Intelligent optimization method for forming process

Abstract

The invention provides a forming process intelligent optimization method, which comprises the following steps of: searching Chinese and English literatures through target forming process keywords, segmenting the literatures, and converting the segmented literatures into vectors which can be identified by a generative large language model through a text embedding model, so as to construct a forming process knowledge base; performing similarity search in a process knowledge base by using a locally deployed generative large language model and giving process parameters, performing stable production by using the process parameters, and meanwhile, testing the quality of a product and feeding back the quality to the generative large language model; and iteratively updating the process parameters by using the generative large language model until the quality of the obtained product reaches an optimization target. The method is high in system stability, high in response efficiency, few in iteration times and high in generalization and can adapt to different production conditions such as forming processes, materials and products, operation suggestions can be generated through a large model, and guidance can be provided for engineering practice.

Description

Technical Field

[0001] This invention belongs to the field of molding process optimization technology, specifically relating to an intelligent optimization method for molding processes based on a generative large model. Background Technology

[0002] Injection molding, die casting, 3D printing, and other manufacturing technologies are widely used in national strategic fields and pillar industries such as aerospace, transportation, energy, and consumer electronics. With the development of intelligent manufacturing, molding technology is moving towards higher precision, automation, and intelligence. Despite some progress, optimizing process parameters remains a major bottleneck affecting product quality stability and improving production efficiency. Because processes such as injection molding and die casting involve complex multi-field couplings (such as shear flow fields, temperature fields, and stress fields), the mechanisms by which these factors affect the quality of the final product are very complex.

[0003] Traditional parameter tuning relies heavily on engineers' experience and repeated trials. While simple and easy to implement, this method is inefficient and struggles to adapt to diverse production needs. To improve optimization efficiency, researchers have introduced experimental design methods (such as orthogonal experiments, Taguchi methods, and response surface methodology) to build empirical models and identify key parameters through a limited number of experiments. Subsequently, numerical simulation techniques (such as Moldflow and Moldex3D software) have been used to predict melt flow behavior and defect formation mechanisms, forming a closed-loop "simulation-optimization" system when combined with optimization algorithms. However, pure simulation is computationally expensive and sensitive to material properties and boundary conditions, limiting its real-time application. In recent years, data-driven intelligent optimization methods have gained attention, with machine learning models (such as artificial neural networks, support vector machines, and random forests) used to establish nonlinear mappings between process parameters and quality indicators. Furthermore, model-free optimization methods are gaining popularity due to their advantage of not requiring precise modeling. Nevertheless, existing methods still face challenges, such as the need for large amounts of labeled data for training and limited generalization ability. More importantly, current intelligent optimization systems lack the ability to understand and reason about process knowledge, failing to achieve knowledge transfer across materials and structures.

[0004] Recently, generative large models have demonstrated powerful knowledge integration and creative generation capabilities in fields such as natural language processing and code generation. Their general representation capabilities offer new possibilities for solving optimization problems in manufacturing processes. However, research on applying this technology to molding process optimization is still in its early stages. How to effectively encode multimodal manufacturing data and utilize large models to generate optimization suggestions that conform to physical laws and engineering constraints is an important scientific question worthy of in-depth research. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides an intelligent optimization method for molding processes. This optimization method includes three parts: a molding process knowledge base, a locally deployed generative large model, and iterative optimization of molding process parameters.

[0006] A method for intelligent optimization of molding process includes the following steps:

[0007] (1) Collect relevant Chinese and English literature from authoritative databases based on the target process keywords; for each literature, perform segmentation and obtain multiple text blocks respectively;

[0008] (2) Use the text embedding model to convert the segmented text blocks into embedding vectors and store them in the local vector database to build a searchable process knowledge base;

[0009] (3) Determine the process materials, product type, optimization objectives, adjustable process parameters and their ranges, and fixed process parameters, and input them into the generative large language model deployed on the local server;

[0010] (4) Generative large language models use the process knowledge base to perform similarity search and find the embedding vectors corresponding to related text blocks;

[0011] (5) Input the embedding vector found in step (4) into the generative large language model to enhance it, and generate the current process parameters by the generative large language model;

[0012] (6) Conduct stable production based on the current process parameters and perform quality testing on the resulting products; if the current product quality is consistent with the optimization target, stop the iteration and output the current process parameters; otherwise, proceed to the next step.

