Source-linked AI summary
Integrating Survival-Based Aging Models with Data-Driven RUL Prognostics
Abhishek Srinivasan, Juan Carlos Andresen, Sepideh Pashami, Anders Holst
TL;DR
Reliable RUL estimates matter for predictive maintenance, but wear-based and sensor-based models capture different, partial aspects of degradation. This paper fuses their failure probability distributions under explicit structural assumptions; on N-CMAPSS, the combined model improves accuracy while maintaining competitive C-index and narrower intervals. The results also show uncertainty-dependent shifts between the components.
Problem
Wear-based aging models and sensor-driven models each capture only partial information, motivating a way to use both for reliable RUL estimation.
Source: §1 · pp. 1–2Method
The framework probabilistically fuses wear-based and sensor-based prognostic components through failure distributions under structural assumptions linking wear, latent health, sensors, and failure.
Source: §1 · p. 2, §8 · p. 11Results
The combined model achieves lower RMSE while maintaining competitive c-index and producing narrower prediction intervals than either component individually.
Source: §7 · p. 9Takeaways & Limitations
Wear-based modeling provides stability in early life, while sensor-based estimation dominates as degradation becomes observable, with switching governed by relative predictive uncertainty.
Source: §8 · p. 11Evaluation uses simulated N-CMAPSS data that may not fully reflect real-world fleets, and the structural assumptions may be violated in some settings.
Source: §7 · pp. 10–11
Abstract
from arXiv · showhide
Predictive maintenance requires reliable remaining useful life (RUL) estimation. Existing methods mainly follow two paradigms: wear-based aging models that capture cumulative degradation and sensor-driven data models that reflect instantaneous health conditions, each providing only partial information. In this work, we propose a probabilistic fusion framework that integrates wear-based and sensor-based prognostic components through failure probability distributions. Based on explicit structural assumptions linking wear, latent health, sensor observations, and failure, we derive a principled combination rule that enables uncertainty-aware integration with adaptive weighting of the components. Experimentally, we assess this combination rule by learning the wear-based component using a parametric survival model and the sensor-based component using a 1D convolutional neural network (1D-CNN) with a post-hoc uncertainty model. Evaluation on multiple N-CMAPSS datasets demonstrates that the fused model improves point accuracy, preserves the C-index, and produces narrower yet well-calibrated prediction intervals compared to either component alone. The results highlight the complementary roles of wear-based survival model and sensor-based deep learning model, and show that their probabilistic integration provides a structured pathway toward more robust and consistent prognostics over the life-time.
1 Introduction
Wear-based models represent long-term aging, while sensor-based models reflect current condition, but each captures only part of system degradation. The paper proposes probabilistic fusion of both sources; on N-CMAPSS, the fused model improves prediction accuracy and yields narrow, well-calibrated intervals compared with either component alone.
1 Introduction
- Predictive maintenance uses condition monitoring to estimate future degradation and RUL for maintenance scheduling, risk management, and resource planning.The introduction places this need in the context of engineering systems including aircraft engines, energy systems, automobiles, and industrial equipment.Source: §1 · p. 1
- Wear-based models represent long-term aging but may miss instantaneous sensor-indicated health, while sensor-based models capture current condition but lack explicit aging structure.Sensor-based predictions may be unstable early in operation when degradation signals are weak, and temporal consistency is not guaranteed without additional constraints.Source: §1 · p. 2
- The study proposes a probabilistic framework that combines wear-based and sensor-based predictions into a failure probability distribution conditioned on both sources.The combination is motivated by an explicit structural assumption linking accumulated wear, latent health, sensor observations, and failure.Source: §1 · p. 2
- The fused model improves prediction accuracy, maintains competitive C-index, and produces well-calibrated narrow prediction intervals compared with either component alone.The authors also report that fusion shifts between wear-driven and sensor-driven predictions according to the components’ relative predictive uncertainty.Source: §1 · p. 2
2 Related Work
Prior work spans survival, RUL, and hybrid prognostics, but commonly treats wear and sensor information within one modeling framework or combines them at intermediate representations. This paper instead separates the two sources and fuses their failure-time distributions under explicit structural assumptions.
