TOSKY Advances AI for Manufacturing Through Sound and Vibration Analysis, Pursuing Real-Time Defect Detection with Physical AI and Edge Devices

AI SaaS company TOSKY Co., Ltd. (TOSKY) is advancing the development of an AI-based defect detection system that utilizes sound and vibration data collected from manufacturing sites.

The project focuses on developing manufacturing AI technology that detects product defects and equipment abnormalities by jointly analyzing acoustic and vibration data generated during heading and rolling processes. As a consortium member, ToSky is participating in the development of AI services applicable to real-world manufacturing environments, leveraging its cloud-based AI technologies and extensive experience in utilizing diverse data sources.

During the first year of the project, directional microphones and vibration sensors were installed on manufacturing equipment to collect data. Based on this data, ToSky developed an AI defect detection model and a real-time monitoring dashboard. The AI model adopts a multimodal sensor-fusion architecture that simultaneously analyzes three-axis acceleration-based vibration data and acoustic data. It was designed to detect various types of abnormal conditions by combining deep learning-based anomaly detection with signal-processing techniques.

In testing, the developed model achieved an F1-score of 0.712 and an inference time of 3.12 seconds, surpassing the target criteria of an F1-score of 0.65 or higher and an inference time of five seconds or less. A dashboard environment was also established to determine whether products were normal or defective using real-time sensor data and to provide immediate alerts when abnormalities were detected. In addition, productivity and AI reliability were verified through the Telecommunications Technology Association (TTA).

From Simple Detection to Physical AI for Real-World Control

In the second year, the system will be expanded beyond conventional AI-based defect detection into Physical AI, where AI decisions are directly translated into physical actions by manufacturing equipment.

When the AI identifies a defective product, an air gun or gate-control device will be used to immediately remove it from the production line. In the event of a critical abnormality, the system can also stop the equipment. Through this approach, ToSky plans to establish an automated quality-control system that connects the entire process from AI decision-making → defect detection → physical sorting and control, moving beyond conventional post-production inspection and manual sorting.

In particular, an edge device-based AI architecture will be introduced to address communication latency associated with cloud-based systems in high-speed manufacturing environments where dozens of products can be produced every second.

The AI models will be lightweighted and configured to perform inference directly on edge devices at manufacturing sites. Technologies such as ONNX, TensorRT, and LiteRT will be utilized to enable real-time defect detection and physical control without requiring a network connection. Through this approach, the project aims to establish a 24/7 quality-monitoring system capable of processing more than 20T/S and operating reliably even in offline environments.

Expanding Commercialization of Manufacturing AI Solutions Beyond Field Validation

During the second year of the project, ToSky will not only improve AI model performance but also pursue cross-site validation and commercialization to expand the technology to other manufacturing environments.

The company plans to develop independent AI defect-detection models using equipment data from each participating manufacturer and conduct cross-validation across different production sites. This approach is intended to secure greater versatility and reliability without making the technology dependent on a specific piece of equipment or manufacturing environment.

Through this project, ToSky plans to expand its accumulated expertise in acoustic- and vibration-based manufacturing AI to a wide range of manufacturing processes, including automotive parts production. The company also aims to develop the technology into AI-based quality-control and process-automation services that manufacturing companies can adopt more easily.

In addition, ToSky plans to continuously expand technology validation cases through participation in domestic and international industry exhibitions and collaboration with relevant organizations and manufacturing companies. By doing so, the company aims to strengthen its competitiveness as an AI service provider supporting AX (AI Transformation) across manufacturing environments.