An industrial design large model-oriented training set construction method

Abstract

The application discloses a training set construction method for an industrial design large model, which takes multi-source text data in the field of industrial design as input, and constructs a high-quality data set suitable for training of the industrial design large model through data acquisition and cleaning, text structuring and semantic vectorization modeling, automatic generation of question and answer samples, and data expansion processing based on metamorphic relations. The method can reduce the cost of manual annotation, improve the structural degree, semantic consistency and expansion capability of the training data, and thus improve the training effect and application performance of the industrial design large model.

Description

Technical Field

[0001] This invention relates to the fields of data processing and industrial design technology, specifically a method for constructing training sets for large-scale industrial design models. Background Technology

[0002] With the rapid development of artificial intelligence technology, deep learning-based natural language processing models have been widely applied in fields such as knowledge-based question answering, text understanding, and intelligent decision-making assistance. Larger models, with more parameters, possess stronger expressive and learning capabilities after thorough training, and even exhibit emergent abilities to comprehensively analyze and solve deeper problems, demonstrating human-like thinking and intelligence. Existing large-scale model training typically relies on large-scale general corpora, acquiring strong language modeling capabilities through pre-training and fine-tuning. However, in specialized fields such as industrial design, design knowledge is highly specialized, with a high proportion of implicit experience, complex semantic dependencies, and significant cross-stage connections in design logic. General corpora and general training methods are insufficient to effectively support the model's accurate understanding of industrial design semantics and constraints. Due to a lack of domain-specific knowledge, large models often exhibit problems such as illusions and digressions when facing such challenges.

[0003] Currently, knowledge in the field of industrial design is mainly scattered across professional textbooks, design specifications, process manuals, and internal corporate design documents and project case studies. This data is diverse in form and lacks uniformity in structure, with a large amount of design experience existing in unstructured text form, lacking a standardized data representation method suitable for training large-scale models. Existing technologies for constructing training sets for specific professional domains often employ simple data cleaning and manual annotation methods, relying on human experience to generate question-answer pairs or labeled data. This approach is not only costly and inefficient but also fails to cover the diverse expressions and potential constraints in design semantics, making it difficult to meet the needs of large-scale industrial design models for high-quality, large-scale training data.

[0004] Furthermore, existing training data construction methods typically focus on organizing and expanding static text samples, lacking systematic modeling of the inherent semantic relationships within designs. This is particularly true in question-and-answer data construction, where it's difficult to guarantee the effectiveness of generated samples in terms of semantic consistency, logical transferability, and robustness. Consequently, the model's generalization ability is insufficient when faced with design variations, changes in design constraints, or adjustments to design conditions. Especially in industrial design scenarios, there are complex evolutionary relationships among design requirements, functional descriptions, material processes, and structural constraints. Traditional methods based on simple data augmentation struggle to reflect the inherent evolutionary patterns of design knowledge.

[0005] Therefore, there is an urgent need for a training set construction method for large-scale industrial design models. This method should start from multi-source data in the industrial design field, use systematic data collection and cleaning techniques to structure and semantically represent design texts, and build high-quality question-and-answer samples based on this. At the same time, it should combine the inherent relationships between design semantics to effectively expand the training data, thereby forming a specialized dataset suitable for training large-scale industrial design models, so as to improve the model's understanding ability, reasoning ability and application reliability in industrial design scenarios. Summary of the Invention

[0006] Therefore, to address the aforementioned shortcomings, this invention provides a method for constructing training sets for large-scale industrial design models. This method solves problems in existing technologies such as difficulty in acquiring training data in the industrial design field, low degree of structure, low efficiency in constructing question-answer samples, and insufficient robustness in data expansion. By systematically processing multi-source text data in the industrial design field, unstructured design knowledge is transformed into a high-quality dataset suitable for training large-scale models, thereby improving the generalization ability and reliability of large-scale industrial design models in professional semantic understanding, design reasoning, and application scenarios.

