applied in the field of materials. Attached Figure Description
[0065] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:
[0066] Figure 1 This is a schematic diagram of the process of the present invention;
[0067] Figure 2 This is a system module diagram of the present invention;
[0068] Figure 3 This is a schematic diagram of the features of the term. Detailed Implementation
[0069] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.
[0070] Example 1:
[0071] In e-commerce platforms, suppliers typically upload product material names and specifications, but lack leaf-level classification attributes for the materials. To facilitate material querying and management, this paper innovatively uses a parallel bidirectional long short-term memory network to extract features from material descriptions, material specifications, material names, and material technical attribute descriptions. The feature vectors are then integrated using a fully connected layer, and finally, a strong classifier is used to obtain the material classification results.
[0072] According to the present invention, a deep learning-based commodity material classification method is provided, such as... Figures 1-3 As shown, it includes:
[0073] A deep learning-based commodity material classification method provided by the present invention includes:
[0074] Step S1: Obtain material data and perform data preprocessing;
[0075] Specifically, in step S1:
[0076] Acquire material data and perform data preprocessing on the material description, material technical attributes, material type and specification, and material name in the material data;
[0077] Step S1.1: Extract text from the material description field, clean and remove irrelevant symbols, obtain the corresponding code for each word in the text by querying a dictionary, and initially convert each line of text into an index vector as a feature item. A , used as input for Bi-LSTM;
[0078] Step S1.2: Based on the material's technical attributes, clean and segment the data, construct a dictionary, and obtain the corresponding code for each word in the processed text by querying the dictionary. This initially transforms each line of text into an index vector Item. B , used as input for Bi-LSTM;
[0079] Step S1.3: Process the text corresponding to the material specification field, remove symbols that cannot represent the text, retain letters and numbers according to the definition based on the writing specifications of the material specifications, obtain the corresponding code for each word in the processed text by querying a dictionary, and initially convert each line of text into a vector as a feature item. C , used as input for Bi-LSTM;
[0080] Step S1.4: The material name text contains existing material specification data. Chinese characters are retained, while redundant letters, numbers, punctuation, and stop words are removed. Each word in the processed text is encoded using a dictionary lookup, and each line of text is initially converted into an index vector Item.D , used as input for Bi-LSTM.
[0081] Step S2: Design and build a deep learning algorithm model, and train it using the processed data;
[0082] Specifically, in step S2:
[0083] Step S2.1: Design a word embedding layer, taking the vectorized short text as input and outputting word vectors;
[0084] Design a word embedding layer, with Item A Item B Item C and Item D As input, it is encoded and transformed into word vectors:
[0085] I A =Embedding(Item) A (1)
[0086] I B =Embedding(Item) B (2)
[0087] I C =Embedding(Item) C (3)
[0088] I D =Embedding(Item) D (4)
[0089] in, Where w ai ∈R L For this data item A The embedded representation of the i-th character of the field, T a For this data item A The length of the text embedded in the field; Where w bi ∈R L For this data item B The embedded representation of the i-th character of the field, T b For this data item B The length of the text embedded in the field; Where w ci ∈R L For this data item C The embedded representation of the i-th character of the field, T C For this data item CThe length of the text embedded in the field; Where w di ∈R L For this data item D The embedded representation of the i-th character of the field, T d For this data item D The length of the text embedded in the field;
[0090] Step S2.2: Design four parallel bidirectional long short-term memory networks Bi-LSTM, using word vectors from the embedding layer as input, to capture long-distance feature dependencies of the text;
[0091] Step S2.3: Design a one-dimensional convolutional layer, using the output of the Bi-LSTM layer as input, to further capture local feature relationships;
[0092] Step S2.4: Design a fully connected layer, output the word vectors trained by Bi-LSTM and one-dimensional convolutional layer to the fully connected layer for classification, and output the classification result.
[0093] Step S3: Optimize the deep learning algorithm model architecture and data features, and conduct comparative experiments to prove the effectiveness of the deep learning algorithm model;
[0094] Specifically, in step S3:
[0095] Step S3.1: For Item B Clean up special characters and calculate the word frequency of all entries. Manually extract the pre-set features. By using whether each material has a certain sub-attribute, use binary classification variables to represent the corresponding text information and convert the material technical attribute description into a matrix composed of 0 and 1 values.
