Autoencoder Ensemble for Time Series Fault Localization

The patent segments the computational task by dividing the ensemble of autoencoder models into multiple independent components that can be trained and executed in parallel. Each autoencoder in the ensemble processes the data independently, allowing the system to avoid the quadratic time complexity of cross-correlation while maintaining anomaly detection capability through aggregated reconstruction errors from multiple models.

The patent replaces the mechanical cross-correlation computation with a neural network-based autoencoder system. Instead of computing pairwise correlations between time series data points (which is computationally expensive), the system uses autoencoders to learn compressed representations and reconstruct the data, with reconstruction errors serving as anomaly indicators. This substitution dramatically reduces computational complexity while preserving detection effectiveness.

The patent merges multiple simple autoencoder models into an ensemble system where each model is relatively simple individually, but their collective output provides robust anomaly detection. The complexity is distributed across multiple independent models rather than concentrated in a single complex model, allowing parallel processing and improved efficiency while maintaining manageable individual model complexity.

The patent changes the parameter of using multiple models with different random seeds rather than a single complex model. By varying the random seeds during training, each autoencoder in the ensemble learns slightly different feature representations, and their aggregated reconstruction errors provide more reliable anomaly detection. This parameter change (from single model to ensemble with varied seeds) improves computational efficiency through parallelization while keeping individual model complexity low.