[0007] This invention is implemented by constructing a training set building method for large-scale industrial design models, including: S100, Data Acquisition and Preprocessing Steps: We will build a private corpus for the industrial design field, systematically collect relevant text resources in the industrial design field, and annotate the documents during the collection stage; we will unify the format and extract the text content of original documents of different formats; we will supplement the text content with data annotation; and we will complete the storage and indexing by data cleaning, terminology standardization and quality classification according to the layered approach of original library and standardized library. S200, Text Cleaning and Normalization Processing Steps: Based on the characteristics of expression, non-design information in the text is filtered out, and design terms, units of measurement, and process descriptions are standardized. The text is semantically segmented, breaking down long texts into the smallest knowledge units. By retaining chapter paths, design stages, and contextual index structure information, the association between adjacent semantic units is established. S300, Data segmentation and standardization steps for training texts: The input text is processed into a standardized ternary embedding structure, including token embedding, segment embedding, and position embedding; the above three types of embedding vectors are linearly summed to generate a composite embedding representation of industrial design text. S400, Text structuring and vectorization embedding steps based on pre-trained models: Semantic extraction is performed using the ALBERT model. Word embedding parameter factorization is used to reduce the number of parameters. Pre-training is optimized using sentence order prediction task. Text vector representation is generated through a multi-layer Transformer encoder. S500, Question and Answer Sample Generation Steps: A two-stream structure of TextCNN and BiLSTM-Attention is used to extract features. The left channel captures local semantics through multi-scale convolution, and the right channel captures global dependencies through a bidirectional long short-term memory network. The concatenation is then passed to a classifier to generate question-answer pairs. S600, Sample expansion steps: Based on the metamorphic relationship, the question-answer pair is systematically expanded. The metamorphic relationship describes the logical consistency that the output answer should maintain when the design input conditions change. By performing property analysis on the functional requirements, ergonomics, material process, structural assembly, regulations and standards and scenario constraints of the design knowledge unit, derivative samples are generated by template replacement, parameter mapping or scenario transfer. After consistency verification, the expanded training set is obtained.

[0008] The invention has the following advantages: This invention is a method for constructing training sets for large-scale industrial design models. By systematically collecting, cleaning, and structuring multi-source text data in the industrial design field, it enables the previously scattered and unstructured design knowledge to be organized and represented in a unified and standardized form. This significantly improves the quality of training data in terms of semantic consistency, logical integrity, and traceability, providing a more stable and reliable data foundation for training large-scale industrial design models. Furthermore, compared to methods that rely heavily on manual experience to generate training samples, this invention effectively reduces human intervention through automated text structuring, semantic vectorization, and question-answering sample generation mechanisms. While ensuring the quality of training data, it significantly improves the efficiency of dataset construction, making it suitable for the continuous expansion and iterative updates of large-scale industrial design corpora. This invention effectively overcomes the problems of insufficient structuring, low efficiency, and limited data generalization ability in existing technologies for constructing training sets in the industrial design field, providing strong support for high-quality training and engineering applications of large-scale industrial design models. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating a method for constructing a training set for a large-scale industrial design model. Figure 2 This is a flowchart of the question-answer pair generation method based on metamorphic relationships proposed in this invention; Figure 3 This invention presents a transformation relationship diagram for industrial scenarios. Detailed Implementation

[0010] The following will be combined with the appendix Figures 1-3This invention will be described in detail, and the technical solutions in the embodiments of this invention will be clearly and completely described. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0011] This invention provides an improved method for constructing training sets for large-scale industrial design models. This method takes multi-source text data from the industrial design field as input and constructs a high-quality dataset suitable for training large-scale industrial design models through data collection and cleaning, text structuring and semantic vectorization modeling, automatic question-answer sample generation, and data augmentation based on metamorphic relationships. This method can improve the structuring, semantic consistency, and scalability of training data while reducing the cost of manual annotation, thereby enhancing the training effect and application performance of large-scale industrial design models.

[0012] In this embodiment, such as Figures 1-3 As shown, a method for constructing a training set for large-scale industrial design models includes: Step S100: Data Acquisition and Preprocessing. A high-quality private corpus for the industrial design field is constructed to provide a stable data foundation for subsequent domain-adaptive pre-training and task fine-tuning of the model.

[0013] First, a systematic inventory and unified collection of relevant text resources in the field of industrial design was conducted, covering various data sources such as professional textbooks, design specifications, process manuals, and internal design cases of enterprises. During the collection phase, the source, version, and permission information of the documents were marked to ensure the traceability and security of the corpus data.

[0014] Subsequently, the original documents of different formats were standardized and the text content was extracted, converting unstructured data into standardized text representations while retaining key information such as chapter hierarchy, heading structure, and table descriptions.

