Pith. sign in

REVIEW 3 major objections 5 minor 39 references

ROAM adapts frozen industrial models to new scenarios without retraining by turning LLM scenario judgments into risk-gated corrections in a five-axis latent space.

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 · grok-4.5

2026-07-11 01:10 UTC pith:NATWCGFC

load-bearing objection Solid systems paper for frozen industrial soft sensors; the headline LLM story is overstated relative to the Bayesian residual corrector that actually moves the numbers. the 3 major comments →

arxiv 2607.06625 v1 pith:NATWCGFC submitted 2026-07-07 cs.LG cs.AIcs.SYeess.SY

Open-Ended Scenario Reasoning for Specialist Model Adaptation

Youcheng Zong , Runda Jia , Ranmeng Lin , Mingxuan Ren , Dakuo He This is my paper
classification cs.LG cs.AIcs.SYeess.SY
keywords large language modelBayesian inferenceonline adaptationprocess industriessoft sensingfrozen specialist modelsrisk gatingsemantic latent space
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.

Process plants already run validated soft-sensor models for quality variables, but sensor drift, feedstock changes, and regime switches make those models drift systematically. Retraining is slow and costly; leaving the model alone leaves lasting bias. ROAM treats adaptation as updating beliefs about the scenario rather than updating model weights. An LLM reads shift logs, maintenance records, and first-window diagnostics and produces a structured prior on five mismatch axes—bias, scale, load, dynamics, and readout—plus a trust score. That prior is fused online with delayed labels and process features through Bayesian posterior updates confined to those axes, then risk-gated so that weak or conflicting evidence simply returns the original frozen prediction. On a mineral thickener and a penicillin fermentation benchmark the method cuts mean absolute error by more than 20 percent on hidden shifts while adding only a few hundred parameters and sub-millisecond overhead, keeping certified models in service.

Core claim

LLM world knowledge and reasoning can be converted into a conservative, low-dimensional adaptation signal for already-deployed specialist models: all corrections live in a five-dimensional semantic latent space, fused with online observations under Bayesian updating and multi-layer risk gating, so that major scenario shifts are corrected without touching frozen parameters and without placing the LLM inside the prediction loop.

What carries the argument

The five-dimensional corrective latent z_r (bias, scale, load, dynamics, readout) together with episode-level observation anomaly score τ_r: LLM priors and heterogeneous online evidence update only this low-dimensional manifold; multi-layer risk gating (prior diffusion, effect-strength, covariance-trace, support-distance, and trust) drives the correction to zero whenever evidence is insufficient, falling back to the frozen backbone.

Load-bearing premise

Deployment can be split into episodes whose dominant operating condition stays roughly stationary, and usable open-ended field records plus first-window diagnostics exist at each episode start so a meaningful prior can be built.

What would settle it

On a new industrial soft-sensor deployment that has genuine hidden shifts but either lacks usable textual logs or changes dominant condition inside short intervals, measure whether ROAM still reduces hidden-shift MAE by roughly 20 percent relative to the frozen backbone while leaving same-family error essentially unchanged; a clear failure on either metric would falsify the central claim.

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

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 proposes ROAM, an adaptation layer that keeps a validated specialist soft-sensor frozen and corrects its outputs in a fixed 5-D semantic latent (bias, scale, load, dynamics, readout) plus an observation-anomaly score. An LLM is used only once per episode as an out-of-loop prior engine that maps open textual evidence, structured context, and first-window diagnostics into a Gaussian prior; delayed labels, observation features, and process diagnostics then update the posterior via Kalman/subspace Gaussian updates, with multi-layer risk gating and fallback to the baseline. Offline, low-capacity linear heads and synthetic anchors define the adaptation manifold. On a plant thickening process and IndPenSim, across seven backbones, ROAM reports ~20% MAE reduction on hidden shifts with ≤839 extra parameters and <0.02 ms/step overhead, with ablations, LLM swaps, and comparisons to fine-tuning/MAML.

