Pith. sign in

REVIEW 3 major objections 5 minor 39 references

This paper argues that remaining useful life and failure-mode probabilities can be learned as temporally consistent value predictions on an absorbing degradation process, using multi-step temporal-difference targets that work even when traj

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 05:16 UTC pith:D6MHMMXZ

load-bearing objection A competent, honestly reported GVF/TD framing for prognostics that deserves refereeing; the main caveat is the unquantified finite-window Markov approximation, though the empirical story is mostly sound. the 3 major comments →

arxiv 2607.22268 v1 pith:D6MHMMXZ submitted 2026-07-24 stat.ML cs.LGstat.AP

General Value Functions for Remaining Useful Life and Failure-Mode Prediction

Hao Yan , Ali Sarabi , Qing Zou , Boyang Xu This is my paper
classification stat.ML cs.LGstat.AP MSC 62N0568T05
keywords remaining useful lifefailure-mode predictiongeneral value functionstemporal-difference learningabsorbing Markov processBellman equationpredictive maintenancebootstrapped targets
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that RUL and failure-mode prediction should be treated as vector general value functions on a Markov degradation process that ends in an absorbing failure set, rather than as independent supervised window-label tasks. The central target is the Bellman recursion: a prediction at the current state must agree with the immediate transition plus the prediction from the successor state. The authors prove this recursion identifies a unique fixed point, characterize the linear projected temporal-difference limit, and show that oracle bootstrapped targets have no larger variance than complete-return Monte Carlo targets. On a synthetic event-triggered simulation and a turbofan benchmark under label-scarce stitch protocols, TD(n,λ) improves RUL and failure-mode prediction relative to a supervised same-backbone Monte Carlo control. The practical payoff is that fragmented, anonymous degradation records can contribute local transitions instead of waiting for complete run-to-failure labels.

Core claim

RUL is modeled as the expected hitting time of an absorbing failure set, and each failure mode as a hitting probability for that mode's terminal subset; both are components of one vector GVF. The Bellman expectation equation ties each state's prediction to the next transition and successor state. TD(n,λ) replaces unavailable complete returns with observed prefixes plus bootstrapped successor predictions. The paper proves that the vector Bellman operator and its truncated TD(n,λ) counterpart are contractions with the same fixed point, that under realizability the linear TD limit equals the Monte Carlo least-squares projection, and that oracle TD targets are Rao–Blackwellized MC returns with n

What carries the argument

The central object is the vector General Value Function (GVF) on an absorbing Markov degradation process: each component is an expected, continuation-gated sum of a user-defined cumulant, with one cumulant counting survived steps for RUL and K cumulants marking entry into each failure-mode terminal set. The carrying identity is the vector Bellman equation V(s) = E[u_{t+1} + Γ_{t+1}V(S_{t+1}) | S_t = s], which TD(n,λ) bootstraps from local transitions; the operator analysis shows its fixed point and its projected linear TD limit.

Load-bearing premise

The fixed-length sensor window used as the predictive state is assumed to be Markov enough for the Bellman recursion to be valid; the paper itself notes that finite windows make the Markov property approximate, so the empirical advantage depends on that approximation being mild.

