This section introduces the characteristics and definition of GIPP. On that basis, a generative AI and digital twins-based framework of intelligent process planning for creating an ideal process planning system is proposed.
3.1 Characteristics of GIPP
In a manufacturing cycle, the traditional process for product manufacturing by the CAPP is illustrated in Fig. 4(a). Initially, product designers create designs using CAD software based on requirements. Subsequently, the CAPP system, which can integrate various extended techniques including Feature Based (FB), Knowledge-Based (KB), Neural Network (NN), Internet Based (IB), Functional Blocks (FBs), Fuzzy Set Theory (FST), Agent-Based (AB), Step-Compliant (STEP), Petri Nets (PN), and Genetic Algorithm (GA) (Yusof and Latif 2013), is employed to generate process plans for the products. These plans are then put into the CAM system to obtain toolpath trajectories and numerical control programs for tool machining. Finally, the products are manufactured using Computer Numerical Control (CNC) machine tool. By drawing an analogy, this paper introduces the process and characteristics of GIPP. Specifically, product designers input their design requirements into GAIM, which can be in the form of text descriptions, 2D CAD drawings, or 3D CAD models. GAIM then directly generates the process plan, presented in structured data, such as text or tables, containing detailed information about the product's machining, including motion paths and numerical control programs. Finally, the process plan is put into the digital twin, enabling real-time interaction between the virtual and physical realms for product manufacturing, as shown in Fig. 4(b).
With the strengths of generative AI and digital twin, GIPP exhibits the ensuing characteristics:
-
Multimodality. The inputs can be multimodal, including text descriptions, 2D drawings, or 3D models, while the outputs can take various forms, such as text or tables, according to on-site needs.
-
Efficient management capability. Providing manufacturing enterprises with a generative method to recognize and extract multimodal data, thus establishing an effective mechanism for obtaining and manipulating high-quality process data or knowledge.
-
Flexibility. Scalability, adaptability, and customizability to suit individual manufacturing companies and novel processes.
-
User-Friendliness. Powerful human-computer interaction functionality, offering a user-friendly interface that enables non-experts to use it effortlessly. Additionally, it should provide users with instant and inspirational feedback.
-
Reliability. Utilizing digital twins, it possesses the capacity to evaluate and validate process knowledge and plans.
-
Explainability. Additionally, it can trace and offer logical inferences for the generated process knowledge and plans.
3.2 Definition and framework of GIPP
Based on the analyses provided above, we propose the following definition for GIPP:
Definition 1
GIPP is an intelligent process planning system that integrates generative AI and digital twin technologies. It leverages the capabilities of generative AI, such as easy access and efficient high-quality content generation, along with the benefits of digital twin technology, including real-time simulation and verification. This integration empowers the process planning system with robust abilities of data and knowledge management, human-computer interaction, process knowledge or plans generation, verification, and feedback optimization. Its primary objective is to optimize enterprise productivity and production quality while ensuring high adaptability to accommodate personalized manufacturing needs and reduce production costs.
With this definition, we introduce the architecture of GIPP, as depicted in Fig. 5, which comprises three primary functional layers, namely the Data Layer, Model Layer, and Application Layer.
Specifically, in the Data Layer, the massive and diverse data from various sources within manufacturing enterprises [48] are effectively managed and utilized by constructing a universal data management method or mechanism. This approach can be knowledge acquisition templates [49], machine learning algorithms [50], or their combination, which enables the extraction and classification of process data to form the pre-training and fine-tuning datasets required for training the GAIM. These datasets are stored and dynamically updated using large-scale databases, such as MySQL, to function as a dynamic knowledge base.
The Model Layer comprises the technologies of generative AI and digital twins. Specifically, the emergence of the Transformer [51] has profound implications for GAIM, serving as the intersection across various domains within AI. Because we used the GPT-2 architecture in the case study, Fig. 5 utilizes the Transformer's decoder component to symbolize the Transformer. Notably, GAIM based on the Transformer architecture can be categorized into three main types: Encoder-only, Encoder-Decoder, and Decoder-only. Regardless of the selected architecture, these GAIMs are inherently complex and resource-intensive. For instance, GPT-2 medium boasts 1.5 billion parameters, while GPT-3 scales up significantly to a staggering 175 billion parameters. Such models demand substantial computational power, thereby leading to elevated research costs. Consequently, it is crucial to integrate pertinent key technologies for lightweighting GAIM, including reinforcement learning from human feedback (RLHF) [52], prompt learning [53], knowledge graph [54], and related lightweight techniques like pruning [55], distilling [56], and data augmentation [57]. The integration aims to reduce the training parameters while preserving the model's functionality, ultimately lowering the training cost. Based on this, the model undergoes pre-training and fine-tuning using the input dataset in the data layer and is then applied to specific process planning scenarios. The digital twin model consists primarily of the physical and virtual spaces of CNC machine tools. The physical space represents the actual carrier of the entire machining process, encompassing CNC, end mill, blank workpiece, force sensor, current sensor, and other components. The virtual space encompasses a geometric model, mechanism model, and behaviour model. These models interact with the physical space by utilizing twin data from a real-time database for perception and control. Subsequently, they integrate and interact with the GAIM through an application programming interface (API). By inputting the decision-making knowledge generated by the GAIM, the virtual space conducts process simulation, machining monitoring, quality prediction, and process evaluation. These functions serve to validate the generated content, thereby improving its reliability. Ultimately, the verification results will be fed back to GAIM as novel data, culminating in the iterative enhancement of the training dataset and conferring upon GAIM a certain degree of explainability.
In the Application Layer, the intelligent process planning system can be accomplished by training it with different fine-tuned datasets, allowing for implementation in various scenarios. Similar to the general model ChatGPT, the system functions as an intelligent question-answering system, providing expert information in the field of manufacturing processes when presented with process planning queries. It can address specific requirements of product designers, act as an experienced process planner, generate comprehensive process plans, and provide iterative feedback and improvements for design proposals. Additionally, the system assists novices in quickly familiarizing themselves with the manufacturing domain and aids employees in acquiring new knowledge and skills. The system serves as an AI coach, guiding novices to become skilled professionals in the enterprise and provides personalized and targeted training programs for employees seeking to learn and improve.