Thank you for the update Arthur. What is the "default" NuPIC model you are using? For anomaly detection we recommend using this set params [1]. You can pull these params directly into your model with the getScalarMetricWithTimeOfDayAnomalyParams() [2], as we've done for the detector in NAB [3].
Are you using anomaly score or anomaly likelihood to flag the anomalies? What values does the model output at the first and second outages? If a NuPIC model encounters the same anomalies repeatedly, it will learn these to be normal and stop detecting them. What is the granularity of the data? Yes it would help to see the code and/or the data. Also, if you're able to share the labeled anomaly data, you should check out the NAB Competition: numenta.org/nab<http://numenta.org/nab/>. You can win a bunch of money for the data! [1] https://github.com/numenta/nupic/blob/master/src/nupic/frameworks/opf/common_models/anomaly_params_random_encoder/best_single_metric_anomaly_params.json [2] https://github.com/numenta/nupic/blob/master/src/nupic/frameworks/opf/common_models/cluster_params.py#L29 [3] https://github.com/numenta/NAB/blob/master/nab/detectors/numenta/numenta_detector.py#L73-L78 Cheers, Alex -- Alexander Lavin Software Engineer Numenta
