A variant design method for electronic equipment based on large-scale language model

Abstract

A method for designing electronic equipment variants based on a large-scale language model belongs to the technical field of electronic equipment variant design. In order to solve the problem of generating a CBB model that meets the requirements of the corresponding field and function, the present invention includes providing a command for generating a target electronic equipment CBB model to the Tongyi Qianwen model; providing a command for adding a prompt word to the Tongyi Qianwen model; generating an electronic equipment CBB model, and then inputting a text command to the Tongyi Qianwen model, and the Tongyi Qianwen model generates an electronic product architecture according to the text command and the electronic equipment CBB model; obtaining text features as an electronic equipment variant data set; using a Chinese version of the BERT model as an electronic equipment variant design Bert model, and then fine-tuning the electronic equipment variant design Bert model based on the Lora method to obtain a trained Chinese version of the BERT model; filtering out irrelevant electronic equipment CBB models according to the relevance of the model text features. The present invention has high correctness.

Description

Technical Field

[0001] The invention belongs to the technical field of electronic equipment variant design, and in particular relates to an electronic equipment variant design method based on a large-scale language model. Background Art

[0002] The CBB (Common Building Block) model is used in system design in electronic equipment and other fields. The core idea is to decompose complex systems into a series of standardized and modular components that can be reused across different products and systems, that is, a collection of parts with closely related functions and a high number of combinations.

[0003] The CBB model decomposes the system into multiple independent modules, so that designers in different fields can focus on the design of their corresponding modules rather than the entire system. This modular approach improves design efficiency and team collaboration efficiency. At the same time, due to the independence of each CBB module, product problems can be limited to a specific range, and only a few specific modules need to be focused on, reducing the cost of inspecting and repairing the product system.

[0004] Using proven CBB modules can avoid repeated design and testing, saving time and resources while ensuring the reliability and stability of new product systems. By combining different CBB modules, multiple system variants can be quickly and flexibly generated to meet different application scenarios and customer needs.

[0005] However, in addition to the electronic information field knowledge required for designing electronic equipment, some electronic products and CBB models often involve knowledge from other disciplines, such as materials, mechanics, etc., due to their specific functions and application scenarios. At present, the market demand for electronic equipment is becoming increasingly diverse, while traditional manual design methods generally have a single product type and are difficult to cope with the complex, flexible and changeable needs of many industries; and the design process of a complete product in traditional methods takes a long time. Summary of the invention

[0006] The problem to be solved by the present invention is to generate a CBB model that meets the requirements of corresponding fields and functions, and propose an electronic equipment variant design method based on a large-scale language model.

[0007] To achieve the above object, the present invention is implemented through the following technical solutions:

[0008] A method for designing electronic equipment variants based on a large-scale language model comprises the following steps:

[0009] S1. Provide a command to generate a CBB model of a target electronic equipment to the Tongyi Qianwen model, including introductory information on the field and purpose of the CBB model of the target electronic equipment;

[0010] S2. Provide a command to add a prompt word to the Tongyi Qianwen model, and the prompt word is used as an example of the target electronic equipment CBB model;

[0011] S3 generates an electronic equipment CBB model, and then inputs a text command to the Tongyi Qianwen model, and the Tongyi Qianwen model generates an electronic product architecture based on the text command and the electronic equipment CBB model;

[0012] S4. Tongyi Qianwen model analyzes the electronic equipment CBB model and electronic product architecture generated in step S3, and the electronic equipment CBB model introduction text and electronic product introduction text to obtain text features as an electronic equipment variant data set;

[0013] S5. Use the Chinese version of the BERT model as the electronic equipment variant design Bert model, then fine-tune the electronic equipment variant design Bert model based on the Lora method, and use the electronic equipment variant data set obtained in step S4 to train the electronic equipment variant design Bert model to obtain a trained Chinese version of the BERT model;

[0014] S6. The trained Chinese version BERT model obtained in step S5 is used to determine the correlation between the text features of the electronic products of the electronic product architecture generated by the Tongyi Qianwen model and their corresponding electronic equipment CBB models, and to filter out irrelevant electronic equipment CBB models.

