One Class Process Anomaly Detection Using Kernel Density Estimation Methods

We present a one-class anomaly detection method that uses time series sensor data to detect anomalies or faults in semiconductor fabrication processes. Critically, this method is trained using only small amounts of known successful run data, making it possible to implement for many processes and recipes without needing example faults. The proposed method uses kernel density estimation (KDE) to create probability distributions for sensor values during nominal processing. When classifying unseen sensor data, we determine the likelihood that it arose from this (often non-Gaussian) nominal distribution, allowing us to classify new signals as nominal, or faulty. We present model extensions that enable adaptation to changes in the underlying process, i.e., concept drift, as well as transfer learning techniques that enable training of anomaly detectors for new process recipes with less data. The proposed methods are tested on historical data from plasma etch and ion implantation processes, outperforming benchmark methods including traditional statistical process control (SPC), one-class support vector machine (OC-SVM), and variational auto-encoder (VAE) based detectors.