[0096] Step S3.2: Convolve the output of the Bi-LSTM layer in the original model with four types of one-dimensional convolutional layers, and finally input the proposed handcrafted feature Gaussian kernel support vector machine to obtain the classification result;
[0097] Step S3.3: Experiments were conducted using machine learning models, including random forest and support vector machine, for comparison. The models before and after optimization were compared, and the models were evaluated using four evaluation metrics: accuracy, precision, recall, and F1 score, to demonstrate the effectiveness of the models.
[0098] Step S4: Use the optimized deep learning algorithm model to classify the given materials.
[0099] Example 2:
[0100] Example 2 is a preferred embodiment of Example 1, and is used to illustrate the present invention in more detail.
[0101] The present invention also provides a commodity material classification system based on deep learning. The commodity material classification system based on deep learning can be implemented by executing the process steps of the commodity material classification method based on deep learning. That is, those skilled in the art can understand the commodity material classification method based on deep learning as a preferred embodiment of the commodity material classification system based on deep learning.
[0102] A deep learning-based commodity material classification system according to the present invention includes:
[0103] Module M1: Acquires material data and performs data preprocessing;
[0104] Specifically, in module M1:
[0105] Acquire material data and perform data preprocessing on the material description, material technical attributes, material type and specification, and material name in the material data;
[0106] Module M1.1: Extracts text from the material description field, cleans and removes irrelevant symbols, obtains the corresponding code for each word in the text by querying a dictionary, and initially transforms each line of text into an index vector as a feature item. A , used as input for Bi-LSTM;
[0107] Module M1.2: Based on the material's technical attributes, the data is cleaned, segmented, and a dictionary is built. Each word in the processed text is coded by querying the dictionary, and each line of text is initially converted into an index vector Item. B , used as input for Bi-LSTM;
[0108] Module M1.3: Processes the text corresponding to the material specification field, removes symbols that cannot represent the text, retains letters and numbers according to the definition based on the writing specifications of the material, obtains the corresponding code for each word in the processed text by querying a dictionary, and initially transforms each line of text into a vector as a feature item. C , used as input for Bi-LSTM;
[0109] Module M1.4: The material name text contains existing material specification data, retaining Chinese characters while removing redundant letters, numbers, punctuation, and stop words. Each word in the processed text is encoded using a dictionary lookup, and each line of text is initially converted into an index vector Item. D , used as input for Bi-LSTM.
[0110] Module M2: Design and build deep learning algorithm models, and train them using processed data;
[0111] Specifically, in module M2:
[0112] Module M2.1: Designs a word embedding layer that takes vectorized short text as input and outputs word vectors;
[0113] Design a word embedding layer, with Item A Item B Item C and Item D As input, it is encoded and transformed into word vectors:
[0114] I A =Embedding(Item) A (1)
[0115] I B =Embedding(Item) B (2)
[0116] I C =Embedding(Item) C (3)
[0117] I D =Embedding(Item) D (4)
[0118] in, Where w ai ∈R L For this data item A The embedded representation of the i-th character of the field, T a For this data item A The length of the text embedded in the field; Where w bi ∈R L For this data item B The embedded representation of the i-th character of the field, T b For this data item B The length of the text embedded in the field; Where w ci ∈R L For this data item C The embedded representation of the i-th character of the field, T C For this data item C The length of the text embedded in the field; Where w di ∈R L For this data item D The embedded representation of the i-th character of the field, T d For this data item DThe length of the text embedded in the field;
[0119] Module M2.2: Design four parallel bidirectional long short-term memory networks Bi-LSTM, which take word vectors from the embedding layer as input to capture long-distance feature dependencies of the text;
[0120] Module M2.3: Design a one-dimensional convolutional layer, using the output of the Bi-LSTM layer as input, to further capture local feature relationships;
[0121] Module M2.4: Design a fully connected layer to output the word vectors trained by Bi-LSTM and one-dimensional convolutional layer to the fully connected layer for classification and output the classification results.
[0122] Module M3: Optimizes the architecture and data features of deep learning algorithm models, and conducts comparative experiments to prove the effectiveness of deep learning algorithm models;
[0123] Specifically, in module M3:
[0124] Module M3.1: For Item B Clean up special characters and calculate the word frequency of all entries. Manually extract the pre-set features. By using whether each material has a certain sub-attribute, use binary classification variables to represent the corresponding text information and convert the material technical attribute description into a matrix composed of 0 and 1 values.
[0125] Module M3.2: Convolve the output of the Bi-LSTM layer in the original model with four types of one-dimensional convolutional layers, and finally input the proposed handcrafted feature into a Gaussian kernel support vector machine to obtain the classification result;
[0126] Module M3.3: Experiments were conducted using machine learning models, including random forest and support vector machine, for comparison. The models before and after optimization were compared, and the models were evaluated using four evaluation metrics: accuracy, precision, recall, and F1 score, to demonstrate the effectiveness of the models.
