The Investigation of AutoEncoder-Based Neural Network for NMOS Circuit Anomaly Detection

This paper investigates the application of a neural network autoencoder for anomaly detection in NMOS circuits within the semiconductor industry. The focus is on the ability of the autoencoder to recognize deviations from normal operation, signifying potential anomalies. The circuits were designed using PySpice, simulating various operating conditions. Anomalies were induced by altering the load resistance and the autoencoder was trained to identify these changes through reconstruction error analysis. The model demonstrated high accuracy in anomaly detection, suggesting effective learning of the normal operational patterns. The reconstruction errors were compared to a threshold set one standard deviation above the mean to determine anomalies. Results indicated that the autoencoder could reliably identify most anomalies, providing a novel tool for circuit health monitoring and enhancing electronic device reliability. However, the model's inability to capture the full complexity of NMOS behavior under diverse conditions and the empirical nature of the threshold suggest areas for improvement. Future work will aim to increase the model's robustness, utilize a more extensive and varied dataset, and explore semi-supervised learning to detect previously unidentified anomalies.