SPICE: Streaming PCA Fault Identification and Classification Engine in Predictive Maintenance

Abstract

Data-driven predictive maintenance needs to understand high-dimensional “in-motion” data, for which fundamental machine learning tools, such as Principal Component Analysis (PCA), require computation-efficient algorithms that operate near-real-time. Despite the different streaming PCA flavors, there is no algorithm that precisely recovers the principal components as the batch PCA algorithm does, while maintaining low-latency and high-throughput processing. This work introduces a novel processing framework, employing temporal accumulate/retract learning for streaming PCA. The framework is instantiated with several competitive PCA algorithms with proven theoretical advantages. We benchmark the framework in a real-world predictive maintenance scenario (i.e. fault classification in a coal coke production line) and prove its low-latency (millisecond level) and high-throughput (thousands events/second) processing guarantees.