Improving Autoencoder-Based Anomaly Detection in Embedded Systems using Data Transformation

Abstract

Abstract Quality control is a critical aspect of modern electronic circuit industry. In addition to being a prerequisite for proper functioning, circuit quality is closely related to safety, security, and economics. Deep learning models have been extensively used in embedded systems testing and anomaly detection. However, performance is heavily dependent on the data available for training. Often, the number of samples or even its quality is limited. This leads to poor training process and low performance. We present a solution to improve anomaly detection in embedded systems by transforming time signals acquired from the printed circuit board under test. The proposed approach is experimentally validated in two autoencoder-based anomaly detection systems. Hence, two types of signals are analyzed: electric current and thermographic signatures. In both cases, electrical or thermographic signals are pre-processed prior to being converted into spectrogram images, which are then used to train and test the autoencoder. The achieved anomaly detection accuracy improvement for the thermographic case is 71%, compared with the raw data. For the electric current case, we show how data transformations enable autoencoder training where, using raw data, training is not feasible. In this case, we find accuracy improvements of up to 98%. The results indicate that, even in a scenario where the available data are limited, it is possible to achieve an acceptable performance using the proposed technique.