Incremental learning of LSTM-autoencoder anomaly detection in three-axis CNC machines

There has been a continual effort to develop smarter, more effective CNC machines, capable of fully autonomous operation. To achieve this goal, the machines must be able to automatically detect operational and process anomalies before they cause serious damage. It has been shown that using Artificial Intelligence techniques, such as LSTM-AutoEncoders is an effective method for anomaly detection of issues such as machine chatter. Transfer learning is a valuable tool to decrease the amount of data required to implement this approach, but has lower accuracy than directly training a network on a large dataset. By implementing an incremental-ensemble of weak learners, we have been able to, not only capture changes in system dynamics over time, but incrementally improve the accuracy of a network trained through transfer learning to be comparable to a network directly trained on a large dataset. This allows us to quickly deploy networks on new systems, and obtain highly accurate anomaly estimates