Non-intrusive Embedded Systems Anomaly Detection using Thermography and Machine Learning

Abstract

Quality control in electronic system manufacturing is achieved mainly through system testing. Device miniaturization and multilayer Printed Circuit Boards have increased the electronic circuit test complexity considerably and processes based on manual inspections have become outdated and inefficient. The concept of Industry 4.0 has enabled the manufacturing of customized products based on customers’ demands, which demands a high degree of flexibility in production processes, with low cost and without placing numerous test points. In this paper, we propose two automated test solutions based on machine learning and thermographic analysis. We propose deploying autoencoders and random forest in two different manners to detect firmware or hardware anomalies based on the circuit board’s temperature signature. We validate our proposal using two firmware versions running independently on the test board. We obtained an anomaly detection rate above 98%. In the random forest approach, we require all data classes for training, whereas the autoencoder only requires the reference class, which is expected in real scenarios.

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

“…Here, implementation details and considerations are discussed. We also summarize the anomaly detection systems based on thermographic and electric current signatures detailed in (de Oliveira et al, 2022;de Oliveira et al, 2021). Additionally, we define quantitative analysis parameters to compare the improvements obtained with pre-processing.…”

“…We validate the proposed transformations in two anomaly detection systems implemented by our group (de Oliveira et al, 2022;de Oliveira et al, 2021). Such systems are based on thermographic and electrical current signatures.…”

“…Fig.8Anomaly detection system: the infrared camera captures the DUT's thermal image; temperature signature is plotted; spectrogram is obtained from this signature and converted to imageA; imageA is used as input of a trained autoencoder, which reconstructs it as imageB; the difference between imageA and imageB is compared to a threshold λ, if the difference > λ  Anomaly detected. Adapted from(de Oliveira et al, 2021) …”