Anomaly detection using an ensemble of models

1. A computer-implemented method comprising:

training a plurality of ensembles of deep learning models using clustered time series training data from numerous components in an Information Technology (IT) infrastructure, the clustered time series training data having been generated through removal of anomalies, using a statistical method, from historical time series data from the numerous components in the Information Technology (IT) infrastructure, and by clustering the historical time series data with removed anomalies to generate the clustered time series training data, wherein each ensemble in the plurality of ensembles is trained using a corresponding cluster in the clustered time series training data;

inputting aggregated time series data to a respective ensemble of deep learning models, the respective ensemble of deep learning models being determined based on matching the inputted aggregated time series data to a corresponding cluster used to train the respective ensemble;

identifying, using a voting method, anomalies in the aggregated time series data based on respective portions of the aggregated time series data that are indicated as anomalous by a majority of deep learning models in the respective ensemble of the deep learning models;

grouping the anomalies according to relationships between the anomalies; and

performing a mitigation action in response to grouping the anomalies.

2. The method of claim 1 , wherein the clustered time series training data are clustered using fuzzy clustering.

3. The method of claim 1 , wherein the clustered time series training data are clustered using fuzzy clustering with dynamic time warping distance.

4. The method of claim 1 , wherein the respective ensemble of deep learning models comprises at least three types of deep learning model.

5. The method of claim 1 , wherein the respective ensemble of deep learning models comprises a plurality of long short-term memory (LSTM) models.

6. The method of claim 5 , wherein one deep learning model of the respective ensemble of deep learning models comprises a LSTM model with autoencoders.

7. The method of claim 5 , wherein one deep learning model of the respective ensemble of deep learning models comprises a LSTM model with uncertainty estimation.

8. The method of claim 5 , wherein one deep learning model of the respective ensemble of deep learning models comprises a LSTM model with dropouts.

9. The method of claim 1 , wherein grouping the anomalies is based on child-parent relationships between the anomalies.

10. The method of claim 1 , wherein grouping the anomalies is based on a cosine similarity between a group of anomalies and a historical group of anomalies.

11. The method of claim 1 , wherein grouping the anomalies is based on similarities between feature sets of anomalies, wherein the feature sets are generated using autoencoders.

12. The method of claim 1 , further comprising:

generating the clustered time series training data by removing anomalies from historical data using an Extreme Studentized Deviate (ESD) test.

13. The method of claim 12 , wherein the Extreme Studentized Deviate (ESD) test comprises a seasonal hybrid generalized ESD test.

14. The method of claim 1 , further comprising:

receiving feedback related to the anomalies;

updating the clustered time series training data based on the feedback; and

re-training the multiple ensembles of deep learning models using the updated clustered time series training data.

15. The method of claim 1 , further comprising:

receiving feedback related to the anomalies; and

suppressing at least one anomaly based on the feedback.

16. The method of claim 1 , wherein the method is performed by an anomaly detection system according to software that is downloaded to the anomaly detection system from a remote data processing system.

17. The method of claim 16 , wherein the method further comprises:

metering a usage of the software; and

generating an invoice based on metering the usage.

18. A system comprising:

one or more processors; and

one or more computer-readable storage media storing program instructions which, when executed by the one or more processors, are configured to cause the one or more processors to perform a method comprising:

training a plurality of ensembles of deep learning models using clustered time series training data from numerous components in an Information Technology (IT) infrastructure, the clustered time series training data having been generated through removal of anomalies, using a statistical method, from historical time series data from the numerous components in the Information Technology (IT) infrastructure, and by clustering the historical time series data with removed anomalies to generate the clustered time series training data, wherein each ensemble in the plurality of ensembles is trained using a corresponding cluster in the clustered time series training data;

inputting aggregated time series data to a respective ensemble of deep learning models, the respective ensemble of deep learning models being determined based on matching the inputted aggregated time series data to a corresponding cluster used to train the respective ensemble;

identifying, using a voting method, anomalies in the aggregated time series data based on respective portions of the aggregated time series data that are indicated as anomalous by a majority of deep learning models in the respective ensemble of the deep learning models;

grouping the anomalies according to relationships between the anomalies; and

performing a mitigation action in response to grouping the anomalies.

19. A computer program product comprising one or more computer readable storage media, and program instructions collectively stored on the one or more computer readable storage media, the program instructions comprising instructions configured to cause one or more processors to perform a method comprising:

training a plurality of ensembles of deep learning models using clustered time series training data from numerous components in an Information Technology (IT) infrastructure, the clustered time series training data having been generated through removal of anomalies, using a statistical method, from historical time series data from the numerous components in the Information Technology (IT) infrastructure, and by clustering the historical time series data with removed anomalies to generate the clustered time series training data, wherein each ensemble in the plurality of ensembles is trained using a corresponding cluster in the clustered time series training data;

inputting aggregated time series data to a respective ensemble of deep learning models, the respective ensemble of deep learning models being determined based on matching the inputted aggregated time series data to a corresponding cluster used to train the respective ensemble;

identifying, using a voting method, anomalies in the aggregated time series data based on respective portions of the aggregated time series data that are indicated as anomalous by a majority of deep learning models in the respective ensemble of the deep learning models;

grouping the anomalies according to relationships between the anomalies; and

performing a mitigation action in response to grouping the anomalies.

20. The computer program product according to claim 19 , wherein the clustered time series training data utilizes fuzzy clustering with dynamic time warping distance; and

wherein the respective ensemble of deep learning models comprises a first long short-term memory (LSTM) model with autoencoders, a second LSTM model with uncertainty estimation, and a third LSTM model with dropouts.