[0015] Furthermore, in the commands for generating the target electronic equipment CBB model in step S1, the first command instruction1 is to give the expected number x of CBB modules or units for electronic equipment manufacturing and electronic equipment variant design, and the second command instruction2 is to give examples of CBB modules or units for electronic equipment manufacturing and electronic equipment variant design for specific fields or work.

[0016] Furthermore, a prompt word command instruction3 is added in step S2, and the prompt word is an example CBB model, including one or more of a drone autonomous navigation system, an automatic driving control system, an encrypted communication module, and a WIFI module.

[0017] Furthermore, the specific implementation method of step S3 includes the following steps:

[0018] S3.1. Generate the CBB model of electronic equipment, the expression is:

[0019] cbb=Qwen(instruction1 or instruction2,instruction3)

[0020] Among them, cbb represents the CBB model set, Qwen represents the Tongyi Qianwen model;

[0021] S3.2. Set the Tongyi Qianwen model input text command instruction4 to generate multiple electronic product architectures based on the electronic equipment CBB model and other related CBB models;

[0022] S3.3. Input text commands to the Tongyi Qianwen model. The Tongyi Qianwen model generates an electronic product architecture based on the text commands and the electronic equipment CBB model. The expression is:

[0023] eframework=Qwen(cbb,instruction4)

[0024] eframework i =(eproduct i ,cbb i,1 ,cbb i,2 ,…,cbb i,n )

[0025] Among them, eframework is a collection of electronic product architectures. i is the i-th electronic product architecture in the set, eproduct i For the architecture eframework i A single electronic product, cbb i,n Represents the i-th electronic product architecture eframework i The nth CBB model individual in the eframework, i is a positive integer ranging from 1 to the total number of electronic product architectures contained in the eframework collection; n represents the architecture eframework i Total number of CBB models included.

[0026] Furthermore, the specific implementation method of step S4 includes the following steps:

[0027] S4.1. Set the fifth command instruction5 to generate an electronic equipment CBB model or an electronic product introduction text;

[0028] S4.2. Tongyi Qianwen Model For the electronic equipment CBB model and electronic product architecture generated in step S3, the expressions of the electronic equipment CBB model introduction text and the electronic product introduction text are analyzed as follows:

[0029] cbbtext=Qwen(cbb,instruction5)

[0030] eproducttext=Qwen(eproduct,instruction5)

[0031] Among them, cbbtext is the collection of introduction texts of the CBB model, and eproducttext is the collection of introduction texts of electronic products.

[0032] Furthermore, the specific implementation method of step S5 includes the following steps:

[0033] S5.1. The introduction text of the electronic product and the introduction text of the CBB model in the electronic equipment variant dataset obtained in step S4 are concatenated into a long text, and then separated by the separator [SEP];

[0034] S5.2. The long text obtained in step S5.1 is first encoded by the Chinese-BERT-wwm model to extract the main text features;

[0035] S5.3. The text encoding obtained in step S5.2 is transformed in dimension by the linear layer of the Chinese-BERT-wwm model and normalized by the sigmoid activation function to obtain the matching probability of the introduction text of the electronic product and the introduction text of the CBB model in the electronic equipment variant dataset, and the Chinese-BERT-wwm model is trained;

[0036] S5.4. After the training is completed, a trained Chinese-BERT-wwm model is obtained, and the trained Chinese-BERT-wwm model is used to judge the correctness of the electronic product architecture data generated by the Tongyi Qianwen model, that is, the correlation between the electronic product and the text features of its subordinate CBB model.

