Applying AI in Supporting Additive Manufacturing Machine Maintenance

Introduction

With the industry‘s expansion on applying big data and machine learning, digitalized manufacturing has generated increasing volumes of real-time data containing valuable information about machine and process conditions [1]. This data can be processed to implement predictive maintenance and optimize operations, improving productivity and ensuring workplace safety [2, 3, 4].

The application of Artificial Intelligence (AI) to support human-operated machine maintenance in manufacturing is an emerging field with the potential to transform the industry. AI technologies, such as Large Language Models (LLMs), have enabled more effective interactions between humans and machines, playing an essential role in their operation, maintenance, and training [5, 6, 7]. LLMs, like GPT-3 [8] and GPT-4 [9], are recent technologies that allow more natural and efficient interaction between humans and machines. These AIs can assist in operation, maintenance, and training, offering personalized and contextualized support to operators. However, significant data protection and integrity challenges remain, especially in industrial settings.

This paper discusses an AI-assisted information architecture focused on supporting the maintenance of an additive manufacturing machine in a learning factory environment. The proposed model uses sensors for real-time monitoring and a gateway for local and cloud processing, facilitating fault detection and supporting predictive maintenance, assisted by an LLM trained with additive manufacturing contextual information. The study addresses the implementation of a secure architecture, combining AI and VPN networks to maintain data privacy and integrity in industrial environments. By integrating emerging technologies, such as intelligent virtual assistants and secure networks, this approach seeks to validate a robust architecture that is practical and innovative with applications for the smart manufacturing industry.

The implemented framework is characterized by remarkable applicability and adaptability across various manufacturing machines by adjusting to and incorporating each machine type’s specific properties and processes. For instance, in the case of a CNC milling machine, the framework can be tailored to monitor spindle speed, tool wear, and precision tolerances, while in a robotic assembly line, it could focus on motor torque, alignment accuracy, and cycle times. This adaptability ensures that the framework can optimize secure manufacturing strategies by integrating machine-specific metrics into its monitoring and predictive maintenance algorithms, generally regardless of the origin of the specific optimization problem.

By leveraging secure data extraction methods through a VPD network and gateway, the framework ensures robust data security while enabling real-time data collection from multiple machine components and additionally integrated sensors. For example, in a smart factory producing automotive components, the AI could analyze vibration data and thermal signatures to detect anomalies, reducing risks of unplanned downtime. This secure and adaptable approach fosters resilience, enhances efficiency, and supports sustainable manufacturing operations across industries.

Architecture

The primary objective of the technological architecture is illustrated in Figure 1, highlighting several essential characteristics of assistive AI applications for machine operation. Traditionally, the Human-Machine Interface (HMI) is implemented as a screen-based device due to the operational conditions of machinery [10]. However, one of the critical innovations introduced by natural language AI is the voice communication capability, which further simplifies interaction within industrial HMIs [11]. Another essential feature is the AI’s interpretative capability regarding a database (typically cloud-based) that can be continually updated, constraining the AI‘s response sample field and reducing the risk of erroneous outputs. Additionally, the architecture supports the AI‘s ability to execute varied interpretations and commands, such as initiating human technical assistance when the user requests.

Figure 1

Functional architecture of the industrial use case

Current architectures for assistive Artificial Intelligence (AI) applications often adopt the model represented in Schematic A of Figure 2. In this architectural model, a web-based service sends requests via APIs to an AI engine hosted in a cloud computing environment. The web-based service is responsible for the interface and communication management, and the AI engine operates within cloud infrastructures, facilitating scalability and remote access [2, 3, 4]. However, this configuration presents significant challenges when applied to industrial networks, particularly concerning data security, latency, and reliability [12].

Figure 2

Schematic model architectures

One of the main challenges to overcome is data security. Protecting sensitive information is crucial in industrial environments, as transmitting data over the Internet to cloud services can expose the system to vulnerabilities and cyber-attacks. Additionally, recurring operational costs and dependence on cloud infrastructure can pose significant obstacles to large-scale adoption of this architecture in the industrial sector.

To address these issues, VPNs are necessary to ensure the security and privacy of transmitted data. Architectures B and C, illustrated in Figure 2, present alternative solutions. In Architecture B, a VPN creates a secure communication tunnel between system components, allowing the AI engine to remain in the cloud while protecting the entire system. In Architecture C, implementing a local LLM eliminates the need for external cloud communication, enhances security, and reduces latency, though it limits response capacity and application scalability. This configuration enables assistive chats to operate more efficiently and securely in industrial environments.

Table 1 compares the different architectures discussed, highlighting their main characteristics and suitability for various industrial scenarios.