Significance. If the claims hold, ROAM addresses a real industrial constraint: certified soft sensors cannot be retrained quickly under feedstock, sensor, or regime shifts, yet field logs contain usable scenario knowledge. The design is practically attractive—frozen backbone, auditable low-dimensional corrections, explicit fallback, and negligible runtime cost—and the evaluation is stronger than typical industrial LLM papers (seven architectures, plant + public process, efficiency tables, paradigm baselines, code release). Even if LLM priors prove secondary, the Bayesian residual corrector on a semantic manifold is a useful contribution for safe online soft-sensor adaptation. The work is therefore significant for process soft sensing and industrial ML deployment, provided the role of open-ended LLM reasoning is stated accurately relative to the non-LLM evidence channels.

major comments (3)
  1. Abstract and §I claim that LLM world knowledge/reasoning is turned into the adaptation signal yielding >20% MAE reduction on major/hidden shifts. Table V (GRU, thickening) contradicts that attribution as the primary driver: frozen baseline hidden-shift MAE 22.957 falls to 18.468 under a flat prior once label residuals + observation + process evidence are enabled (Var. E), while adding the LLM prior only moves the result to 18.329 (Var. I). The text itself states that E→I “mainly provides directional initialization” and that “label-residual and evidence channels already deliver most of the adaptation benefit.” The distinctive claim (open-ended scenario reasoning as the adaptation mechanism) is therefore overstated relative to a non-LLM Bayesian residual corrector. Please either (i) reframe title/abstract/contributions around the full ROAM stack with LLM as optional prior, or (ii) add expe
  2. §III-D and the online loop rely on delayed labels (Kalman update with residual e_t = y_t − ŷ⁰_t after delay Δ) as “the most direct evidence source.” The headline industrial narrative (rapid response when relabeling/retraining is costly; Note to Practitioners) needs a clearer statement of label availability: how large is Δ in each environment, what fraction of the hidden-shift gain remains with longer delays or sparse labels, and whether ROAM still helps in the pure open-loop soft-sensing regime where labels arrive only after hours/days. Without this, it is hard to separate “adaptation without retraining” from standard online residual correction once labels stream in.
  3. §III-A/Table I fix industrial mismatch to five semantic axes plus τ_r by construction. This is a load-bearing modeling axiom for the “semantically interpretable” claim and for subspace decoupling (Table V H vs I). The paper does not show that these axes are necessary or sufficient across processes, nor what happens when multiple shift sources co-occur beyond the 5-D span (acknowledged only briefly in §V). Please justify the axis set more carefully (e.g., residual explained variance per axis on held-out shifts) and report failure cases when the true mismatch is outside {bias, scale, load, dynamics, readout}.
minor comments (5)
  1. Table III units: MAE/RMSE reported as ×10³ for thickening; state the physical unit of underflow concentration so absolute errors are interpretable for practitioners.
  2. Fig. 5 posterior strips: “serrated ends mark out-of-range values” is hard to read in print; consider clipping markers or a separate scale for τ_r.
  3. §III-B: six semantic axes are mentioned for LLM output (five corrective + trust) while Table I lists five; keep terminology consistent.
  4. Hyperparameter defaults (σ²_min/max, η, β, Δu, λg) are discussed in sensitivity (§IV-F) but not tabulated in one place; a small default table would aid reproduction.
  5. Related work on test-time adaptation and industrial soft-sensor recalibration could cite a few more non-LLM online bias-correction baselines for completeness.

Circularity Check

0 steps flagged

No significant circularity: ROAM is an empirical residual-correction system whose offline manifold and online Kalman updates are trained/evaluated on held-out shifts, not a first-principles derivation that reduces to its inputs by construction.

full rationale

The paper’s load-bearing chain is engineering design plus external test error, not a closed derivation. Offline, ridge maps and synthetic anchors (controlled bias/scale/dynamics perturbations on training data) define a 5-D residual manifold; online, a Gaussian prior from the LLM is fused with delayed labels and subspace diagnostics via Kalman/subspace updates, then risk-gated (Eqs. 1–18, §III). Test scenarios (hidden/visible/observation shifts; IndPenSim faults) are held out from training and provide independent MAE/RMSE/R² (Tables III–IV). Using delayed labels within an episode is standard online adaptation with lag, not fitting the reported metric by construction. Ablation (Table V) even shows most hidden-shift gain under a flat prior once label/evidence channels are on—so the headline result is not forced by the LLM prior. Self-citations (soft sensing, meta-learning, LLM decision support) appear only as related work and do not supply uniqueness theorems or ansatzes that force the method. No self-definitional loop, fitted-input-as-prediction of the same quantity, or renaming of a known closed-form result is present. Attribution of gains to “LLM reasoning” may be overstated relative to residual Bayesian correction, but that is an attribution/correctness issue, not circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 3 invented entities