[0037] Furthermore, the specific implementation method of step S5.3 includes the following steps:

[0038] S5.3.1. The Chinese-BERT-wwm model is trained using the Lora method, and the expression is:

[0039] W0+W’=W0+AB T

[0040] Among them, W0 represents the original parameters of the model, which is an n×d matrix and is not updated during the training process. W’ is the change in parameters during the training process, which has the same size as W0. A and B represent the matrix sets of Lora fine-tuning parameters. A is an n×l matrix and B is a d×l matrix. The positive integers l, d, and n represent the dimensions of the parameter matrices, and l<<d,n. The computational amount during the training process is reduced through matrix decomposition;

[0041] S5.3.2. The process of fine-tuning the Chinese-BERT-wwm model is expressed in formula form as follows:

[0042] Wlora i =Wusei +A i B i T

[0043] p i,j =CBERT(Wlora U Wnotuse,(eproducttext i ,[SEP],cbbtext j ))

[0044] p i,j =sigmoid(Linear(p i,j ))

[0045] Among them, CBERT is the abbreviation of Chinese-BERT-wwm model, Wuse i , Wnotuse represent the parameter sets of the Chinese-BERT-wwm model that participate in fine-tuning and those that do not participate in fine-tuning, respectively. i is the result after parameter fine-tuning. U represents the set union operation, Linear represents the linear layer that transforms the dimensionality of the BERT model output, and p i,j Represents the matching probability between electronic product i and electronic equipment CBB model j.

[0046] Furthermore, in step S6, for electronic equipment CBB models whose matching probability between electronic products and electronic equipment CBB models is less than 0.5, the electronic equipment CBB models are determined to be irrelevant electronic equipment CBB models and are filtered and deleted.

[0047] Beneficial effects of the present invention:

[0048] The electronic equipment variant design method based on a large-scale language model described in the present invention automatically generates a CBB model that meets the corresponding field and functional requirements based on the semantic understanding, retrieval and generation capabilities of the Tongyi Qianwen model, and on this basis uses the Tongyi Qianwen model to summarize and generate the electronic product architecture.

[0049] The electronic equipment variant design method based on a large-scale language model described in the present invention is based on the CBB model and electronic product architecture data generated by the Tongyi Qianwen model, and uses the Lora method to fine-tune the Bert language model that can effectively judge the degree of correlation between electronic products and the CBB model.

[0050] The electronic equipment variant design method based on a large-scale language model described in the present invention combines two models. While using the Tongyi Qianwen model to flexibly generate electronic equipment data in various fields and functions, a fine-tuned Bert model is used to verify the correctness of the results, thereby effectively reducing the error of the results. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A flowchart of a method for designing electronic equipment variants based on a large-scale language model according to the present invention;

[0052] Figure 2 It is a schematic diagram of the overall process of the present invention;

[0053] Figure 3 Generating a flow chart for an electronic equipment variant data set of the present invention;

[0054] Figure 4 Schematic diagram of BERT structure;

[0055] Figure 5 The flowchart of the calculation method of the matching probability between the electronic product of the present invention and the CBB model;

[0056] Figure 6 It is the CBB model and introductory text;

[0057] Figure 7 For electronic products and introduction texts;

[0058] Figure 8 It is the correspondence between electronic products and CBB model. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the present invention more clear, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention, that is, the specific embodiments described are only part of the embodiments of the present invention, rather than all of the specific embodiments. The components of the specific embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations, and the present invention can also have other embodiments.

[0060] Therefore, the following detailed description of the specific embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents the selected specific embodiments of the present invention. Based on the specific embodiments of the present invention, all other specific embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0061] In order to further understand the content, features and effects of the present invention, the following specific implementation methods are given as examples, and the attached Figure 1 -Attached Figure 8 The detailed instructions are as follows:

[0062] Embodiment 1:

[0063] A method for designing electronic equipment variants based on a large-scale language model comprises the following steps:

[0064] S1. Provide a command to generate a CBB model of a target electronic equipment to the Tongyi Qianwen model, including introductory information on the field and purpose of the CBB model of the target electronic equipment;

[0065] Further, in the command for generating the target electronic equipment CBB model in step S1, the first command instruction1 is to give an expected number x of CBB modules or units for electronic equipment manufacturing and electronic equipment variant design, and the second command instruction2 is to give an example of CBB modules or units for electronic equipment manufacturing and electronic equipment variant design for a specific field or work;

[0066] Further, specific fields include the automotive manufacturing field or the communication work field;

[0067] Furthermore, large language models (LLMs) are usually based on the Transformer model and its encoder or decoder, and learn common language representations from massive corpora in various fields based on a large number of parameters. There are two main pre-training methods: gap completion and autoregressive prediction generation. They have good results in different fields and types of downstream tasks such as classification and question answering. By fine-tuning various types of downstream tasks based on the pre-trained large model, the pre-trained knowledge can be used to accelerate the convergence of downstream tasks, reduce their training time, improve training efficiency, and provide more knowledge for downstream tasks.

