An intelligent hybrid monitoring framework for robust defect detection in solder paste printing

Annals of Forest Science · Springer Nature · 2026-11-13 · DOI: 10.1007/s10845-026-03032-w

Аннотация

Solder paste printing (SPP) is reported to account for approximately 60% of soldering defects in surface mount technology assembly, yet predictive modeling and statistical process monitoring have developed largely in isolation. This study proposes a closed-loop LSTM-RF-EWMA framework that unifies temporal prediction and statistical surveillance, distinguishing expected systematic variation from assignable-cause deviations in SPP. A long short-term memory (LSTM) network extracts temporal features from sequential solder paste inspection (SPI) data, capturing paste thixotropy, stencil and squeegee wear, and aperture clogging. A random forest (RF) combines these features with the printing parameters to predict deposit volume, area, and height. Exponentially weighted moving average (EWMA) charts then monitor the prediction residuals, and periodic retraining adapts the model to process drift. Evaluated on high-mix production data supplemented by an L18 designed experiment and Monte Carlo simulation, the framework achieved a mean absolute percentage error of 4.80% and reduced volume RMSE by 25% relative to a standalone RF model. With residual autocorrelation accounted for, the EWMA chart produced 18% fewer false alarms than a Shewhart chart with the same in-control run length (0.55% versus 0.68%) and detected 1.0σ shifts within 20 boards twice as often (86% versus 42%). Squeegee pressure was the dominant factor in transfer efficiency. After deployment, with equipment, stencil, and paste unchanged, defects per million opportunities decreased from 2508 to 1014 (59.6%), raising the sigma level from 4.3 to 4.6. Requiring only existing in-line SPI, the framework offers a practical route to proactive SPP quality control.

Читать у издателя →

Похожие статьи