An online algorithm for streaming anomaly detection
The data come from a PHM challenge, the objective was Remaining Useful Life (RUL) prediction in the context of controlled ball bearing stress tests. We used that data to test our online anomaly detector. The detector needs a minimal number of points to set up its internal parameters, then it works in real time in a completely unsupervised way. The detector outputs raw values which are then adaptively filtered and evaluated in a time-aware algorithm for change detection. A lower anomaly score implies a more anomalous point.
Raw datapoints are 2560 acceleration values for both the horizontal and vertical directions, these are sampled every 10 seconds in 0.1 seconds (25.6 kHz). The raw values are preprocessed and spectral features are extracted, allowing to collapse the information in 66 features.
You can test different settings with the slider, higher values tend to give more alarms. Note that all the experiments are ball bearing stress tests at the end of which the bearing under examination is broken.