[0068] Furthermore, the pre-trained Tongyi Qianwen model provided by Alibaba automatically constructs an electronic equipment dataset: without providing the source data of electronic equipment variants, or only giving a few samples, the Tongyi Qianwen model is used to convert simple short texts representing user needs into a certain scale of general type or specific field electronic equipment dataset: electronic equipment products, subordinate CBB models, and the corresponding relationship between the two - electronic equipment product architecture model. The Qianwen model is a generative model based on Transformer-decoder, which uniformly regards different types of generation tasks as text sequence regression prediction. Tongyi Qianwen is fully functional, ranks at the forefront in various performance rankings, and is more suitable for domestic use. Using simple text commands, or providing prompt information such as a small amount of sample data in the input, it can meet the design needs of electronic equipment in multiple fields and for various purposes.

[0069] S2. Provide a command to add a prompt word to the Tongyi Qianwen model, and the prompt word is used as an example of the target electronic equipment CBB model;

[0070] Furthermore, a prompt word command instruction3 is added in step S2, and the prompt word is an example CBB model, including one or more of a drone autonomous navigation system, an automatic driving control system, an encrypted communication module, and a WIFI module.

[0071] S3 generates an electronic equipment CBB model, and then inputs a text command to the Tongyi Qianwen model, and the Tongyi Qianwen model generates an electronic product architecture based on the text command and the electronic equipment CBB model;

[0072] Furthermore, the specific implementation method of step S3 includes the following steps:

[0073] S3.1. Generate the CBB model of electronic equipment, the expression is:

[0074] cbb=Qwen(instruction1 or instruction2,instruction3)

[0075] Among them, cbb represents the CBB model set, Qwen represents the Tongyi Qianwen model;

[0076] S3.2. Set the Tongyi Qianwen model input text command instruction4 to generate multiple electronic product architectures based on the electronic equipment CBB model and other related CBB models;

[0077] S3.3. Input text commands to the Tongyi Qianwen model. The Tongyi Qianwen model generates an electronic product architecture based on the text commands and the electronic equipment CBB model. The expression is:

[0078] eframework=Qwen(cbb,instruction4)

[0079] eframework i =(eproduct i ,cbb i,1 ,cbb i,2 ,…,cbb i,n )

[0080] Among them, eframework is a collection of electronic product architectures. i is the i-th electronic product architecture in the set, eproduct i For the architecture eframework i A single electronic product, cbb i,n Represents the i-th electronic product architecture eframework iThe nth CBB model individual in the eframework, i is a positive integer ranging from 1 to the total number of electronic product architectures contained in the eframework collection; n represents the architecture eframework i Total number of CBB models included.

[0081] S4. Tongyi Qianwen model analyzes the electronic equipment CBB model and electronic product architecture generated in step S3, and the electronic equipment CBB model introduction text and electronic product introduction text to obtain text features as an electronic equipment variant data set;

[0082] Furthermore, the specific implementation method of step S4 includes the following steps:

[0083] S4.1. Set the fifth command instruction5 to generate an electronic equipment CBB model or an electronic product introduction text;

[0084] S4.2. Tongyi Qianwen Model For the electronic equipment CBB model and electronic product architecture generated in step S3, the expressions of the electronic equipment CBB model introduction text and the electronic product introduction text are analyzed as follows:

[0085] cbbtext=Qwen(cbb,instruction5)

[0086] eproducttext=Qwen(eproduct,instruction5)

[0087] Among them, cbbtext is the introduction text collection of the CBB model, and eproducttext is the introduction text collection of electronic products.

