State Discovery and Prediction from Multivariate Sensor Data

The advent of cloud computing and autonomous data centers operating fully without human supervision has highlighted the need for fault-tolerant architectures and intelligent software tools for system parameter optimization. Demands on computational throughput have to be balanced with environmental concerns, such as energy consumption and waste heat. Using multivariate time series data collected from an experimental data center, we build a state model using clustering, then estimate the states represented by the clusters using both a hidden Markov model and a long-short term memory neural net. Knowledge of future states of the system can be used to solve tasks such as reduced energy consumption and optimized resource allocation in the data center.