A Unified Industrial Large Knowledge Model Framework in Industry 4.0 and Smart Manufacturing

The initial step in constructing an ILKM involves the creation of an LKL. This library is pivotal for accommodating the breadth and diversity of industrial data, thus serving as a foundational resource for subsequent analytical tasks. During this phase, it is essential to categorize the data into domain-specific categories systematically. Such an organization enables researchers and data scientists to streamline their efforts, allowing for efficient retrieval of domain-specific data to inform the development of ML models tailored to address distinct industry-related challenges. Within these categories, based on the usage and nature of the data, industrial data can be divided into two primary types: human-interpretable and structured machine-generated. Human-interpretable data, inherently designed for human cognition, comprise elements such as text documents, annotated images, coding scripts crafted by programmers, and multimedia content. This type of data can be seen as insightful information or knowledge and is used for the later development of domain-specific knowledge ML models. On the other hand, structured machine-generated data comprises sensor readings, machine logs, operational parameters, and more. This data type can be leveraged for analytical purposes in technical and industrial contexts.

In the second step of the ILKM framework, the focus shifts to transforming domain-specific data (human-interpretable data from LKL) into structured domain instruction sets. This transformation is crucial for enhancing the performance of LLMs in targeted domains by generating domain-centric knowledge and achieving multi-modal data fusion [8]. These structured instructions, vital for finetuning the LLM and retrieving domain knowledge, ensure that the model is proficient in addressing domain-specific challenges and enhancing problem-solving capabilities [25, 5]. As depicted in Figure 2, the domain instruction data are organized into three parts: first, the domain instruction, which identifies the problem’s domain and may include sub-tags for refined categorization; second, the input, which clearly outlines the current problem; and third, the output, which presents the corresponding solution.

The third step of the ILKM framework entails an initial pre-training of the base LLM with domain-specific data sourced from LKL. This pre-training imbues the LLM with rich domain-specific knowledge. Following this, the pre-trained LLM undergoes a fine-tuning process, guided by domain instructions, transforming it into a domain knowledge LLM proficient in the designated field. To refine the LLM’s expertise, several common techniques to enhance and train LLMs can be summarized as follows: reinforcement learning from human feedback [4], instruction tuning [12, 21], mixture of experts [17], prompt engineering [1, 14], retrieval-augmented generation [5], and leveraging attention mechanism [19], The objective of this step is to build a robust LLM that possesses extensive domain knowledge and expertise. This LLM can then guide the development of new ML models capable of addressing complex challenges and real industrial problems.

Upon the successful training of the domain knowledge LLM, the fourth step involves utilizing it as a domain expert for subsequent specialized model development. In this step, domain instruction data serves as the prompt, propelling the LLM to address specific analytical problems. The domain-specific knowledge LLM, acting on these instruction inputs, proposes targeted solutions. In addition, human experts may interact and intervene, offering strategic guidance to refine the LLM’s outputs. These solutions are then transferred to a coding-focused LLM [3, 22], which incrementally develops code aligned with the domain knowledge LLM’s insights, thereby creating a new ML model for specific problems. In addition, the structured machine-generated data serves as the dataset for new ML model training and testing. Finally, this step culminates in the generation of actionable solutions, ready to be integrated into decision-making workflows.

The main difference between ILKMs and LLMs lies in their purpose and functionality. ILKMs are designed to handle specific industrial tasks, utilizing relevant structured industrial data and domain-specific knowledge to provide precise, expert-level solutions. In contrast, LLMs are more generalized, leveraging extensive training on diverse textual data to solve language-related tasks, such as text generation, conversation, and language translation. To better illustrate the characteristics of ILKMs, a detailed comparison between ILKM and LLM is presented in Table 1. They are compared and explained from eight perspectives: “Data,” “Purpose,” “Applications,” “Data Privacy and Security,” “Domain-Specific Knowledge,” “Integration and Customization,” “Scalability,” and “RealTime Decision-Making.”

As illustrated in Figure 3, the “6S Principle” is proposed as a guideline for the future development of ILKMs. The “6S Principle” encompasses six key components: “Specialized Domain Knowledge,” “Scrutability,” “Safety,” “Scalability,” “Sustainability,” and “Systematization and Standardization.” The details of the purpose, challenges, and opportunities for each principle are presented in Figure 3. All of these principles are crucial for the successful application of ILKMs in industrial settings, ensuring that ILKMs can address specific needs and challenges faced in Industry 4.0 and smart manufacturing.

There are several opportunities for developing ILKMs in the future of Industry 4.0 and smart manufacturing. In the field of material discovery and synthesis, ILKMs can analyze historical data from the literature to summarize design guidelines and principles as domain-specific instructions. These instructions can then guide the design of new experiments, ultimately facilitating the discovery of new materials. In the area of engineering design, ILKM can assimilate knowledge from multimodalities (such as text, 2D images, 3D shapes, and sound) gathered from historical products and provide possible optimization directions for designers and engineers to enhance new product performance. In the realm of prognostics and health management, ILKMs can analyze data from historical failures and maintenance strategies to aid in the diagnosis and prognosis of complex industrial machine systems, ultimately enabling predictive maintenance and lifecycle management. Furthermore, intelligent QA platforms can be developed across various industrial sectors with the revolution of ILKM. Instead of manual information searches performed by humans, ILKM can automatically retrieve relevant information and generate responses, thus assisting employees with their queries.