[0015] Based on this, a structured reconstruction strategy for the industrial design field is introduced to supplement the text content with metadata annotations such as design stage, content type and constraint attributes, so that the corpus can have a clear design logic structure while maintaining the original design semantics.

[0016] Finally, the corpus was screened and optimized through governance methods such as data cleaning, terminology standardization and quality grading, and the data was stored and indexed in a hierarchical manner according to the original library and the standardized library, so as to provide a high-quality data foundation for subsequent text segmentation, embedding modeling and domain pre-training based on design semantics.

[0017] Step S200: Text cleaning and normalization process. Based on the completion of the private corpus of industrial design, and considering the characteristics of complex semantics, large design logic span, and frequent cross-chapter associations in industrial design texts, this embodiment further cleans and semantically segments the normalized text.

[0018] First, based on the expressive characteristics of the industrial design field, non-design information in the text is filtered out, and design terms, units of measurement, and process descriptions are standardized to reduce the interference of semantic noise on model learning.

[0019] Secondly, taking "design intent - functional description - structural and process constraints" as the core, the text is semantically driven to break it down into the smallest knowledge units that maintain the semantic integrity of the design, thereby taking into account both the model input length limit and the continuity of design logic.

[0020] Finally, by preserving structural information such as chapter paths, design stages, and context indexes, the association between adjacent semantic units is established, providing a continuous and reconfigurable semantic input sequence for subsequent embedding modeling and domain pre-training of industrial design texts.

[0021] To enhance the ability of existing data cleaning models to capture long-distance information in long texts, this embodiment proposes a contextual enhancement method based on a dual-block attention mechanism. This method uses an attention mechanism and adjusts the attention weights using a partial rotation embedding parameter adjustment method while adding a long short-term memory layer. This effectively captures the relative positional and content information of long texts, thereby achieving targeted and high-precision data redundancy removal and correction tasks.

[0022] Intra-block attention is used to compute the inner product of queries and keys within the same block. For a length of... l Long sequences, divide the sequence into Each block ensures that the position index within each block does not exceed the block size. s This embodiment uses a key index and a query intra-block position index. ,in This represents the position index of the query during intra-block attention. During intra-block attention, the position indexes of the query and key are adjusted according to the following formula: ; absolute index within the same block i and j That is, satisfying ( i / s = j / s And (0≤) j ≤ i < l ), element (M[ i ][ j The difference between the positional encoding of the query and the key is defined as follows: ; when( i / s = j / s When calculating (M), follow the formula above. i ][ j Then, the first i The first query and the first j The intra-block attention score between each key is calculated as follows: ; To aggregate information from other blocks, inter-block attention is introduced. Since the positional index of the query is typically greater than the positional index of the key, this reflects the left-to-right flow of information, i.e., when ( i ≥ j When ), satisfying ( P q [ i ]≥ P k [ j This embodiment preserves the position index of the key. P k Considering KV caching, and searching for a new set of query location indexes in inter-block attention, denoted as... Because the key's position index repeats cyclically, the maximum position index is ( max ( P k )= s -1). To ensure that the position index of the query is greater than the key of all previous blocks, a simple strategy is to allocate a fairly large position index for long-distance queries, such as the maximum position index during pre-training ( c -1> max ( P k )),in c This refers to the length of the pre-trained context. , when( i / s ≠ j / s In this embodiment, it is possible to provide data from different blocks. q i and k j relative position matrix M as follows: ; However, using only inter-block attention may not maintain the precise relative positions between adjacent tags, leading to performance degradation. This embodiment introduces a contiguous block attention mechanism to adjust... The front of the middle w Use position indexes to maintain wLocality of adjacent tags. Formally, given a block size s Pre-training size c and local windows w ,have: ; in w This represents the local window size, which can be directly set as the difference between the pre-training length and the block size. For data from consecutive blocks... i,j ),use and P k Calculated M [ i ][ j This formula ensures that adjacent w The key is closest to the current query.

[0023] This method effectively enhances the model's ability to understand and analyze long texts, significantly improves data cleaning accuracy, and provides a high-quality data foundation for the construction of subsequent datasets.

