Authors - Megha Potdar, Andhe Dharani, Ch.Ram Mohan Reddy Abstract - Semiconductor fabrication processes suffer significant yield losses, often exceeding 20%, due to equipment anomalies in critical stages like plasma etching and lithography, where traditional Statistical Process Control fails to detect subtle, non-linear drifts in multivariate sensor data such as temperature, pressure, and gas flow. This paper proposes a novel hybrid AI framework combining Long Short-Term Memory Autoencoder for unsupervised reconstruction-based anomaly detection with Isolation Forest for robust outlier scoring and severity ranking, enabling real-time predictive maintenance and Remaining Useful Life estimation. The LSTM-AE compresses temporal sequences into a latent space and flags anomalies via elevated Mean Squared Error thresholds (>95th percentile), while Isolation Forest filters multivariate errors to minimize false positives. RUL prediction employs linear regression on error trends for proactive scheduling. Implemented in a Keras/TensorFlow MLOps pipeline with
Open Zoom