Pith. sign in

REVIEW 1 major objections 7 minor 32 references

LLM reads plant docs to fix sensor errors before prediction

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 · glm-5.2

2026-07-08 16:12 UTC pith:P5DXZ3G6

load-bearing objection MCC uses LLM-extracted measurement semantics to build independent process references and corrects supported local measurement conflicts before prediction — a genuinely new pipeline idea with strong empirical results and a clean, conservative design. the 1 major comments →

arxiv 2607.06111 v1 pith:P5DXZ3G6 submitted 2026-07-07 eess.SY cs.AIcs.SY

LLM-Guided Measurement Credibility Correction for Trustworthy Industrial Process Inference

Youcheng Zong , Runda Jia , Dakuo He This is my paper
classification eess.SY cs.AIcs.SY
keywords measurement credibilityindustrial soft sensingLLM semantic extractionsensor fault correctionpre-inference correctionmeasurement semanticsprocess inference
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.

Industrial process prediction models fail not only when the forecasting model is weak, but when the incoming sensor readings are already wrong — biased, drifted, stale, or derived from shared instruments. This paper proposes Measurement Credibility Correction (MCC), a method that uses a large language model to read process documentation and variable metadata offline, extract measurement semantics (what each variable measures, how, and whether it is independent of other variables), and then use those semantics online to build an independent reference for each correctable measurement. When a measurement conflicts with its independent reference beyond a calibrated threshold, MCC nudges it toward the reference before the predictor ever sees it. The LLM is never called at inference time; its output is frozen into a semantic matrix that guides a lightweight numerical observer. Across three industrial datasets and eight forecasting backbones, MCC reduces mean absolute error by 30.7% on real test splits and 80.3% on splits with injected sensor corruption, while adding only 0.5–2.0k parameters and at most 0.089 ms/step of inference latency.

Core claim

The central object is the measurement-semantic direction: an embedding vector derived offline by an LLM from process documents, encoding each variable's measurement source, process role, and independence relationships. These directions form a frozen matrix that a numerical observer uses to route external evidence and build an independent reference for each target variable — without the target variable's own reading entering the reference. The key mechanism is supported local inconsistency: a measurement is corrected only when it deviates from its reference more than the background inconsistency across all variables (so genuine plant-wide operating changes are not mistaken for sensor faults),

What carries the argument

measurement-semantic direction

Load-bearing premise

The entire method depends on the LLM correctly identifying which variables are genuinely independent references for each target variable — meaning they do not share instruments, derived formulas, or control actions. If the LLM misclassifies a dependent variable as independent, the reference it builds is not independent, and correction would amplify the distortion rather than fix it. The paper's ablation with shuffled semantics shows degradation, but does not test the case of

What would settle it

Inject correlated corruption into two variables that the LLM classified as mutually independent references; if MCC amplifies rather than reduces error, the independence classification was wrong.

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

If this is right

  • Process plants with existing documentation but no explicit fault-detection infrastructure could deploy MCC as a pre-inference filter that catches sensor drift before it corrupts downstream predictions, without requiring labeled fault data or process balance equations.
  • The offline-LLM / online-numerical split means the semantic reasoning cost is paid once at deployment setup, not per inference step, making the approach viable for real-time control loops with minute-scale sampling intervals.
  • If measurement semantics can be kept current as plants are retrofitted or instruments are replaced, the same framework could serve as a continuous credibility layer across changing measurement chains — though the paper identifies this as an open problem.
  • The background-subtraction design implies that correlated multi-variable faults (where several sensors drift together) would not be caught, which sets a boundary on the class of degradations MCC can address.

Where Pith is reading between the lines

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

  • The framework could be extended to domains beyond process industries — such as building energy management or medical monitoring — wherever heterogeneous sensor metadata exists and individual channels can independently drift.
  • A natural test of the independence assumption would be to inject correlated corruption into two variables that the LLM classified as independent references for each other; if the LLM was wrong about independence, MCC would amplify error rather than correct it, and this failure mode would be distinguishable from the ablation results currently reported.
  • The sensitivity analysis showing stability across different LLM generators and embedding models suggests the gain comes from the semantic structure itself rather than any single model's reasoning quality — but it also implies that a simpler, non-LLM knowledge source (such as a curated ontology) might achieve comparable results if one existed for the plant.
  • The conservative correction design (calibrated coverage at 95%, background subtraction, support weighting) means MCC trades maximum correction power for safety; a plant with well-characterized fault modes might benefit from a less conservative gate, but this would require fault labels the current method deliberately avoids.