The central claim rests on a modeling choice that industrial mismatch is well captured by five semantic axes plus a trust score, linear residual maps from frozen hidden states, Gaussian conjugacy for online fusion, and episode-level stationarity with usable open text. Free parameters are the usual noise/gate hyperparameters and offline ridge fits; invented entities are the 5-D corrective latent and ROAM itself as an engineering construct rather than a physical discovery.

free parameters (7)
  • prior variance bounds σ²_min, σ²_max
    Map LLM confidence/uncertainty into Gaussian prior variance; sensitivity analysis shows category trade-offs.
  • label observation noise σ²_y
    Kalman residual update noise; hand/default-tuned with visible sensitivity.
  • subspace observation noises σ²_obs, σ²_dyn, σ²_read, σ²_load
    Control strength of observation and process evidence channels in subspace updates.
  • risk-gate coefficients η1, η2, β1, β2 and gate scale λg
    Exponential decay of corrections from posterior trace and support distance; defaults chosen for cross-category compromise.
  • update interval Δu
    Batching interval for observation/process evidence; short intervals raise same-family error in sensitivity plots.
  • offline ridge λ and alternating iterations N_iter
    λ=0.001 and 3 iterations fit linear axis maps and heads on anchors plus training episodes.
  • axis coefficient mapping W,a and linear diagnostic heads
    Fitted offline by ridge/alternation so residuals are explained by the 5-D latent; frozen online.
axioms (6)
  • ad hoc to paper Industrial mismatch for soft sensors is adequately represented by five semantic axes (bias, scale, load, dynamics, readout) plus an observation-anomaly trust score.
    Table I and §III-A define the latent; success depends on this taxonomy covering dominant shifts.
  • domain assumption Deployment can be partitioned into episodes that are approximately scenario-stationary, with external boundaries resetting the prior.
    §III-A and Note to Practitioners; frequent within-interval changes would break the prior/posterior design.
  • domain assumption Linear maps from standardized backbone hidden states and diagnostic features suffice to produce residual corrections.
    §III-C offline learning uses only linear ridge heads; nonlinear mismatch is deferred to future work.
  • domain assumption LLM outputs can be treated as structured direction/strength/confidence judgments convertible to a Gaussian prior without entering the prediction loop.
    §III-B; risk gates are required because this conversion can be unreliable.
  • standard math Standard linear-Gaussian Kalman / subspace Gaussian updates correctly fuse delayed labels and diagnostic latent observations.
    §III-D equations (6)–(11).
  • domain assumption Frozen validated specialist models must not have parameters modified after deployment for certification/regression reasons.
    Introduction and related work; motivates out-of-loop correction rather than fine-tuning.
invented entities (3)
  • 5-D semantic corrective latent z_r (bias/scale/load/dynamics/readout) no independent evidence
    purpose: Confine all adaptation degrees of freedom to an interpretable low-dimensional manifold for residual correction.
    Defined in Table I and §III-A; independent evidence is only empirical performance under shifts, not an external physical measurement of the axes.
  • ROAM risk-gated Bayesian adaptation layer (LLM prior engine + multi-layer gates) no independent evidence
    purpose: Turn open-ended scenario text and online evidence into conservative corrections of a frozen model.
    Framework-level construct; validated only by the paper's experiments and ablations.
  • Episode-level observation anomaly score τ_r no independent evidence
    purpose: Gate observation-side corrections when the observation pathway is judged anomalous.
    Fused from LLM trust and diagnostic head (§III-B); no external ground-truth anomaly label required by the method.