[0013] (7) Feed the test quality indicators back to the generative large language model. Based on the current product quality indicators and the gap between them and the optimization target, repeat steps (4) to (7) until the product quality is consistent with the optimization target.

[0014] In step (1) above:

[0015] The authoritative databases mentioned include CNKI, IEEE Xplore, and Web of Science.

[0016] As a preferred option, the target process is selected from one of the following: injection molding, die casting, additive manufacturing, forging, and rolling.

[0017] The target process keywords include the Chinese and English names of the target process, process parameters, and process parameters.

[0018] Specifically, keywords for injection molding include "injection molding process", "process parameters", "injection molding", and "process parameter"; keywords for die casting include "die casting process", "process parameters", "die casting", and "process parameter".

[0019] As a preferred method, each document is segmented according to a fixed number of characters, using the following segmentation formula:

[0020]

[0021] in, This represents the i-th text block of any document, where i ≥ 1; Represents a character; This is the window size, i.e., the maximum text length retained in each segmentation; Step size refers to the distance between the starting positions of two adjacent text blocks during text segmentation, used to control the degree of overlap between blocks; and .when When adjacent text blocks overlap, it helps maintain the integrity of the context.

[0022] In step (2) above:

[0023] As a preferred approach, the following preprocessing is performed before inputting the text block into the text embedding model:

[0024] First, each sentence in the text block is segmented to obtain a word sequence (a word sequence contains one or more characters). Then, special markers indicating the start and end are inserted at the beginning and end of each word sequence. Finally, padding or truncation is performed to ensure that all word sequences have a fixed length. At the same time, an attention mask is constructed for each word, and the resulting word sequence and the corresponding attention mask are used as input to the text embedding model.

[0025] Specifically, before the text blocks are input into the text embedding model, they need to undergo the following preprocessing:

[0026] 1) Tokenization. Tokenization is the process of converting a raw string into basic units that the model can recognize, namely, tokens. For example, the sentence "Injection speed too fast will cause fly edges." might be converted into: ["inject", "shoot", "speed", "degree", "too", "fast", "will", "cause", "start", "fly", "edge", "."]. Text blocks are processed by tokenization to obtain multiple token sequences.

[0027] 2) Add special markers. Special symbols indicating the start and end of the sentence, namely "[CLS]" and "[SEP]", need to be inserted at the very beginning and end of each word sequence. The resulting word sequence is: ["[CLS]", "note", "shoot", "speed", "degree", "pass", "fast", "will", "lead", "rise", "fly", "side", ".", "[SEP]"];

[0028] 3) Padding or truncation. Since neural networks require batch inputs to have a uniform length, but actual text blocks vary in length, they must be adjusted to a fixed length. If a text block is too long, redundant words are removed from the end or middle. If it is too short, padding with "[PAD]" is added at the end.

[0029] The processed lexical sequence can be represented as:

[0030]

[0031] In the formula, It is a sequence of lexical terms; Let i represent the i-th word element, where i ∈ [1, L]; The length of the sequence or the number of words.

[0032] To mask out invalid padding information in subsequent calculations, an attention mask also needs to be constructed for each lexical:

[0033]

[0034] In the formula, For attention masking; is the attention mask value corresponding to the i-th word, where 1 indicates valid content or valid word, and 0 indicates that the position is invalid padding information; is the length of the sequence.

[0035] The word sequence obtained after the above processing and the corresponding attention mask together constitute the input of the text embedding model, so as to ensure that the text embedding model only focuses on the real text content (effective words) and ignores the padding part.

[0036] As a preferred approach, the text embedding model includes an embedding layer, multiple Transformers, and an average pooling layer;

[0037] For each word in a text block, the embedding layer converts each word into a vector and adds positional information to it to obtain the embedding vector for each word. The embedding vectors of all words in the text block are combined into a matrix and used as the input of the multi-layer Transformer.

