Shoe product management system for big data processing shoe material customization

Abstract

The invention relates to the technical field of shoe product customization, in particular to a shoe product management system for big data processing shoe material customization, which comprises a data collection and preprocessing unit, a natural language processing core unit, a quantitative index generation unit and a customization process fusion unit, a data collection and preprocessing unit collects and standardizes expected text data of a user for shoe materials, a natural language processing core unit identifies user intentions, keywords and modifiers and analyzes demand priorities, a quantitative index generation unit converts functional demands and style preferences into quantitative indexes and feature vectors, and the quantitative indexes and the feature vectors are stored in a database. The customization process fusion unit inputs quantitative indexes into a computer aided design model, screens shoe materials, adjusts design parameters and generates a design scheme, defines a design direction by using a quantification means, reduces errors between user expectations and design results, considers shoe material cost and supply conditions, meets diversified requirements of users, and improves design efficiency. And the shoe product customization quality and the user satisfaction are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of shoe product customization, and more specifically, to a shoe product management system for big data - processed shoe material customization. Background Art

[0002] Shoe product customization is an important technology. In the traditional shoe material customization design process, due to problems in communication or description, there are often large errors between the user's expectations and the final design results.

[0003] When a user expresses their expected shoe material style or style to a designer, due to the lack of accurate demand understanding and quantification means, the designer mainly relies on experience to interpret the user's description. However, the user's expression may be vague, unprofessional, and contain a large number of non - standardized expressions, making it difficult for the designer to accurately grasp the key information. There are prone to deviations in the understanding of color matching, material texture, decorative details, etc. At the same time, in terms of functional requirements, such as functions like breathability and wear resistance, there is a lack of scientific quantification standards and bases. It is difficult for the designer to convert the user's functional requirements into specific design indicators, and thus there is blindness in selecting shoe materials and determining design parameters during the design process, resulting in low design process efficiency and difficulty in meeting the diverse needs of users, seriously affecting the quality of shoe product customization and the user experience. To solve this technical problem, we provide a shoe product management system for big data - processed shoe material customization. Summary of the Invention

[0004] The purpose of the present invention is to provide a shoe product management system for big data - processed shoe material customization to solve the problems raised in the above - mentioned background art.

[0005] To achieve the above - mentioned purpose, a shoe product management system for big data - processed shoe material customization is provided, including a data collection and pre - processing unit, a natural language processing core unit, a quantification index generation unit, and a customization process integration unit;

[0006] The data collection and pre - processing unit collects the user's expected text data and historical expected text data of shoe materials through a data acquisition module, and performs standardization processing on the collected data through a standardization module;

[0007] The natural language processing core unit includes a built - in intention recognition and classification module, which uses a pre - trained model based on deep learning to perform intention recognition on the pre - processed text to judge the user's intention, and uses a sequence labeling model to identify keywords and modifiers in the text. It also includes a priority parsing module, which constructs a priority model through a rule engine combined with a machine learning algorithm. For the text sorted by the user's determined priority, the priority is extracted, and for the text without clear sorting, the priority is calculated based on word frequency statistics and semantic association analysis;

[0008] The quantization index generation unit establishes the mapping relationship from functional requirements to quantization indices through the functional quantization mapping module and sets the weight coefficients. At the same time, it obtains the final breathability quantization value through weighted calculation, and quantizes various fashion styles through the style quantization coding module to construct a style feature vector, quantizing and composing a multi-dimensional vector from the dimensions of color matching, material texture, and decorative details;

[0009] The customization process integration unit transmits the generated quantization design indices to the computer-aided design model through the design software docking module. The computer-aided design model filters the materials in the shoe material library according to the functional quantization indices, adjusts the shoe style design parameters according to the style vector, and generates the final design scheme.

[0010] As a further improvement of this technical solution, the standardization module adopts a standardization algorithm based on the word vector space, specifically as follows:

[0011] Perform word segmentation on the collected text data to obtain a series of words, and then map each word to the pre-trained word vector space to obtain the corresponding word vectors. For synonyms and near-synonyms, classify them into the same standard word by calculating the cosine similarity between the word vectors. For misspelled words, use the edit distance algorithm combined with the semantic information of the word vectors to correct the errors and convert them into standard words.

