Introduction
Artificial Intelligence is a strategic technology leading this round of technological revolution and industrial transformation, exhibiting a “leading goose” effect, with strong spillover and driving force. The general artificial intelligence large model (hereinafter referred to as “large model”) represents a new stage in the development of artificial intelligence from specialization to generalization. It integrates functions such as intelligent perception, intelligent analysis, intelligent decision-making, and intelligent execution, achieving the optimization of production factor allocation through the deep integration of data, computing power, and algorithms.
From the perspective of production structure, artificial intelligence technology has become a core component of modern industrial production, significantly improving production efficiency, optimizing resource allocation, and reducing production costs, thus providing strong technical support and impetus for new industrialization development. The deep integration of general artificial intelligence with manufacturing can further accelerate the industrial system’s advancement towards high-end, intelligent, and green development.
1. Overview of Large Models and Their Industrial Application Development
On one hand, there is a global upsurge in the development of large models. The emergence of large models has pushed artificial intelligence into a new stage of development, with leading domestic and foreign enterprises becoming the technological high ground for large model competition. AI large models are artificial intelligence models with large-scale parameters and complex structures. With the increase in model size, the volume of training data, and improvements in computing power, AI large models have achieved significant breakthroughs in natural language processing, image recognition, speech recognition, and multimodal recognition. Since 2020, the large model market has rapidly grown globally, entering an explosive period.
In terms of international progress, OpenAI, as a leading institution in the industry, has released language large models ChatGPT (2022) and GPT-4 (2023), speech large model Whisper (2022), and visual large model DALL-E (2021). Microsoft has integrated GPT-4 capabilities into key products such as Windows 11, Office365, and Bing, forming the Copilot series of applications. At the same time, Google launched the multimodal large model Gemini (2023), and Meta released the language large model LLaMA (2023), with foreign tech giants joining the large model competition.
Domestically, various tech companies are actively following the global trend of large model development. Baidu has released the language large model “Wenxin Yiyan”, Alibaba has launched the language large model “Tongyi Qianwen”, iFlytek has released the language large model “Xinghuo Cognition”, Baichuan Intelligence has launched “Baichuan Large Model”, and Zhiyun AI has released the ChatGLM series of language large models, while the Chinese Academy of Sciences has introduced the cross-modal large model “Zidong Taichu”.
On the other hand, the application of large models to the B-end, especially in the industrial sector, has become an industry consensus. Large models have shown a trend of development characterized by a foundational large model as the technical base and industrial applications as the entry point, leading to the emergence of the industrial large model concept. The foundational model enhances the parameter quantity and structural generality, integrating and expressing more domain knowledge and modal knowledge, forming an all-knowing, all-capable general model.
The industrial large model relies on the structure and knowledge of the foundational model, integrating data and expert experience from specific industrial sectors to form vertical, scenario-specific, and specialized application models. Compared to foundational models, industrial large models have advantages such as fewer parameters, higher specialization, and stronger feasibility, offering low-cost solutions for technological breakthroughs, product innovations, and production transformations in vertical industrial fields.
2. Seven Major Models of Industrial Applications of Large Models
From the perspective of the entire lifecycle of industrial products, it can be divided into stages such as research and development, design, simulation, production, testing, operation and maintenance, and after-sales. This report analyzes the characteristics and elements of each stage and preliminarily proposes the following potential application models of large models in the industrial field, combined with the capabilities of large models themselves.
(1) Principle-Based R&D Large Model
This model can analyze product structural data to explore the configuration and mechanisms of products at a microscopic level and generate new structures and characteristics of products through the model’s emergent capabilities. For example, in drug development, AI large models can analyze a large amount of known drug molecular data to identify the optimal drug candidates and generate a new drug molecule design scheme, significantly shortening the time and cost of drug development and increasing the success rate of drug development.
Target discovery and drug structure design. Target discovery is a core link in the drug development process, where targets are the binding sites of drugs in the body. The relationship between drugs and targets can be likened to that of a key and its corresponding lock. Traditional target discovery requires extensive scientific work, making numerous biological hypotheses about targets, and designing a series of experiments for verification, which takes a long time. Large models can analyze known drug molecular structure data and the knowledge graph of the correlation between drug molecules and diseases to identify molecular features that interact with disease targets, and then automatically generate new drug molecule design schemes using these molecular features. Drug design scheme evaluation and optimization. After generating several drug molecule design schemes, the model can evaluate the generated designs under human prompts, filtering out the most promising drug candidates. Additionally, it can optimize drug efficacy and toxicity by predicting drug metabolism pathways and drug concentrations, thus optimizing drug dosing and administration plans.