[0038] Each layer in a multi-layer Transformer consists of a multi-head self-attention module and a position-feedforward neural network. The multi-head self-attention module of the first-layer Transformer calculates the self-attention output using a matrix composed of the embedding vectors of all words as input. The position-feedforward neural network independently performs non-linear transformations on the vector at each position in the word sequence based on the self-attention output and the matrix composed of the embedding vectors of all words. The output of the position-feedforward neural network is added to the input and then normalized to obtain the output of the first-layer Transformer. The second-layer Transformer takes the output of the first-layer Transformer as input, repeats the calculation process of the first-layer Transformer, and uses its output as the input of the third-layer Transformer. This process is repeated until the output of the last-layer Transformer is obtained.

[0039] The average pooling layer takes the arithmetic mean of the output of the last Transformer layer to obtain the embedding vector representing the entire text block.

[0040] As a further preferred option, the embedding layer converts individual tokens into vectors and adds positional information using the following formula:

[0041]

[0042] In the formula, For the first Embedding vectors corresponding to each term; This represents the i-th word element in a certain text block; It is a word element The corresponding word embedding vector; It is a word element The corresponding position embedding vector.

[0043] After the transformation by the embedding layer, the subsequent generative large language model can not only understand the meaning of each word, but also perceive the order of the word in the word sequence.

[0044] As a further optimization, the multi-head self-attention module has multiple parallel self-attention heads, each independently calculating the self-attention output. First, the query, key, and value of the self-attention are calculated using a matrix composed of the embedding vectors of all tokens:

[0045]

[0046] In the formula, , , They are the first A query, key, and value matrix of self-attention heads, i∈[1,h], where h is the number of self-attention heads; The matrix consists of the embedding vectors of all words. The number of lexical units; , , The first Each self-attention head is an independent learnable weight matrix.

[0047] Next, the self-attention output of each self-attention head is calculated using the following formula:

[0048]

[0049] In the formula, This is a self-attention output; It is a scaling factor used to stabilize the training process; This is the transpose of the key matrix; It is a nonlinear function; It is The mask matrix constructed from attention masks, The number of lexical units, its first... The value of the row is:

[0050]

[0051] In the formula, j∈[1,L].

[0052] Then, the outputs of all self-attention heads are concatenated, and a linear transformation is applied to obtain the final output of the multi-head self-attention module:

[0053]

[0054] In the formula, For multi-head self-attention modules, the input is The output result at that time; For the first The output of each self-attention head, i∈[1,h], where h is the number of self-attention heads; This indicates splicing based on the last dimension; This is the output projection matrix, i.e., the matrix obtained after initialization or update.

[0055] As a further preferred option, the calculation process of the position feedforward neural network is as follows:

[0056]

[0057] In the formula, This is the output of the position feedforward neural network; , This is the learnable weight matrix of the feedforward neural network; , is the learnable bias vector of the feedforward neural network; It is the ReLU activation function;

[0058] The input to the position-based feedforward neural network is obtained using the following formula:

[0059]

[0060] in, It is a matrix consisting of the embedding vectors of all words; This is the output of the multi-head self-attention module; This is the layer normalization function.

[0061] As a further optimization, the calculation formula for adding the output and input of the position feedforward neural network and then performing layer normalization is as follows:

[0062]

[0063] In the formula, This is the output of a Transformer layer corresponding to the position feedforward neural network.

[0064] In step (3) above:

[0065] The determination of adjustable process parameters is based on their impact on the key molding mechanism. Generally, parameters with a greater impact are selected as adjustable process parameters, i.e. parameters to be optimized, while other parameters remain unchanged.

[0066] Generative large language models can be deployed using local deployment architectures such as Ollam. One of the generative large language models to choose from is Deepseek, Qwen, or similar.

[0067] Preferably, in step (4), cosine similarity calculation is used for similarity search:

[0068]

[0069] in, For similarity calculation functions, , These represent the input to the generative large language model and the embedding vector representing a text block in the injection molding process knowledge base, respectively.