What would settle it

On a degradation process with a hidden state that affects RUL beyond a 20–30 step window, measure the Bellman residual of one-step TD targets on held-out transitions and compare TD(n,λ) against complete-return Monte Carlo regression when full labels are available; if TD is systematically biased and loses to MC under abundant labels, the finite-window Markov assumption fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Fragmented and anonymous trajectory segments, which supervised RUL regression discards until complete run-to-failure labels exist, become usable training data through local Bellman transitions.
  • Under label scarcity, TD(n,λ) maintains RUL accuracy and failure-mode F1 in settings where a supervised Monte Carlo control collapses because it requires terminal-containing segments.
  • Under full run-to-failure supervision, TD remains comparable to supervised regression, so the GVF reformulation does not cost accuracy.
  • Because both heads regress a common vector Bellman target, RUL and failure-mode probabilities are trained to be temporally consistent with the same successor state.
  • Oracle TD(n) targets have no larger variance than complete Monte Carlo targets, so when the successor-value model is accurate, bootstrapping should be preferred to waiting for terminal returns.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's theory assumes a finite sensor window is Markov enough; a natural extension the authors leave implicit is using recurrent or latent-state encoders to make the Markov assumption closer to exact, which would likely strengthen the empirical gains.
  • Because discounted mode scores tilt toward modes with shorter expected time to failure, practitioners should use γ_mode = 1 or calibrate when they need actual failure-mode probabilities rather than urgency-weighted rankings.
  • A testable extension is to apply the same GVF recursion to distributional RUL targets such as quantiles or survival curves by changing the cumulants, since the Bellman framework does not depend on the specific scalar target.
  • The target-variance decomposition suggests a model-selection heuristic: if bootstrapped targets are noisy early in life, increase n or λ near the beginning of trajectories; this could be verified by per-life-stage variance studies.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper reformulates remaining-useful-life (RUL) prediction and failure-mode classification as vector General Value Function (GVF) prediction on an absorbing degradation process. RUL is represented as a survival-time GVF and each failure mode as a terminal-event GVF, with targets connected through a vector Bellman equation. The authors propose a TD(n,λ) estimator with bootstrapped, stop-gradient targets, and provide theory identifying the Bellman fixed point, projected linear-TD limits, and a variance decomposition showing that oracle TD targets are Rao–Blackwellized versions of Monte Carlo returns. Experiments on an event-triggered simulation and on NASA C-MAPSS compare TD against a same-backbone Monte Carlo (MC) control and a health-index (HI) baseline, under full trajectories and under a label-scarce, identity-free 'stitch' protocol. The central claims are that TD improves RUL and failure-mode prediction relative to supervised MC, especially when complete terminal labels are scarce, and that fragmented records can contribute local Bellman transitions instead of being discarded.

Significance. If the claims hold, the practical contribution is genuine: identity-free, partially observed degradation segments, which are normally unusable for supervised RUL regression, can be converted into Bellman-consistent training data. The framework also cleanly separates the choice of predictive state from the definition of the prognostic target, and the stitch protocol is a well-designed stress test for the label-scarce setting. The theory is largely standard (and is correctly attributed to Tsitsiklis & Van Roy, and to Cheikhi & Russo), but the adaptation to vector GVFs is useful: the fixed-point identification, the soft-horizon bound, and the mode-tilting factorization are clear and non-trivial. However, the paper’s strongest evidence is weakened by two issues: the finite-window predictive state violates exact Markovity on which the theory rests, and the C-MAPSS hyperparameters are selected on test-endpoint RMSE, so the empirical advantage on that benchmark may be optimistic. The failure-mode evaluation also relies on pseudo-labels that serve both as training targets and as the evaluation reference. These are substantial but local problems, and the manuscript already demonstrates unus