2.1 Survival Analysis for Time-to-Failure Modeling
- Parametric survival models estimate lifetime distributions from failure data, while Kaplan–Meier and Cox approaches relax distributional assumptions but retain their own assumptions.The Cox model assumes proportional hazards, which may not hold when degradation dynamics evolve.Source: §2.1 · p. 3
2.2 Remaining Useful Life (RUL) Prediction
- RUL approaches include reliability-based, physics-based, and data-driven models, with deep learning mapping sensor time series to RUL but often using indirect wear proxies.Reliability models may require strong degradation assumptions; physics-based models need domain knowledge and may not scale to interacting degradation modes.Source: §2.2 · p. 3
2.3 Bridging Survival Modeling and RUL Estimation
- Prior hybrid methods combine survival and RUL learning or fuse physical information with sensor inputs, states, or representations; this work instead combines distinct sources at the failure-distribution level.The stated contribution is a probabilistic combination mechanism rather than a particular hybrid model, potentially applicable to existing industrial prediction models.Source: §2.3 · p. 4
2.4 Relation to the Present Work
- The proposed rule treats wear and sensor observations as structurally distinct partial views of failure and combines P(F|W) and P(F|S) into P(F|W, S) under conditional independence assumptions.The paper makes the data-generating assumptions and probabilistic fusion mechanism explicit.Source: §2.4 · p. 4
3 Model Assumptions
The model links accumulated wear to latent health, which affects both sensor observations and failure. It treats health as remaining equivalent operating hours and uses that equivalence to express and factorize failure probabilities.
3 Model Assumptions
- The assumed data-generating structure has wear factors influence latent health, which in turn affects sensor observations and failure occurrence.The model treats health as hidden and not directly observable.Source: §3 · p. 4, Figure 1 · p. 5
- The model assumes a deterministic connection between health and failure, representing health as the remaining equivalent operating hours until failure.Under this equivalence, the distribution of health is treated as the distribution of time to failure.Source: §3 · pp. 4–5
- This health–failure equivalence allows the joint likelihood to be expressed using observable quantities and factorized using Bayes’ rule.The baseline P(F) represents failure probability without wear or sensor information and uses a constant hazard based on MTBF, yielding an exponential distribution.Source: §3 · p. 5
4 Method
The method estimates wear- and sensor-conditioned failure distributions, adds a baseline distribution, and combines the components probabilistically. The wear component uses a conditional Weibull survival model, while the sensor component derives a truncated Gaussian RUL distribution from CNN predictions and an uncertainty estimate.
4 Method
- The combined failure distribution uses wear-conditioned, sensor-conditioned, and baseline failure distributions as its three components.Each component supplies a failure probability density over future time.Source: §4 · p. 5
4.1 Wear-based failure model (P(F | W))
- The wear-based component uses a Weibull survival model and conditions its failure density on survival until the current time.It renormalizes by the survival probability at the current time to obtain a consistent conditional distribution.Source: §4.1 · p. 5
4.2 Sensor-based failure model (P(F | S))
- The sensor-based component uses sliding-window multivariate time series in a 1D-CNN to estimate mean RUL, then a separate network estimates its conditional standard deviation.The resulting Gaussian time-to-failure distribution is truncated to non-negative remaining life.Source: §4.2 · p. 6
4.3 Baseline failure distribution (P(F))
- Without wear or sensor information, the baseline assumes constant hazard λ0 = 1/MTBF, yielding a memory-less exponential failure distribution.The baseline predicts the same distribution irrespective of time, wear, or sensor information.Source: §4.3 · p. 6
4.4 Fusion of components
- The fusion equation combines the wear-, sensor-, and baseline-conditioned terms as an unnormalized density, then normalizes it numerically over the remaining-life grid.The implementation evaluates the terms as probability density functions over future remaining life; the grid runs from 0 to 128 cycles in 1-cycle increments.Source: §4.4 · p. 6
5 Experiments
The experiments evaluate wear-based, sensor-based, and fused prognostics on selected N-CMAPSS datasets using point-estimate and probabilistic metrics. The setup combines a Weibull survival model with a 1D-CNN and a post-processing uncertainty model.
