CN-122019559-A - Electronic component library management method based on AI

CN122019559ACN 122019559 ACN122019559 ACN 122019559ACN-122019559-A

Abstract

The invention relates to the technical field of artificial intelligence, and provides an AI-based electronic component library management method, which can carry out continuous mask language model training on a BERT layer in BERT-BiLSTM-CRF, reduce the difficulty of subsequent training while enhancing the data adaptability of the field, improve the context semantic understanding capability, carry out multi-granularity feature fusion based on a multi-granularity priori knowledge fusion attention mechanism, enable a model to be more comprehensive and focus on key features, improve the model accuracy, utilize an electronic component entity extraction model to carry out entity extraction on a target material description text, get rid of the dependence on fixed separators and hard coding rules, process the material description text in any format, solve the problem of fragmented and term-intensive adaptability of component description, and update library field increment to an electronic component library, thereby realizing the automatic and efficient management of the electronic component library.

Inventors

LIU HAODONG

Assignees

深圳市多酷科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260410

Claims (10)

1. The AI-based electronic component library management method is characterized by comprising the following steps: constructing a character level training sample and a tag vocabulary according to the historical material description text, and packaging the character level training sample and the tag vocabulary into a training example by using a BERT word segmentation device; Using BERT-BiLSTM-CRF deep learning architecture as an infrastructure, and using enterprise history component description data to carry out continuous mask language model training on a BERT layer in the infrastructure to obtain a pre-training model; performing multi-granularity feature fusion on the training examples based on a multi-granularity priori knowledge fusion attention mechanism to obtain input features; Performing iterative loop training on the pre-training model by using the input features to obtain an electronic component entity extraction model; Responding to a batch update instruction of an electronic component library triggered based on a target material description text, and performing entity extraction on the target material description text by using the electronic component entity extraction model to obtain a structured parameter list; invoking a predefined format template according to the material type included in the target material description text, and generating a library field according to the predefined format template and the structural parameter list; and incrementally updating the library field to the electronic component library.
2. The AI-based electronic component library management method of claim 1, wherein constructing a character-level training sample and tag vocabulary from historical material description text comprises: converting the historical material description text into a character level list to obtain the character level training sample; Generating an initial tag list in BIO format for the character-level training sample; traversing the initial tag list to obtain all unique tags; And adding preset labels at preset positions in all the unique labels, and numbering the rest labels from 1 according to alphabetical order to obtain the label vocabulary with a label-digital identification mapping dictionary and a digital identification-label mapping dictionary.
3. The AI-based electronic component library management method of claim 2, wherein the encapsulating the character-level training samples and the tag vocabulary into training examples with a BERT segmenter comprises: The BERT word segmentation device is utilized to segment the character level training sample as a whole sentence to obtain an input identification sequence, an attention mask sequence, a sentence type identification sequence and a word index sequence, wherein in the word segmentation process, each split word is marked by a word splitting identification; Traversing the word index sequence, and distributing label identifications for each traversed sub word to obtain a label pair Ji Suoyin sequence, wherein for each traversed current word index, when the current word index is null, the label identification of the current word index is configured as the preset label, and when the current word index is different from the traversed last word index, the label identification of the current word index is configured as the corresponding unique label; Filling and cutting the input identification sequence, the attention mask sequence, the sentence type identification sequence and the tag alignment index sequence according to a preset field length to obtain an intermediate sequence; converting the attention mask sequence obtained after the processing into a Boolean mask to obtain a Boolean type attention mask sequence; and integrating the intermediate sequence and the Boolean attention mask sequence to obtain the training example.
4. The AI-based electronic component library management method as recited in claim 3, wherein the infrastructure comprises a BERT layer, a random inactivation layer, a bidirectional long and short time memory network layer, a full connection layer and a conditional random field layer, wherein the performing continuous mask language model training on the BERT layer in the infrastructure by using enterprise historical component description data with the BERT-BiLSTM-CRF deep learning architecture as the infrastructure comprises: constructing a domain term sample set according to the enterprise historical component description data; loading only the BERT layer, and performing domain adaptive pre-training on the BERT layer by utilizing the domain term sample set; And accessing the BERT layer obtained after training into the random inactivation layer, the bidirectional long short-time memory network layer, the full-connection layer and the conditional random field layer to obtain the pre-training model.
5. The AI-based electronic component library management method of claim 1, wherein the multi-granularity feature fusion is performed on the training examples based on a multi-granularity prior knowledge fusion attention mechanism, and obtaining input features comprises: Inputting the historical material description text to the BERT layer to obtain a BERT word vector; Extracting the characteristics of the character-level training sample through a character-level characteristic extraction model to obtain character-level characteristics; splicing the BERT word vector and the character-level feature according to the subwords to obtain an intermediate fusion feature; Acquiring a domain dictionary, calculating the attention weight of the intermediate fusion feature by using the domain dictionary, and generating the attention feature of the fusion domain knowledge according to the attention weight; and splicing the intermediate fusion feature with the attention feature to obtain the input feature.
6. The AI-based electronic component library management method of claim 4, wherein the entity extracting the target material description text using the electronic component entity extraction model to obtain a structured parameter list comprises: processing the target material description text based on the multi-granularity priori knowledge fusion attention mechanism to obtain initial characteristics; inputting the initial features to the bidirectional long-short-term memory network layer to extract long-distance context-dependent features; mapping the long-distance context dependent features by using the full connection layer to obtain the emission score of the label corresponding to each position; Inputting the emission fraction of the corresponding tag at each position to the conditional random field layer to obtain a BIO format tag sequence; Restoring each sub word level tag in the BIO format tag sequence into an original word level tag, and merging continuous tags of the same type according to BIO rules to extract an entity fragment; calculating a confidence score of each entity fragment; Filtering the entity fragments according to the entity types and the confidence scores of the entity fragments to obtain a plurality of candidate entities; Acquiring a confidence threshold; And processing the plurality of candidate entities according to the confidence threshold to obtain the structured parameter list.
7. The AI-based electronic component library management method of claim 6, wherein processing the plurality of candidate entities according to the confidence threshold to obtain the structured parameter list comprises: for each candidate entity of the plurality of candidate entities, converting the candidate entity into a structured parameter and then adding to the list of structured parameters when the confidence score of the candidate entity is greater than the confidence threshold, or And when the confidence score of the candidate entity is smaller than or equal to the confidence threshold, marking the candidate entity as a state to be confirmed, and pushing the marked candidate entity to a queue to be processed of a manual input interface.
8. The AI-based electronic component library management method of claim 1, wherein the incrementally updating the library field to the electronic component library comprises: Loading all library files of the electronic component library; traversing each row of original fields of the library file one by one according to the material number corresponding to the library field; For each library field traversed to the corresponding material number, when the library field is different from the corresponding original field, replacing the corresponding original field with the library field, wherein when the library field is a printed circuit board packaging field, multi-selection packaging is executed according to the material type corresponding to the library field; for each bin field that is not traversed to the corresponding bin number, the bin field is added to the last row of the bin file.
9. The AI-based electronic component library management method of claim 1, further comprising: Responding to the temporary material number generation instruction, and inquiring the maximum material number in the electronic component library; acquiring a temporary material number identifier and a preset step length; calculating the sum of the maximum material number and the preset step length to obtain a target value; Splicing the temporary material number identifier with the target value to obtain a temporary material number; And adding the temporary material number to the electronic component library.
10. The AI-based electronic component library management method of claim 1, further comprising: And in response to a priority updating instruction triggered based on the priority file, emptying column data corresponding to the priority in the electronic component library, and adding the priority correspondence in the priority file to a column corresponding to the priority.