major comments (3)
  1. [§3.1, §4.2.3, Eq. (7); Theorem 1; Proposition 3] The theoretical guarantees (Theorem 1, Proposition 3) are explicitly conditional on Markovity of the predictive state S_t. Section 4.2.3 admits that a finite window 'generally makes the Markov property approximate rather than exact.' When the true physical degradation state is not recoverable from W=20–30 sensor steps, the bootstrap term sg(V_θ(S_{t+n})) in Eq. (7) estimates E[return | S_{t+n}], not E[return | full history], so TD converges to the fixed point of the wrong Bellman operator. This is a systematic bias, not just added variance, and it attacks the mechanism (bootstrapped Bellman consistency) that the headline label-scarce gains are attributed to. No bound, diagnostic, or sensitivity analysis for this approximation error is provided. I request one of: (a) a controlled simulation experiment where the true latent degradation state is known and the window length can be varied, sh
  2. [§6.2 and Table 3] The authors disclose that C-MAPSS hyperparameters are selected by minimizing test-endpoint RMSE, because no disjoint split exists for that comparison, and they call the results 'benchmark-tuned and may be optimistic.' This is a load-bearing weakness for the empirical claim 'TD improves RUL and failure-mode prediction' on C-MAPSS, especially because TD has more tunable components (n, λ, γ, W) than the MC control. Although the rule is uniform across methods, a method with more hyperparameters can benefit more from test-set selection. Please provide a nested validation procedure (e.g., train/validation/test splits of the training fleet) or a sensitivity analysis over (n, λ, W, γ) that shows the reported improvements are not an artifact of test-endpoint model selection.
  3. [§6 and Tables 2–3] For FD003/FD004, failure-mode labels are not native to C-MAPSS; the paper derives reference pseudo-labels by DTW-k-means on held-out sensor-trajectory embeddings. These same pseudo-labels are used as MC's supervised mode target and as the evaluation reference for both methods, while TD's terminal/near-terminal anchors also use the same pseudo-mode labels. The classification comparison therefore measures agreement with a clustering of the same sensor data, not with known physical failure modes. The paper reports no validation of the pseudo-labels (e.g., against a physics-based label, or sensitivity to the number of clusters). I ask the authors to either provide such validation or explicitly mark the F1 columns as pseudo-label-relative and temper the failure-mode claims accordingly. This is particularly important because simulation label-scarce classification results are omitted (Section 5
minor comments (5)
  1. [Abstract and §6.3] The abstract claims TD 'improves RUL and failure-mode prediction,' but in Table 2 the FD003/FD004 full-trajectory minority-class F1 values are ties within noise (e.g., 0.912±0.009 vs 0.921±0.000 on FD003). The failure-mode claim should be qualified to the label-scarce setting unless more evidence is provided.
  2. [Eq. (19)] Please define T_B explicitly for C-MAPSS test engines: is it the post-window observed lifetime or the full run-to-failure length that is held out? The term 'per-step NAE' is standard but the denominator should be unambiguous, particularly for truncated test trajectories.
  3. [§5.4 and Figure 4] Figure 4 uses λ=0.5 only for 'visual contrast' while canonical results use λ=1.0 (full-trajectory) and λ=0.7 (stitch). Please clarify in the caption that these are separately trained models, and state how λ was selected on validation data for the reported runs.
  4. [Appendix G.5, Proposition 4] The uniqueness argument in Proposition 4 requires the uniform bound sup_{s∈C} E_s[τ]<∞, not merely almost-sure absorption. The text does mention this, but the proposition statement should carry the assumption explicitly to avoid a reader applying it to finite-state chains only by accident.
  5. [Algorithm 1] The line 'if |D| ≥ B then' outside the mini-batch loop is confusing; training should occur whenever at least one mini-batch is available. The pseudocode should be restructured so the mini-batch loop is unconditional inside the training loop.

Circularity Check

0 steps flagged

No load-bearing circularity: the core GVF/TD derivation is self-contained and rests on external TD theory; only minor self-referential evaluation caveats and non-load-bearing self-citations.