[0012] As a further improvement of this technical solution, the intention recognition and classification module adopts an adaptive adjustment strategy of the pre-trained model based on deep learning that introduces a multi-head self-attention mechanism. According to different text lengths and complexities, dynamically adjust the number of heads and the distribution of attention weights in the multi-head self-attention mechanism. At the same time, add a multi-classifier to the output layer of the model, and use the Softmax function to convert the output of the model into the probability distribution of different intention categories, and determine the user's intention according to the category with the largest probability value.

[0013] As a further improvement of this technical solution, the sequence labeling model adopts a model that combines a bidirectional long short-term memory network and a conditional random field based on a deep learning architecture, specifically as follows:

[0014] Use the deep learning architecture to extract features from the input text to capture the long-distance dependencies in the text, then input the extracted features into the bidirectional long short-term memory network to learn the context information of the text, and finally perform sequence labeling on the output of the bidirectional long short-term memory network through the conditional random field layer to identify the keywords and modifiers in the text. During the training process, adopt an adversarial training method to introduce a discriminator to distinguish between real labels and labels generated by the model.

[0015] As a further improvement of this technical solution, the priority model constructed by the priority parsing module adopts a method based on graph neural network, which is specifically as follows:

[0016] Take the words in the text as the nodes of the graph, and the semantic associations between the words as the edges to construct a semantic graph. Use the graph neural network to learn the semantic graph and update the feature representations of the nodes. For the text with the priority sorting determined by the user, assign initial priorities to the nodes according to the sorting information in the text. For the text without clear sorting, calculate the priorities of the nodes by combining the degree centrality of the nodes in the graph with the word frequency statistics results.

[0017] As a further improvement of this technical solution, when the functional quantization mapping module establishes the mapping relationship from functional requirements to quantization indicators, it adopts a method combining case-based reasoning and fuzzy logic, which is specifically as follows:

[0018] Establish a historical case library to store historical functional requirements and corresponding quantization indicators. When there is a new functional requirement, retrieve the most similar case from the case library through similarity calculation, and then use fuzzy logic to adjust the retrieved case and set the weight coefficient. Obtain the final breathability quantization value through weighted calculation.

[0019] As a further improvement of this technical solution, when the style quantization encoding module constructs the style feature vector, it adopts a method combining principal component analysis and clustering analysis, which is specifically as follows:

[0020] Collect fashion style sample data from the dimensions of color matching, material texture, and decorative details, extract features from the sample data to obtain high-dimensional feature vectors, then use principal component analysis to reduce the dimensionality of the high-dimensional feature vectors, and finally use the clustering analysis algorithm to cluster the feature vectors after dimensionality reduction. Classify the styles beyond the similarity threshold into one category, and assign a unique code to each category of styles for constructing the style feature vector.

[0021] As a further improvement of this technical solution, when the computer-aided design model screens the materials in the shoe material library according to the functional quantization indicators, it adopts a multi-objective optimization algorithm, which is specifically as follows:

[0022] Take the functional quantization indicators as multiple optimization objectives, and at the same time take the cost and supply situation of the shoe materials as constraints. Search for the optimal shoe material combination that meets multiple objectives in the shoe material library through the multi-objective optimization algorithm. When adjusting the shoe design parameters according to the style vector, use the genetic algorithm to optimize the design parameters, and continuously iterate and update the design parameters according to the requirements of the style vector to generate a shoe design scheme that meets the user's style requirements.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] In a shoe product management system for customizing shoe materials through big data processing, the natural language processing core unit has an intention recognition and classification module built therein. By means of a pre-trained model based on deep learning and an adaptive adjustment strategy introducing a multi-head self-attention mechanism, it can dynamically adjust the number of heads and attention weights according to the length and complexity of different texts, accurately recognize the user's intention, add a multi-classifier and a Softmax function in the output layer to convert the model output into a probability distribution of different intention categories, thereby determining the user's intention, greatly improving the pertinence and efficiency of the service. The sequence labeling model adopts an architecture combining a bidirectional long short-term memory network and a conditional random field, effectively capturing the long-distance dependence relationship and context information of the text, accurately identifying keywords and modifiers, and making the system's understanding of the user's needs more precise. The priority parsing module constructs a priority model using a method based on a graph neural network. Whether it is text with explicit sorting by the user or text without explicit sorting, it can reasonably calculate the priority, further enhancing the system's processing ability for the user's needs and the quality of the customization service. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is the overall block diagram of the present invention.