5.1 Dataset and Experimental Setup
- The study evaluates DS01, DS03, and DS04, excluding DS02 because some time series were shorter than the 50-cycle window used by the 1D-CNN.N-CMAPSS provides simulated run-to-failure engine data with 14 sensors, 14 virtual sensors, and four operating conditions per cycle.Source: §5.1 · p. 6
- Two training units were randomly selected for validation, while the dataset’s predefined train and test splits were retained.Source: §5.1 · p. 7
- The wear-based component uses a Weibull survival model estimated by maximum likelihood from training failure times; its conditional RUL distribution is truncated and renormalized at the current cycle.Source: §5.1 · p. 7
- The sensor component uses a three-layer 1D-CNN, with a separate uncertainty network trained post hoc on validation data to produce a Gaussian predictive distribution.Source: §5.1 · p. 7
5.2 Evaluation Metrics
- Evaluation covers point accuracy, interval coverage and width, and ranking consistency using RMSE, PICP, NIPW, and C-index.The metrics assess point estimates and probabilistic predictions; PICP measures true-value inclusion and NIPW summarizes normalized interval width.Source: §5.2 · p. 7
6 Results
Across three N-CMAPSS datasets, the combined model performs favorably against the Weibull and 1D-CNN components on accuracy and interval sharpness while retaining competitive coverage and ranking consistency. Visual, error, and uncertainty-scaling analyses illustrate how the fusion behavior varies with lifecycle stage and component uncertainty.
6.1 Empirical Evaluation
- Across datasets, the combined model has the lowest RMSE, best NIPW, and comparable C-index, with a favorable overall tradeoff in accuracy, calibration, and sharpness.Source: §6.1 · pp. 7–8, Table 1 · p. 8
- The Weibull model attains the highest PICP, with reported values of 1.00 reflecting conservative but comparatively wide intervals.Source: §6.1 · p. 8
6.2 Visual Evaluation
- For DS01 unit 7, the Weibull prediction follows true RUL closely, while the combined model is stable early and incorporates improving sensor estimates later; DS04 illustrates poorer 1D-CNN performance.Source: §6.2 · p. 8, Figure 2 · p. 9
6.3 Error Analysis of Combined Model
- For DS01, the combined model follows the Weibull error profile at large true RUL, then shifts toward sensor-based behavior as failure approaches and 1D-CNN error decreases.Source: §6.3 · p. 8
6.4 Ablation Study
- Deflating the 1D-CNN standard deviation moves the combined estimate toward the 1D-CNN prediction, while inflating it shifts the estimate toward Weibull.The uncertainty-scaling experiment shows that the combined outcome varies systematically with the sensor component’s relative uncertainty.Source: §6.4 · pp. 8–9
7 Discussion
Across datasets, fusion improves RMSE while maintaining competitive C-index performance and narrower prediction intervals. Its behavior adapts with predictive uncertainty and system life, subject to assumptions about the relationship between health and failure and conditional independence of wear and sensor information.