[0024] Step S300: Data Segmentation and Standardization of Training Text. The input layer segments and standardizes the text data from professional books, providing a foundation for multi-batch training of the model. Industrial design texts often contain heterogeneous information such as design concept descriptions, functional requirements, material and process constraints, and human-computer interaction and aesthetic semantics. Therefore, the input layer needs to explicitly model the structure and positional relationships of the text while maintaining semantic continuity. To this end, this paper processes the input text into a standardized ternary embedding structure: token embedding, segment embedding, and position embedding. First, the industrial design-related text is tokenized, introducing special tokens [CLS] and [SEP] at the beginning and end of phrases, respectively. The [CLS] token is used to aggregate the semantic information of the entire design and can be used for downstream tasks such as design classification, design style identification, or functional attribute prediction. The [SEP] token is used to distinguish different semantic units, separating the design requirement text (TextA) from the design description or solution text (TextB) in tasks such as design question answering, design scheme comparison, or requirement-solution matching. Segmented embeddings are used to distinguish design information from different semantic sources in the input text, and only include two value forms: each tag in the first segment of text is assigned a value of 0, and each tag in the second segment of text is assigned a value of 1. When the input contains only a single design description (such as a single product structure description or process paragraph), its segmented embedding is uniformly set to 0, thus maintaining the consistency of design semantics. Finally, positional embeddings are introduced to describe the relative position of each tag in the text sequence, especially used to characterize the sequential relationship of design logic in industrial design texts, such as the progression from user requirements to functional definitions, and from structural schemes to manufacturing processes. The above three types of embedding vectors are linearly summed to generate a composite embedding representation of industrial design text, providing a basic representation for the subsequent model to learn the relationships, constraints, and implicit design rules between design knowledge.

[0025] Step S400: Text structuring and vectorization embedding based on the pre-trained model. This step transforms the complex embedding representation of industrial design text into a vector representation, deeply mining key feature information and optimizing its vector representation to improve the quality of the training input data and achieve faster model training convergence. This step uses the ALBERT model for semantic extraction of the text, fusing the semantic information of the words themselves and the context at the sentence level to create dynamic word vector representations. First, the text format category information [CLS] is serialized... E [CLS] Similarly, the ending text [SEP] is processed into vector form. E [SEP]Because excessively long input text increases the complexity of model training and thus raises training costs, the input text is segmented into multiple token fragments, represented as follows: ; The text fragment is vectorized into... ,in E i Indicates the first in the text i The feature vector of a text serialization vector of 1 character.

[0026] In the feature extraction process, the optimized Eneoder structure in Transformer is used to remove redundant computations. The word embedding parameters are factored into two small matrices. Instead of directly mapping the one-hot vectors to a hidden space of size H, they are first mapped to a low-dimensional word embedding space and then back to the hidden space. Through this decomposition, the word embedding parameters are reduced from O(n0) to O(n0). V × H ) decreased to o( V × E + V × H When the hidden layer size H is much larger than E, the number of parameters is significantly reduced. Furthermore, since the large-scale question answering task is a binary classification task predicting whether two sentences appear consecutively in the original text, positive examples are two consecutive sentences in the article, while negative examples are constructed by selecting one sentence from each of two documents. Because the construction of negative examples makes the model more inclined to predict sentence topic relationships rather than sentence continuity, parameters used across multiple layers of the model are shared during training. The positive example selection method remains unchanged, the same as in the next sentence inference task. Negative examples are constructed by selecting two consecutive sentences from one document and swapping their order. This allows for the expression of rich semantic information with fewer model parameters, effectively reducing the model size.

[0027] Subsequently, 10 layers of the aforementioned Transformer bidirectional encoder are stacked to fully mine and combine deep semantic features and shallow texture features. T rm This represents the output of a multi-layer bidirectional Transformer module. Each network layer consists of two sub-layers: a multi-head self-attention mechanism layer and a feedforward network layer, connected by a residual network module. The multi-head self-attention mechanism layer calculates the relationships between words, while the feedforward network layer fuses the positional information of the words. The input and output of this sub-layer are then summed and normalized. The input layer of the multi-head self-attention mechanism network forms input vectors Q, K, and V based on the query, key, and value of each word in the text sequence. Utilizing the multi-head attention mechanism... ; ; The output matrices of multiple network layers are concatenated into a large word vector matrix. ; Here, W is an additional weight matrix that compresses the dimensions of the concatenated matrix to a uniform size equal to the sequence length. , , This represents the weight matrix for each Q, K, V vector. d t This represents the dimension of each Q, K, and V vector. The mechanism described above calculates the relationship between each word and all other words in the sentence at the sentence level, and adjusts the weight of each word within the sentence based on these relationships to obtain a new vector representation, thus achieving an efficient word vector representation of the text sequence.

