Anomaly Detection in Semiconductor Manufacturing using Hybrid Autoencoder Models

Abstract

Semiconductor fabrication processes generate highdimensional, multi-sensor telemetry streams that are critical for quality control but challenging to analyze in real time. Traditional anomaly detection methods rely on static thresholding or shallow statistical models, which struggle with non-linear sensor correlations, high noise levels, and severe class imbalance. This paper proposes a hybrid deep learning approach that combines an autoencoder for non-linear feature extraction with an isolation forest for robust anomaly scoring. The methodology is validated on the publicly available UCI SECOM dataset, comprising 1,567 production samples with 591 sensor features and binary pass/fail labels. Extensive experiments demonstrate that the proposed hybrid model achieves an F1-score of 0.873 and ROC-AUC of 0.921, outperforming standalone autoencoder, isolation forest, and traditional statistical baselines by +9.2% in F1-score and -13.7% in false positive rate. The hybrid approach also reduces inference latency by 31% compared to end-to-end neural classifiers, making it suitable for edge deployment in industrial IoT environments. Results confirm that coupling unsupervised representation learning with ensemble-based anomaly scoring significantly improves defect detection reliability in high-dimensional manufacturing processes. The proposed pipeline can be integrated into existing SCADA and IIoT platforms and scaled to complex process industries.