Predictive Maintenance - Training
This workflow trains an autoregressive model on data from a properly functioning rotor to compute error statistics. These statistics serve as a baseline, enabling early detection of anomalies during deployment.
The input data includes 313 time series of spectral amplitudes from 28 sensors located on a rotor machine. The workflow first filters the data to training data covering only normal functioning.
Then, it loops over each frequency column at a time, trains an auto-regressive model using 10 past values as predictors and calculates in-sample prediction error statistics. Lastly, it saves the model and prediction error statistics for deployment.