[0028] Step S500: Question and answer sample generation step.

[0029] The left channel employs the TextCNN text classification model, using sliding windows of different sizes to perform convolutional and pooling operations on the input text vector. This captures and combines local features of the text sequence, extracting semantic information at different levels of abstraction to obtain a high-level feature vector representation of the text. It mainly consists of four parts: an input layer, a convolutional layer, a pooling layer, and an output layer. For an input length of... n Chinese text, convolutional layers employ h Sliding windows of different sizes perform convolution operations on the text input vector to learn text features, at position... i The convolutional feature values ​​are obtained by using a convolution kernel.

[0030] ; in, k This represents the dimension of the word vector corresponding to each word in the text sequence. w Represents the convolution kernel, with a dimension of . hxk , T i:i+h-1 Represents the first element of the input matrix. i Arrive at the i+h A sliding window consisting of -1 rows. b Indicates the bias parameter. f This represents a non-linear mapping function. The pooling layer uses a 1-MaxPool maximum pooling strategy to select the largest feature value from each sliding window.

[0031] ; The right-channel BiLSTM layer selectively memorizes the input based on the LSTM gate structure, remembering important information and forgetting less important information. It determines which new information to retain in the current state. Previous state output. h t-1 and current input information x t The input, used as the sigmoid number, generates a value between 0 and 1 to determine how much new information needs to be retained. Through the forget gate and input gate, the complete state for the next time step is obtained, which is used to generate the hidden layer output for the next time step. h t The output gate determines what information is output from the cell state. Similar to the input gate, it is generated by a sigmoid function. o t This determines how much cell status information needs to be output. When the cell status information is multiplied by... o t First, it is activated through the tanh layer, and then the output information of this LSTM structure is obtained. h t When using LSTM to obtain semantic information, it is necessary to consider not only the information content before the word but also the information content after the word. BiLSTM, by obtaining and fusing information from both directions of the sentence, can better meet this requirement. Therefore, it can better capture bidirectional semantic dependencies. Assume that at time... t The hidden layer state is output from the forward LSTM. The hidden layer state is then output to the LSTM. Then output .

[0032] To enhance feature representation capabilities, the feature vector output from the LSTM module is input into the multi-head attention module. Multi-head self-attention is essentially composed of multiple self-attention modules, and the self-attention mechanism computes three new vectors. Q , K , V Each is obtained by multiplying the embedding vector by a randomly initialized matrix. Then, 0 is multiplied by... K The transpose of represents the encoding of a word, indicating the level of attention given to other parts of the input. This level of attention is then divided by a constant and a softmax operation is performed to represent the relevance of other parts to the word. Finally, is used... V Multiplying the value obtained by sofmax, the result is as follows: ; Multi-head self-attention consists of multiple self-attention sets, initialized with multiple groups. Q , K , VThen, reduce these matrices to a single matrix, and multiply it by a randomly initialized matrix. The calculation formula is as follows: ; In multi-head self-attention mechanisms, the positional encoding dimension and embedding dimension are the same. The positional encoding is added to the embedding value and passed to the next layer instead of the original embedding value. The position vector represents the position of the current word. There are three commonly used calculation methods. In model training, this method adopts a hybrid training strategy, based on the transfer learning, fine-tuning, and other aspects of model training. The first method uses sinusoidal positional encoding. The positional encoding must have the same dimension as the word vector. A sine function is used when the position is even, and a cosine function is used when the position is odd. ; The second method is relative position expression. When the relative position exceeds the absolute value of a certain threshold, the threshold is used as the substitute.

[0033] The third method employs positional encoding. An independent vector is learned for each position, and the method is roughly the same as generating word vectors. After comparison, this paper adopts the positional encoding method. A feedforward neural network provides non-linear transformation. The dimension of the attention mechanism's output is determined by the product of the input batch size and the sentence length, and the product of the number of convolutional kernel layers in the discriminator and the total number of convolutional kernels.

[0034] The concatenate feature merging layer directly concatenates the feature vectors output from the TextCNN layer and the BiLSTM+multi-head attention layer to form a comprehensive feature input to the fully connected layer. Finally, the classification result is obtained through a Softmax classifier.

[0035] Step S600: Sample expansion step.

