William Conner - Director of Applications, INFICON
Time series anomaly detection, i.e., detecting faulty data by examining data vs. time during the process, requires the use of multiple advanced algorithms to find small, but important variations in the data without an increase in false positive rates. To address the growing need for automated anomaly detection, INFICON is focusing on combining proven statistical envelope analyses with machine learning methodology to enable more precise fault detection with less user configuration.
An envelope model is best suited for user interpretation. Unlike models that can only determine whether or not there is a fault, models that calculate time series envelopes are able to plot the variation in the time series data, and show when the fault occurred.
Statistical Envelope models use historical data to determine the expected variation for each variable within each process step and for each time point across all processes. The model calculates upper and lower limits based on what is determined to be the "perfect" path of the variable in time and the observed variation in the data (Figure 1).
The Machine Learning component of the model automatically examines the data and modifies the algorithm to optimize fault detection. The models automatically:
To calculate the most applicable time path for each parameter, the model classifies each parameter. The classifiers are:
The modeling infrastructure learns the processes running on every chamber of every tool, spawning new models, performing automated training, and tracking faults. Users no longer have to decide which processes need monitoring or remember to update analyses as new recipes are developed and new products are introduced into a fab.
The model is designed for continuous learning, and accepts user input about detected and missed faults to continuously adjust the calculated envelopes. If a fault is later identified, the model will compare external information to internal findings, and adjust itself for future anomaly detection.
INFICON is currently partnering with several customers to beta test the time series anomaly detection model in a production environment and is planning to make it available for purchase in the first quarter of 2021.