[0127] Module M4: Uses an optimized deep learning algorithm model to classify a given material.
[0128] According to the present invention, a computer-readable storage medium storing a computer program is provided, wherein when the computer program is executed by a processor, it implements the steps of any of the deep learning-based commodity material classification methods described herein.
[0129] A commodity material classification device based on deep learning according to the present invention includes: a controller;
[0130] The controller includes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the deep learning-based commodity material classification methods; or, the controller includes any of the deep learning-based commodity material classification systems.
[0131] Example 3:
[0132] Example 3 is a preferred example of Example 1, and is used to illustrate the present invention in more detail.
[0133] Step 1: Obtain material data.
[0134] Step 2: Perform data preprocessing on the "Material Description", "Material Technical Attributes", "Material Specifications" and "Material Name" in the material data.
[0135] Step 3: Design and build a deep learning algorithm model, and train it using the processed data.
[0136] Step 4: Further optimize the model architecture and data features, and conduct comparative experiments to prove the effectiveness of the model.
[0137] Step 5: Use the optimized model to complete the intelligent classification of the given materials.
[0138] Step 2 includes the following steps:
[0139] Step 2.1: In intelligent material classification, relevant material information needs to be provided, such as: "Material Description", "Material Technical Attribute Description", "Material Type Specification", "Material Name", "Intermediate Category Code", "Major Category Code", etc. These attributes are the fields referred to in the text, and the text is the specific content of the fields.
[0140] Extract the "Material" field and the text following it, clean and remove irrelevant symbols, obtain the corresponding code for each word in the text by querying a dictionary, and initially transform each line of text into an index vector as a feature item. A For the "Material Description" field, using the word "material" as a delimiter, the text following it is extracted and subjected to data cleaning, word segmentation, etc., to construct a {word-index} dictionary. Each word in the processed text is coded by querying the dictionary, thus initially transforming each line of text into an index vector Item. A , used as input for Bi-LSTM;
[0141] Step 2.2: Perform character-level word segmentation similar to "Material Description", and establish separate word lists and corresponding index vectors as feature items. BFor the "material technical attributes," the data is cleaned, segmented, and a {word-index} dictionary is constructed. Each word in the processed text is coded by querying the dictionary, initially transforming each line of text into an index vector Item. B , used as input for Bi-LSTM;
[0142] Step 2.3: Process the text corresponding to the "Material Specification" field, removing symbols that do not represent the text, such as "\|-,;.()", etc. Based on the writing standards for industrial material specifications, some letters represent special meanings, such as "WJ41H-16C", where "W indicates bellows seal, J indicates gate valve, 4 indicates flange connection, 1 indicates straight flow channel, H indicates the sealing surface is a Cr-containing alloy (stainless steel), and 16C indicates a nominal pressure of 1.6 MPa". Based on these definitions, retain only the letters and numbers. Construct a {word-index} dictionary, and obtain the corresponding code for each word in the processed text by querying the dictionary, initially converting each line of text into an index vector Item. C This is used as input to the Bi-LSTM and a {word-index} dictionary is constructed. Each word in the processed text is coded by querying the dictionary, thus initially transforming each line of text into an index vector Item. C , used as input for Bi-LSTM.
[0143] Step 2.4: The "Material Name" text contains existing "Material Specification" data. To avoid duplication, only the Chinese characters are retained, and redundant letters, numbers, punctuation, and stop words are removed. A {word-index} dictionary is constructed, and the corresponding code for each word in the processed text is obtained by querying the dictionary, initially converting each line of text into an index vector Item. D , used as input for Bi-LSTM.
[0144] Step 3 includes the following steps:
[0145] Step 3.1: Design a word embedding layer, taking the vectorized short text as input and outputting word vectors;
[0146] Design the word embedding layer, based on the previously mentioned Item. A Item B Item C Item D As input, it is encoded and transformed into word vectors of appropriate dimensions, which can describe the relationship between words at a higher semantic level, allowing the Bi-LSTM model to better utilize the contextual relationships in the text. The specific formulas are shown in (1)-(4):
[0147] I A =Embedding(Item) A(1)
[0148] I B =Embedding(Item) B (2)
[0149] I C =Embedding(Item) C (3)
[0150] I D =Embedding(Item) D (4)
[0151] in Where w ai ∈R L For this data item A The embedded representation of the i-th character of the field, T a For this data item A The text embedding length of the field can be obtained similarly.
