The lay abstract featured today (for Condition assessment and predictive maintenance for contact probe using health index and encoder-decoder LSTM model by Shun-Sun Luk, Yanwen Jin, Xiaoge Zhang, Vincent To-Yee Ng, Jingyuan Huang, Chak-Nam Wong) is from Quality and Reliability Engineering International with the full article now available to read here.

How to cite

Lay Abstract

A contact probe is a device that interfaces with the testing point of an electrical component and is widely used to continuously monitor the quality of microelectronic components during manufacturing. Degradation of the contact probe arising from repetitive testing necessitates occasional maintenance actions, such as cleaning or replacement. A defective contact probe can lead to misleading test results and cause the false rejection of quality products, significantly reducing the production yield. Therefore, assessing the condition of the contact probe and implementing predictive maintenance are critical for enhancing measurement performance.

Traditional defect detection approaches primarily monitor control limits of specific parameters that are determined based on engineers’ experience. These subjective limits tend to be less precise and lack adaptability in accounting for random noise in the manufacturing process.

This paper addresses the defect detection issue of an operating contact probe by developing an informative condition assessment and predictive maintenance strategy. The proposed model can forecast the health condition of the contact probe for the next half-hour and trigger maintenance alerts when necessary. Instead of relying on subjective limits for condition monitoring, a health index indicating the probability of abnormal conditions over a specific time range is utilized. This health index is defined based on statistical results obtained from features selected through principal component analysis. Afterwards, two Long Short-Term Memory (LSTM) encoder-decoder models are developed, and they achieve an accuracy above 75% for predicting maintenance actions within the next 15 minutes. Furthermore, an explainable artificial intelligence technique is employed to discuss and compare the features of the two proposed models. The model with a convolutional neural network (CNN) as an encoder is found to outperform the model with an LSTM encoder in extracting the inner-feature correlation.

This article serves as a starting point for real-time predictive maintenance of operating contact probes. Numerous valuable research topics for further study are identified, such as optimizing maintenance actions based on real-time feedback, integrating additional sensors to refine the accuracy and timeliness of maintenance alerts, and balancing model preparation time with prediction time.