[0026] The meanings of the various reference numerals in the figure are as follows:

[0027] 1. Data collection and preprocessing unit; 11. Data acquisition module; 12. Standardization module; 2. Natural language processing core unit; 21. Intention recognition and classification module; 22. Priority parsing module; 3. Quantification index generation unit; 31. Functional quantification mapping module; 32. Style quantification coding module; 4. Customization process integration unit; 41. Design software docking module. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0028] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0029] The present invention provides a shoe product management system for customizing shoe materials through big data processing. Please refer to Figure 1 as shown, which includes a data collection and preprocessing unit 1, a natural language processing core unit 2, a quantification index generation unit 3, and a customization process integration unit 4;

[0030] The data collection and preprocessing unit 1 collects the expected text data and historical expected text data of shoe materials through the data acquisition module 11, and performs standardization processing on the collected data through the standardization module 12. The collected text data is a continuous string. In order to process and analyze individual words subsequently, it is necessary to split it into a series of independent words. Use a word segmentation tool to perform word segmentation on the collected text data. For example, for the Chinese text "I want breathable sports shoes", after word segmentation by the word segmentation tool, words such as "I", "want", "breathable", "of", and "sports shoes" are obtained, providing clear word inputs for subsequent word vector mapping and standardization operations. Mapping words into a pre-trained word vector space can convert words into numerical vector representations. For each word obtained after word segmentation, look up its corresponding word vector in the pre-trained word vector model. For example, if "breathable" is set as , then the word vector of the word "breathable" in the word vector model is , where is the dimension of the word vector, making the standardization operation more accurate and intelligent. There may be a large number of synonyms and near-synonyms in the text. They express similar semantics, but will increase the redundancy of the data. For any two words and , their corresponding word vectors are and respectively. Calculate the cosine similarity between them, and set a similarity threshold. If the cosine similarity is greater than or equal to the similarity threshold, then it is considered that and are synonyms or near-synonyms, and group them into the same standard word. One of the words can be selected as the standard word, and other words are replaced with this standard word, which helps to better mine the key information in the text. There may be misspelled words in the collected text data, and these errors will affect the subsequent analysis and processing results. For each word , first look up the candidate word set with the smallest edit distance from in the vocabulary of the pre-trained word vector model. The edit distance refers to the minimum number of edit operations required to convert one string into another string between two strings. Calculate the cosine similarity between the word vector of the word and the word vectors of each word in the candidate word set . Select the word with the largest cosine similarity and an edit distance less than a certain threshold as The error correction result provides a more accurate input for subsequent analysis and processing. The processed words are combined into a new text in the original order. For example, after the above processing, the original text "I want breathable sports shoes" is standardized to "I want ventilated sports shoes", providing high-quality input for subsequent processing.

[0031] The core unit 2 of natural language processing includes a built-in intent recognition and classification module 21. It uses a pre-trained model based on deep learning to perform intent recognition on the preprocessed text, judge the user's intent, input the text processed by the normalization module 12 into the pre-trained model to obtain the feature representation of the text. Different text lengths and complexities have different requirements for the attention mechanism. Define a head number adjustment function , where is the text length, is the text complexity. In the multi-head attention mechanism, the attention weight of each head is adjusted by a learnable weight matrix . The elements of the matrix are dynamically updated according to the text length and complexity. Specifically, after calculating the attention scores of each head, calculate the adjusted attention weights through to improve the accuracy of intent recognition. Add a fully connected layer as a multi-classifier to the output layer of the pre-trained model, and the number of its neurons is equal to the number of intent categories . Let the output of the pre-trained model be , the weight matrix of the multi-classifier be , and the bias vector be . Then the output of the multi-classifier . Then use the Softmax function to convert into the probability distribution of different intent categories . The formula is ; where , is the probability of the th intent category, is the th element of the output vector of the multi-classifier, so that the model can clearly give the probability of each intent category. Select the category with the largest probability value as the user's intent, that is, find the largest element in , and its corresponding category is the user's intent, which can be expressed as , improving the pertinence and efficiency of the service.