full rationale

I walked the derivation chain and found no equation, fitted parameter, or uniqueness claim that reduces to its own inputs. RUL and failure-mode probabilities are defined as hitting-time and hitting-probability expectations on an absorbing Markov process (Definitions 5-6), and Eq. (5) is the Bellman expectation equation specialized to those definitions; the target quantities are not re-introduced as fitted outputs. Theorem 1's fixed-point and projected-TD convergence claims are imported from Tsitsiklis & Van Roy (1997) after explicitly verifying the standard hypotheses, and Proposition 3's variance decomposition is a conditional-expectation/Rao-Blackwell argument whose Markov assumption is stated rather than hidden. The empirical TD-versus-MC comparison uses the same CNN1D backbone, preprocessing, and splits, so TD's advantage is not forced by normalization or by using MC's fitted values. The finite-window Markov approximation is honestly disclosed (Section 3.1, Section 4.2.3) and is a correctness risk, not circularity. Two caveats keep this from a clean 0: the C-MAPSS failure-mode pseudo-labels are used both as training anchors and as the evaluation reference (Section 6: 'The same reference labels score both methods, and TD's terminal/near-terminal anchors use the same pseudo-mode labels'), making that F1 comparison self-referential rather than an independent external benchmark; and the two self-citations (Xu et al. 2025 in Remark 1; Yan et al. 2016 in the literature review) are contextual and not load-bearing. Neither caveat constitutes a specific reduction of a prediction to an input, so I report no circular steps, only a minor self-referential evaluation caveat.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central method rests on standard reinforcement-learning assumptions plus a practical Markov approximation. Hand-chosen hyperparameters (discount, truncation, trace, loss weights, window length) and the segment-sampling representativeness assumption are load-bearing for the empirical claims. The C-MAPSS mode pseudo-labels are an operationalized entity, but no new physical quantities are introduced.

free parameters (6)
  • Continuation γ_bar = 0.995 (RUL; shared soft horizon on C-MAPSS)
    Sets the exponential soft-horizon for the RUL GVF; the TD head trains on a discounted survival-time value, which is later inverted deterministically to report RUL.
  • TD truncation n = 5 (simulation), 4 (C-MAPSS)
    Controls bias-variance tradeoff; chosen on validation for simulation and by benchmark endpoint RMSE for C-MAPSS.
  • Trace parameter λ = 1.0 full-trajectory, 0.7 stitch, 0 for C-MAPSS mode head
    Determines the temporal span of the TD target; canonical values are selected on validation data or by test-endpoint RMSE.
  • Loss weights = λ_rul=1, λ_cls=5000
    Hand-set weights balancing RUL MSE and mode-head MSE scales in the joint loss.
  • Window length W = 30 (FD001/3/4), 20 (FD002)
    Defines the predictive state S_t; fixed following prior C-MAPSS practice, not tuned per method.
  • Mode pseudo-label k = 2 clusters (DTWk-means)
    Reference failure-mode labels for C-MAPSS FD003/FD004 are generated by clustering; the same labels anchor TD terminal returns, supervise MC, and score both methods.
axioms (5)
  • domain assumption Predictive state S_t is Markov for the prognostic targets; Bellman equations are exact only then
    Stated in Section 3.1; with fixed-window states the property is approximate, as admitted in Section 4.2.3.
  • domain assumption Almost-sure absorption and/or uniform bounded expected absorption time for undiscounted results
    Used in Propositions 2 and 4 and stated as required in Section 4.3; not guaranteed in all practical settings.
  • standard math On-policy ergodic sampling, bounded cumulants/features, Robbins–Monro step sizes, and well-posed projected equations
    Appendix G Assumptions 1–6; imported from Tsitsiklis and Van Roy (1997) for the linear projected-TD convergence result.
  • domain assumption Observed partial segments represent the transition distribution (non-informative fragmentation)
    Remark 2 notes that target computability is not consistency under arbitrary informative censoring; the stitch protocol assumes the segment pool is representative.
  • domain assumption DTWk-means pseudo-labels on C-MAPSS approximate true failure modes
    Section 6 derives reference labels because FD003/FD004 lack native failure-mode labels; both training supervision and evaluation rely on these approximate labels.

pith-pipeline@v1.3.0-alltime-deepseek · 25570 in / 14274 out tokens · 114308 ms · 2026-08-01T05:16:35.999657+00:00 · methodology