[0070] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0071] The intelligent optimization method for molding processes of this invention searches for Chinese and English literature using keywords related to the target molding process. After segmenting the literature, it converts the segments into vectors recognizable by a generative large language model using a text embedding model, thereby constructing a molding process knowledge base. Then, a locally deployed generative large language model performs a similarity search within the process knowledge base and provides process parameters. These parameters are used for stable production, while product quality is tested and fed back to the generative large language model. The generative large language model iteratively updates the process parameters until the product quality reaches the optimization target. This method exhibits strong system stability, high response efficiency, and few iterations. It also demonstrates high generalization ability, adapting to different molding processes, materials, products, and other production conditions. Furthermore, the large model can generate operational suggestions, providing guidance for engineering practice. Attached Figure Description

[0072] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0073] Figure 2 Photograph of the injection molding machine used in the embodiments;

[0074] Figure 3 This is a photograph of the final product that meets the requirements in the embodiments of the present invention. Detailed Implementation

[0075] Taking injection molding as an example, the molding process involves the coupled influence of multiple fields such as shear flow field, temperature field, and stress field, making the mechanism by which process parameters affect the quality of the final product extremely complex. Therefore, the optimization of process parameters is a key bottleneck restricting the stability of product quality and the improvement of production efficiency. Traditional manual trial-and-error methods rely on engineering experience and have low optimization efficiency, while data-driven methods have high data requirements and poor generalization, neither of which are suitable for the precision manufacturing of high-end products.

[0076] like Figure 1 As shown, taking injection molding as an example, an intelligent optimization method for molding processes mainly consists of three processes: the construction of a molding process knowledge base, the deployment of a generative large model, and intelligent process optimization.

[0077] Step 1: Collect relevant Chinese and English literature from authoritative databases based on the target process keywords; for each literature, segment it and obtain multiple text blocks.

[0078] 1.1 First, collect relevant Chinese and English literature from authoritative academic databases such as CNKI, IEEE Xplore, and Web of Science using the keywords "injection molding process" and "process parameter".

[0079] 1.2. After the preliminary collection of the literature, in order to adapt to the input length limit of the embedding model and improve the relevance of subsequent retrieval, it is necessary to perform document segmentation on the long-form literature, generally segmenting the text by a fixed number of characters. At the same time, an overlapping area is set to avoid semantic fragmentation.

[0080] Let the character length of the text material of a piece of literature be , and its character sequence be , then the th text block can be expressed as:

[0081]

[0082] In the formula, represents a character; is the window size, that is, the maximum text length retained for each segmentation; is the step size, which refers to the distance between the starting positions of two adjacent text blocks during text segmentation and is used to control the overlap degree between blocks; is the serial number of the segmented text block ( ).

[0083] And , to ensure that there is overlap between adjacent text blocks, which helps to maintain the integrity of the context.

[0084] Step 2. Use the text embedding model to convert the segmented text blocks into vectors and store them in the local vector database to build a retrievable process knowledge base.

[0085] This step is to use the text embedding model to convert unstructured natural language text into a structured vector representation, enabling the subsequent generative large language model to judge semantic similarity based on the distance of the text in the vector space, thereby supporting semantic retrieval, similarity calculation, and artificial intelligence tasks.

[0086] The input of the text embedding model is the text blocks obtained by segmentation in Step 1, in the form of natural language strings. Before the text blocks are input into the text embedding model, they need to go through the following preprocessing:

[0087] 1) Word segmentation. First, segment each sentence in the text block to obtain a sequence of tokens (corresponding to one or more characters); word segmentation is the process of converting the original string into the basic units recognizable by the model, that is, tokens. For example, the sentence "Excessive injection speed will cause flash." may be converted to: ["注", "射", "速", "度", "过", "快", "会", "引", "起", "飞", "边", "。"]. The text block undergoes word segmentation to obtain multiple sequences of tokens.

[0088] 2) Add special markers. Special symbols indicating the start and end of the sentence, namely "[CLS]" and "[SEP]", need to be inserted at the very beginning and end of each word sequence. The resulting word sequence is: ["[CLS]", "note", "shoot", "speed", "degree", "pass", "fast", "will", "lead", "rise", "fly", "side", ".", "[SEP]"];

[0089] 3) Padding or truncation. Since neural networks require batch inputs to have a uniform length, but actual text blocks vary in length, they must be adjusted to a fixed length. If a text block is too long, redundant words are removed from the end or middle. If it is too short, padding with "[PAD]" is added at the end.