pith-pipeline@v1.1.0-grok45 · 22242 in / 3786 out tokens · 41338 ms · 2026-07-11T01:10:39.742150+00:00 · methodology

0 comments
read the original abstract

Process industries have accumulated validated specialist models, yet sensor drift, feedstock variation, and regime switching cause these models to degrade systematically in new scenarios. Collecting new labeled data and retraining is costly, while continuing with the original model incurs persistent bias. Existing adaptation methods require modifying model parameters with sufficient labeled data, making rapid response on deployed systems difficult. Using LLMs as direct predictors risks hallucinations and uncontrollable outputs. Such predictors also cannot incorporate unstructured scenario knowledge from the field. To address these limitations, this article proposes Reasoning-Driven Open Adaptation for Specialist Models (ROAM), a framework that uses LLM world knowledge and reasoning to adapt frozen specialist models to unseen scenarios without retraining. ROAM confines all corrections to a low-dimensional, semantically interpretable latent space. LLM-generated scenario judgments and online observations are fused under a unified probabilistic framework. A risk-constrained mechanism suppresses corrections under unreliable LLM evidence or abrupt scenario shifts and falls back to the original frozen model when evidence is insufficient. Experiments on a mineral thickening process and the public IndPenSim penicillin fermentation dataset show that ROAM reduces MAE by over 20\% in major shift settings such as hidden shifts with only 839 additional parameters and under 0.02\,ms per-step overhead. These results indicate that LLM reasoning can be turned into a conservative adaptation signal for industrial models already in service.

Figures

Figures reproduced from arXiv: 2607.06625 by Dakuo He, Mingxuan Ren, Ranmeng Lin, Runda Jia, Youcheng Zong.

Figure 1
Figure 1. Figure 1: Adaptation paradigms for industrial specialist models. (a) Parameter [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the ROAM framework. The left half shows the offline phase, which synthesizes anchors from training-episode residuals and alternates [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Online posterior update and conservative correction. (A) The LLM [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: No. 7 tailings thickening dewatering site at a plant in Nanjing, China, showing system input signals, the digital-twin interface, and on-site deployment [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Prediction curves on eight representative IndPenSim anomaly scenarios comparing GRU (Base) and GRU+ROAM. Lower strips report the final [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Relative MAE of LLM prior engines in the thickening environment [PITH_FULL_IMAGE:figures/full_fig_p009_7.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 · 39 canonical work pages

  1. [1]

    Data-driven soft sensors in the process industry,

    P. Kadlec, B. Gabrys, and S. Strandt, “Data-driven soft sensors in the process industry,”Computers & Chemical Engineering, vol. 33, no. 4, pp. 795–814, 2009

    work page 2009

  2. [2]

    Nonlinear dynamic soft sensor modeling with supervised long short-term memory network,

    X. Yuan, L. Li, and Y . Wang, “Nonlinear dynamic soft sensor modeling with supervised long short-term memory network,”IEEE Transactions on Industrial Informatics, vol. 16, no. 5, pp. 3168–3176, 2020

    work page 2020

  3. [3]

    Novel transformer based on gated convolutional neural network for dynamic soft sensor modeling of industrial processes,

    Z. Geng, Z. Chen, Q. Meng, and Y . Han, “Novel transformer based on gated convolutional neural network for dynamic soft sensor modeling of industrial processes,”IEEE Transactions on Industrial Informatics, vol. 18, no. 3, pp. 1521–1529, 2022

    work page 2022

  4. [4]

    Zero-shot capillary segmentation in dermoscopy images via sam2: A case study on oral mucosa,

    W. Su, Y . Zong, R. Jia, J. Qin, and M. Li, “Zero-shot capillary segmentation in dermoscopy images via sam2: A case study on oral mucosa,”IEEE Journal of Biomedical and Health Informatics, vol. 30, no. 5, pp. 4376–4387, 2026

    work page 2026

  5. [5]

    FIGAN: A missing industrial data imputation method customized for soft sensor application,

    Z. Yao and C. Zhao, “FIGAN: A missing industrial data imputation method customized for soft sensor application,”IEEE Transactions on Automation Science and Engineering, vol. 19, no. 4, pp. 3712–3722, 2022

    work page 2022

  6. [6]

