Short-term trend of generative AI large model:

Expanding new scenarios does not replace small models

FIG. 3 Application of generative AI large model in manufacturing industry

Generative AI large model capabilities cover the generation of structured data, text, images, audio and video and other fields, but the exploration in the manufacturing field is still focused on the processing and generation of structured data, natural language and image data. The formation of this situation is mainly because there is no basic model of audio and video field with strong ability, so there is no industrial case related to small model field such as equipment diagnosis based on voice print analysis and safety production based on video analysis. Generative AI exploration also covers the full life cycle of manufacturing research and development design and planning, production process control, operation management optimization, and product service optimization.

In the R&D, design and planning stage, on the one hand, the interaction ability of natural language is used to realize the expansion of CAD software functions. For example, Back2CAD launched CADGPT™ based on the support of Elaine CAD Bot, ChatGPT and Amazon AWS. Supports intelligent recommendation, document generation, code production and other functions. On the other hand, design efficiency is improved based on image data generation ability. For example, Haier Design works with Amazon Cloud Technology and partner Nolibox to create AIGC solution, which introduces AIGC image generation ability into product design, UI design, CMF design, brand design and other links. It covers the business scenarios of industrial design such as new product design, revamping and upgrading, and channel customization.

In the manufacturing process, the ability around knowledge question answering and code generation has become an important research hotspot. For example, Siemens and Microsoft are also collaborating on code generation tools for programmable logic controllers (PLCS), and ChatGPT is used to generate PLC code from natural language input. Authentise introduces 3DGPT for additive manufacturing technology Q&A by fine-tuning a general purpose large language model using 12,000 scientific additive manufacturing papers. Users can get answers to professional questions such as “How to reduce the possibility of defects when using powder stainless steel”. For example, Innovation Qizhi launched AInno-15B industrial large model, which supports generative AI applications through large model service engine, and realizes industrial robot control, enterprise private domain data analysis, enterprise private domain knowledge base and other applications.

SprutCAM X builds CAM virtual assistants with ChatGPT apis that enable engineers to operate machine tools, such as drilling a 10mm diameter hole at points (100, 25) “and the AI assistant executes code to generate the corresponding CAM. C3iot also builds Generative AI services for multiple industries and fields based on a large language model, and provides equipment operation and maintenance services for a large manufacturing enterprise based on generative AI. With the help of C3 Generative AI, operators can diagnose the root causes of equipment failures using simplified workflows. When an operator finds a production issue, it can go straight to the C3 Generative AI and search through troubleshooting guides and textbooks to identify potential causes.

Figure 4 Architecture diagram of C3IOT generated AI service

In the operation and management process, the ability to add intelligent question-and-answer and data analysis based on the large language model has become the mainstream. For example, in the field of ERP, Yonyou builds yongpt on the base of large models such as ChatGPT, Wenxinyi and Llama. In the infrastructure of the large model, the deterministic matters are returned to the original product functions of Yonyou BIP. The development of uncertain matters, inferential matters and matters assigned by human mind consciousness to large models can support business insight, intelligent order generation, supplier risk control, dynamic inventory optimization and other applications. In the field of CRM, Salesforce and Microsoft have strengthened the integration and application of generative AI in their products.

Figure 5. yongpt architecture of UF

In the process of product service optimization, the ability to integrate large models into products has become the focus of exploration to improve the intelligent ability of products in consumer electronics, automobiles and other fields. For example, the smart speaker Vifa ChatMini launched by Guogang Electric has built-in ChatGPT and Wenxin Yiyi dual models. On the basis of maintaining professional acoustic standards, compared with traditional smart speakers, Vifa ChatMini has significant advantages in natural language generation and emotional expression. It can be applied to specific user groups such as the elderly and children for emotional support and intelligent learning companionship, and can also be used as an intelligent assistant in daily work and planning.

In summary, at present, the exploration path of generative AI large model in the manufacturing industry is initially presented as three paths:

The first is to improve efficiency by directly integrating general capabilities such as question answering and code generation of basic large models. For example, Haier, Siemens, etc. CAD, PLC code generation; Salesforce, Microsoft, ABB, UF and other software in CRM, ERP, production management and other access to large models, improve the professional software data analysis, document management, knowledge and other auxiliary capabilities.

The second is to achieve scene innovation and add new functions by focusing on the field through fine tuning and external knowledge base. For example, Authentise introduced 3DGPT for additive manufacturing technology Q&A by using the fine-tuning of 12,000 scientific additive manufacturing papers to fine-tune a general purpose large language model.

The third is to build a large industrial model from pre-training. For example, the innovative AInno-15B industrial large model distills part of knowledge from Llama 2, Falcon, Bloom and other open source large models, and then combines the parameter structure designed by itself and the accumulated industrial knowledge data for training. Through the three training steps of Pretrain, SFT and RLHF, the model can get more understanding of industry, support question and answer interaction and more standard answers.