[0036] In this embodiment, to enhance the coverage of training data in scenarios involving changes in industrial design semantics, alterations in design conditions, and evolution of solutions, the initial question-answer pair set obtained in step S500 is systematically expanded using metamorphic relations (MR) oriented towards industrial design semantics. These metamorphic relations describe the logical consistency or predictable change pattern that the output answer should maintain when design input conditions undergo controllable changes. Their design and selection follow the following basic principles and are implemented in conjunction with constraints imposed by the industrial design scenario: (1) Correctness criteria: The metamorphic relationship should cover the semantic properties or design specifications of the design to ensure that the semantics of the derived samples are consistent and conform to the design logic.

[0037] When constructing the metamorphic relationship, the first step is to analyze the properties of the design knowledge units corresponding to the question-and-answer pairs, clarifying their source of constraints and applicable boundaries. These properties include, but are not limited to: functional requirement constraints (e.g., "must be met / must not occur"), ergonomic constraints (e.g., size range, accessibility, comfort), material and process constraints (e.g., surface treatment, wear and corrosion resistance, manufacturability), structural assembly constraints (e.g., clearances, tolerances, assembly sequence), regulatory and standard constraints (e.g., safety regulations, environmental protection requirements), and scenario constraints (e.g., outdoor / medical / children / high temperature and humidity). Based on this, the questions or conditions are transformed according to preset metamorphic rules, and the changes in the answers are simultaneously constrained to ensure that the generated derivative question-and-answer pairs satisfy: Semantic consistency and logical derivability; It does not violate design specifications or common sense boundaries (e.g., materials and processes are not feasible, regulations do not allow it, etc.). The transformation-response relationship is maintained with the original question-answer pair, thereby ensuring the correctness and usability of the derived samples.

[0038] (2) The principle of easy generation of derivative test cases: When constructing metamorphic relationships, the complexity of sample generation algorithm should be considered comprehensively, and metamorphic relationships with many changing elements or complex expressions but which can be implemented efficiently should be given priority.

[0039] To improve the efficiency of large-scale expansion, this embodiment controls the time complexity of the derived sample generation algorithm during the metamorphic relation construction stage. It prioritizes metamorphic relations that can generate derived question-answer pairs through template replacement, parameter mapping, synonym rewriting, and constraint dictionary-driven methods to avoid high-cost global reasoning or extensive manual intervention. Simultaneously, while ensuring correctness, it prioritizes metamorphic relations with "many changing elements" or "complex expressions" to enhance the semantic coverage of training samples. For example: Changes in requirements: Combining and replacing requirements such as "portable / drop-resistant / waterproof / easy to clean"; Parameters and thresholds: Perform range transformations or threshold adjustments on quantifiable parameters such as size, weight, strength, lifespan, and temperature range; Constraint replacement category: Feasible alternatives to constraints such as materials, processes, surface treatments, and environmental protection levels (e.g., material grade replacement, process route replacement); Scene transfer type: Transferring scene conditions such as "indoor → outdoor", "adult → child", "normal environment → high temperature and high humidity" and driving the corresponding adjustment of the answer.

[0040] Using the above methods, derivative samples with significant semantic differences and broad coverage can be obtained with low computational overhead.

[0041] (3) Prioritize metamorphic relationships that can generate unique derivative test cases: reduce ambiguity and multiple solutions to a problem, and improve data consistency.

[0042] To reduce the risk of ambiguity in generated question-answer pairs and improve the consistency of training labels, metamorphic relations that produce unique derived results are prioritized. That is, for the same original question-answer pair, given transformation rules and parameters, only one explicit derived question and answer correspond to it. For transformations that may lead to multiple solutions (e.g., simultaneously changing multiple coupled constraints, or having multiple feasible material / process route options), uniqueness can be achieved by limiting the transformation range, introducing priority rules, or fixing decision-making strategies. For example: Prioritize alternative material supply chains (e.g., "prioritize manufacturability, then cost, and finally appearance"). Single-variable control is used for parameter transformation (only one constraint or one parameter range is changed at a time). Limit scene migration to a single scene dimension change (e.g., only change the user group or only change the environmental conditions).

[0043] The above uniqueness strategy ensures that the derived question-answer pairs have stable supervision signals in semantic annotation and model training.

[0044] (4) Select metamorphic relationships with a wide range of applicability: The set of metamorphic relationships covers the largest possible industrial design semantic space and task scenarios.