    Semi-supervised probabilistic learning network for soft sensor mod- eling with partially labeled data,

    X. Shi, R. Li, H. Morales, A. Amicarelli, W. Huang, and W. Xiong, “Semi-supervised probabilistic learning network for soft sensor mod- eling with partially labeled data,”IEEE Transactions on Automation Science and Engineering, vol. 22, pp. 16 309–16 321, 2025

    work page 2025

  7. [7]

    Spatio-temporal consistent soft sensor modeling and monitoring of thermal power plants based on physical knowledge,

    Q. Wang, P. Sha, L. Heistrene, Y . Ding, and Y . Du, “Spatio-temporal consistent soft sensor modeling and monitoring of thermal power plants based on physical knowledge,”IEEE Transactions on Automation Science and Engineering, vol. 23, pp. 7682–7693, 2026

    work page 2026

  8. [8]

    A survey on deep learning for data-driven soft sensors,

    Q. Sun and Z. Ge, “A survey on deep learning for data-driven soft sensors,”IEEE Transactions on Industrial Informatics, vol. 17, no. 9, pp. 5853–5866, 2021. PREPRINT, JULY 2026 11

    work page 2021

  9. [9]

    Hybrid grid search and bayesian optimization-based random forest regression for predicting material compression pressure in manufacturing processes,

    Y . Zong, Y . Nian, C. Zhang, X. Tang, L. Wang, and L. Zhang, “Hybrid grid search and bayesian optimization-based random forest regression for predicting material compression pressure in manufacturing processes,”Engineering Applications of Artificial Intelligence, vol. 141, p. 109580, 2025. [Online]. Available: https://www.sciencedirect.com/ science/artic...

    work page 2025

  10. [10]

    A deep probabilistic transfer learning framework for soft sensor modeling with missing data,

    Z. Chai, C. Zhao, B. Huang, and H. Chen, “A deep probabilistic transfer learning framework for soft sensor modeling with missing data,”IEEE Transactions on Neural Networks and Learning Systems, vol. 33, no. 12, pp. 7598–7609, 2022

    work page 2022

  11. [11]

    Deep subdomain learning adaptation network: A sensor fault-tolerant soft sensor for industrial processes,

    X. Zhang, C. Song, J. Zhao, Z. Xu, and X. Deng, “Deep subdomain learning adaptation network: A sensor fault-tolerant soft sensor for industrial processes,”IEEE Transactions on Neural Networks and Learn- ing Systems, vol. 35, no. 7, pp. 9226–9237, 2024

    work page 2024

  12. [12]

    Model-agnostic meta-learning for fast adaptation of deep networks,

    C. Finn, P. Abbeel, and S. Levine, “Model-agnostic meta-learning for fast adaptation of deep networks,” inProceedings of the International Conference on Machine Learning (ICML), 2017, pp. 1126–1135

    work page 2017

  13. [13]

    A domain knowledge- guided industrial large model framework: A case study in battery health estimation and recycling,

    B. Chen, H. Shao, Y . Qin, Y . Jin, and X. Hu, “A domain knowledge- guided industrial large model framework: A case study in battery health estimation and recycling,”IEEE Transactions on Industrial Informatics, vol. 21, no. 10, pp. 8080–8090, 2025

    work page 2025

  14. [14]

    Socially aware load forecasting utilizing large language models,

    W. Chen, X. Zhang, L. Zhu, J. Shi, Z. Chang, Z. Han, and Y . Zhang, “Socially aware load forecasting utilizing large language models,”IEEE Transactions on Industrial Informatics, vol. 22, no. 1, pp. 231–242, 2026

    work page 2026

  15. [15]

    Explainable artificial intelligence for fault diagnosis of industrial processes,

    K. Jang, K. E. S. Pilario, N. Lee, I. Moon, and J. Na, “Explainable artificial intelligence for fault diagnosis of industrial processes,”IEEE Transactions on Industrial Informatics, vol. 21, no. 1, pp. 4–11, 2025

    work page 2025

  16. [16]

    Fault-prototypical adapted network for cross- domain industrial intelligent diagnosis,