[0152]
[0153] Step 3.2: Design four parallel bidirectional long short-term memory (Bi-LSTM) networks, using word vectors from the embedding layer as input, to capture long-range feature dependencies in the text. Bi-LSTM can capture dependencies over longer distances. Using a parallel quadruple Bi-LSTM network to extract features from four fields—material description, material specifications, material name, and material technical attribute description—can better capture contextual semantic information. However, directly merging the four extracted results fails to highlight important information, leading to a degree of information decay. Using Conv1D convolution operations to merge the results extracts important features from local information, compensating for the shortcomings of Bi-LSTM in capturing local text features and obtaining the key features of the text information.
[0154] Step 3.3: Design a one-dimensional convolutional layer, using the output of the Bi-LSTM layer as input, to further capture local feature relationships.
[0155] Step 3.4: Design a fully connected layer, output the word vectors trained by Bi-LSTM and one-dimensional convolutional layer to the fully connected layer for classification, and output the classification results.
[0156] Step 4 includes the following steps:
[0157] Step 4.1: For Item BThe process involves cleaning up special characters and calculating the word frequency of all entries. The first 115 features are manually extracted. By using binary categorical variables to represent the corresponding text information based on whether each material possesses a specific sub-attribute, the material's technical attribute description is converted into a matrix of 0-1 values. Specific features are as follows: Figure 3 As shown.
[0158] Step 4.2: Convolve the output of the Bi-LSTM layer in the original model with four types of one-dimensional convolutional layers (Conv1D), and finally obtain the classification result by inputting the handcrafted feature Gaussian kernel support vector machine proposed in 4.1.
[0159] Step 4.3: Experiments were conducted using common machine learning models such as random forest and support vector machine for comparison. The models before and after optimization were compared. The model was evaluated using four evaluation metrics: accuracy, precision, recall, and F1 score, thereby proving the effectiveness of the model.
[0160] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.
[0161] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.
Claims
1. A commodity material classification method based on deep learning, characterized in that, include: Step S1: Obtain material data and perform data preprocessing; Step S2: Design and build a deep learning algorithm model, and train it using the processed data; Step S3: Optimize the deep learning algorithm model architecture and data features, and conduct comparative experiments to prove the effectiveness of the deep learning algorithm model; Step S4: Use the optimized deep learning algorithm model to classify the given materials; In step S1: Acquire material data and perform data preprocessing on the material description, material technical attributes, material type and specification, and material name in the material data; Step S1.1: Extract text from the material description field, clean and remove irrelevant symbols, obtain the corresponding code for each word in the text by querying a dictionary, and initially convert each line of text into an index vector as a feature. , used as input for Bi-LSTM; Step S1.2: Based on the material's technical attributes, clean and segment the data, construct a dictionary, and obtain the corresponding code for each word in the processed text by querying the dictionary, initially converting each line of text into an index vector. , used as input for Bi-LSTM; Step S1.3: Process the text corresponding to the material specification field, remove symbols that cannot represent the text, retain letters and numbers according to the definition based on the writing specifications of the material specifications, obtain the corresponding code for each word in the processed text by querying a dictionary, and initially convert each line of text into a vector as a feature. , used as input for Bi-LSTM; Step S1.4: The material name text contains existing material specification data. Chinese characters are retained, while redundant letters, numbers, punctuation, and stop words are removed. Each word in the processed text is encoded using a dictionary lookup, and each line of text is initially converted into an index vector. , used as input for Bi-LSTM; In step S2: Step S2.1: Design a word embedding layer, taking the vectorized short text as input and outputting word vectors; Design a word embedding layer, to , , and As input, it is encoded and transformed into word vectors: （1） （2） （3） （4） in, ,in For this data The field's first Embedded representation of a character, For this data The length of the text embedded in the field; ,in For this data The field's first Embedded representation of a character, For this data The length of the text embedded in the field; ,in For this data The field's first Embedded representation of a character, For this data The length of the text embedded in the field; ,in For this data The field's first Embedded representation of a character, For this data The length of the text embedded in the field; Step S2.2: Design four parallel bidirectional long short-term memory networks Bi-LSTM, using word vectors from the embedding layer as input, to capture long-distance feature dependencies of the text; Step S2.3: Design a one-dimensional convolutional layer, using the output of the Bi-LSTM layer as input, to further capture local feature relationships; Step S2.4: Design a fully connected layer, output the word vectors trained by Bi-LSTM and one-dimensional convolutional layer to the fully connected layer for classification, and output the classification result; In step S3: Step S3.1: For Clean up special characters and calculate the word frequency of all entries. Manually extract the pre-set features. By using whether each material has a certain sub-attribute, use binary classification variables to represent the corresponding text information and convert the material technical attribute description into a matrix composed of 0 and 1 values. Step S3.2: Convolve the output of the Bi-LSTM layer in the original model with four types of one-dimensional convolutional layers, and finally input the proposed handcrafted feature Gaussian kernel support vector machine to obtain the classification result; Step S3.3: Experiments were conducted using machine learning models, including random forest and support vector machine, for comparison. The models before and after optimization were compared, and the models were evaluated using four evaluation metrics: accuracy, precision, recall, and F1 score, to demonstrate the effectiveness of the models.
