Source-linked AI summary
Multi-Sensor Prognostics using an Unsupervised Health Index based on LSTM Encoder-Decoder
Pankaj Malhotra, Vishnu TV, Anusha Ramakrishnan, Gaurangi Anand, Lovekesh Vig, Puneet Agarwal, Gautam Shroff
TL;DR
Fixed degradation assumptions can be unreliable for RUL estimation from machine sensor data. The paper learns an unsupervised LSTM-ED health index from healthy-state reconstruction error and uses it for RUL estimation, reporting promising results across turbofan and milling datasets and some cases outperforming assumption-based models.
Problem
RUL methods often assume exponential or linear degradation, although real-world degradation may not follow a fixed shape.
Method
An LSTM encoder-decoder reconstructs healthy multivariate sensor time series, and pointwise reconstruction error provides an unsupervised health index for RUL estimation.
Results
The health index captures degradation, and the approach achieves comparable performance without domain-specific or exponential/linear degradation assumptions.
Takeaways & Limitations
The approach shows promising overall results and in some cases outperforms models relying on assumed health-degradation behavior.
Takeaways & Limitations
The benchmark comparison with other turbofan-dataset methods is not exhaustive.
Abstract
from arXiv · showhide
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.
1. INTRODUCTION
Sensor data enables condition-based maintenance and RUL prediction, but real-world degradation often violates fixed-shape assumptions. The paper proposes an unsupervised LSTM-ED health index based on reconstruction error rather than prescribed degradation models.
- Motivation: Sensor readings can support condition-based maintenance and prediction of a machine’s remaining useful life.Condition-based maintenance can yield financial savings compared with time-based maintenance.
- Challenges: Traditional RUL methods often assume exponential or linear degradation curves, but these assumptions may not hold in real-world datasets.Machines can differ in degradation timing, initial wear, and degradation shape.
- Challenges: Practical prognostics is complicated by noisy sensors and limited sensor data through end-of-life because machines undergo periodic maintenance.
- Proposed approach: LSTM-ED learns normal multivariate time-series behavior and uses pointwise reconstruction error to compute an unsupervised health index.The approach avoids assuming a particular degradation-curve shape.
- Findings: The learned health index decreases as the system degrades and can support RUL estimation without domain-specific or exponential/linear degradation assumptions.The paper reports comparable performance while avoiding those assumptions.
- Paper organization: The paper covers linear-regression and LSTM-ED health-index estimation, followed by RUL estimation from health-index curves.
2. APPROACH OVERVIEW
The approach represents each machine as normalized multivariate sensor time series, optionally reduces sensor correlations with PCA, and accommodates operating regimes separately. RUL is estimated from the resulting sequence representation.
- Problem formulation: The task is to estimate a currently operating instance’s RUL from its sensor history using historical instances that extend to end-of-life.
- Data representation: Each instance is represented by a multivariate time series with one vector of readings for each cycle and sensor.The sensor data is z-normalized before further processing.
- Data representation: Principal Components Analysis produces derived sensors with reduced linear correlations, with p denoting the number of retained components.The best p can be selected using a validation set.
- Operating regimes: The basic formulation assumes one operating regime across the system’s life, while separate models can handle multiple regimes.The paper describes this extension for the milling-machine case study.
3. LINEAR REGRESSION BASED HEALTH INDEX ESTIMATION
The linear-regression formulation maps derived sensor readings to a health index curve, whose target values are specified through degradation assumptions. The fitted mapping is then applied to train and test instances for RUL estimation.
- Health-index formulation: The health-index curve assigns a scalar health value to each cycle of an instance across its total life.The formulation uses values of 1 for healthy behavior and 0 near an unacceptable or failing condition.
- Regression model: A mapping fθ from derived sensor readings to health-index values is fitted using Ordinary Least Squares on target curves for training instances.
- Target-curve assumptions: The conventional target curve assumes exponential health degradation, with β controlling the fraction of cycles assigned endpoint values.The starting and ending β fractions of cycles receive health-index values of 1 and 0, respectively.
- Target-curve assumptions: An alternative target assigns health-index values of 1 to healthy-condition data and 0 to failure-condition data.This alternative uses only sensor readings from those endpoint conditions to learn the regression model.