    Z. Chai and C. Zhao, “Fault-prototypical adapted network for cross- domain industrial intelligent diagnosis,”IEEE Transactions on Automa- tion Science and Engineering, vol. 19, no. 4, pp. 3649–3658, 2022

    work page 2022

  17. [17]

    A meta- contrastive learning hybrid model for adaptive temperature trend prediction in variable ladle preheating,

    Y . Zong, R. Jia, S. Wu, L. Zhang, and D. He, “A meta- contrastive learning hybrid model for adaptive temperature trend prediction in variable ladle preheating,”Engineering Applications of Artificial Intelligence, vol. 162, p. 112750, 2025. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0952197625027812

    work page 2025

  18. [18]

    Tent: Fully test-time adaptation by entropy minimization,

    D. Wang, E. Shelhamer, S. Liu, B. Olshausen, and T. Darrell, “Tent: Fully test-time adaptation by entropy minimization,” inProceedings of the International Conference on Learning Representations (ICLR), 2021

    work page 2021

  19. [19]

    A 2RA-NSMTSllm: Adversarially aligning retrieval-augmented LLMs for nonstationary multivariate time series forecasting,

    J. Chu, C. Liu, X. Bai, and J. Tan, “A 2RA-NSMTSllm: Adversarially aligning retrieval-augmented LLMs for nonstationary multivariate time series forecasting,”IEEE Transactions on Industrial Informatics, vol. 22, no. 3, pp. 1805–1816, 2026

    work page 2026

  20. [20]

    Zero-shot fault diagnosis via LLM-guided complexity-aware fuzzy boundary learning,

    H. Han, X. Gao, H. Han, H. Gao, and Y . Qi, “Zero-shot fault diagnosis via LLM-guided complexity-aware fuzzy boundary learning,”IEEE Transactions on Industrial Informatics, pp. 1–12, 2026

    work page 2026

  21. [21]

    Causality-aware LLM-enhanced graph representation learning for adaptive power system control,

    F. Yao, J. Liu, Y . Tao, J. Qiu, H. H.-C. Iu, G. Chen, and Z. Y . Dong, “Causality-aware LLM-enhanced graph representation learning for adaptive power system control,”IEEE Transactions on Industrial Informatics, pp. 1–12, 2026

    work page 2026

  22. [22]

    Joint knowledge graph and large language model for fault diagnosis and its application in aviation assembly,

    P. Liu, L. Qian, X. Zhao, and B. Tao, “Joint knowledge graph and large language model for fault diagnosis and its application in aviation assembly,”IEEE Transactions on Industrial Informatics, vol. 20, no. 6, pp. 8160–8169, 2024

    work page 2024

  23. [23]

    Llm-driven human-ai collaborative decision support system for complex industrial processes: A case study in metallurgy,

    Y . Zong, R. Jia, K. Li, D. Xue, L. Zhang, and D. He, “Llm-driven human-ai collaborative decision support system for complex industrial processes: A case study in metallurgy,”Neural Networks, vol. 202, p. 109055, 2026

    work page 2026

  24. [24]

    Are language models actually useful for time series forecasting?

    M. Tan, M. A. Merrill, V . Gupta, T. Althoff, and T. Hartvigsen, “Are language models actually useful for time series forecasting?” inAdvances in Neural Information Processing Systems 37, vol. 37, 2024, pp. 60 162–60 191. [Online]. Available: https://proceedings.neurips.cc/paper files/paper/2024/hash/ 6ed5bf446f59e2c6646d23058c86424b-Abstract-Conference.html

    work page 2024

  25. [25]

    Adaptive multimodal industrial fault diagnosis with attention-driven fusion boosting unimodal performance,

    D. Wang, X. Pan, Y . Gao, F. Tsung, and F. Li, “Adaptive multimodal industrial fault diagnosis with attention-driven fusion boosting unimodal performance,”IEEE Transactions on Automation Science and Engineer- ing, vol. 23, pp. 4454–4476, 2026

    work page 2026

  26. [26]

    Dynamic dependencies of fault-related variables based interpretable fault diagnosis framework for industrial processes,

