Convolutional Recurrent Reconstructive Network for Spatiotemporal Anomaly Detection in Solder Paste Inspection.

Area of Science:

• Manufacturing Process Monitoring
• Artificial Intelligence in Quality Control
• Computer Vision for Defect Detection

Background:

• Surface Mount Technology (SMT) is crucial for printed-circuit board (PCB) production.
• Solder Paste Printer (SPP) defects can lead to defective PCBs, detected by Solder Paste Inspection (SPI).
• Existing methods struggle with accurately decomposing and identifying anomaly patterns from printer defects.

Purpose of the Study:

• To propose a novel deep learning model, the Convolutional Recurrent Reconstructive Network (CRRN), for anomaly detection in SPI data.
• To effectively decompose anomaly patterns originating from SPP defects.
• To enhance the accuracy and efficiency of defect detection in SMT processes.

Main Methods:

• Development of a CRRN model comprising a spatial encoder, a spatiotemporal encoder-decoder with Convolutional Spatiotemporal Memories (CSTMs) and a spatiotemporal attention (ST-Attention) mechanism, and a spatial decoder.
• Training the CRRN exclusively on normal data to detect anomalies via reconstruction error.
• Utilizing CSTMs for efficient spatiotemporal pattern extraction and ST-Attention to address long-term dependencies.

Main Results:

• The proposed CRRN significantly outperforms conventional models in anomaly detection accuracy.
• The CRRN successfully decomposes anomaly patterns, enabling effective identification of printer defects.
• The anomaly maps generated by CRRN demonstrate strong discriminative power for defect classification.

Conclusions:

• The CRRN is a highly effective tool for anomaly detection in SPI data, crucial for SMT quality control.
• The model's ability to decompose anomaly patterns provides valuable insights into printer defects.
• This approach offers a promising solution for automated defect detection and classification in PCB manufacturing.