[0045] To enhance the generalization ability of the training set, when constructing the metamorphic relationship set, priority is given to metamorphic relationships with broad applicability and cross-product category transferability, so that the metamorphic relationship set covers the widest possible range of industrial design semantics, including but not limited to key dimensions such as "requirement—function—structure—process—cost—regulation—scenario". Specifically, this embodiment covers, but is not limited to, the following typical industrial design scenarios: Scenarios of changing needs: changes in user demands, changes in target audience, and changes in usage habits; Scenarios involving changing constraints: material / process limitations, changes in cost ceilings, and changes in the supply chain; Environmental changes: changes in working conditions such as temperature and humidity, corrosion, impact, and outdoor exposure; Standards and compliance scenarios: Upgraded safety standards and changes in environmental protection requirements; Solution evolution scenarios: design iterations such as structural simplification, reduction of assembly steps, and improvement of maintainability.

[0046] By expanding the coverage of the metamorphic relation set, the training data can still have strong transferability when faced with new product categories or new design conditions.

[0047] After the transformation and expansion are completed, the generated derivative question-answer pairs are subjected to consistency verification and quality screening. The verification includes semantic conflict detection, constraint feasibility check and duplicate sample removal, and finally the expanded training sample set is obtained.

[0048] In another embodiment, the dual-block attention mechanism in step S200 can be replaced by modules such as local / global attention mechanism, inter-block attention mechanism, and self-attention mechanism. Its main purpose is to extract contextual information of long text targets to enhance the model's ability to understand text, thereby better completing tasks such as data cleaning, data deduplication, and data supplementation.

[0049] In another embodiment, the word embedding model ALBERT in step S400 can be replaced by other language preprocessing models such as BERT and DERT, wherein the position encoding strategy and embedding method can be replaced by dynamic position encoding, mask encoding, dynamic mask embedding, etc.

[0050] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing a training set for large-scale industrial design models, characterized in that, include: S100, Data Acquisition and Preprocessing Steps: We will build a private corpus for the industrial design field, systematically collect relevant text resources in the industrial design field, and annotate the documents during the collection stage; we will unify the format and extract the text content of original documents of different formats; we will supplement the text content with data annotation; and we will complete the storage and indexing by data cleaning, terminology standardization and quality classification according to the layered approach of original library and standardized library. S200, Text Cleaning and Normalization Processing Steps: Based on the characteristics of expression, non-design information in the text is filtered out, and design terms, units of measurement and process descriptions are standardized; the text is semantically driven to break it down into the smallest knowledge units. By preserving chapter paths, design phases, and context index structure information, the association between adjacent semantic units is established; S300, Data segmentation and standardization steps for training texts: The input text is processed into a standardized ternary embedding structure, including token embedding, segment embedding, and position embedding; the above three types of embedding vectors are linearly summed to generate a composite embedding representation of industrial design text. S400, Text structuring and vectorization embedding steps based on pre-trained models: Semantic extraction is performed using the ALBERT model. Word embedding parameter factorization is used to reduce the number of parameters. Pre-training is optimized using sentence order prediction task. Text vector representation is generated through a multi-layer Transformer encoder. S500, Question and Answer Sample Generation Steps: A two-stream structure of TextCNN and BiLSTM-Attention is used to extract features. The left channel captures local semantics through multi-scale convolution, and the right channel captures global dependencies through a bidirectional long short-term memory network. The concatenation is then passed to a classifier to generate question-answer pairs. S600, Sample expansion steps: Based on the metamorphic relationship, the question-answer pair is systematically expanded. The metamorphic relationship describes the logical consistency that the output answer should maintain when the design input conditions change. By performing property analysis on the functional requirements, ergonomics, material process, structural assembly, regulations and standards and scenario constraints of the design knowledge unit, derivative samples are generated by template replacement, parameter mapping or scenario transfer. After consistency verification, the expanded training set is obtained.

2. The method for constructing a training set for a large-scale industrial design model according to claim 1, characterized in that, Step S200 also includes a contextual capability enhancement method based on a two-block attention mechanism, which includes: An attention mechanism is used, and attention weights are adjusted by a partial rotation embedding parameter adjustment method. A long short-term memory layer is added to capture the relative position and content information of long texts. Intra-block attention is used to calculate the inner product of the query and key within the same block, dividing the sequence into n= l / s blocks, ensuring that the position index within each block does not exceed the block size s, according to the formula. Adjust the position of the query and key indexes; for absolute indexes within the same block. i and j Element M[ i ][ j The difference is defined as the positional encoding between the query and the key: Inter-block attention is used to aggregate information from other blocks and assign location indexes for long-distance queries. Where c is the pre-trained context length; a continuous block attention mechanism is introduced to adjust... The first w position indices are used to maintain the locality of w neighboring labels. Given the block size s, the pre-training size c, and the local window w, we have .