- RUL estimation: The fitted regression parameters generate health-index curves for training and test instances, which are later compared to estimate test-instance RUL.
4. LSTM-ED BASED TARGET HI CURVE
The method trains an LSTM encoder-decoder on healthy subsequences and converts reconstruction errors across overlapping windows into a target health-index curve. This unsupervised HI supports RUL estimation without requiring a predefined degradation shape.
- Reconstruction model: Healthy subsequences train an LSTM encoder-decoder to reconstruct multivariate time-series data.The encoder produces a fixed-length representation, while the decoder reconstructs the sequence.
- Reconstruction error: Pointwise reconstruction error is used as an unnormalized health measure, with higher error indicating poorer health.The model is trained only on subsequences corresponding to perfect health, commonly taken from an instance’s first operational cycles.
- Reconstruction model: During inference, the decoder reconstructs subsequences and averages predictions for each point across all overlapping windows.Each point may be predicted multiple times because subsequences overlap; their average becomes the final point prediction.
- Reconstruction error: The reconstruction errors are normalized into target HI values for each training instance, and squared errors can make large errors produce much smaller HI values.These target HI curves are then used to estimate the linear-regression parameters for subsequent RUL modeling.
5. RUL ESTIMATION USING HI CURVE MATCHING
RUL is estimated by matching a test instance’s HI curve to time-shifted HI curves from training instances. Similarity-weighted remaining-life estimates are combined while smoothing noise and limiting predictions when early-life RUL is difficult to estimate.
- HI curve matching: The test HI curve is matched against training HI curves using different time-lags to account for differing degradation timelines.The selected lag minimizes Euclidean distance between the curves.
- RUL aggregation: Multiple high-similarity time-lags are combined rather than selecting only the single minimum-distance match.This produces a final estimate based on the total life of the matched training instances.
- Practical adjustments: Moving-average smoothing reduces non-monotonic HI noise, while a maximum RUL bound addresses uncertainty during very early healthy operation.The paper notes that early-life RUL is difficult to estimate before a fault has appeared.
- RUL estimation: For a matched training instance and time-lag, the RUL estimate equals the training instance’s remaining cycles after the test instance’s observed segment.The estimate is expressed as L(u) − L(u∗) − t.
- RUL aggregation: Similarity scores weight the candidate RUL estimates, and only sufficiently similar matches contribute to the final weighted average.The threshold α controls how many candidate estimates are retained.
6. EXPERIMENTAL EVALUATION
The evaluation applies the approach to turbofan engines using multiple RUL-related metrics and compares several health-index models. Reconstruction error generally increases through engine life, while LR-ED2 performs strongly against the evaluated alternatives.
- Evaluation setup: The evaluation uses Timeliness Score, Accuracy, MAE, MSE, and MAPE to assess RUL prediction performance.For test instance u*, the RUL error is defined as estimated RUL minus actual RUL; lower Timeliness Score indicates better performance.
- Evaluation setup: The turbofan dataset contains 24 sensor readings for 100 training and 100 test engines, with test trajectories truncated before failure.The task is to predict RUL for the 100 test engines using provided actual RUL values.
- Turbofan results: Average LSTM-ED reconstruction error increases as the fraction of engine life passed increases, supporting its use as a health indicator.Initial-life reconstruction error is small, while later-life error is higher; reconstruction error is interpreted as poorer health when larger.
- Turbofan results: RUL estimates from the LSTM-ED health index are fairly accurate, although reconstruction error at a given life fraction varies substantially across engines.The evaluation also examines RUL estimates, prediction-error histograms, and error relative to health index at the final cycle.
- Model comparison: LR-ED2 performs significantly better than the other three evaluated models and achieves performance comparable to RULCLIPPER without assuming exponential degradation.The comparison includes LSTM-ED, LR-Exp, LR-ED1, and LR-ED2; RULCLIPPER is reported as a literature benchmark.
- Model comparison: The worst test instance contributes 23% for LR-Exp and LR-ED2, versus 17% for LR-ED1, of the Timeliness Score.For LR-Exp and LR-ED2, the worst instance contributes nearly one-quarter of the total score, while the other 99 instances have very good timeliness scores.
6.3 Milling Machine Dataset
The milling-machine study uses multivariate sensor and tool-wear data across two material regimes to evaluate LSTM-ED reconstruction errors as health indicators and RUL inputs. Reconstruction error increases with life passed, while the resulting models provide cycle-level RUL estimates and show a relationship between error and maintenance cost in the pulverizer study.
