Fast Adaptive RNN Encoder⁻Decoder for Anomaly Detection in SMD Assembly Machine.

Area of Science:

• Manufacturing Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Surface Mounted Device (SMD) assembly lines require adaptable anomaly detection models for diverse manufacturing environments.
• Existing anomaly detection methods may lack the speed necessary for rapid adaptation in flexible manufacturing.

Purpose of the Study:

• To propose a fast adaptive anomaly detection model for SMD assembly machines.
• To leverage Recurrent Neural Network (RNN) Encoder-Decoder architecture for efficient anomaly detection using machine sounds.

Main Methods:

• Developed a Recurrent Neural Network (RNN) Encoder-Decoder model, noting its parameter efficiency compared to Auto-Encoders (AE).
• Utilized operating machine sounds as input for the anomaly detection system.
• Determined abnormality by calculating the Euclidean distance between generated and observed sound sequences.

Main Results:

• The proposed RNN Encoder-Decoder model demonstrated a significantly reduced training process for fast adaptation.
• Experimental evaluation on SMD assembly machine datasets showed cutting-edge performance.
• The model achieved high accuracy in detecting anomalies with rapid adaptation capabilities.

Conclusions:

• The RNN Encoder-Decoder based anomaly detection model offers a fast and effective solution for flexible manufacturing environments.
• The model's ability to adapt quickly to changing conditions is crucial for maintaining high-quality SMD assembly.
• Machine sound analysis provides a viable data source for real-time anomaly detection in industrial settings.