    L. Ma, Q. Yang, and K. Peng, “Dynamic dependencies of fault-related variables based interpretable fault diagnosis framework for industrial processes,”IEEE Transactions on Automation Science and Engineering, pp. 1–1, 2026

    work page 2026

  27. [27]

    Causal disen- tangled graph neural network for fault diagnosis of complex industrial process,

    R. Liu, Q. Zhang, D. Lin, W. Zhang, and S. X. Ding, “Causal disen- tangled graph neural network for fault diagnosis of complex industrial process,”IEEE Transactions on Industrial Informatics, vol. 21, no. 1, pp. 386–395, 2025

    work page 2025

  28. [28]

    Bayesian hierarchical graph neural networks with uncertainty feedback for trust- worthy fault diagnosis of industrial processes,

    D. Chen, Z. Xie, R. Liu, W. Yu, Q. Hu, X. Li, and S. X. Ding, “Bayesian hierarchical graph neural networks with uncertainty feedback for trust- worthy fault diagnosis of industrial processes,”IEEE Transactions on Neural Networks and Learning Systems, vol. 35, no. 12, pp. 18 635– 18 648, 2024

    work page 2024

  29. [29]

    Deep bayesian slow feature extraction with application to industrial inferential modeling,

    C. Jiang, Y . Lu, W. Zhong, B. Huang, D. Tan, W. Song, and F. Qian, “Deep bayesian slow feature extraction with application to industrial inferential modeling,”IEEE Transactions on Industrial Informatics, vol. 19, no. 1, pp. 40–51, 2023

    work page 2023

  30. [30]

    Online fault diagnosis for industrial processes with bayesian network-based probabilistic ensemble learning strategy,

    W. Yu and C. Zhao, “Online fault diagnosis for industrial processes with bayesian network-based probabilistic ensemble learning strategy,”IEEE Transactions on Automation Science and Engineering, vol. 16, no. 4, pp. 1922–1932, 2019

    work page 1922

  31. [31]

    Twofold weighted-based statistical feature KECA for nonlinear industrial process fault diagnosis,

    T. Li, Y . Han, X. Hu, B. Ma, and Z. Geng, “Twofold weighted-based statistical feature KECA for nonlinear industrial process fault diagnosis,” IEEE Transactions on Automation Science and Engineering, vol. 22, pp. 3901–3910, 2025

    work page 2025

  32. [32]

    A tutorial on support vector regression,

    A. J. Smola and B. Sch ¨olkopf, “A tutorial on support vector regression,” Statistics and Computing, vol. 14, no. 3, pp. 199–222, 2004

    work page 2004

  33. [33]

    Xgboost: A scalable tree boosting system,

    T. Chen and C. Guestrin, “Xgboost: A scalable tree boosting system,” inProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016, pp. 785–794

    work page 2016

  34. [34]

    Learning phrase representations using RNN encoder–decoder for statistical machine translation,

    K. Cho, B. van Merri ¨enboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y . Bengio, “Learning phrase representations using RNN encoder–decoder for statistical machine translation,” in Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). Doha, Qatar: Association for Computational Linguistics, 2014, pp...

    work page 2014

  35. [35]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,”Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997

    work page 1997

  36. [36]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” inAdvances in Neural Information Processing Systems 30 (NeurIPS 2017), 2017, pp. 5998–6008. [Online]. Available: https://papers.neurips.cc/paper/7181-attention-is-all-you-need

    work page 2017

  37. [37]

    Informer: Beyond efficient transformer for long sequence time-series forecasting,

    H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang, “Informer: Beyond efficient transformer for long sequence time-series forecasting,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 12, 2021, pp. 11 106–11 115

    work page 2021

  38. [38]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” inThe First Conference on Language Modeling (COLM), 2024. [Online]. Available: https://openreview.net/forum?id= tEYskw1VY2

    work page 2024

  39. [39]

    The development of an industrial-scale fed-batch fermentation simulation,

    S. Goldrick, A. S ¸tefan, D. Lovett, G. Montague, and B. Lennox, “The development of an industrial-scale fed-batch fermentation simulation,” Journal of Biotechnology, vol. 193, pp. 70–82, 2015

    work page 2015