Assisting clinical trial design. Experimental verification is an indispensable part of drug development, but experiments often consume a lot of time and resources. AI large models can help researchers design more effective clinical trial plans, such as predicting drug safety and efficacy, optimizing clinical trial sample sizes and timelines, etc. By predicting and filtering, the number and complexity of experiments can be reduced, thereby improving the efficiency and quality of drug development.
(2) Prospective Design Large Model
This model can generate innovative product design schemes, better assisting technicians in quickly transforming design ideas and intentions into specific implementation plans. For example, in traditional industrial design, large models can achieve rapid generation of engineering drawings and design schemes, and assist in layout optimization and parameter verification, greatly reducing the time consumed in industrial design and enhancing product development efficiency.
Mathematical modeling and code writing. In the industrial design process, many problems require the establishment of specific mathematical models for analysis and solution, such as mechanical product design frequently encountering issues like strength verification, size optimization, and standard part selection, which can be costly when inviting experts for analysis and modeling. Large models can provide reference solutions for these issues in a short time based on specific design requirements and automatically generate corresponding program codes to guide specific industrial design practices. Diversified design scheme generation. For example, in CAD, existing massive standardized material libraries provide a wealth of engineering drawing and layout planning data. Large models can utilize this data, combined with the designer’s creative ideas and specific requirements, to generate diversified design schemes for designers’ reference. They can also quickly optimize and adjust design schemes, helping engineers create layouts faster and with fewer errors.
(3) Efficient Simulation Large Model
This model can utilize its generative capabilities to provide virtual simulation test scenarios or environments that meet design requirements, addressing issues such as insufficient test data and singular testing environments in industrial product design, thus enhancing product reliability. For instance, in automotive manufacturing, AI large models can generate simulation testing environments for crash simulations and safety assessments, significantly improving simulation accuracy and efficiency, promoting advancements in vehicle design and safety performance.
Diverse simulation scenario generation. Automotive companies accumulate a large amount of real data during actual vehicle testing and simulation experiments, including vehicle structures, collision data, and material properties. Large models can learn the nonlinear relationships between vehicle structures, material properties, and collision responses using this data, understanding the impact of different parameters (such as collision speed, angle, vehicle construction, etc.) on collision responses, and generating new and diverse collision scenarios through their emergent capabilities. This allows vehicle models to perform simulation predictions under various conditions, including previously unencountered scenarios, and compensates for data gaps in special scenarios, enhancing the comprehensiveness and accuracy of automotive simulation testing.
Parameter optimization and rapid prediction. In traditional vehicle collision simulation, different parameter combinations must be tested multiple times to obtain the best results, with each test requiring significant time for manual parameter adjustment. Large models can leverage zero-shot knowledge analysis capabilities to quickly predict collision responses for different parameter combinations through question-and-answer formats. This helps find the optimal parameter combination in a short time, reducing the cycle of vehicle design and testing. Additionally, it can provide innovative design suggestions based on historical collision cases, potentially involving material selection, structural adjustments, and other aspects.
(4) Refined Detection
This model can leverage the zero-shot learning capabilities of large models, combined with AR/VR and other virtual reality technologies, to achieve rapid and efficient visual detection of various industrial scenarios such as product quality defects, personnel violations, and assembly errors. For example, in quality inspection and safety monitoring in industrial production, large models can quickly detect abnormal information in specified areas or among personnel based on external visual sensing devices and simple instructions, significantly reducing costs related to manual inspection, sample collection, and model training.
High-efficiency industrial quality inspection. Taking PCB defect detection as an example, general visual large models can leverage strong generalization capabilities to perform semantic segmentation directly on raw PCB images without relying on factory sample data and localized fine-tuning training, and implement the identification of various defects such as lack of solder, solder bridging, and pinholes by combining simple rules and methods. This addresses practical problems of difficulty in obtaining PCB sample data and labeling, avoiding high costs associated with training and parameter tuning, and enhancing the efficiency of industrial product defect detection.