[0090] The processed lexical sequence can be represented as:

[0091]

[0092] In the formula, It is a sequence of lexical terms; Let i represent the i-th word element, where i ∈ [1, L]; The length of the sequence or the number of words.

[0093] To mask out invalid padding information in subsequent calculations, an attention mask also needs to be constructed for each lexical:

[0094]

[0095] In the formula, For attention masking; is the attention mask value corresponding to the i-th word, where 1 indicates valid content or valid word, and 0 indicates that the position is invalid padding information; is the length of the sequence.

[0096] The word sequence obtained after the above processing and the corresponding attention mask together constitute the input of the text embedding model, so as to ensure that the text embedding model only focuses on the real text content (effective words) and ignores the padding part.

[0097] The text embedding model consists of the following parts: 1) an embedding layer; 2) a multi-layer Transformer; and 3) an average pooling layer. It takes a sequence of words and their corresponding attention masks as input, processes them, and finally outputs the text vectorization result, i.e., the embedding vector. The specific processing steps are as follows:

[0098] 2.1 The embedding layer converts each word into a vector and adds positional information to it to obtain the embedding vector corresponding to that word.

[0099] Specifically, each lexical unit Convert to:

[0100]

[0101] In the formula, For the first Embedding vectors corresponding to each term; This represents the i-th word element in a certain text block; It is a word element The corresponding word embedding vector; It is a word element The corresponding position embedding vector.

[0102] In this way, subsequent generative large language models can not only understand the meaning of each word, but also perceive the order of that word in the word sequence.

[0103] The embedding vectors of all words corresponding to the text block are assembled into a matrix. This serves as the input to a multi-layered Transformer. Among them, The number of lexical units.

[0104] 2.2 In a multi-layer Transformer, each layer contains a multi-head self-attention module and a position feedforward neural network.

[0105] For each Transformer layer, the multi-head self-attention module has multiple parallel self-attention heads, each of which independently computes the self-attention output.

[0106] First, the multi-head self-attention module of the first-layer Transformer utilizes a matrix composed of the embedding vectors of all words. Calculate the query, key, and value for self-attention:

[0107]

[0108] In the formula, , , They are the first A query, key, and value matrix of self-attention heads, i∈[1,h], where h is the number of self-attention heads; , , The first Each self-attention head is an independent learnable weight matrix.

[0109] Next, the self-attention output of each self-attention head is calculated using the following formula:

[0110] (6)

[0111] In the formula, This is a self-attention output; It is a scaling factor used to stabilize the training process; This is the transpose of the key matrix; It is a nonlinear function; It is The mask matrix constructed from attention masks, The number of lexical units, its first... The value of the row is:

[0112]

[0113] In the formula, j∈[1,L].

[0114] Then, the outputs of all self-attention heads are concatenated, and a linear transformation is applied to obtain the final output of the multi-head self-attention module:

[0115] (7)

[0116] In the formula, For multi-head self-attention modules, the input is The output result at that time; For the first The output of each self-attention head, i∈[1,h], where h is the number of self-attention heads; This indicates splicing based on the last dimension; This is the output projection matrix, i.e., the matrix obtained after initialization or update.

[0117] The output of the multi-head self-attention module A matrix is ​​formed with the embedding vectors of all words. Add them together and perform layer normalization to obtain the input of the first layer of the Transformer's position-forward neural network, i.e.:

[0118]

[0119] In the formula, This is the input to the position-fedforward neural network. This is the layer normalization function.

[0120] Following the multi-head self-attention module is a position-feedforward neural network, which independently performs non-linear transformations on the vector at each position in the word sequence.

[0121] The calculation process of the position feedforward neural network is as follows:

[0122] (9)

[0123] In the formula, This is the output of the position feedforward neural network; , This is the learnable weight matrix of the feedforward neural network; , is the learnable bias vector of the feedforward neural network; This is the ReLU activation function.

[0124] The output of the first Transformer layer's position-forward neural network is added to the input, and layer normalization is performed to obtain the final output of the first Transformer layer:

[0125]

[0126] In the formula, This is the output of a Transformer layer.