[0088] S5. Use the Chinese version of the BERT model as the electronic equipment variant design Bert model, then fine-tune the electronic equipment variant design Bert model based on the Lora method, and use the electronic equipment variant data set obtained in step S4 to train the electronic equipment variant design Bert model to obtain a trained Chinese version of the BERT model;

[0089] Table 1 Commands required to generate electronic equipment variant design data using the Tongyi Qianwen model

[0090]

[0091]

[0092] The electronic equipment dataset is generated using the Tongyi Qianwen model in the form of a website or API. The number of various data generated after calling the model once is shown in Table 2.

[0093] The generated data sample is as follows:

[0094] (1) CBB model: GPS positioning module, Bluetooth communication module, artificial intelligence processing unit;

[0095] (2) CBB model in specific fields (taking automobile as an example): in-vehicle infotainment system, power integrated control system, and autonomous driving control unit;

[0096] (3) Electronic product architecture: UAV—>[autonomous navigation system, GPS positioning module, wireless charging module, WIFI communication module, video transmission module];

[0097] The Tongyi Qianwen model was used many times to generate electronic equipment data sets, among which 23 electronic product architectures and 281 CBB models with relatively reasonable structures and close to real electronic equipment were selected as training data sets. Each electronic product architecture consists of 4 to 6 CBB models, which are used to fine-tune the Bert model for judging the correctness of the electronic product architecture.

[0098] Table 2 Data generation results

[0099]

[0100]

[0101] Furthermore, the specific implementation method of step S5 includes the following steps:

[0102] S5.1. The electronic product introduction text and the CBB model introduction text in the electronic equipment variant data set obtained in step S4 are concatenated into a long text and then separated by a separator [SEP];

[0103] S5.2. The long text obtained in step S5.1 is first encoded by the Chinese-BERT-wwm model to extract the main text features;

[0104] Furthermore, BERT is an unsupervised pre-trained language model proposed by Google in 2018. Its basic architecture is the encoder part of the transformer. By pre-training basic tasks such as missing word prediction in various general massive corpora, it can give a reasonable row vector encoding of length 768 for each word in the sequence text and the entire sequence, corresponding to its spatial structure and attribute information in the general corpus. Figure 2-3 However, it cannot obtain enough text sequence features. The Chinese version of BERT used in this project can encode and process more than 20,000 Chinese characters, English words, numbers, and special characters, with Chinese characters accounting for the majority, and can handle Chinese NLP tasks in most fields.

[0105] S5.3. Transform the dimension of the text encoding obtained in step S5.2 through the linear layer of the Chinese-BERT-wwm model and normalize it with the sigmoid activation function to obtain the matching probability of the electronic product introduction text and the CBB model introduction text in the electronic equipment variant dataset, and train the Chinese-BERT-wwm model;

[0106] Furthermore, the goal of the LoRA method is to obtain better model training results with lower computational complexity. The original parameters of the large model are not updated during the fine-tuning process; the difference between the weights after LoRA fine-tuning and the original weights is decomposed into the inner product of two matrices with smaller dimensions, and are initialized using Gaussian distribution and all-zero matrix respectively. It is a simple and feasible method and performs well in the fine-tuning of large models such as GPT.

[0107] Furthermore, the specific implementation method of step S5.3 includes the following steps:

[0108] S5.3.1. Train the Chinese-BERT-wwm model using the LoRA method, and the expression is:

[0109] W0 + W’ = W0 + AB T

[0110] Among them, W0 represents the original parameters of the model, which is an n×d matrix and is not updated during the training process. W’ is the change in parameters during the training process, which has the same size as W0. A and B represent the matrix sets of LoRA fine-tuning parameters. A is an n×l matrix and B is a d×l matrix. The positive integers l, d, and n represent the dimensions of the parameter matrices, and l << d, n. The computational complexity during the training process is reduced through matrix decomposition;

[0111] S5.3.2. The process of fine-tuning the Chinese-BERT-wwm model is expressed in formula form as follows:

