Manufacturers face difficulties in aligning product specification documents with Bills of Materials, particularly after cross-border mergers and acquisitions. A new framework proposed by researchers Yu-Chi Lin and Jr-Fong Dang and published on Science Direct hopes to refine this complex process.
There are several problems commonly encountered by manufacturing and engineering firms in this context, including the fact that part numbering systems can differ between an organisation’s existing entities, making organisations need to optimise ‘on the fly’ as the acquisition’s product details are added and digitised. Acquired factories may rely on informal or poorly-documented codes, or may have wildly-different naming and cataloguing schema.
Inconsistencies of any type affect Bills of Materials (documents that link design data, procurement records, and production planning covering a product’s full lifecycle). When part numbers and definitions do not conform to shared rules, errors propagate through multi-level assemblies and can cause data entry errors in systems that are designed, ironically, to unify processes. The process of integrating disparate systems is therefore slow and costly.
Conventional Bill of Materials creation depends on manual interpretation of specification documents such as data sheets and compliance certificates. These are often unstructured and multilingual. Engineers review such paperwork line by line, extract what they believe to be relevant attributes, and assign part numbers according internal rules. Data is then added into the controlling organisation’s Product Lifecycle Management (PLM) or ERP systems.
The process typically involves the segmentation of documents into classes, manual copy and paste extraction, field validation, and finally, system entry. Errors are made and can be caught too late, affecting prototyping or, in the worst case scenario, production. Duplicate part numbers, inconsistent naming, and missing attributes are commonplace, and usually, a good deal of reworking is required before acceptable efficiency levels are attained.
The framework proposed by Yu-Chi Lin and Jr-Fong Dang introduces a generative AI pipeline designed to replace manual workflows with a controlled process with an large language model at its core. It integrates an LLM with RAG processes, with the aim of converting unstructured spec docs into semantically-consistent part numbers and, therefore, structured Bill of Materials records. The proposed design reflects typical governance requirements, where random text generation is clearly unacceptable.
The proposed pipeline operates in stages. Firstly, multi-format document parsing would extract text and tabular content from specification documents. The framework described uses a multimodal Qwen model, in the researchers’ case to process bilingual Chinese and English materials. The model is capable of identifying relevant attributes like electrical ratings, dimensions, compliance standards, and supplier.
Next, RAG ensures the model’s outputs are tuned to local domain knowledge. It retrieves the relevant corporate standards and part records from the controlling organisation at the point of inference. This reduces the risk of hallucinated or otherwise inconsistent attributes.
Third, schema validation governs the generation of part numbers and Bill of Materials entries. Outputs are structured in JSON where foreign-key relationships may be defined.
The validation steps help ensure generated part numbers conform to predefined coding logic and that references to suppliers, categories, and assemblies match existing master data.
Importantly, if validation fails, the system prompts for human intervention before persisting its data. Human-in-the-loop verification is still required before final approval records, which helps traceability and accountability. Human review, it is to be noted, is selective rather than comprehensive, focusing on cases flagged by the model where its confidence is low or if new component categories emerge.
Once validated, output is placed in a digital bill of materials, termed an x-EDBOM, which acts as the canonical dataset. This can then be parsed by other systems such as the relevant ERP. Data synchronisation in this manner reduces discrepancies between the activities of engineering, procurement, and production functions.
The x-EDBOM also enables comparisons of seemingly-identical components from different suppliers, with automated analysis flagging similarities in price, lead time, and technical specifications, for example. Procurement teams can therefore assess alternative choices (or potential duplicates) from the same data.
Empirical validation of the research was conducted at a “mid-sized electronics manufacturer”. In a proof-of-concept evaluation using ten-fold cross-validation, the researchers said their framework reduced average processing time per document from 25 minutes to 2 minutes, a 12.5x improvement, at an average accuracy of 91%. Estimated monthly cost savings were calculated on the basis of labour costs. The figures, therefore, indicate efficiency gains, although, to coin a phrase, your mileage may vary, according to the complexity of the existing supply chain, cost of labour, number of additional component sources, etc.
(Image source: “Component Parts for a Custom Slitting Saw Holder” by tudedude is licensed under CC BY-NC-SA 2.0.)
Want to experience the full spectrum of enterprise technology innovation? Join TechEx in Amsterdam, California, and London. Covering AI, Big Data, Cyber Security, IoT, Digital Transformation, Intelligent Automation, Edge Computing, and Data Centres, TechEx brings together global leaders to share real-world use cases and in-depth insights. Click here for more information.
TechHQ is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
