Limitations of EWMA and CUSUM for Early Detection:
How LSTM and GRU Complement Plant Control

Introduction

Classical statistical methods like EWMA and CUSUM work well for monitoring individual plant variables — they're proven, interpretable, and computationally lightweight. But they struggle with a specific and increasingly common challenge: complex predictive patterns hidden in multivariate noise.

As plants add more sensors and data streams, the question is no longer just "has this variable drifted?" It becomes "what combination of signals, across time, predicts a failure three hours from now?" That's a different problem — and it requires different tools.

The Technical Concept

EWMA and CUSUM monitor univariate variables — vibration, temperature, pressure — for mean and dispersion changes. They are excellent at detecting when a single variable drifts from its expected behaviour. Statistical process control built on these methods has decades of validated industrial application.

Recurrent neural networks (RNNs), specifically LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Unit) architectures, complement these by identifying patterns across multivariate time series that traditional methods cannot detect. They learn temporal dependencies — the sequence of events that precedes a failure — across multiple simultaneous data streams.

The Real Problem

Successful deployment of LSTM/GRU in industrial settings requires three things that are rarely all present at once:

• Synchronised, labelled multivariate data — all sensor streams aligned to a common timeline and linked to actual failure events, not just alarm triggers.
• Computational capacity — real-time inference from recurrent models demands more than a standard PLC. Edge hardware selection matters.
• Interpretability — industrial operators and maintenance teams need to understand and trust the output. A black-box prediction is often rejected regardless of its accuracy.

Practical Implications

Implementation requires integrating multiple sensors into a unified time-series pipeline, building a labelled event history that connects process data to real failures (not just alarms), and deploying appropriate edge hardware for real-time analysis.

The strategic framing: LSTM and GRU enrich classical methods rather than replacing them. EWMA and CUSUM remain the backbone of univariate process control. RNNs add a predictive layer for multivariate failure patterns. Both have a role — the mistake is assuming one makes the other obsolete.

Start with the data infrastructure. Without synchronised, labelled, contextualised multivariate history, even the best model architecture will underperform.