[0112] Wlora i = Wuse i + A i B i T

[0113] p i,j = CBERT(Wlora U Wnotuse, (eproducttext i , [SEP], cbbtext j ))

[0114] p i,j = sigmoid(Linear(p i,j ))

[0115] Among them, CBERT is the abbreviation of Chinese-BERT-wwm model, Wuse i , Wnotuse represent the parameter sets of the Chinese-BERT-wwm model that participate in fine-tuning and those that do not participate in fine-tuning, respectively. i is the result after parameter fine-tuning. U represents the set union operation, Linear represents the linear layer that transforms the dimensionality of the BERT model output, and p i,j represents the matching probability between electronic product i and electronic equipment CBB model j;

[0116] Furthermore, the BERT model has 12 layers, and the structure of each layer is roughly the same, including the calculation of the attention mechanism and the merging of multiple attention calculation results. The steps involving matrix operations mainly include two types of parameters: one is the parameter for matrix inner product operation with the input, which is usually a large matrix with hundreds of rows and columns; the other is the parameter for matrix addition as a bias after the matrix inner product operation, which is a single-row matrix; the Lora method only fine-tunes the first type of parameters. During the training process, each multi-row and column matrix parameter corresponds to two low-dimensional matrices with the same inner product size. The training results of the low-dimensional matrix are added to the corresponding original parameters as changes after the inner product operation, which has no effect on the structure of the trained model.

[0117] S5.4. After the training is completed, the trained Chinese-BERT-wwm model is obtained, and the trained Chinese-BERT-wwm model is used to judge the correctness of the electronic product architecture data generated by the Tongyi Qianwen model, that is, the correlation between the electronic products and the text features of their subordinate CBB models.

[0118] S6. The trained Chinese version BERT model obtained in step S5 is used to determine the correlation between the text features of the electronic products of the electronic product architecture generated by the Tongyi Qianwen model and their corresponding electronic equipment CBB models, and to filter out irrelevant electronic equipment CBB models.

[0119] Furthermore, in step S6, for electronic equipment CBB models whose matching probability between electronic products and electronic equipment CBB models is less than 0.5, the electronic equipment CBB models are determined to be irrelevant electronic equipment CBB models and are filtered and deleted.

[0120] Furthermore, on this basis, only the first K CBB models with the highest probability of matching with the corresponding electronic products are retained. The number of retained models can be adjusted by the user. In this embodiment, K=6.

[0121] Since the product parts and component architecture information required for the electronic equipment variant design task is relatively unpopular, it is difficult to find public data sets in the corresponding field. Therefore, this embodiment mainly uses a method based on a generative large model, using the retrieval, reasoning, and generation capabilities of the Tongyi Qianwen model to obtain the CBB model of electronic equipment in the corresponding field, and further analyzes the electronic products that can be generated by these CBB models, and constructs an electronic equipment variant design and architecture knowledge graph. This embodiment also organizes some electronic equipment variant CBB models generated by the Tongyi Qianwen model and the product architecture given on its basis into a data set. The Lora method is used locally to fine-tune the Chinese-BERT-wwm model, which has a relatively high proportion of Chinese in the training data and is better at processing Chinese information, so that it can judge whether the introduction of each electronic equipment CBB model meets the corresponding requirements based on the given target product introduction text, and further filter out the wrong product architecture, that is, the CBB model that is not related to the target electronic product.

[0122] Furthermore, the Chinese-BERT-wwm model was fine-tuned using 23 electronic product architectures and 281 CBB models generated by the Tongyi Qianwen model, and compared with two methods: full parameter training and training only the linear layer. Each electronic product corresponds to 4-6 CBB models. Sample training data is as follows Figure 6-Figure 8 shown.

[0123] The test results of the model are shown in Tables 3 and 4. The evaluation indicators are the following two: prec, precision, is the proportion of samples that actually belong to the positive class (TP) among all samples predicted to be positive (TP+FP); rec, recall, is the proportion of samples that are correctly identified as positive (TP) among all positive samples (TP+FN).

