[Paper Review] Multi-Sensor Prognostics using an Unsupervised Health Index based on LSTM Encoder-Decoder

Pankaj Malhotra, Vishnu Tv+5 more|arXiv (Cornell University)|Aug 22, 2016

Machine Fault Diagnosis TechniquesEngineering39 references165 citations

TL;DR

The paper introduces an unsupervised health index (HI) derived from a LSTM encoder-decoder to estimate RUL from multi-sensor time-series, without assuming degradation shape, and evaluates on Turbofan, Milling, and Pulverizer datasets.

ABSTRACT

Many approaches for estimation of Remaining Useful Life (RUL) of a machine, using its operational sensor data, make assumptions about how a system degrades or a fault evolves, e.g., exponential degradation. However, in many domains degradation may not follow a pattern. We propose a Long Short Term Memory based Encoder-Decoder (LSTM-ED) scheme to obtain an unsupervised health index (HI) for a system using multi-sensor time-series data. LSTM-ED is trained to reconstruct the time-series corresponding to healthy state of a system. The reconstruction error is used to compute HI which is then used for RUL estimation. We evaluate our approach on publicly available Turbofan Engine and Milling Machine datasets. We also present results on a real-world industry dataset from a pulverizer mill where we find significant correlation between LSTM-ED based HI and maintenance costs.

Motivation & Objective

• Motivate condition-based maintenance by predicting RUL without assuming a fixed degradation form.
• Propose an unsupervised HI learned from normal-operation reconstruction errors.
• Demonstrate HI-based RUL estimation via curve matching across training instances.

Proposed method

• Train an LSTM-Encoder-Decoder to reconstruct healthy multivariate time-series subsequences.
• Compute reconstruction error as an unnormalized health indicator; derive a target HI per instance.
• Learn a linear regression mapping from derived sensor features to a HI in [0,1].
• Use HI curves from training instances to estimate RUL of a test instance via curve matching with time-lags.
• Optionally use LR models with linear (LR-Lin) or exponential (LR-Exp) HI targets for comparison.
• Match HI curves between test and train instances with weighted averaging over similar time-lags.

Experimental results

Research questions

• RQ1 Can an unsupervised HI learned from reconstruction error capture degradation without assuming a degradation form?
• RQ2 How well can HI-based RUL estimation perform relative to domain-knowledge-based degradation models?
• RQ3 Does HI curve matching across instances provide robust RUL estimates under varying initial health and operating conditions?
• RQ4 What is the impact of using LR versus LSTM-ED-derived HI targets on RUL accuracy?

Key findings

• LSTM-ED based HI tracks degradation as reconstruction error increases over life.
• LR-ED 2 (normalized squared reconstruction error) outperforms LR-Exp, LR-ED 1, and LR-ED 2 variants on Turbofan data.
• LR-ED 2 achieves performance comparable to or better than the RUL-CLIPPER baseline without relying on exponential degradation assumptions.
• RUL estimates improve as HI at the last cycle decreases, with lower variance closer to end-of-life.
• HI-based methods show competitive results across datasets, including milling and pulverizer mill, with meaningful correlation to maintenance costs.