Single-event-latchup detection based on machine learning

In space, the radiation effects on electronic devices may lead to anomalies referred to as Single-Event-Effects (SEEs). The Single-Event-Latchup (SEL) is a type of SEE arising from heavy-ions striking semiconductor devices, and is characterized as a high-current abnormality. This abnormality causes the loss of functionality and may result in permanent device damage. In addition to SELs, there are micro-SELs where the induced current (abnormality) is lower and more localized than SELs. They are consequently more difficult to detect as their characteristics are often masked by the normal operating current of the semiconductor device. The conventional detection of SELs is rather simplistic – it is based on the magnitude of their elevated current thresholds where the elevation is typically set at 4x-5x of the usual operating current. However, this method would be less effective in detecting micro-SELs because the typical operation currents can mask micro-SEL-induced currents due to the low magnitude of micro-SELs. Specifically, the current of the micro-SEL may be below the said operating current. To alleviate this shortcoming, Machine learning (ML) algorithms have reportedly been utilized in place of thresholds-based detection as a more precise method of detecting SELs and micro-SELs. Although ML methods can identify micro-SELs, the dataset used always restricts their preferences. In this dissertation, we apply ML algorithms to detect SELs and micro-SELs, particularly to improve the dataset over reported methods for the model’s training. Through a series of laser tests, the SEL current profiles are captured at high sampling rates to record the transient characteristics for the ML algorithm learning. These SEL-induced current data are thereafter resampled by sliding windows. The newly acquired samples are transferred into feature vectors through feature extraction. The vectors are used for training and testing for five types of ML models. The experimental results show that different operation states of semiconductor devices can lead to a degraded classifier performance. The highest accuracy is 86.6% in our experiments. We propose a new data processing step encompassing the removal of the base current from the original current data. The new data is thereafter used to train classifiers for recognition to solve this problem. The accuracy of the newly trained classifier is improved from 86.6% to 93.1%, i.e., a worthy improvement of 6.5%.Master of Science (Electronics