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

1 major / 7 minor

Summary. The paper proposes LLM-Guided Measurement Credibility Correction (MCC), a pre-inference correction module for industrial soft sensing and forecasting. MCC uses an LLM offline to extract measurement semantics from process documentation, builds leave-one-out independent process references from semantically qualified external variables, and applies a calibrated, conservative correction gate before the downstream predictor receives the input window. The method is evaluated on three industrial datasets (ladle preheating, thickener dewatering, IndPenSim) across eight backbones under both real-test and controlled-corruption protocols, reporting average relative MAE reductions of 30.7% and 80.3% respectively, with 0.5–2.0k additional online parameters and ≤0.089 ms/step inference time. The mathematical formulation is internally consistent: the observer is leave-one-out by construction (Eq. 11 excludes x_i), calibration uses a clean set (Eq. 18), and the correction gate is conservative (Eqs. 20–21). Code is reported as publicly available.

Significance. The paper addresses a practically important problem: industrial predictors can receive biased, stale, or derived measurements that look plausible but no longer represent the real process. The key insight—that a correlated variable is not necessarily a credible reference when variables share instruments, derived formulas, or control actions—is well-motivated and distinguishes MCC from correlation-based sensor reconstruction. The offline-LLM/online-numerical architecture is a sensible design choice that avoids LLM latency at inference time. The ablation study (Table VII) systematically tests key components, and the computational efficiency analysis (Table VIII) demonstrates genuine deployability. The method is backbone-agnostic and evaluated across multiple real industrial datasets. These are meaningful contributions to the soft-sensing and industrial inference literature.

major comments (1)
  1. §IV.E, Table VII: The ablation study tests 'Shuffled Semantic' and 'Correlation' variants, but neither addresses the most dangerous failure mode: a single confident LLM misclassification of a shared-instrument or derived variable as independent. If the LLM misclassifies such a variable, the observer (Eq. 22) learns the spurious correlation on nominal data, and when the shared instrument drifts, both target and reference move together, making the inconsistency undetectable (Eq. 15 yields near-zero residual). Worse, if only the reference is affected by the shared instrument fault, MCC would actively correct the target toward the corrupted reference (Eq. 21), amplifying error. The paper acknowledges this boundary in §V but provides no experiment constructing a known shared-instrument or derived relationship to verify the LLM identifies it correctly. This is load-bearing because the entire校正
minor comments (7)
  1. §III.B, Eq. (3): The LLM prompt or interaction protocol used to generate semantic cards is not described. A brief example of the prompt structure and a sample output would improve reproducibility.
  2. §IV.A, Table I: The LLM is listed as 'DeepSeek-v4-pro' and embedding model as 'text-embedding-v4'. These model versions should be verified for accuracy; if they are placeholders or internal names, the actual model identifiers should be provided.
  3. Tables IV–VI: No standard deviations or significance tests are reported across the three independent runs mentioned in §IV.A. Adding error bars or at minimum noting whether improvements are statistically significant would strengthen the claims.
  4. §IV.D, Fig. 5: The figure caption could specify which corruption types and magnitudes are shown for the Corrupted Test panels, as these details are not immediately clear from the axis labels.
  5. §III.D, Eq. (17): The background subtraction uses median over M. The sensitivity to the choice of median versus mean or other robust estimators is not discussed. A brief note on this design choice would be helpful.
  6. References [2], [18], [19], [20], [22]: Several references are dated 2026. Please verify these are correctly cited and publicly available at the time of submission.
  7. §IV.G, Fig. 7: The sensitivity analysis figure is dense and some labels overlap. Improving readability would help readers assess the robustness claims.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful and constructive assessment. The referee correctly identifies that the most dangerous failure mode for MCC—a single confident LLM misclassification of a shared-instrument or derived variable as independent—is not experimentally tested in the current manuscript. We agree this is a load-bearing concern and will add a targeted experiment in revision. Below we address the major comment point by point.

read point-by-point responses

  1. Referee: §IV.E, Table VII: The ablation study tests 'Shuffled Semantic' and 'Correlation' variants, but neither addresses the most dangerous failure mode: a single confident LLM misclassification of a shared-instrument or derived variable as independent. If the LLM misclassifies such a variable, the observer (Eq. 22) learns the spurious correlation on nominal data, and when the shared instrument drifts, both target and reference move together, making the inconsistency undetectable (Eq. 15 yields near-zero residual). Worse, if only the reference is affected by the shared instrument fault, MCC would actively correct the target toward the corrupted reference (Eq. 21), amplifying error. The paper acknowledges this boundary in §V but provides no experiment constructing a known shared-instrument or derived relationship to verify the LLM identifies it correctly. This is load-bearing because the entire校正