2. A commodity material classification system based on deep learning, characterized in that, include: Module M1: Acquires material data and performs data preprocessing; Module M2: Design and build deep learning algorithm models, and train them using processed data; Module M3: Optimizes the architecture and data features of deep learning algorithm models, and conducts comparative experiments to prove the effectiveness of deep learning algorithm models; Module M4: Uses an optimized deep learning algorithm model to classify a given material; In module M1: Acquire material data and perform data preprocessing on the material description, material technical attributes, material type and specification, and material name in the material data; Module M1.1: Extracts text from the material description field, cleans and removes irrelevant symbols, obtains the corresponding code for each word in the text by querying a dictionary, and initially transforms each line of text into an index vector as a feature. , used as input for Bi-LSTM; Module M1.2: Based on the material's technical attributes, the data is cleaned, segmented, and a dictionary is built. Each word in the processed text is coded by querying the dictionary, and each line of text is initially converted into an index vector. , used as input for Bi-LSTM; Module M1.3: Processes the text corresponding to the material specification field, removes symbols that cannot represent the text, retains letters and numbers according to the definition based on the writing specifications of the material, obtains the corresponding code for each word in the processed text by querying a dictionary, and initially transforms each line of text into a vector as a feature. , used as input for Bi-LSTM; Module M1.4: The material name text contains existing material specification data, retaining Chinese characters while removing redundant letters, numbers, punctuation, and stop words. Each word in the processed text is encoded using a dictionary lookup, and each line of text is initially converted into an index vector. , used as input for Bi-LSTM; In module M2: Module M2.1: Designs a word embedding layer that takes vectorized short text as input and outputs word vectors; Design a word embedding layer, to , , and As input, it is encoded and transformed into word vectors: （1） （2） （3） （4） in, ,in For this data The field's first Embedded representation of a character, For this data The length of the text embedded in the field; ,in For this data The field's first Embedded representation of a character, For this data The length of the text embedded in the field; ,in For this data The field's first Embedded representation of a character, For this data The length of the text embedded in the field; ,in For this data The field's first Embedded representation of a character, For this data The length of the text embedded in the field; Module M2.2: Design four parallel bidirectional long short-term memory networks Bi-LSTM, which take word vectors from the embedding layer as input to capture long-distance feature dependencies of the text; Module M2.3: Design a one-dimensional convolutional layer, using the output of the Bi-LSTM layer as input, to further capture local feature relationships; Module M2.4: Design a fully connected layer to output the word vectors trained by Bi-LSTM and one-dimensional convolutional layer to the fully connected layer for classification, and output the classification results; In module M3: Module M3.1: For Clean up special characters and calculate the word frequency of all entries. Manually extract the pre-set features. By using whether each material has a certain sub-attribute, use binary classification variables to represent the corresponding text information and convert the material technical attribute description into a matrix composed of 0 and 1 values. Module M3.2: Convolve the output of the Bi-LSTM layer in the original model with four types of one-dimensional convolutional layers, and finally input the proposed handcrafted feature into a Gaussian kernel support vector machine to obtain the classification result; Module M3.3: Experiments were conducted using machine learning models, including random forest and support vector machine, for comparison. The models before and after optimization were compared, and the models were evaluated using four evaluation metrics: accuracy, precision, recall, and F1 score, to demonstrate the effectiveness of the models.
3. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the deep learning-based commodity material classification method as described in claim 1.
4. A commodity material classification device based on deep learning, characterized in that, include: Controller; The controller includes a computer-readable storage medium storing a computer program as described in claim 3, wherein the computer program, when executed by a processor, implements the steps of the deep learning-based commodity material classification method as described in claim 1; or, the controller includes the deep learning-based commodity material classification system as described in claim 2.