[0127] The second-layer Transformer takes the output of the first-layer Transformer as input, repeats the computation process of the first-layer Transformer, and uses its output as input for the third-layer Transformer. This process is repeated until the output of the last-layer Transformer is obtained.

[0128] 2.3 After processing by multiple Transformers, the average pooling layer takes the arithmetic mean of the output of the last Transformer layer to obtain an embedding vector representing the entire text block.

[0129] Finally, the embedding vectors and their metadata for all generated representative text blocks are stored together to form an injection molding process knowledge base. Stored entries include embedding vectors, original text blocks, and metadata such as document titles and authors.

[0130] Step 3: Regarding the generative large language model, the Deepseek generative large language model is deployed on a local server using the Ollam local deployment architecture, serving as the core inference engine for the injection molding process intelligent optimization system. Through local deployment, the system can directly process sensitive information such as process parameters and product quality data generated during production, without uploading to third-party platforms, significantly improving data security and system autonomy. Simultaneously, the local operating mode reduces dependence on external services, enhancing system stability and response efficiency, making it particularly suitable for closed-loop optimization scenarios requiring rapid feedback.

[0131] Step 4: In terms of intelligent process optimization, clarify the process materials, product types, optimization objectives, adjustable process parameters and their ranges, and fixed process parameters;

[0132] This embodiment uses injection molding process optimization as an example. The selected material is polymethyl methacrylate (PMMA), and the product is a flat part. The optimization index is set as product weight, and the optimization goal is to achieve a product weight of 16.60g while ensuring a defect-free product and stable production. Injection speed, holding pressure, and cooling time are selected as optimization variables, mainly based on their dominant role in the key molding mechanism. Meanwhile, other process parameters remain unchanged during the optimization process.

[0133] Step 5: First, input the optimization task and related information into the locally deployed generative large language model. The input information (user input) is as follows:

[0134] The current injection molding material is PMMA, the product is a flat part, and the quality indicator is the product weight, with a target value of 16.60g. The fixed holding time is 5 seconds, the injection time is 5 seconds, the injection pressure is 100MPa, and the barrel temperature from the inlet to the nozzle is 185℃, 200℃, 215℃, and 220℃ respectively. The initial process parameters are generated by adjusting three process parameters: injection speed (optimized range 20-180mm / s), holding pressure (range 10~100MPa), and cooling time (8~25s).

[0135] After obtaining the optimization task, the generative large language model first utilizes the process knowledge base to search for relevant information (vectors corresponding to text blocks) in the knowledge base through similarity search. The vector database (process knowledge base) supports efficient data similarity search, and can quickly find the most semantically relevant text blocks when facing user queries. The similarity search uses cosine similarity calculation:

[0136] (11)

[0137] In the formula, For similarity calculation functions, , These represent the user input (the input to the generative large language model) and the embedding vector representing the text block in the injection molding process knowledge base, respectively.

[0138] Step 6: After retrieving the knowledge base content (relevance vector) most relevant to the current user input, send this content along with the original input (user input) to the generative large language model. This allows the generative large language model to refer to the relevant content (relevance vector) in the knowledge base when responding, thereby enhancing the knowledge base of the generative large language model. Subsequently, through analysis and reasoning, based on engineering experience in injection molding, the generative large language model recommends the optimal process parameters as follows: injection speed 120mm / s, holding pressure 25MPa, and cooling time 15s.

[0139] Step 7: Perform injection molding using the obtained process parameters (such as the injection molding machine used). Figure 2 As shown, stable production was achieved, and the average weight of the product was measured to be 16.52g, with an error of only 0.08g from the optimization target of 16.60g, which is very close to the optimization target.

[0140] Step 8: Next, feed the current product quality information back to the generative large language model. At the same time, based on the current product quality and the gap between it and the optimization target, perform a similarity search and repeat steps 6 to 8 until the obtained product quality is consistent with the optimization target.

[0141] This embodiment achieved process optimization through three iterations, demonstrating the effectiveness and superiority of the method described in this embodiment.

