The data come from Backblaze, it involves more than two million records from more that a year of measurings of hard disk S.M.A.R.T. values. All hard disks are from a particular vendor and of the same model. The algorithm works in a completely unsupervised way and learns the normal hard disk behaviour without any assumptions about a parametric model. In this demo we simulate test results that would be obtained day by day for a period of four months from September to the end of December 2014, assuming no corrective action to be taken. The model has never "seen" these records before.
You can test different levels for the anomaly score threshold and see how it impacts the detection rate and false alarms. You can also choose the day to simulate.