0 comments
read the original abstract

Remaining useful life (RUL) prediction and failure-mode classification are central tasks in predictive maintenance. Many data-driven pipelines use fixed-window supervised learning with complete terminal labels; such routes do not naturally encode the temporal recursion linking successive degradation-state predictions when observations are partial or unit identities are unavailable. We formulate prognostics as vector General Value Function (GVF) prediction on an absorbing degradation process, treating RUL and failure-mode probabilities as temporally consistent targets rather than independent window-level labels, and estimate them with a multi-step temporal-difference estimator, TD($n,\lambda$). Supporting theory identifies the Bellman fixed point of the vector GVFs, characterizes the linear projected-TD limit and its relation to complete-return Monte Carlo regression under realizability, and explains when bootstrapped TD targets are less variable than Monte Carlo returns. On an event-triggered multimode simulation and NASA C-MAPSS label-scarce stitch data, TD improves RUL and failure-mode prediction relative to a supervised same-backbone Monte Carlo control, especially under scarce complete labels. Practically, fragmented, identity-free degradation records can contribute local Bellman transitions instead of being discarded until complete run-to-failure labels are available.

Figures

Figures reproduced from arXiv: 2607.22268 by Ali Sarabi, Boyang Xu, Hao Yan, Qing Zou.

Figure 1
Figure 1. Figure 1: Dynamic prognosis as value prediction. A unit evolves until absorption in F = S k Fk; RUL is the expected time to failure, and mode prediction is the hitting probability of each Fk. Successive forecasts satisfy a Bellman recursion along the absorbing degradation process. This viewpoint clarifies the distinction between Monte Carlo (MC) and temporal-difference (TD) estimation. Conventional supervised RUL re… view at source ↗
Figure 2
Figure 2. Figure 2: Simulation: example trajectories for both modes. Modes differ in both the late-phase [PITH_FULL_IMAGE:figures/full_fig_p019_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Simulation: label-scarce stitch RUL NAE ( [PITH_FULL_IMAGE:figures/full_fig_p022_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Simulation: λ effect on prediction smoothness (λ=0.5 shown for contrast; canonical full￾trajectory model uses λ=1.0). TD(n, λ) gives smoother RUL and failure-risk trajectories than MC. to C-MAPSS; following Fu et al. (2025), we derive reference pseudo-labels by DTW k-means (k=2) on held-out sensor-trajectory embeddings, with no RUL or endpoint information. These labels score failure-mode F1 and provide MC’… view at source ↗
Figure 5
Figure 5. Figure 5: C-MAPSS full-data RUL NAE by remaining-life bucket (TD( [PITH_FULL_IMAGE:figures/full_fig_p025_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: C-MAPSS full-data failure-mode classification accuracy by remaining-life bucket (FD003 [PITH_FULL_IMAGE:figures/full_fig_p033_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: C-MAPSS label-scarce stitch: RUL NAE by remaining-life bucket across fractions [PITH_FULL_IMAGE:figures/full_fig_p035_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: C-MAPSS stitch minority-class F1 for FD003/FD004, mean [PITH_FULL_IMAGE:figures/full_fig_p035_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Simulation UMAP embeddings colored by remaining life. Panel purities: TD( [PITH_FULL_IMAGE:figures/full_fig_p036_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: C-MAPSS UMAP embeddings on FD003/FD004, colored by failure mode and remaining [PITH_FULL_IMAGE:figures/full_fig_p037_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references

  1. [1]

    IEEE Transactions on Industrial Electronics , volume=

    Direct remaining useful life estimation based on support vector regression , author=. IEEE Transactions on Industrial Electronics , volume=. 2017 , publisher=

    2017

  2. [2]

    Journal of Quality Technology , volume=

    A data-level fusion approach for degradation modeling and prognostic analysis under multiple failure modes , author=. Journal of Quality Technology , volume=. 2018 , publisher=

    2018

  3. [3]

    Reliability Engineering & System Safety , volume=

    A survey of the application of gamma processes in maintenance , author=. Reliability Engineering & System Safety , volume=. 2009 , publisher=

    2009

  4. [4]

    2017 IEEE International Conference on Prognostics and Health Management (ICPHM) , pages=

    Long Short-Term Memory Network for Remaining Useful Life Estimation , author=. 2017 IEEE International Conference on Prognostics and Health Management (ICPHM) , pages=. 2017 , organization=

    2017

  5. [5]

