MTS datasets for anomaly detection. (With data feature annotation and anomaly type annotation)

In order to enhance research and development efforts, we have made the labeled datasets available, comprising of two privately collected datasets from partner companies as well as six publicly accessible datasets.

Datasets

Detailed information of the eight datasets is shown below.

Dataset	#Entities	#Metrics	#Train	#Test	Anomalies (%)
D1	26	49	14400	23040	0.05
SMD	28	38	28479	28479	0.04
ASD	12	19	8640	4320	0.05
SMAP	54	25	2818	7331	0.13
MSL	27	55	4308	6100	0.11
SWaT	1	51	496800	449919	0.12
WADI	1	123	1048571	172801	0.06

These datasets with anomaly labels can be downloaded from GoogleDrive

For each dataset, there is a XXX-train.json and a XXX-test.json. For each file, the format is as follows:

Key	Values
'data'	a list of length #entities, each item is #metrics * #timepoints
'label'	a list of length #entities,each item is #timepoints

Data Feature Annotation

We label three key features that influence the performance of anomaly detection algorithms: smoothness, periodicity, and periodic metric pattern correlation. The annotation results are palced in DataFeatureAnnotation. For each dataset, the format is as follows:

Entity Index (omitted)	Periodicity	Smoothness	Metric_Pattern_Correlation
0	xx	xx	xx
1	xx	xx	xx
...	xx	xx	xx
n	xx	xx	xx

Anomaly Type Annotation

Based on existing research and empirical analysis, we classify anomalies into six main types: global anomaly, contextual anomaly, pattern anomaly, frequency anomaly, trend anomaly, and others. We code the each anomaly type into a integer:

Anomaly type	Value
Global	2
Contextual	6
Pattern	3
Frequency	4
Trend	5
Others	1
The annotation results are palced in AnomalyTypeAnnotation.
The annotation of each dataset is stored as a JSON format file.
Each file is a contains #entities list, and each list is of length #timepoints.
For each time point, the value is the anomaly type of the corresponding time point and '0' denotes normal data.