Hybrid clustering-partitioning techniques that optimizes accuracy and compute cost for prognostic surveillance of sensor data
The disclosed embodiments relate to a system for performing prognostic surveillance operations on sensor data. During operation, the system obtains a group of signals from sensors in a monitored system during operation of the monitored system. Next, if possible, the system performs a clustering operation, which divides the group of signals into groups of correlated signals. Then, for one or more groups of signals that exceed a specified size, the system randomly partitions the groups of signals into smaller groups of signals. Next, for each group of signals, the system trains an inferential model for a prognostic pattern-recognition system based on signals in the group of signals. Then, for each group of signals, the system uses a prognostic pattern-recognition system in a surveillance mode and the inferential model to detect incipient anomalies that arise during execution of the monitored system.
1. A method for performing prognostic surveillance on a monitored system, comprising: obtaining a group of signals from sensors in the monitored system during operation of the monitored system; performing a cluster operation, which divides the group of signals into groups of correlated signals if possible; for one or more groups of signals that exceed a specified size, randomly partitioning the groups of signals into smaller groups of signals; for each group of signals, training an inferential model for a prognostic pattern-recognition system based on signals in the group of signals; for each group of signals, using a prognostic pattern-recognition system in a surveillance mode and the inferential model to detect incipient anomalies that arise during execution of the monitored system; and obtaining from a user a pre-specified balance point between compute cost and prognostic accuracy; wherein randomly partitioning the groups of signals into smaller groups of signals comprises using a dual-objective iterative partitioning procedure, which optimizes a tradeoff according to the pre-specified balance point between prognostic accuracy and compute cost during subsequent prognostic surveillance operations.
2. The method of claim 1 , wherein the dual-objective iterative partitioning procedure iteratively:
partitions the groups of signals into smaller groups of signals in a random manner, without considering correlations between signals in the groups of signals;
trains and runs the inferential model on the smaller groups of signals;
determines a compute cost and an associated prognostic accuracy involved in training and running the inferential model on the smaller groups of signals; and
terminates the iterative procedure when a balance point is reached between compute cost and prognostic accuracy.
3. The method of claim 1 , wherein prior to performing the dual-objective iterative partitioning procedure, the method further comprises allowing a user to pre-specify the balance point between compute cost and prognostic accuracy.
4. The method of claim 1 , wherein performing the clustering operation involves using one or more of the following techniques:
k-means clustering; and
tri-point clustering.
5. The method of claim 1 , wherein the inferential model is trained using a nonlinear, nonparametric (NLNP) regression technique.
6. The method of claim 5 , wherein the NLNP regression technique comprises a Multivariate State Estimation Technique (MSET).
7. The method of claim 1 , wherein using the prognostic pattern-recognition system to detect an incipient anomaly comprises detecting one or more of the following:
an impending failure of the monitored system; and
a malicious-intrusion event in the monitored system.
8. The method of claim 1 , wherein obtaining the group of signals comprises:
obtaining the group of signals from the sensors during operation of the monitored system;
storing the group of signals in a time-series database; and
subsequently retrieving the group of signals from the time-series database for use in a dual-objective iterative partitioning procedure.
9. A non-transitory, computer-readable storage medium storing instructions that when executed by a computer cause the computer to perform a method for performing prognostic surveillance on a monitored system, the method comprising: obtaining a group of signals from sensors in the monitored system during operation of the monitored system; performing a cluster operation, which divides the group of signals into groups of correlated signals if possible; for one or more groups of signals that exceed a specified size, randomly partitioning the groups of signals into smaller groups of signals; for each group of signals, training an inferential model for a prognostic pattern-recognition system based on signals in the group of signals; and for each group of signals, using a prognostic pattern-recognition system in a surveillance mode and the inferential model to detect incipient anomalies that arise during execution of the monitored system; and obtaining from a user a pre-specified balance point between compute cost and prognostic accuracy; wherein randomly partitioning the groups of signals into smaller groups of signals comprises using a dual-objective iterative partitioning procedure, which optimizes a tradeoff according to the pre-specified balance point between prognostic accuracy and compute cost during subsequent prognostic surveillance operations.
10. The non-transitory, computer-readable storage medium of claim 9 , wherein the dual-objective iterative partitioning procedure iteratively:
partitions the groups of signals into smaller groups of signals in a random manner, without considering correlations between signals the groups of signals;
trains and runs the inferential model on the smaller groups of signals;
determines a compute cost and an associated prognostic accuracy involved in training and running the inferential model on the smaller groups of signals; and
terminates the iterative procedure when a balance point is reached between compute cost and prognostic accuracy.
11. The non-transitory, computer-readable storage medium of claim 9 , wherein prior to performing the dual-objective iterative partitioning procedure, the method further comprises allowing a user to pre-specify the balance point between compute cost and prognostic accuracy.
12. The non-transitory, computer-readable storage medium of claim 9 , wherein performing the clustering operation involves using one or more of the following techniques:
k-means clustering; and
tri-point clustering.
13. The non-transitory, computer-readable storage medium of claim 9 , wherein the inferential model is trained using a nonlinear, nonparametric (NLNP) regression technique.
14. The non-transitory, computer-readable storage medium of claim 13 , wherein the NLNP regression technique comprises a Multivariate State Estimation Technique (MSET).
15. The non-transitory, computer-readable storage medium of claim 9 , wherein using the prognostic pattern-recognition system to detect an incipient anomaly comprises detecting one or more of the following:
an impending failure of the monitored system; and
a malicious-intrusion event in the monitored system.
16. The non-transitory, computer-readable storage medium of claim 9 , wherein obtaining the group of signals comprises:
obtaining the group of signals from the sensors during operation of the monitored system;
storing the group of signals in a time-series database; and
subsequently retrieving the group of signals from the time-series database for use in a dual-objective iterative partitioning procedure.
17. A system that performs prognostic-surveillance operations on a monitored system, comprising: at least one processor and at least one associated memory; and a prognostic-surveillance mechanism that executes on the at least one processor, wherein during operation, the prognostic-surveillance mechanism: obtains a group of signals from sensors in the monitored system during operation of the monitored system; performs a cluster operation, which divides the group of signals into groups of correlated signals if possible; for one or more groups of signals that exceed a specified size, randomly partitions the groups of signals into smaller groups of signals; for each group of signals, trains an inferential model for a prognostic pattern-recognition system based on signals in the group of signals; for each group of signals, uses a prognostic pattern-recognition system in a surveillance mode and the inferential model to detect incipient anomalies that arise during execution of the monitored system; and obtains from a user a pre-specified balance point between compute cost and prognostic accuracy; wherein while randomly partitioning the groups of signals into smaller groups of signals, the prognostic-surveillance mechanism uses a dual-objective iterative partitioning procedure, which optimizes a tradeoff according to the pre-specified balance point between prognostic accuracy and compute cost during subsequent prognostic surveillance operations.