    Machine Learning , volume=

    Learning to predict by the methods of temporal differences , author=. Machine Learning , volume=. 1988 , publisher=

    1988

  6. [6]

    Machine Learning , volume=

    Multitask Learning , author=. Machine Learning , volume=. 1997 , publisher=

    1997

  7. [7]

    IEEE Transactions on Industrial Informatics , volume=

    Joint Learning of Degradation Assessment and RUL Prediction for Aeroengines via Dual-Task Deep LSTM Networks , author=. IEEE Transactions on Industrial Informatics , volume=. 2019 , publisher=

    2019

  8. [8]

    IEEE Transactions on Instrumentation and Measurement , volume=

    A Deep Branched Network for Failure Mode Diagnostics and Remaining Useful Life Prediction , author=. IEEE Transactions on Instrumentation and Measurement , volume=. 2022 , publisher=

    2022

  9. [9]

    IEEE Transactions on Automatic Control , volume=

    An Analysis of Temporal-Difference Learning with Function Approximation , author=. IEEE Transactions on Automatic Control , volume=. 1997 , publisher=

    1997

  10. [10]

    IIE Transactions , volume=

    Condition monitoring and remaining useful life prediction using degradation signals: Revisited , author=. IIE Transactions , volume=. 2013 , publisher=

    2013

  11. [11]

    IISE Transactions , volume=

    Statistical degradation modeling and prognostics of multiple sensor signals via data fusion: A composite health index approach , author=. IISE Transactions , volume=. 2018 , publisher=

    2018

  12. [12]

    IISE Transactions , volume =

    Changyue Song and Kaibo Liu and Xi Zhang , title =. IISE Transactions , volume =. 2019 , publisher =

    2019

  13. [13]

    IIE Transactions , volume=

    Remaining useful life prediction of individual units subject to hard failure , author=. IIE Transactions , volume=. 2014 , publisher=

    2014

  14. [14]

    IIE transactions , volume=

    Failure Event Prediction using the Cox Proportional Hazard Model Driven by Frequent Failure Signatures , author=. IIE transactions , volume=. 2007 , publisher=

    2007

  15. [15]

    IISE Transactions , volume=

    Remaining useful life prediction based on the mixed effects model with mixture prior distribution , author=. IISE Transactions , volume=. 2017 , publisher=

    2017

  16. [16]

    IISE Transactions , volume=

    Remaining useful life prediction based on degradation signals using monotonic B-splines with infinite support , author=. IISE Transactions , volume=. 2020 , publisher=

    2020

  17. [17]

    IEEE Transactions on Automation Science and Engineering , volume=

    A data-level fusion model for developing composite health indices for degradation modeling and prognostic analysis , author=. IEEE Transactions on Automation Science and Engineering , volume=. 2013 , publisher=

    2013

  18. [18]

    2018 , publisher=

    Reinforcement learning: An introduction , author=. 2018 , publisher=

    2018

  19. [19]

    Proceedings of the 10th International Conference on Autonomous Agents and Multiagent Systems , pages=

    Horde: A Scalable Real-time Architecture for Learning Knowledge from Unsupervised Sensorimotor Interaction , author=. Proceedings of the 10th International Conference on Autonomous Agents and Multiagent Systems , pages=

  20. [20]

    Developing a predictive approach to knowledge , author=

  21. [21]

    Proceedings of the 40th International Conference on Machine Learning , pages=

    On the statistical benefits of temporal difference learning , author=. Proceedings of the 40th International Conference on Machine Learning , pages=. 2023 , volume=

    2023

  22. [22]

    The Journal of Machine Learning Research , volume=

    Policy evaluation with temporal differences: A survey and comparison , author=. The Journal of Machine Learning Research , volume=. 2014 , publisher=

    2014

  23. [23]

    Technometrics , volume=

    Joint models for event prediction from time series and survival data , author=. Technometrics , volume=. 2021 , publisher=

    2021

  24. [24]