Implementation

The implementation described was evaluated within the controlled environment of the Factory of the Future Lab, a Learning Factory at the University of São Paulo (USP). In this context, latency was not a critical parameter. Thus, the experimental focus was assessing the feasibility of different application scenarios and their scalability potential. Consequently, Architecture B (Figure 2) was selected for deployment due to its balanced approach to security, cost efficiency, and operational scalability. The implemented system architecture is illustrated in Figure 3.

Figure 3

Implemented system architecture and utilized elements

The architecture is currently being implemented on a Fused Deposition Modeling (FDM) Additive Manufacturing machine, which serves as a source of experimental data and a demonstrator of the proposed maintenance strategy. Given that FDM processes are characterized by high variability across several parameters, finding an optimal configuration presents a complex and challenging scenario [13].

The technology was implemented by integrating a tablet as the main interface, a gateway equipped with a 5G sim card, and OpenAI’s AI engine. A voltage and current sensor were attached to the additive manufacturing machine, allowing the AI system to detect and report electrical issues in real time by comparing them with the training dataset. The tablet is the primary interaction between the human operator and the assistive system. It provides a Human-Machine Interface for monitoring and controlling the machine. By interacting with the assistive chat, operators can receive immediate notifications about electrical failures or power interruptions and obtain guidance on operational and maintenance procedures.

Based on Schematic B (Figure 2), the implemented architecture uses a Virtual Private Network (VPN) to ensure secure communication between local devices and the cloud-based AI engine. The gateway collects data from the voltage and current sensors and transmits it through the VPN to the AI system in the cloud. This configuration allows the AI to analyze real-time data and promptly alert the operator about detected anomalies, enhancing operational efficiency and equipment safety. The VPN addresses security concerns associated with industrial networks, creating a secure tunnel for transmitting sensitive data [14]. Additionally, by maintaining the AI engine in the cloud and ensuring secure communication, the system leverages cloud services‘ computational power and scalability without compromising data integrity.

The role of the VPN in data security is fundamental in this context. A VPN protects sensitive data transmitted between the machine, the gateway, and the cloudbased AI engine against interception and unauthorized access. The VPN creates an encrypted tunnel for data transmission, guaranteeing the confidentiality and integrity of information [14]. This is especially critical in industrial environments, where data security is essential to prevent cyber threats and protect intellectual property. By implementing the VPN, the system architecture meets the necessary security requirements to operate reliably and securely within an industrial setting. Figure 4 shows the solution implementation in the selected machine.

Figure 4

Photos of the local installation

The technology was integrated and tested using the OpenAI AI engine to process received data and generate query responses. Initially, we extracted the content from the technical manual and reformulated significant portions of the material, organizing it into a structured dataset in FAQs (frequently asked questions). During this process, we emphasized critical information on operational procedures, such as startup and shutdown, and maintenance guidelines, including troubleshooting for operational failures or interruptions during functioning, such as a printing stoppage.

Subsequently, we consolidated these two documents into a specialized database to enhance efficient responses to technical queries. This approach resulted in the ChatGPT model delivering more precise and contextually accurate responses, effectively addressing specific technical demands. Various query scenarios were conducted during troubleshooting to validate the system‘s ability to identify relevant content within technical materials such as manuals and datasheets. Figure 5 shows the output of the chat.

Figure 5

AI-assisted chat. The left figure represents a question made by the user and the answer given by the system (The right figure illustrates the training proposed by the system)

Results and Conclusions

Implementing a technical assistant copilot for operating and maintaining an additive manufacturing machine and utilizing a VPN for secure data transmission proved practical and innovative. By integrating a gateway equipped with a 5G sim card and connected to a voltage and current sensor, the virtual assistant can monitor the machine’s status in real time, identifying electrical issues and providing operational support. The combination of the tablet interface, the gateway with a sensor, and advanced AI constitutes an innovative solution that enhances human-machine interaction within the industrial context while validating an architecture that offers elevated levels of security and reliability.

A key innovation of this project was transforming the assistive chat from a passive helper into an active agent. The assistant anticipates the operators‘ needs and takes action by proactively notifying them of detected electrical issues, potentially reducing downtime and preventing possible failures. Moreover, a VPN ensures secure communication between devices and the virtual assistant, allowing private networks to be used within the industrial environment without compromising data security.

The results highlight the substantial potential of integrating emerging technologies, such as intelligent virtual assistants and secure networks, in industrial applications. This approach can be extended to other equipment and processes, fostering more effective human-machine interaction. Future research may further explore expanding the assistant‘s functionalities, incorporating additional sensor parameters, and optimizing human-machine interaction to improve operational efficiency and safety.