- Dataset: The dataset contains 10 variables across 167 runs, with six dependent sensors and two material-specific operating regimes.There are 109 runs for material 1 and 58 for material 2; each run is treated as one tool-life cycle.
- Method: The LSTM-ED is trained on the first run of each case, and average run reconstruction error becomes that cycle’s target health index.Mean and standard deviation features are derived from 9000-point sequences for six sensors.
- Model selection: The best configurations use PCA1 for material 1 and LR-ED1 for material 2, with model parameters selected using minimum MAPE1.The selected settings are p = 1, λ = 0.025, α = 0.98, τ = 15 for material 1 and p = 2, λ = 0.005, α = 0.87, τ = 13, c = 45 for material 2.
- Health-index behavior: Average LSTM-ED reconstruction error increases with the fraction of tool life passed for both materials, making it an appropriate health indicator.Figure 7 reports reconstruction-error trends against cycles passed and error histograms.
- RUL results: Around 70% of RUL prediction errors fall within [-4, 6] for material 1 and [-3, 1] for material 2.The results use interpolated cycle-level data, with Figure 8 showing predicted and actual RULs after interpolation.
- Industrial validation: For the pulverizer mill, days with reconstruction error E > tE include a large fraction (>0.61) with high ongoing maintenance cost C > tC.The authors interpret this significant correlation as evidence that reconstruction error captures mill health.
7. RELATED WORK
The related work covers physical, reliability, statistical, Bayesian, regression, and neural approaches to prognostics and anomaly detection. The paper distinguishes its LSTM-ED approach by learning health-related reconstruction behavior from normal multivariate sequences without imposing a fixed degradation-curve form.
- Statistical and Bayesian methods: Bayesian methods model sensor-variable distributions or recursively estimate health indices without assuming a specific health-index curve form.The paper also contrasts these methods with composite health indices based on exponential assumptions.
- Trajectory similarity: Trajectory-similarity methods estimate RUL by matching health-index or PCA-derived curves, with this paper instead obtaining a univariate HI from LSTM-ED reconstruction errors.The comparison concerns how health trajectories are represented before RUL estimation.
- Neural prognostics: Deep CNNs have been proposed to learn abstract features from raw sensor data and directly estimate RUL, with comparisons reported on the Turbofan Engine dataset.The cited work is also compared with multilayer perceptron, support-vector, and relevance-vector regression methods.
- LSTM and anomaly detection: LSTM-based anomaly methods predict future normal time-series and use prediction errors as health or novelty measures, whereas this work applies LSTM-ED reconstruction to RUL estimation.Related reconstruction models have primarily been evaluated for novelty or anomaly detection.
8. DISCUSSION
The discussion presents an unsupervised LSTM-ED health index learned from healthy multivariate time-series and used for RUL estimation. Results are promising across the Turbofan Engine and Milling Machine datasets, while pulverizer-mill evidence links reconstruction error with maintenance cost and fault severity.
- Approach: The proposed method learns a reconstruction model from healthy system behavior and uses its output to estimate an unsupervised health index.The health index is derived from multi-sensor time-series data.
- RUL estimation: The unsupervised HI supports RUL estimation without relying on domain knowledge or assumed exponential degradation models.The discussion frames this as an alternative to degradation models whose fixed forms may not capture instance-specific behavior.
- Benchmark findings: The approach shows promising overall results and in some cases outperforms models that rely on health-degradation assumptions on the Turbofan Engine and Milling Machine datasets.The paper-level conclusion concerns both benchmark datasets rather than a single metric.
- Industrial case study: The pulverizer-mill case study shows signs of correlation between LSTM-ED reconstruction error and maintenance cost, suggesting that the method captures fault severity.This is presented as evidence from a real-world industry dataset.
A. BENCHMARKS ON TURBOFAN ENGINE DATASET
The supplied passage identifies Table 4 as a performance comparison of approaches on the Turbofan Engine dataset. It does not provide the table’s numerical results or a supported comparison outcome.
- Benchmark comparison: Table 4 compares the performance of various approaches on the Turbofan Engine dataset.NR denotes “Not Reported” in the table.