Intelligent safety production supervision. In the coal mining industry, leveraging the machine vision recognition technology of large models, combined with equipment operational status data, can identify dangerous scenarios such as personnel entering hazardous areas, personnel falling, cutting ground, roof bolting, drilling depth, anchor agent usage, stirring time, secondary fastening, and anchor cable tensioning, transitioning from manual supervision of tunneling operations to automated monitoring, thereby improving the standardization of tunneling operations and enhancing the safety factor of coal mining production.
Personalized detection scenario expansion. By combining language large models and visual large models for multimodal perception and interaction, the application scope of large models in industrial visual inspection can be expanded, improving application flexibility. For example, in industrial quality inspection, voice commands can control large models to conduct defect detection across different types, regions, and levels, meeting the inspection needs of various products.
(5) Intelligent Regulation
In large modern production lines, intelligent scheduling and control of multiple key nodes are required to enhance production line operational efficiency. AI large models can analyze diverse historical data to better understand the complex relationships in industrial scheduling tasks, such as production demand, resource availability, and task priority, thereby optimizing task allocation and scheduling at each node, enhancing production efficiency and flexibility. For instance, in industrial robotics, large models can automatically integrate and analyze various production data, enabling rapid task assignment and dynamic task adjustments for robots, becoming the “nerve center” of large industrial production lines.
Complex pattern learning and rapid task allocation. In large-scale production lines such as automotive manufacturing, multiple industrial robots are typically configured to handle sub-tasks like assembly, welding, and painting. AI large models can collect performance data of robots, workstation statuses, and production plans, learning complex information such as robot skills, task complexity, and transfer times between workstations, and predict the efficiency of different robots executing different tasks. When new tasks arrive, the model can quickly decide which robot to allocate to minimize task waiting times and production cycles.
Dynamic task allocation adjustment. Large models can analyze robot performance and efficiency from historical data and dynamically adjust task allocation strategies based on different task requirements. This helps optimize robot utilization, ensuring they perform at their best across various tasks. If issues arise such as robot failures, workstation malfunctions, or changes in production plans, large models can quickly respond and readjust task allocations to cope with unforeseen circumstances.
Motion control code generation. From the perspective of individual robot movements, production personnel can interact via text, voice, etc., to rapidly generate customized motion control codes through large models based on different task requirements. For example, inputting the instruction “Please write a PLC program to control the robot to transfer parts from point A to point B” into the large model. This model of generating motion control instructions based on large models can significantly enhance the flexibility of industrial robots, enabling flexible production line control.
(6) Scientific Operation and Maintenance Large Model
This model can leverage its powerful reasoning capabilities to analyze and predict various data during the production process, thereby enhancing intelligent operation and maintenance levels and improving production management mechanisms. For example, in warehouse management, large models can manage and integrate various categories and modalities of data within the supply chain, enhancing the operational efficiency of product information flow and creating a more convenient and faster industrial product supply chain system.
Intelligent warehouse management. Using robots embedded with multimodal large models for shelf management, inventory management, and order picking can enable autonomous shelf positioning, inventory management, and item transportation through the strong visual generalization capabilities of large models, improving warehouse operational efficiency. At the same time, large models can predict inventory levels and develop restocking strategies based on sales velocity and inventory turnover rates, helping to timely replenish stock, avoid stock shortages that affect sales, and prevent overstocking that could lead to inventory backlog and capital occupation issues.
Efficient data management. The supply chain of industrial production involves a vast amount of data, such as raw material data, processing technology data, equipment status information, environmental information, personnel information, logistics information, etc. Large models can quickly organize, classify, and analyze data from different formats and sources, reducing data organization costs. Furthermore, this classified and organized data can be used for further fine-tuning of the large model, achieving a beneficial interaction between data and models.
(7) Customized After-Sales Large Model
This model can leverage its significant advantages in natural language dialogue to make after-sales service no longer confined to fixed Q&A databases, enabling more natural, smooth, and effective conversations with customers, thereby helping industrial enterprises achieve customized after-sales services that meet diverse user needs, further enhancing customer loyalty and user growth, and expanding business scope. For example, in mechanical equipment after-sales, large models can combine multimodal and digital human technologies, allowing customers to describe equipment faults or issues in natural language, with the system accurately understanding and providing detailed, personalized solutions.