[0032] Use a sequence labeling model to identify keywords and modifiers in the text. The input text contains rich semantic information, but the original text form is not conducive to direct processing by the model. A pre-trained word embedding model is selected to convert each word in the input text into a word vector of a fixed dimension, providing more representative input features for the subsequent bidirectional long short-term memory network and enhancing the model's expressive ability. The meanings of keywords and modifiers in the text are often closely related to their context. The extracted feature sequence is input into the bidirectional long short-term memory network, which consists of a forward long short-term memory network and a backward long short-term memory network. The forward long short-term memory network processes the input sequence from left to right, and the backward long short-term memory network processes the input sequence from right to left, improving the recognition accuracy of keywords and modifiers and enabling a more accurate analysis of the relationships between words in the text. The output of the bidirectional long short-term memory network is the feature representation at each moment. However, directly labeling based on these features may ignore the dependencies between labels. The output of the bidirectional long short-term memory network is input into the conditional random field layer, which defines a conditional probability. During training, the parameters of the conditional random field are learned by maximizing the log-likelihood function of the training data. During prediction, the Viterbi algorithm is used to find the label sequence with the highest score to facilitate a more accurate identification of keywords and modifiers in the text, and the labeling result is more reasonable semantically. During the training process, the model may suffer from overfitting or generate low-quality labels. The method of adversarial training is introduced by adding a discriminator, which is a binary classifier used to distinguish between true labels and labels generated by the model. The input of the discriminator is the feature sequence and the corresponding label sequence, and the output is a probability value. The goal of the generator is to generate a label sequence that can deceive the discriminator, and the goal of the discriminator is to accurately distinguish between true labels and generated labels. During training, the parameters of the generator and the discriminator are alternately updated. By continuously iteratively updating the parameters of the generator and the discriminator, the label sequence generated by the generator is made closer to the true label, enabling the model to perform well when facing different types of texts and enhancing the robustness and accuracy of the sequence labeling model.

[0033] It also includes a priority parsing module 22, which builds a priority model through a rule engine combined with a machine learning algorithm, extracts the priority of the text according to the priority order determined by the user, and calculates the priority of the text that is not clearly ordered based on word frequency statistics and semantic association analysis. There are semantic associations between words in the text. Representing these relationships in the form of a graph can more intuitively reflect the structure and interaction between words. The input text is segmented to obtain a series of words, and each word is used as a node in the graph. If there is a semantic association between two words, an edge is added between them. The semantic similarity between the words is calculated using a pre-trained word vector model. When the similarity exceeds a certain threshold, , it is believed that there is a semantic association between the two words, so that the model can analyze the text from the perspective of the semantic association of the words, laying the foundation for accurate priority calculation. The graph neural network can automatically learn the feature representation of the nodes in the graph, and through the information transmission between the nodes, it can explore the potential relationship between the nodes, and select the graph neural network model to update each node in the graph to obtain the final feature representation of the node, so that the characteristics of the node can better reflect its importance in the semantic graph, providing a more accurate basis for subsequent priority calculation. For the priority sorting text determined by the user, it contains clear sorting information, and the initial priority is assigned to the node based on this information. Assuming that the user's priority sorting text is " ",in It is the words in the text. According to this sorting information, the initial priority is assigned to the corresponding node For other nodes that are not in the sorted text, the initial priority can be set to 0, which helps the model converge to a reasonable priority allocation result more quickly. For texts that are not clearly sorted, it is necessary to comprehensively evaluate the importance of the node by calculating the degree centrality and word frequency statistics of the node in the graph. The degree centrality of is defined as the number of its neighbor nodes. The degree centrality reflects the degree of connection between nodes in the semantic graph. The larger the degree, the more extensive the connection between the node and other nodes. The frequency of each word in the text is counted. The higher the word frequency, the more important the word is in the text. The degree centrality and word frequency statistics are combined to calculate the node Priority ,For texts that are not clearly sorted, the priority of the nodes can be reasonably calculated, which improves the versatility and applicability of the model.