    Technometrics , volume=

    The inverse Gaussian process as a degradation model , author=. Technometrics , volume=. 2014 , publisher=

    2014

  25. [25]

    Technometrics , volume=

    A class of hierarchical multivariate Wiener processes for modeling dependent degradation data , author=. Technometrics , volume=. 2024 , publisher=

    2024

  26. [26]

    Multiple Sensor Data Fusion for Degradation Modeling and Prognostics Under Multiple Operational Conditions , year=

    Yan, Hao and Liu, Kaibo and Zhang, Xi and Shi, Jianjun , journal=. Multiple Sensor Data Fusion for Degradation Modeling and Prognostics Under Multiple Operational Conditions , year=

  27. [27]

    IISE Transactions , volume=

    A Bayesian deep learning framework for interval estimation of remaining useful life in complex systems by incorporating general degradation characteristics , author=. IISE Transactions , volume=. 2020 , publisher=

    2020

  28. [28]

    Mechanical Systems and Signal Processing , volume=

    Deep learning and its applications to machine health monitoring , author=. Mechanical Systems and Signal Processing , volume=. 2019 , publisher=

    2019

  29. [29]

    Reliability Engineering & System Safety , volume=

    Remaining Useful Life Estimation in Prognostics Using Deep Convolution Neural Networks , author=. Reliability Engineering & System Safety , volume=. 2018 , issn=. doi:10.1016/j.ress.2017.11.021 , keywords=

    doi:10.1016/j.ress.2017.11.021 2018

  30. [30]

    Neurocomputing , volume=

    A recurrent neural network based health indicator for remaining useful life prediction of bearings , author=. Neurocomputing , volume=. 2017 , publisher=

    2017

  31. [31]

    Fault diagnosis and remaining useful life estimation of aero engine using LSTM neural network , year=

    Yuan, Mei and Wu, Yuting and Lin, Li , booktitle=. Fault diagnosis and remaining useful life estimation of aero engine using LSTM neural network , year=

  32. [32]

    , journal=

    Zhang, Yongzhi and Xiong, Rui and He, Hongwen and Pecht, Michael G. , journal=. Long Short-Term Memory Recurrent Neural Network for Remaining Useful Life Prediction of Lithium-Ion Batteries , year=

  33. [33]

    Domain Adaptive Remaining Useful Life Prediction With Transformer , year=

    Li, Xinyao and Li, Jingjing and Zuo, Lin and Zhu, Lei and Shen, Heng Tao , journal=. Domain Adaptive Remaining Useful Life Prediction With Transformer , year=

  34. [34]

    Advances in Neural Information Processing Systems , volume=

    The surprising efficiency of temporal difference learning for rare event prediction , author=. Advances in Neural Information Processing Systems , volume=

  35. [35]

    2008 International Conference on Prognostics and Health Management , pages=

    Damage Propagation Modeling for Aircraft Engine Run-to-Failure Simulation , author=. 2008 International Conference on Prognostics and Health Management , pages=. 2008 , organization=

    2008

  36. [36]

    IIE transactions , volume=

    Residual-life distributions from component degradation signals: A Bayesian approach , author=. IIE transactions , volume=. 2005 , publisher=

    2005

  37. [37]

    IEEE Transactions on Automation Science and Engineering , volume=

    Degradation Modeling and Prognostic Analysis Under Unknown Failure Modes , author=. IEEE Transactions on Automation Science and Engineering , volume=. 2025 , publisher=

    2025

  38. [38]

    Journal of Quality Technology , year=

    Partially Observable Markov Decision Process Framework for Operating Condition Optimization Using Real-Time Degradation Signals , author=. Journal of Quality Technology , year=

  39. [39]

    IEEE Transactions on Neural Networks and Learning Systems , volume=

    A Shape-Constrained Neural Data Fusion Network for Health Index Construction and Residual Life Prediction , author=. IEEE Transactions on Neural Networks and Learning Systems , volume=. 2021 , doi=

    2021