I am doing an Anomaly Detection with time series clustering of a real life manufacturing process. I used window slider und clustered these windows with k-means (high dimensions >50). After that I did a PCA to plot the found cluster centroids. In the next step I am struggeling with getting an anomaly score. I want to reconstruct the signal with the centroids and overlay them with real life signal to get a reconstruction error. Any ideas how I can implement this in KNIME? Already found a good description with python here: http://amid.fish/anomaly-detection-with-k-means-clustering
Would be great if somebody can give me some hints or other usefull tips about my problem or anomaly detection in general thx
I am working on the same task. I read the work in - http://amid.fish/anomaly-detection-with-k-means-clustering. I am trying to do the same think with multiple time series. Have you done it with only 1 time series?
I’ve done it with only one big long time series describing a real manufacturing drilling process. The solution is in another post of me here: k-means clustering: distance from a certain datapoint to his next centroid. The solution is about doing a distance measurement of an anomalous cluster window from it’s nearest cluster centroid (=reconstruction error). However I still have big problems with the distance measurement because the distances are way to big. Maybe you can report here how you are doing I would appreciate this