[0034] The quantitative index generation unit 3 establishes a mapping relationship from functional requirements to quantitative indicators and sets weight coefficients through the functional quantitative mapping module 31, and obtains the final air permeability quantitative value through weighted calculation. It collects historical functional requirement information related to housekeeping services, such as the requirements of different types of cleaning tasks for the air permeability of tools, and the corresponding air permeability quantitative indicators, such as air permeability and air permeability, and represents each case as a tuple , where is a description of functional requirements, is the corresponding quantified index value. Store these cases in a database or file system to form a historical case library , which provides rich reference data for the quantification of new functional requirements, helps to obtain quantification results quickly and accurately. For new functional requirements , calculate its similarity with the functional requirements of each case in the case library , find the case with the highest similarity , and its corresponding quantified index is , quickly find the most relevant historical cases through similarity calculation, utilize existing experience and data, reduce the difficulty and time cost of quantifying new functional requirements, define fuzzy rules and fuzzy sets. For example, for the breathable functional requirement, define fuzzy sets "high breathable requirement", "medium breathable requirement", "low breathable requirement", and determine the membership function of each fuzzy set. According to the differences between the new functional requirement and the functional requirements of the retrieved cases, obtain an adjustment coefficient through fuzzy inference, improve the accuracy and adaptability of the quantification result, and better meet the characteristics of the new functional requirement. In addition to the most similar case, the top cases with relatively high similarity can also be selected, set weight coefficients for each case, for the quantified index of each case , combine the adjustment coefficient to obtain the final breathable quantification value through weighted calculation , so that the final quantification value can more comprehensively and accurately reflect the new functional requirement.

[0035] ​And through the style quantization encoding module 32, various fashion styles are quantized and encoded to construct style feature vectors. A multi-dimensional vector is quantized and composed from the dimensions of color matching, material texture, and decorative details. A large number of sample data of different fashion styles in the form of pictures and text descriptions are collected through web crawlers, professional fashion databases, and offline research. For each picture or text description, data annotation and recording are carried out respectively from three dimensions: color matching, such as the number of colors, color ratio, and color contrast; material texture, such as softness, glossiness, roughness; and decorative details, such as the type of decorative elements, decorative position, and decorative density. This enables subsequent feature extraction and analysis to be based on more comprehensive and accurate data. For the color matching dimension, color histogram and color moment methods are used to extract color features and convert the color information into vector representation. For the material texture dimension, texture analysis algorithms can be used to extract texture features. For the decorative details dimension, target detection and image segmentation techniques are used to identify decorative elements and extract features such as their quantity and position. The features extracted from different dimensions are combined to form a high-dimensional feature vector, which is convenient for computer processing and analysis and lays the foundation for style quantization encoding. The high-dimensional feature vector may have problems of information redundancy and dimensionality disaster, increasing the computational complexity and analysis difficulty. First, the high-dimensional feature vector is standardized to obtain a standardized feature vector, making the mean of each feature 0 and the variance 1. Calculate the covariance matrix of the feature vector, perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues and corresponding eigenvectors, and select the eigenvectors corresponding to the first largest eigenvalues to form a projection matrix. Project the standardized feature vector onto the matrix to obtain a feature vector with reduced dimensions, making the data more compact and easier to process, and improving the efficiency and performance of style quantization encoding. Although the dimension of the feature vector with reduced dimensions is reduced, it still contains the feature information of different fashion styles. Through the clustering analysis algorithm, first initialize the cluster centers. For each sample in the feature vector with reduced dimensions, calculate its distance from each cluster center, and assign each sample to the category where the nearest cluster center is located. Then update the center of each cluster until the cluster centers no longer change. For the clustering result, set a similarity threshold. If the average distance between two categories is less than the similarity threshold, then merge these two categories into one category, making the boundaries between different styles clearer and facilitating subsequent encoding operations. In order to represent the classified fashion styles in a quantized form, a unique code needs to be assigned to each style category. Assign a unique integer code to each style category obtained by clustering. For each sample, according to the style category it belongs to, combine the corresponding code as an element with the proportion of this style in the sample to construct a style feature vector, realizing the quantization encoding of fashion styles and providing a basis for subsequent applications such as fashion style matching and recommendation.