Description

Electronic component library management method based on AI Technical Field The invention relates to the technical field of artificial intelligence, in particular to an electronic component library management method based on AI. Background In the development process of electronic products, a hardware engineer needs to maintain a huge library of electronic components, usually in the form of an Excel file or a dedicated database, and the library contains information such as material numbers, descriptions, specification models, PCB (Printed Circuit Board ) packages, schematic diagram packages, device values, data manual paths, procedures, material types, priorities, unit prices and the like. Moreover, this information is distributed among enterprise resource planning systems (EnterpriseResourcePlanning, ERP), product lifecycle management systems (ProductLife-CYCLEMANAGEMENT, PLM), and procurement systems (e.g., material unit price, supplier priority), which are not well-integrated with the multi-source data of existing solutions. When an enterprise introduces new materials or changes the material information, engineers still need to manually update the component library. In order to solve this problem, the electronic component library management scheme adopted in the prior art still needs to rely on fixed separators or hard coding rules for parameter extraction in terms of material description analysis. The method has extremely strong dependence on the description format, unstructured natural language description cannot be processed, and once suppliers or internal personnel change description habits, analysis rules can be immediately disabled, and manual continuous maintenance and updating are still needed. In addition, the existing analysis rules are static, a new description mode cannot be learned from historical data, and dynamic changes of enterprise material descriptions cannot be adapted. The exception handling mechanism of the existing scheme is not perfect, and for descriptions which cannot be resolved, the existing scheme usually reports errors or skips directly, and lacks intelligent shunt and manual feedback closed loops. In view of this, how to effectively and automatically manage the electronic component library is a problem to be solved. Disclosure of Invention In view of the foregoing, it is desirable to provide an AI-based electronic component library management method, which aims to solve the problem that effective automated management of an electronic component library is not possible. An AI-based electronic component library management method, the AI-based electronic component library management method comprising: constructing a character level training sample and a tag vocabulary according to the historical material description text, and packaging the character level training sample and the tag vocabulary into a training example by using a BERT word segmentation device; Using BERT-BiLSTM-CRF deep learning architecture as an infrastructure, and using enterprise history component description data to carry out continuous mask language model training on a BERT layer in the infrastructure to obtain a pre-training model; performing multi-granularity feature fusion on the training examples based on a multi-granularity priori knowledge fusion attention mechanism to obtain input features; Performing iterative loop training on the pre-training model by using the input features to obtain an electronic component entity extraction model; Responding to a batch update instruction of an electronic component library triggered based on a target material description text, and performing entity extraction on the target material description text by using the electronic component entity extraction model to obtain a structured parameter list; invoking a predefined format template according to the material type included in the target material description text, and generating a library field according to the predefined format template and the structural parameter list; and incrementally updating the library field to the electronic component library. An AI-based electronic component library management apparatus, the AI-based electronic component library management apparatus comprising: The construction unit is used for constructing a character level training sample and a tag vocabulary according to the historical material description text, and packaging the character level training sample and the tag vocabulary into a training example by utilizing a BERT word segmentation device; The training unit is used for using the BERT-BiLSTM-CRF deep learning architecture as an infrastructure and utilizing the enterprise history component description data to carry out continuous mask language model training on the BERT layer in the infrastructure so as to obtain a pre-training model; the fusion unit is used for carrying out multi-granularity feature fusion on the training examples based on a multi-granularity priori knowledge f