[0142] During the optimization process, the process parameters obtained in the second iteration were: injection speed 150 mm / s, holding pressure 30 MPa, and cooling time 12 s. It can be seen that the optimization logic for this large-scale model is to slightly increase the injection speed and holding pressure to fill more melt during the injection and holding stages, thereby increasing the product quality, which aligns with engineering experience. Then, using the updated process parameters, stable injection molding production was conducted, and the average weight of the product was measured to be 16.53 g, with an error of 0.07 g. In the third iteration, the process parameters obtained were: injection speed 160 mm / s, holding pressure 38 MPa, and cooling time 9 s. The product obtained from stable injection molding production (e.g.) Figure 3 The mass (as shown) is 16.60g, which meets the optimization target and can be produced stably.

Claims

1. A method of intelligent optimization of a forming process, characterized by, The method comprises the following steps: (1) Collecting relevant Chinese and English literatures from authoritative databases according to target process keywords; for each literature, performing segmentation processing and obtaining multiple text blocks respectively; (2) Converting the segmented text blocks into embedding vectors by using a text embedding model and storing the embedding vectors in a local vector database to construct a searchable process knowledge base; (3) Determining process materials, product types, optimization targets, adjustable process parameters and their ranges, and fixed process parameters, and inputting them into a generative large language model deployed on a local server; (4) The generative large language model performs similarity search on the process knowledge base to find the embedding vectors corresponding to the relevant text blocks; (5) Inputting the embedding vectors found in step (4) into the generative large language model for enhancement, and generating the current process parameters by the generative large language model; (6) Producing stably according to the current process parameters, and testing the quality of the obtained products; if the quality of the current product is consistent with the optimization target, stop iteration and output the current process parameters; if not, go to the next step; (7) Feedback the test quality indicators to the generative large language model, and repeat steps (4) to (7) according to the gap between the quality indicators of the current product and the optimization target until the quality of the product is consistent with the optimization target.

2. The method of claim 1, wherein, In step (1), the target process is selected from one of injection molding process, die casting process, additive manufacturing process, forging process and rolling process; The target process keywords include the Chinese and English names of the target process, process parameters and process parameters.

3. The method of claim 1, wherein, In step (1), each literature is segmented according to a fixed number of characters, and the segmentation formula is as follows: ； wherein, represents the i-th text block of any document, i ≥ 1; represents a character; is a window size; is a step size; and .

4. The method of claim 1, wherein, In step (2), the following preprocessing is performed before inputting the text block into the text embedding model: First, perform word segmentation on each sentence in the text block to obtain a token sequence, then insert special markers indicating the beginning and end at the beginning and end of each token sequence, and finally pad or truncate to ensure that all token sequences have the same length; at the same time, construct an attention mask for each token, and use the obtained token sequence and corresponding attention mask as the input of the text embedding model.

5. The method of claim 4, wherein, In step (2), the text embedding model comprises an embedding layer, a multi-layer Transformer and an average pooling layer; For all tokens corresponding to each text block, the embedding layer converts each token into a vector and adds position information to obtain an embedding vector corresponding to each token; the embedding vectors of all tokens are combined into a matrix and used as the input of the multi-layer Transformer; Each layer of the multi-layer Transformer contains a multi-head self-attention module and a position feedforward neural network. The multi-head self-attention module of the first layer of the Transformer calculates a self-attention output with a matrix composed of embedding vectors of all word tokens as input; a position feedforward neural network independently performs a nonlinear transformation on a vector of each position in the sequence of word tokens according to the self-attention output and the matrix composed of embedding vectors of all word tokens; the output of the position feedforward neural network is added to the input, and then layer normalization processing is performed to obtain the output of the first layer of the Transformer; the second layer of the Transformer takes the output of the first layer of the Transformer as input, repeats the calculation process of the first layer of the Transformer, and takes the output as input of the third layer of the Transformer, and so on, until the output of the last layer of the Transformer is obtained; The average pooling layer takes an arithmetic mean of the output of the last layer of the Transformer to obtain an embedding vector representing the entire text block.

6. The method of claim 1, wherein, In step (4), a cosine similarity calculation is performed for the similarity search: ； wherein, is a similarity computation function, , respectively represent the input of the generative large language model and the embedding vectors representing the text blocks in the injection molding process knowledge base.

Images