[0036] The customized process integration unit 4 transmits the generated quantified design indicators to the computer-aided design model through the design software docking module 41. The computer-aided design model selects materials from the shoe material library according to the functional quantification indicators, adjusts the shoe design parameters based on the style vector, and generates the final design scheme. When selecting shoe materials in the shoe material library, multiple functional quantification indicators need to be considered simultaneously, such as breathability, wear resistance, and comfort. Also, the cost and supply situation of the shoe materials will limit the selection. Let the functional quantification indicators be in number, and they are respectively . For example, is the breathability, is the wear resistance. The cost of the shoe material is , and the supply situation is represented by the supply quantity . Let there be kinds of shoe materials in the shoe material library. The selection variable for each kind of shoe material is , , being a binary variable. represents the selection of the th kind of shoe material, and represents non-selection. Then the optimization objective can be expressed as maximizing the linear combination of each functional quantification indicator, that is, , where is the weight of the nd functional indicator. , and the constraint conditions are the cost constraint and the supply constraint , , where is the maximum cost limit, and is the minimum supply quantity requirement for the th kind of shoe material. Randomly generate a certain number of shoe material combinations as the initial population , and the certain number is . Each individual is a -dimensional binary vector . For each individual in the population, calculate its values of each functional indicator , and sort the individuals according to the non-dominated relationship, dividing the population into different non-dominated levels. Select parent individuals from the current population through the tournament selection method, and perform crossover and mutation operations to generate the offspring population . The crossover operation can adopt single-point crossover or multi-point crossover, and the mutation operation can randomly change the value of a certain . Combine the parent population and the offspring population into a new population . Perform non-dominated sorting on , and select the first individuals as the next-generation population , continuously iterate until the maximum number of iterations is reached. The selected shoe material combinations can better meet the various functional requirements of users for shoes, while meeting the cost and supply constraints, improving the design quality and practicality of shoes. When adjusting the shoe design parameters according to the style vector, the design parameter space is usually relatively complex, with many possible combinations. Suppose the design parameters of the shoe style are in number, which are respectively , such as the toe shape, heel height, and shoelace style. The style vector is , where represents a certain dimensional feature of the style. Randomly generate number of design parameter combinations as the initial population , and each individual is a -dimensional vector . The fitness function is used to measure the matching degree between the design parameter combination and the style vector , and is defined as , where is the distance between the style feature vector corresponding to the design parameter combination and the style vector . Use the roulette wheel selection method to select parental individuals from the current population, perform crossover and mutation operations to generate the offspring population , calculate the fitness values of the offspring population, and select individuals with higher fitness values to form the next generation population . Repeat the above steps until the maximum number of iterations is satisfied. The genetic algorithm has strong global search ability, can find better solutions in the complex design parameter space, and can continuously adjust the design parameters according to the requirements of the style vector to generate a shoe design scheme that meets the user's style needs. The designed shoe style can better meet the user's personalized needs for style, improving the fashionability and user satisfaction of the shoe style.

[0037] In the present invention, the data collection and preprocessing unit 1 collects and standardizes the expected text data of users for shoe materials. The natural language processing core unit 2 identifies the user's intentions, keywords, and modifiers, analyzes the priority of requirements, and the quantization index generation unit 3 converts the functional requirements and style preferences into quantization indexes and feature vectors. The customized process integration unit 4 inputs the quantization indexes into the computer-aided design model, screens shoe materials, adjusts design parameters, and generates design schemes, using quantization means to clarify the design direction, reducing the error between the user's expectations and the design results. At the same time, considering the cost and supply situation of shoe materials, it meets the diverse needs of users, improving the customization quality of shoe products and user satisfaction.

[0038] The foregoing has shown and described the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and the above embodiments and the descriptions in the specification are only preferred examples of the present invention, and are not used to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of the present invention claimed is defined by the appended claims and their equivalents.

Claims

1. A shoe product management system for shoe material customization using big data processing, characterized in that: It includes a data collection and preprocessing unit (1), a natural language processing core unit (2), a quantitative indicator generation unit (3), and a customized process fusion unit (4); The data collection and preprocessing unit (1) collects the user's expected text data and historical expected text data on shoe materials through a data collection module (11), and performs standardization processing on the collected data through a standardization module (12); The natural language processing core unit (2) includes a built-in intent recognition and classification module (21), which uses a pre-trained model based on deep learning to perform intent recognition on the pre-processed text, judge the user's intent, and use a sequence annotation model to identify keywords and modifiers in the text. It also includes a priority analysis module (22), which uses a rule engine combined with a machine learning algorithm to build a priority model, extracts the priority of the text according to the priority ranking determined by the user, and calculates the priority of the text that is not clearly ranked based on word frequency statistics and semantic association analysis; The quantitative index generation unit (3) establishes a mapping relationship between functional requirements and quantitative indicators and sets weight coefficients through a functional quantitative mapping module (31), and obtains a final breathability quantitative value through weighted calculation, and quantitatively encodes various fashion styles through a style quantitative encoding module (32), constructs a style feature vector, and quantifies and forms a multi-dimensional vector from the dimensions of color matching, material texture, and decorative details; The customization process fusion unit (4) transmits the generated quantitative design index to the computer-aided design model through the design software docking module (41); the computer-aided design model selects materials from the shoe material library according to the functional quantitative index, adjusts the shoe design parameters according to the style vector, and generates a final design solution.