7 Discussion
- Across datasets, the combined model achieves lower RMSE, maintains competitive C-index, and produces narrower prediction intervals than either component alone.The discussion describes this as a favorable balance between accuracy, calibration, and interval width.Source: §7 · p. 9
- As failure approaches, the combined estimate shifts from following the Weibull model toward the sensor-based estimate as sensor degradation information improves.In early operation, when true RUL is large, the combined model closely follows the Weibull model; Figure 4 describes the corresponding error transition for DS01.Source: §7 · p. 9, Figure 4 · p. 10
- Changing the sensor model’s estimated variance systematically changes component influence: lower variance favors the sensor estimate, while higher variance favors Weibull.The ablation attributes this behavior to relative predictive uncertainty rather than heuristic blending.Source: §7 · p. 9, Figure 5 · p. 11
- Fusion effectiveness depends on well-calibrated predictive uncertainty, since miscalibrated variances can cause unreliable predictions to receive excessive weight.The discussion identifies uncertainty calibration as important for practical deployment.Source: §7 · pp. 9–10
- The evaluation uses a Weibull survival model and a 1D-CNN, so performance and generality across alternative component models remain to be evaluated.The framework is described as model-agnostic, but the study tests only these specific model choices.Source: §7 · p. 10
- The experiments use simulated N-CMAPSS data, which may not capture the complexity and variability of real-world fleets.The discussion names maintenance actions, usage histories, operating regimes, sensor drift, missing data, and interacting degradation mechanisms as possible sources of real-world variability.Source: §7 · p. 10
- The fusion framework assumes a deterministic relationship between latent health and failure, and conditional independence between wear and sensor information given health.The discussion notes these assumptions may be violated, including when wear information is encoded in sensor measurements.Source: §7 · p. 11
8 Conclusion
The conclusion presents probabilistic fusion of wear-based survival modeling and sensor-based RUL estimation under explicit structural assumptions. On N-CMAPSS, the combined model improves prediction accuracy while retaining competitive C-index performance and narrower intervals, and offers a way to integrate data-driven models with existing survival approaches.
8 Conclusion
- The framework combines wear-based survival modeling and sensor-based RUL estimation at the level of failure probability distributions under explicit structural assumptions.The authors characterize this as principled uncertainty-aware fusion rather than heuristic model averaging.Source: §8 · p. 11
- On N-CMAPSS, the combined model improves prediction accuracy while maintaining competitive C-index performance and producing narrower prediction intervals.The conclusion also reports a transition from wear-based stability early in life toward sensor-based estimation as degradation becomes observable, governed by relative predictive uncertainty.Source: §8 · p. 11
- The framework offers a way to integrate advanced data-driven models with existing survival-based approaches rather than replacing them.The authors frame this as support for a gradual transition toward data-driven and AI-enabled maintenance strategies.Source: §8 · pp. 11–12
Motivation on the standard deviation Model
The passage motivates modeling heteroskedastic uncertainty because predicted and true RUL do not align perfectly and errors spread around the diagonal. It describes modeling the uncertainty term as a function of the predicted mean.
Motivation on the standard deviation Model
- True and predicted RUL do not perfectly align on the diagonal, and the scatter has a spread around it.Source: Motivation on the standard deviation… · p. 13
- That spread motivates modeling heteroskedastic uncertainty rather than relying on a point estimate alone.Source: Motivation on the standard deviation… · p. 13
- The uncertainty term is modeled as a function of the predicted mean, written as σ̂(μ̂(S)).Source: Motivation on the standard deviation… · p. 13
Validating the results across individual Units
Unit-level evaluation preserves the overall trends: the combined model performs best on NIPW, with Unit 9 showing lower interval coverage.
Validating the results across individual Units
- Across individual units, the combined model consistently achieves the best NIPW while maintaining comparable RMSE and PCIP, except on Unit 9, where PCIP is lower.The figure caption describes the per-unit evaluation as verifying consistent trends for DS01.Source: Validating the results across… · p. 13, Figure 7 · p. 14
- For Unit 9, raising the PICP confidence level from α = 0.95 to α = 0.99 raises PICP to 0.81, indicating its 95% predictive interval is too narrow.The passage attributes this result to local uncertainty miscalibration.Source: Validating the results across… · p. 13
Additional Information about the data
- Experiments use DS01, DS03, and DS04, with 20-step tumbling-window averages and length-50 sliding windows.The subsets contain 4, 6, and 4 test units, respectively.Source: Additional Information about the data · p. 13
- The 1D-CNN scatter plot shows increasing prediction spread at larger RUL values, motivating heteroskedastic uncertainty modeling.Source: Figure 6 · p. 14