    Authors: The referee is correct on all counts. This is a genuine gap in the experimental validation, and we will address it in revision. revision: yes

Circularity Check

0 steps flagged

No significant circularity; one minor self-citation that is not load-bearing

full rationale

The paper's derivation chain is largely self-contained. The observer is fitted on nominal training data to minimize (x*_i - x_i)^2 (Eq. 22), and the calibration distribution F̂_i is estimated on a clean calibration set (Eq. 18). The correction strength a_i(t) is then determined by how far the current inconsistency exceeds the 95th percentile of this calibration distribution (Eq. 19-20). While the reader's take raises a valid concern that the observer learns to reproduce nominal data and thus the 'correction' is triggered by deviation from a reference fitted on clean data, this is a standard calibration approach, not a circularity: the observer parameters (ω_τ, β_{ij,τ}) are fitted on training data, while the calibration distribution and residual scales are estimated on a separate clean calibration set C. The correction is applied to test-time data that may contain corruptions not present in training or calibration. The central claim—that LLM-derived semantics improve measurement credibility correction—is tested against external benchmarks (three industrial datasets, eight backbones, ablations in Table VII, sensitivity analysis in Fig. 7). The self-citations (refs [9], [19], [22], [29]) are to prior work on related but distinct topics (material pressure prediction, human-AI decision support, capillary segmentation, ladle preheating trend prediction) and do not form a load-bearing chain for the present method's core derivation. The LLM (DeepSeek-v4-pro) and embedding model (text-embedding-v4) are external commercial models, not the authors' own. No equation reduces to its own inputs by construction, and no 'prediction' is a renamed fit. The reader's concern about the observer being fitted to nominal data is a correctness risk (will the reference generalize to corrupted test data?) rather than a circularity (is the result equivalent to the input by construction?). The paper acknowledges this boundary explicitly in §V. Score 2 reflects the minor self-citations present but not load-bearing for the central derivation.

Axiom & Free-Parameter Ledger

8 free parameters · 4 axioms · 2 invented entities

The axiom ledger captures the key fitted parameters (κ, H, sparsity, k, ω_τ, β_{ij,τ}, s_i, η_i) and the domain assumptions the method depends on. The most critical axiom is that LLM-extracted semantics accurately reflect the real measurement chain—this is the load-bearing premise. The invented entities (semantic cards and directions) have partial independent evidence through ablations but lack external validation against ground-truth measurement topology.

free parameters (8)
  • κ (nominal coverage) = 0.95
    Sets the fraction of clean calibration states that should not trigger correction. Default value chosen by the authors; sensitivity analysis (Fig. 7) shows stability around this point.
  • H (observer lags) = 3
    Number of lags used by the observer. Chosen as default; sensitivity analysis shows stability for H∈{1,2,3,5}.
  • Support sparsity (top-k) = 6
    Limits the number of candidate variable–lag pairs in routing. Chosen as default; sensitivity analysis shows stability for k∈{3,6,all}.
  • k (semantic dimension) = 128
    Dimension of the embedding space for semantic directions. Set by the embedding model choice.
  • ω_τ (semantic weight matrix per lag) = learned
    Learned from nominal training data via Eq. 22. Shared across all variables at the same lag.
  • β_{ij,τ} (readout coefficients) = learned
    Learned from nominal training data via Eq. 22. Aligns qualified sources to the standardized scale of x_i.
  • s_i (nominal residual scale) = median of |x̄_i − x̄*_i| on calibration set
    Estimated on clean calibration set C. Not a free parameter in the traditional sense but a data-driven quantity.
  • η_i (reference reliability) = 1/(1+s_i)
    Derived from s_i. Limits correction magnitude.
axioms (4)
  • domain assumption The LLM can accurately extract measurement semantics (source, role, independence, derived relations) from semi-structured process documents.
    Invoked in §III.B (Eq. 3): z_i = Φ_LLM(D, i). The entire method depends on the LLM's semantic cards being correct. The paper does not validate LLM output against ground-truth measurement chains.
  • domain assumption Variables identified as semantically independent references are not jointly distorted by unmodeled shared causes (e.g., shared instruments, control actions not captured in documentation).
    Invoked in §III.C: the reference x̄*_i is built from variables j≠i that the LLM qualifies as independent. If a shared instrument or control action distorts both x_i and x_j, the reference is not independent.
  • domain assumption The clean calibration set C is representative of nominal operating conditions and does not contain hidden measurement distortions.
    Invoked in §III.D (Eq. 18): F̂_i is estimated on C. If C contains hidden distortions, the calibration distribution is contaminated and the correction gate fires incorrectly.
  • domain assumption The background inconsistency median (Eq. 17) correctly distinguishes local measurement conflicts from real shared operating changes.
    Invoked in §III.D: χ_i = d_i − median_{ℓ≠i} d_ℓ. This assumes that when many variables deviate together, it reflects a real operating change rather than correlated measurement corruption.
invented entities (2)
  • Measurement-semantic card (z_i) independent evidence
    purpose: Records measurement source, process role, control/observation attribute, sampling relation, upstream/downstream position, and derived relations for each variable, plus redundant variables that can provide independent references.
    The ablation with shuffled semantics (Table VII) shows that the semantic match matters for performance. The sensitivity analysis (Fig. 7) shows stability across different LLM generators and embedding models. However, no external validation of semantic card accuracy against ground-truth measurement chains is provided.
  • Semantic direction (u_i) independent evidence
    purpose: Frozen embedding vector representing the measurement semantics of each variable, used to compute compatibility and routing weights.
    Derived from the semantic card via a fixed renderer and embedding model (Eq. 4). The ablation replacing LLM cards with metadata (Table VII) shows degradation, providing evidence that the richer semantic content matters.