[0124] Under the condition that the test set accounts for 20% of the total and the positive and negative sample ratio is 1:1, the precision and recall of the Lora fine-tuning method are relatively optimal, but the advantage is not obvious compared with full parameter training, which may be due to the small amount of training data. In addition, due to the small number of parameters, the Lora method has a slight advantage in model operation speed and result convergence speed compared with conventional full parameter training.

[0125] Table 3 Chinese-BERT-wwm model test results

[0126]

[0127] Table 4 Chinese-BERT-wwm model full dataset running results

[0128]

[0129] Running on the entire data set, combined with the TOP-K (K=6) method, only the top K CBB models with the highest matching probability are retained, and obviously irrelevant negative samples are filtered out to improve the accuracy. The Lora method has a more obvious advantage than other methods, and can identify more than half of the correct targets in nearly 300 CBB models for each electronic product.

[0130] It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0131] Although the present application has been described above with reference to specific embodiments, various modifications may be made thereto and parts thereof may be replaced with equivalents without departing from the scope of the present application. In particular, as long as there is no structural conflict, the various features in the specific embodiments disclosed in the present application may be used in combination with each other in any manner, and the fact that these combinations are not exhaustively described in this specification is only for the sake of omitting space and saving resources. Therefore, the present application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for designing electronic equipment variants based on a large-scale language model, characterized in that: The steps include: S1. Provide a command to generate a CBB model of a target electronic equipment to the Tongyi Qianwen model, including introductory information on the field and purpose of the CBB model of the target electronic equipment; S2. Provide a command to add a prompt word to the Tongyi Qianwen model, and the prompt word is used as an example of the target electronic equipment CBB model; S3 generates an electronic equipment CBB model, and then inputs a text command to the Tongyi Qianwen model, and the Tongyi Qianwen model generates an electronic product architecture based on the text command and the electronic equipment CBB model; S4. Tongyi Qianwen model analyzes the electronic equipment CBB model and electronic product architecture generated in step S3, and the electronic equipment CBB model introduction text and electronic product introduction text to obtain text features as an electronic equipment variant data set; S5. Use the Chinese version of the BERT model as the electronic equipment variant design Bert model, then fine-tune the electronic equipment variant design Bert model based on the Lora method, and use the electronic equipment variant data set obtained in step S4 to train the electronic equipment variant design Bert model to obtain a trained Chinese version of the BERT model; S6. The trained Chinese version BERT model obtained in step S5 is used to determine the correlation between the text features of the electronic products of the electronic product architecture generated by the Tongyi Qianwen model and their corresponding electronic equipment CBB models, and to filter out irrelevant electronic equipment CBB models.

2. The electronic equipment variant design method based on a large-scale language model according to claim 1, characterized in that: In the commands for generating the target electronic equipment CBB model in step S1, the first command instruction1 gives the expected number x of CBB modules or units for electronic equipment manufacturing and electronic equipment variant design, and the second command instruction2 gives examples of CBB modules or units for electronic equipment manufacturing and electronic equipment variant design for specific fields or work.

3. The electronic equipment variant design method based on large-scale language model according to claim 2, characterized in that: In step S2, a prompt word command instruction3 is added, and the prompt word is an example CBB model, including one or more of a drone autonomous navigation system, an automatic driving control system, an encrypted communication module, and a WIFI module.

4. The electronic equipment variant design method based on a large-scale language model according to claim 3, characterized in that: The specific implementation method of step S3 includes the following steps: S3.

1. Generate the CBB model of electronic equipment, the expression is: cbb=Qwen(instruction1 or instruction2,instruction3) Among them, cbb represents the CBB model set, Qwen represents the Tongyi Qianwen model; S3.

2. Set the Tongyi Qianwen model input text command instruction4 to generate multiple electronic product architectures based on the electronic equipment CBB model and other related CBB models; S3.