3. The method for constructing a training set for a large-scale industrial design model according to claim 2, characterized in that, In the context enhancement step based on the dual-block attention mechanism, the score of the intra-block attention is calculated as follows: ； in, f This represents the position encoding function. q Represents the query vector. k This represents the key vector.

4. The method for constructing a training set for a large-scale industrial design model according to claim 2, characterized in that, In the context enhancement step based on the dual-block attention mechanism, for ( i , j ),use and P k Calculated M[ i ][ j Ensure that the w adjacent keys are closest to the current query, thus maintaining the precise relative positions between adjacent tags.

5. The method for constructing a training set for a large-scale industrial design model according to claim 1, characterized in that, In step S400, the ALBERT model replaces the next sentence prediction task with a sentence order prediction task. Positive examples are two consecutive sentences in an article, and negative examples are constructed by selecting two consecutive sentences in a document and swapping their order.

6. The method for constructing a training set for a large-scale industrial design model according to claim 1, characterized in that, The method for the text structuring and vectorization embedding steps based on the pre-trained model, as described in S400, is as follows: The ALBERT model is used to extract semantics from the text, fusing semantic information from the words themselves and the context at the sentence level to create dynamic word vector representations; the text format category information [CLS] is serialized into... In vector form, the ending text [SEP] is processed as follows: ; The input text is segmented into multiple token fragments, represented as follows: The text fragments are vectorized into Using the optimized Encoder structure in Transformer, the word embedding parameters are factored, reducing the word embedding parameters from O(V×H) to O(V×E+E×H); a multi-layer Transformer bidirectional encoder is stacked to mine and combine deep semantic features and shallow texture features. Each network layer consists of a multi-head self-attention mechanism layer and a feedforward network layer, which are connected by a residual network module. A multi-head self-attention mechanism is used to calculate the relationships between words. A feedforward network layer integrates the positional information of words. The input and output of this network layer are added together and then normalized.

7. The method for constructing a training set for a large-scale industrial design model according to claim 6, characterized in that, The calculation of the multi-head self-attention mechanism includes: ; ； ； in, It is an additional weight matrix. , , This represents the weight matrix for each Q, K, V vector. d t This represents the dimension of each Q, K, and V vector.

8. The method for constructing a training set for a large-scale industrial design model according to claim 1, characterized in that, The specific method for generating question-and-answer samples in step S500 is as follows: The left channel employs the TextCNN text classification model, using sliding windows of different sizes to perform convolutional pooling operations on the input text vector to capture local features of the text sequence. The right channel uses a BiLSTM layer, selectively memorizing input information based on the LSTM gate structure. By acquiring and fusing information from both directions of the sentence, it captures bidirectional semantic dependencies. The feature vector output from the LSTM module is input into a multi-head attention module. The multi-head self-attention module consists of multiple self-attention points, calculating three new vectors Q, K, and V using the formula... Calculate attention; concatenate the feature vectors output by the TextCNN layer and the BiLSTM+ multi-head attention layer to form a comprehensive feature input to the fully connected layer, and obtain the classification result through the Softmax classifier.

9. The method for constructing a training set for a large-scale industrial design model according to claim 8, characterized in that, The convolution operation of TextCNN is located at... i The convolutional feature values ​​are obtained by using a convolution kernel: ； in, k This represents the dimension of the word vector corresponding to each word in the text sequence. w Represents the convolution kernel, with a dimension of . h × k , T i:i+h-1 Represents the first element of the input matrix. i Arrive at the i + h A sliding window consisting of -1 rows, b Indicates the bias parameter. f This represents a nonlinear mapping function; the pooling layer employs a 1-MaxPool maximum pooling strategy to select the largest eigenvalue from each sliding window. 。 10. The method for constructing a training set for a large-scale industrial design model according to claim 1, characterized in that, The BiLSTM at time t The output is ,in The hidden layer state is the output of the forward LSTM. This represents the hidden layer state output by the backward LSTM.

Images