pith-pipeline@v1.1.0-glm · 20080 in / 3210 out tokens · 586332 ms · 2026-07-08T16:12:13.222396+00:00 · methodology

0 comments
read the original abstract

Industrial prediction and soft sensing depend on credible input measurements. In field deployment, a predictor may receive biased, delayed, stale, or derived measurements that still look plausible. Prediction can then fail before the forecasting backbone becomes the main limitation, because the input window no longer represents the real process. Sensor reconstruction, data reconciliation, and fault-tolerant soft sensing reduce this risk, but they often rely on numerical correlation, alarms, fault labels, or explicit process equations. These assumptions are not always available. A correlated variable can also be an unsafe reference when variables share instruments, derived formulas, soft-sensing chains, or control actions. The key issue is to decide before prediction which external measurements can credibly support the current measurement. To address this issue, this article proposes LLM-Guided Measurement Credibility Correction (MCC). MCC converts measurement meanings in process documents into measurement semantics usable by numerical models. It builds independent process references from semantically qualified external measurements and corrects local measurement conflicts before prediction. The predictor therefore receives a more credible input window. Across multiple complex industrial forecasting and soft-sensing tasks, +MCC achieves average relative MAE reductions of 30.7% on real-test protocols and 80.3% on controlled-corruption protocols. It adds only 0.5--2.0k online parameters, with the slowest +MCC inference time at 0.089 ms/step. These results show that measurement semantics can turn process documents into lightweight pre-inference credibility correction and improve prediction accuracy.

Figures

Figures reproduced from arXiv: 2607.06111 by Dakuo He, Runda Jia, Youcheng Zong.