3. Input text commands to the Tongyi Qianwen model. The Tongyi Qianwen model generates an electronic product architecture based on the text commands and the electronic equipment CBB model. The expression is: eframework=Qwen(cbb,instruction4) eframework i =(eproduct i ,cbb i,1 ,cbb i,2 ,…,cbb i,n ) Among them, eframework is a collection of electronic product architectures. i is the i-th electronic product architecture in the set, eproduct i For the architecture eframework i A single electronic product, cbb i,n Represents the i-th electronic product architecture eframework i The nth CBB model individual in the eframework, i is a positive integer ranging from 1 to the total number of electronic product architectures contained in the eframework collection; n represents the architecture eframework i Total number of CBB models included.

5. The electronic equipment variant design method based on large-scale language model according to claim 4, characterized in that: The specific implementation method of step S4 includes the following steps: S4.

1. Set the fifth command instruction5 to generate an electronic equipment CBB model or an electronic product introduction text; S4.

2. The Tongyi Qianwen model analyzes the electronic equipment CBB model and the electronic product architecture generated in step S3. The expressions for analyzing the introduction text of the electronic equipment CBB model and the introduction text of the electronic product are as follows: cbbtext = Qwen(cbb, instruction5) eproducttext = Qwen(eproduct, instruction5) Among them, cbbtext is the set of introduction texts of the CBB model, and eproducttext is the set of introduction texts of electronic products.

6. The electronic equipment variant design method based on large-scale language model according to claim 5, characterized in that: The specific implementation method of step S5 includes the following steps: S5.

1. The introduction text of the electronic product and the introduction text of the CBB model in the electronic equipment variant dataset obtained in step S4 are concatenated into a long text, and then separated by the delimiter [SEP]. S5.

2. The long text obtained in step S5.1 is first encoded by the Chinese-BERT-wwm model to extract the main text features. S5.

3. The text encoding obtained in step S5.2 is transformed in dimension through the linear layer of the Chinese-BERT-wwm model and normalized by the sigmoid activation function to obtain the matching probabilities of the introduction text of the electronic product and the introduction text of the CBB model in the electronic equipment variant dataset, and the Chinese-BERT-wwm model is trained. S5.

4. After the training is completed, a trained Chinese-BERT-wwm model is obtained, and the trained Chinese-BERT-wwm model is used to judge the correctness of the electronic product architecture data generated by the Tongyi Qianwen model, that is, the relevance between the electronic product and the text features of its subordinate CBB model.

7. The electronic equipment variant design method based on large-scale language model according to claim 6, characterized in that: The specific implementation method of step S5.3 includes the following steps: S5.3.

1. The Chinese-BERT-wwm model is trained using the Lora method. The expression is as follows: W0+W'=W0+AB T Among them, W0 represents the original parameters of the model, which is an n×d matrix and is not updated during the training process. W’ is the change in parameters during the training process, which has the same size as W0. A and B represent the matrix sets of Lora fine-tuning parameters. A is an n×l matrix and B is a d×l matrix. The positive integers l, d, and n represent the dimensions of the parameter matrices, and l << d, n. The computational amount during the training process is reduced through matrix factorization. S5.3.

2. The process of fine-tuning the Chinese-BERT-wwm model is expressed in formula form as follows: Wlora i =Wuse i +A i B i T p i,j =CBERT(Wlora U Wnotuse,(eproducttext i ,[SEP],cbbtext j )) p i,j =sigmoid(Linear(p i,j )) Among them, CBERT is the abbreviation of Chinese-BERT-wwm model, Wuse i , Wnotuse represent the parameter sets of the Chinese-BERT-wwm model that participate in fine-tuning and those that do not participate in fine-tuning, respectively. i is the result after parameter fine-tuning. U represents the set union operation, Linear represents the linear layer that transforms the dimensionality of the BERT model output, and p i,j Represents the matching probability between electronic product i and electronic equipment CBB model j.

8. The electronic equipment variant design method based on large-scale language model according to claim 7, characterized in that: In step S6, for the electronic equipment CBB model with a matching probability less than 0.5 between the electronic product and the electronic equipment CBB model, it is judged as an irrelevant electronic equipment CBB model and filtered and deleted.

Images