Data analytic engine towards the self-management of complex physical systems

1. A method for anomaly detection in complex physical systems, comprising:

extracting, using physical sensors deployed in a complex physical system, features representative of a temporal evolution of the complex physical system;

analyzing, in real time, the extracted features, the analyzing further comprising:

deriving one or more vector trajectories by performing sliding window segmentation of one or more time series;

applying a linear test to determine whether the one or more vector trajectories are linear or nonlinear; and

performing linear or nonlinear subspace decomposition on the vector trajectory based on the linear test;

generating, using a processor, a system evolution model from an ensemble of models based on the analysis of the extracted features of the complex physical system, wherein each of the ensemble of models corresponds to different data properties of the complex physical system;

determining a fitness score by analyzing a plurality of the different data properties of the complex physical system based on specific data dependency relationships, the type of analysis being determined based on particular properties of the features extracted and corresponding aspects of the complex physical system;

generating an alarm if the fitness score exceeds a predetermined number of threshold violations for any of the plurality of different data properties; and

modeling system dynamics of the complex physical system, the modeling comprising:

generating one or more time series models based on at least one of the fitness score or the alarm; and

for each selected time series, generating a profile {θ, F, Δ}, with F representing a feature space, to enable checking future observations based on the profile with learned model parameters θ and a maximum error Δ to be obtained when fitting data to the model based on θ, predicting new observations based on the profile and past observation, and determining if a deviation is caused by a system fault in the complex physical system rather than noise based on the checking and the predicting.

2. The method as recited in claim 1 , further comprising controlling operations of the complex physical system using a controller located remotely from the complex physical system.

3. The method as recited in claim 1 , wherein a window size is increased during the sliding window segmentation to improve detection precision.

4. The method as recited in claim 1 , further comprising controlling operations of the complex physical system remotely, the controlling including implementing corrective action responsive to an indication of a determined system fault.

5. The method as recited in claim 1 , wherein the system evolution model is constructed using a density based approach for a time series which lack smoothness in the one or more vector trajectories.

6. The method as recited in claim 1 , wherein the system evolution model monitors a current system status of the complex physical system, and compares the current system status with the system evolution model to detect anomalies.

7. A system for anomaly detection in complex physical systems, comprising:

a processor coupled to a memory, the processor being configured for:

extracting, using physical sensors deployed in a complex physical system, features representative of a temporal evolution of the complex physical system;

analyzing, in real time, the extracted features, the analyzing further comprising:

deriving one or more vector trajectories by performing sliding window segmentation of one or more time series;

applying a linear test to determine whether the one or more vector trajectories are linear or nonlinear; and

performing linear or nonlinear subspace decomposition on the vector trajectory based on the linear test;

generating a system evolution model from an ensemble of models based on the analysis of the extracted features of the system, wherein each of the ensemble of models corresponds to different data properties of a dynamical system;

determining a fitness score by analyzing, using an analytic engine, a plurality of the different data properties of the dynamical system based on specific data dependency relationships, the type of analysis being determined based on particular properties of the features extracted and corresponding aspects of the complex physical system;

generating an alarm if the fitness score exceeds a predetermined number of threshold violations for any of the plurality of different data properties; and

modeling system dynamics of the complex physical system, the modeling comprising:

generating one or more time series models based on at least one of the fitness score or the alarm; and

for each selected time series, generating a profile {θ, F, Δ}, with F representing a feature space, to enable checking future observations based on the profile with learned model parameters θ and a maximum error Δ to be obtained when fitting data to the model based on θ, predicting new observations based on the profile and past observation, and determining if a deviation is caused by a system fault in the complex physical system rather than noise based on the checking and the predicting.

8. The system as recited in claim 7 , wherein a window size is increased during the sliding window segmentation to improve detection precision.

9. The system as recited in claim 7 , the processor being further configured for controlling operations of the complex physical system remotely, the controlling including implementing corrective action responsive to an indication of a determined system fault.

10. The method as recited in claim 1 , wherein the system evolution model is constructed using a density based approach for a time series which lack smoothness in the one or more vector trajectories.

11. The method as recited in claim 1 , wherein the system evolution model monitors a current system status of the complex physical system, and compares the current system status with the system evolution model to detect anomalies.

12. A computer program product for anomaly detection in one or more complex physical systems, the computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions executable by a processor to cause the processor to perform a method comprising:

extracting, using physical sensors deployed in a complex physical system, features representative of a temporal evolution of the complex physical system;

analyzing, in real time, the extracted features, the analyzing further comprising:

deriving one or more vector trajectories by performing sliding window segmentation of one or more time series;

applying a linear test to determine whether the one or more vector trajectories are linear or nonlinear; and

performing linear or nonlinear subspace decomposition on the vector trajectory based on the linear test;

generating, using a processor, a system evolution model from an ensemble of models based on the analysis of the extracted features of the complex physical system, wherein each of the ensemble of models corresponds to different data properties of the complex physical system;

determining a fitness score by analyzing a plurality of the different data properties of the complex physical system based on specific data dependency relationships, the type of analysis being determined based on particular properties of the features extracted and corresponding aspects of the complex physical system;

generating an alarm if the fitness score exceeds a predetermined number of threshold violations for any of the plurality of different data properties; and

modeling system dynamics of the complex physical system, the modeling comprising:

generating one or more time series models based on at least one of the fitness score or the alarm; and

for each selected time series, generating a profile {θ, F, Δ}, with F representing a feature space, to enable checking future observations based on the profile with learned model parameters θ and a maximum error Δ to be obtained when fitting data to the model based on θ, predicting new observations based on the profile and past observation, and determining if a deviation is caused by a system fault in the complex physical system rather than noise based on the checking and the predicting.

13. The computer-readable storage medium as recited in claim 12 , wherein the method further comprises controlling operations of the complex physical system using a controller located remotely from the complex physical system.

14. The computer-readable storage medium as recited in claim 12 , wherein a window size is increased during the sliding window segmentation to improve detection precision.

15. The computer-readable storage medium as recited in claim 12 , wherein the method further comprises controlling operations of the complex physical system remotely, the controlling including implementing corrective action responsive to an indication of a determined system fault.

16. The computer-readable storage medium as recited in claim 12 , wherein the system evolution model is constructed using a density based approach for a time series which lack smoothness in the one or more vector trajectories.

17. The computer-readable storage medium as recited in claim 12 , wherein the system evolution model monitors a current system status of the complex physical system, and compares the current system status with the system evolution model to detect anomalies.