Figure 1
Figure 1. Figure 1: MCC problem setting and neighboring routes. (a) Raw mea￾surement windows are fed directly to the predictor. (b) Neighboring numerical routes handle inconsistency through alarms, reconstruction, or reconciliation. (c) MCC builds measurement semantics offline and uses LLM-free external references online to decide whether to correct local conflicts. qualified external evidence for each correctable variable an… view at source ↗
Figure 2
Figure 2. Figure 2: MCC method overview. Left: offline LLM processing builds the frozen measurement semantics U from D. Middle: nominal training fits the observer, and clean calibration fixes si, ηi, and Fbi. Right: online correction computes independent references and correction strengths without LLM calls before sending W¯ c t to Gθ. The coefficient βij,τ is an unbiased numerical readout co￾efficient. Semantic compatibility… view at source ↗
Figure 3
Figure 3. Figure 3: Measurement-semantic reference and local correction path. The target measurement x¯i(t) is excluded from the evidence set, so x¯ ∗ i (t) is built only from external variables and their lags. The right side shows how support weighting, background subtraction, and calibration convert ri(t) into a mixing weight. For i /∈ M, MCC keeps x¯ c i (t) = ¯xi(t). These variables may enter the observer as evidence, but… view at source ↗
Figure 4
Figure 4. Figure 4: Industrial processes used in the experiments. (a) Tailings thickening dewatering site in Nanjing, China. (b) Ladle preheating site at a steel plant in Maanshan, China. TABLE V QUANTITATIVE COMPARISON ON THE THICKENER DEWATERING DATASET. BASE AND +MCC ARE EVALUATED UNDER THE SAME REAL TEST AND CORRUPTED TEST PROTOCOLS. Backbone Setting Real Test Corrupted Test MAE↓ RMSE↓ R2↑ MAE↓ RMSE↓ R2↑ GRU Base 0.0147 0… view at source ↗
Figure 5
Figure 5. Figure 5: Representative Transformer prediction trajectories. The panels compare Base, +MCC, and target curves under the same Real Test and Corrupted Test protocols as the main tables. soft-sensing offset under fault operation. On Corrupted Test, the difference follows the injected channel errors. Temperature drift and temperature bias shift the ladle baseline prediction upward, and MCC reduces this erroneous offset… view at source ↗
Figure 6
Figure 6. Figure 6: MCC correction mechanism on IndPenSim with Transformer. (a) Event replay for a local volume bias. (b) Gate activation versus background inconsistency for clean, local-conflict, and shared-shift windows. (c) Replay spectrum for local volume corruptions, with remaining error in the heatmap and gate activation/recovery on the side rails. (d) Evidence circuit for volume. (e) Controlled-corruption response whee… view at source ↗
Figure 7
Figure 7. Figure 7: Sensitivity of MCC to main design choices. Values are normalized by the default setting; horizontal and vertical dashed lines mark default performance and configurations. TABLE VIII ONLINE DEPLOYMENT COST OF MCC. THE TABLE COUNTS ONLY ONLINE PARAMETERS, INFERENCE TIME, AND PEAK RSS; OFFLINE LLM SEMANTIC GENERATION AND EMBEDDING CONSTRUCTION ARE EXCLUDED FROM PER-STEP INFERENCE. Model Size Deployment Cost B… 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

32 extracted references · 32 canonical work pages · 3 internal anchors

  1. [1]

    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

    work page 2021

  2. [2]

    Transfer learning for soft sensors in process industries: A review and future perspectives,

    Y . Liu, J. Zhu, C. Yang, T. Chen, D. S. H. Wong, and Y . Yao, “Transfer learning for soft sensors in process industries: A review and future perspectives,”Industrial & Engineering Chemistry Research, vol. 65, no. 16, pp. 8103–8125, 2026

    work page 2026

  3. [3]

    Reconstruction-based fault identification using a combined index,

    H. H. Yue and S. J. Qin, “Reconstruction-based fault identification using a combined index,”Industrial & Engineering Chemistry Research, vol. 40, no. 20, pp. 4403–4414, 2001

    work page 2001

  4. [4]

    Fault-tolerant soft sensor modeling based on a two-dimensional group distributionally robust optimization framework,

    X. Zhang, C. Song, J. Zhao, and B. Huang, “Fault-tolerant soft sensor modeling based on a two-dimensional group distributionally robust optimization framework,”IEEE Transactions on Automation Science and Engineering, vol. 22, pp. 14 396–14 406, 2025

    work page 2025

  5. [5]

    Time-LLM: Time series forecasting by reprogramming large language models,

    M. Jin, S. Wang, L. Ma, Z. Chu, J. Y . Zhang, X. Shi, P.-Y . Chen, Y . Liang, Y .-F. Li, S. Pan, and Q. Wen, “Time-LLM: Time series forecasting by reprogramming large language models,” inProceedings of the International Conference on Learning Representations (ICLR),

    work page

  6. [6]

    Time-LLM: Time Series Forecasting by Reprogramming Large Language Models

    [Online]. Available: https://arxiv.org/abs/2310.01728

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    AutoTimes: Autoregressive Time Series Forecasters via Large Language Models

    Y . Liu, G. Qin, X. Huang, J. Wang, and M. Long, “AutoTimes: Autoregressive time series forecasters via large language models,” in Advances in Neural Information Processing Systems 37, 2024. [Online]. Available: https://arxiv.org/abs/2402.02370

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  8. [8]

    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

  9. [9]

    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

  10. [10]

    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, ZONGet al.: LLM-GUIDED MEASUREMENT CREDIBILITY CORRECTION 11 p. 109580, 2025. [Onl...

    work page 2025

  11. [11]