Multidimensional interactive after-sales service. The operation and maintenance of mechanical equipment require complex operations, and traditional knowledge-based customer service systems cannot provide precise guidance to users. Utilizing large language models as backend logical reasoning support and virtual digital humans as frontend interaction figures, the system can accurately understand customer needs, rapidly provide detailed and targeted solutions based on its knowledge reserves and specific issues. Additionally, it can assist customers in operating equipment through the gestures and voice interactions of virtual digital humans, enhancing the efficiency of after-sales service while providing a more intuitive and personalized service experience.
3. Outlook for Industrial Large Models
At present, due to challenges such as fragmented industrial scenarios, insufficient computational resources, difficulties in collecting and organizing training data in the industrial field, and issues regarding the safety and reliability of large models, the integration of large models with industry in China is still in its initial exploratory stage and faces certain challenges.
Firstly, foundational large models still dominate the application market, and have not yet penetrated into vertical industrial fields to form specialized industrial large models. Secondly, the current application of large models in industrial production is relatively scattered, and has not yet formed a standardized, systematic industrial application paradigm for large models. Finally, constructing industrial pre-trained large models from the ground up has a high threshold, with only a few leading enterprises capable of conducting R&D on industrial large models.
In response to the above issues, the following suggestions are made.
Advance large model technology research focused on industrial scenarios. Identify common technical issues for large models aimed at industrial applications, focusing on safety, reliability, and real-time performance, and encourage collaboration among universities, enterprises, and research institutions in specialized fields to tackle these challenges. This can be achieved through the expansion of large model industrial datasets, the construction of typical industrial scenario rule sets, and the optimization of model training algorithms, strengthening the technological R&D of domestic large models and enhancing the industrial application capabilities of large models.
Build a large-scale industrial data resource pool for large models. Organize enterprises and research institutions on the supply and demand sides of large models to develop training data standards and testing standards for industrial large models. Relying on national industrial internet big data centers and other standardized platforms, establish an industrial corpus for large models, forming a management mechanism for industrial data. Through financial subsidies, tax reductions, and policy incentives, guide domestic medium and large manufacturing enterprises to share operational data in industrial production, creating a data resource pool covering key domestic industrial sectors to provide data support for the training and testing of industrial large models.
Improve the performance evaluation mechanism for large models in the industrial field. Based on national authoritative institutions, collaborate with various industry demand sides to establish standardized testing sets for large model industrial knowledge Q&A, ensuring evaluation efficiency and result credibility. At the same time, a long-term performance evaluation mechanism for large models in the industrial field should be established, periodically evaluating key performance indicators such as knowledge capability, stability, and safety of large models, and dynamically adjusting evaluation metrics based on changes in industrial structure and data element distribution to promote the continuous empowerment of large models for new industrialization.
Promote pilot demonstrations of typical applications of large models in the industrial field. Integrate fragmented industrial scenarios, distilling typical business scenarios represented by product-assisted design, refined quality detection, and intelligent supply chain management, clarifying the quantitative demand indicators for large models in each scenario, and promoting the establishment of relevant industry standards. Establish a bidirectional interactive mechanism between the supply side of large models and the enterprise application side, facilitating the formation of several characteristic industrial clusters for the collaborative development of large model R&D and manufacturing industries, and promoting the construction of benchmark and demonstrative applications of industrial large models.
Author: General Artificial Intelligence and Industrial Integration Innovation Center, China Industrial Internet Research Institute
Source: New Industrial Network
Content sourced from public channels on the internet; this article is for reference only (for information exchange purposes), and all information should be based on official materials published through official channels.Copyright belongs to the original author or publisher. If there is any infringement, please contact for deletion.
Related Articles
“Implementation Plan for Accelerating the Development of the Artificial Intelligence Industry in Shaanxi Province (2024-2026)” Issued
Academician Li Huajun et al.: Status, Challenges, and Development Suggestions for China’s Deep-Sea Underwater Technology Equipment Development
New Version of the “National Science and Technology Award Regulations” Announced