2. A shoe product management system for shoe material customization based on big data processing according to claim 1, characterized in that: The standardization module (12) adopts a standardization algorithm based on word vector space, which is as follows: The collected text data is segmented to obtain a series of words, and each word is then mapped to the pre-trained word vector space to obtain the corresponding word vector. For synonyms and near-synonyms, they are classified as the same standard word by calculating the cosine similarity between the word vectors. For misspelled words, the edit distance algorithm is used in combination with the semantic information of the word vector to correct the errors and convert them into standard words.

3. A shoe product management system for shoe material customization based on big data processing according to claim 1, characterized in that: The intention recognition and classification module (21) adopts an adaptive adjustment strategy of the multi-head self-attention mechanism when using the deep learning-based pre-training model. According to the different text lengths and complexities, the number of heads and the distribution of attention weights in the multi-head self-attention mechanism are dynamically adjusted. At the same time, a multi-classifier is added to the output layer of the model, and the Softmax function is used to convert the output of the model into a probability distribution of different intent categories, and the user's intention is determined according to the category with the largest probability value.

4. A shoe product management system for shoe material customization based on big data processing according to claim 1, characterized in that: The sequence labeling model adopts a model combining a bidirectional long short-term memory network and a conditional random field based on a deep learning architecture, as follows: The deep learning architecture is used to extract features from the input text to capture the long-distance dependencies in the text. The extracted features are then input into the bidirectional long short-term memory network to learn the contextual information of the text. Finally, the output of the bidirectional long short-term memory network is sequence labeled through the conditional random field layer to identify the keywords and modifiers in the text. During the training process, an adversarial training method is used to introduce a discriminator to distinguish between real annotations and annotations generated by the model.

5. A shoe product management system for shoe material customization based on big data processing according to claim 4, characterized in that: The priority model constructed by the priority parsing module (22) adopts a method based on a graph neural network, which is as follows: The words in the text are used as nodes of the graph, and the semantic associations between words are used as edges to construct a semantic graph. The graph neural network is used to learn the semantic graph and update the feature representation of the nodes. For the priority-ranked text determined by the user, the initial priority is assigned to the node according to the ranking information in the text. For the text that is not clearly ranked, the priority of the node is calculated by calculating the degree centrality of the node in the graph combined with the word frequency statistics.

6. A shoe product management system for shoe material customization based on big data processing according to claim 5, characterized in that: The function quantification mapping module (31) establishes a mapping relationship between function requirements and quantitative indicators by using a method combining case-based reasoning and fuzzy logic, as follows: Establish a historical case library to store historical functional requirements and corresponding quantitative indicators. When there are new functional requirements, retrieve the most similar cases from the case library through similarity calculation, and then use fuzzy logic to adjust the retrieved cases and set weight coefficients. The final air permeability quantitative value is obtained through weighted calculation.

7. A shoe product management system for shoe material customization based on big data processing according to claim 6, characterized in that: The style quantization encoding module (32) uses a method combining principal component analysis and cluster analysis when constructing a style feature vector, as follows: Fashion style sample data is collected from the dimensions of color matching, material texture, and decorative details, and features are extracted from the sample data to obtain high-dimensional feature vectors. Principal component analysis is then used to reduce the dimensionality of the high-dimensional feature vectors. Finally, a clustering analysis algorithm is used to cluster the reduced-dimensional feature vectors. Styles that exceed the similarity threshold are classified into one category, and a unique code is assigned to each style category to construct a style feature vector.

8. A shoe product management system for shoe material customization based on big data processing according to claim 7, characterized in that: The computer-aided design model uses a multi-objective optimization algorithm when screening shoe material library materials according to functional quantitative indicators, as follows: Functional quantitative indicators are used as multiple optimization objectives, and the cost and supply of shoe materials are used as constraints. The optimal shoe material combination that meets multiple objectives is searched in the shoe material library through a multi-objective optimization algorithm. When adjusting the shoe design parameters according to the style vector, a genetic algorithm is used to optimize the design parameters, and the design parameters are continuously iterated and updated according to the requirements of the style vector to generate a shoe design solution that meets the user's style needs.

Images