    Reliable soft sensors with an inherent process graph constraint,

    R. Zhai, J. Zheng, Z. Song, and Z. Ge, “Reliable soft sensors with an inherent process graph constraint,”IEEE Transactions on Industrial Informatics, vol. 20, no. 6, pp. 8798–8806, 2024

    work page 2024

  12. [12]

    A review of process fault detection and diagnosis,

    V . Venkatasubramanian, R. Rengaswamy, K. Yin, and S. N. Kavuri, “A review of process fault detection and diagnosis,”Computers & Chemical Engineering, vol. 27, no. 3, pp. 293–311, 2003

    work page 2003

  13. [13]

    Survey on data-driven industrial process monitoring and diagnosis,

    S. J. Qin, “Survey on data-driven industrial process monitoring and diagnosis,”Annual Reviews in Control, vol. 36, no. 2, pp. 220–234, 2012

    work page 2012

  14. [14]

    Narasimhan and C

    S. Narasimhan and C. Jordache,Data Reconciliation and Gross Error Detection: An Intelligent Use of Process Data. Gulf Professional Publishing, 1999

    work page 1999

  15. [15]

    Graph neural network-based anomaly detection in multivariate time series,

    A. Deng and B. Hooi, “Graph neural network-based anomaly detection in multivariate time series,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 5, 2021, pp. 4027–4035

    work page 2021

  16. [16]

    Adversarial graph neural network for multivariate time series anomaly detection,

    B. Zheng, L. Ming, K. Zeng, M. Zhou, X. Zhang, T. Ye, B. Yang, X. Zhou, and C. S. Jensen, “Adversarial graph neural network for multivariate time series anomaly detection,”IEEE Transactions on Knowledge and Data Engineering, vol. 36, no. 12, pp. 7612–7626, 2024

    work page 2024

  17. [17]

    TimeCMA: Towards LLM-empowered multivariate time series forecasting via cross-modality alignment,

    C. Liu, Q. Xu, H. Miao, S. Yang, L. Zhang, C. Long, Z. Li, and R. Zhao, “TimeCMA: Towards LLM-empowered multivariate time series forecasting via cross-modality alignment,” inProceedings of the AAAI Conference on Artificial Intelligence, 2025, pp. 18 780–18 788. [Online]. Available: https://mlanthology.org/aaai/2025/ liu2025aaai-timecma/

    work page 2025

  18. [18]

    TEST: Text prototype aligned embedding to activate LLM’s ability for time series,

    C. Sun, H. Li, Y . Li, and S. Hong, “TEST: Text prototype aligned embedding to activate LLM’s ability for time series,” inProceedings of the International Conference on Learning Representations (ICLR),

    work page

  19. [19]

    TEST: Text Prototype Aligned Embedding to Activate LLM's Ability for Time Series

    [Online]. Available: https://arxiv.org/abs/2308.08241

    work page internal anchor Pith review Pith/arXiv arXiv

  20. [20]

    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, vol. 22, no. 5, pp. 3681–3692, 2026

    work page 2026

  21. [21]

    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. [Online]. Available: https: //www.sciencedirect.com/science/article/pii/S0893608026005150

    work page 2026

  22. [22]

    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

  23. [23]

    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

  24. [24]

    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

  25. [25]

    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

  26. [26]

    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

  27. [27]

    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

  28. [28]

    iTransformer: Inverted transformers are effective for time series forecasting,

    Y . Liu, T. Hu, H. Zhang, H. Wu, S. Wang, L. Ma, and M. Long, “iTransformer: Inverted transformers are effective for time series forecasting,” inProceedings of the International Conference on Learning Representations (ICLR), 2024. [Online]. Available: https://openreview.net/forum?id=JePfAI8fah

    work page 2024

  29. [29]

    A time series is worth 64 words: Long-term forecasting with transformers,

    Y . Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam, “A time series is worth 64 words: Long-term forecasting with transformers,” inProceedings of the International Conference on Learning Representations (ICLR), 2023. [Online]. Available: https://openreview.net/forum?id=Jbdc0vTOcol

    work page 2023

  30. [30]

    ModernTCN: A modern pure convolution structure for general time series analysis,

    D. Luo and X. Wang, “ModernTCN: A modern pure convolution structure for general time series analysis,” inProceedings of the International Conference on Learning Representations (ICLR), 2024. [Online]. Available: https://openreview.net/forum?id=vpJMJerXHU

    work page 2024

  31. [31]

    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

  32. [32]

    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