pith. sign in

arxiv: 2606.26861 · v1 · pith:BTYGCXULnew · submitted 2026-06-25 · 💻 cs.CL

Cascaded Multi-Granularity Pruning for On-Device LLM Inference in Industrial IoT

Jinghan Wang , Yanjun Chen , Wei Zhang , Xiaotong Huang , Tianchen Liu , Gaoliang Peng This is my paper

Pith reviewed 2026-06-26 04:55 UTC · model grok-4.3

classification 💻 cs.CL
keywords LLM pruningmodel compressionindustrial IoTedge inferenceStructural Independence Assumptionmulti-granularity pruningon-device deploymentbearing fault diagnosis
0
0 comments X

The pith

Cascaded multi-granularity pruning reaches 13.8 times compression on MHA+GELU LLMs for IIoT edge devices.

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

The paper presents a cascaded pruning framework that removes LLM components in stages from layers to attention heads to feed-forward channels, inserting low-rank recovery between stages to update importance estimates. An information-theoretic ordering is justified by the Structural Independence Assumption, a checkable condition that holds for MHA+GELU architectures but fails for GQA+SwiGLU ones. On fault-diagnosis models ranging from 88 million to 6.25 billion parameters, the approach delivers 13.8 times compression with 83.82 percent accuracy on satisfying architectures while producing a 74-point accuracy collapse on violating ones. The results matter because one-shot structured pruning methods lose reliability at the extreme ratios required for industrial edge hardware.

Core claim

The cascaded multi-granularity pruning framework removes layers, attention heads, and feed-forward channels in coarse-to-fine order with lightweight low-rank recovery between stages; an information-theoretic analysis supplies the ordering, and the Structural Independence Assumption predicts that per-component pruning criteria remain reliable precisely when the architecture satisfies SIA, which MHA+GELU models do and GQA+SwiGLU models do not, yielding 13.8 times compression at 83.82 percent accuracy on the former class and a roughly 74 percentage-point accuracy drop on the latter.

What carries the argument

The cascaded multi-granularity pruning procedure ordered by information-theoretic analysis and conditioned on the Structural Independence Assumption, which determines whether component-wise importance scores stay reliable across pruning stages.

If this is right

  • Higher compression ratios become usable on SIA-satisfying architectures without the accuracy collapse seen in one-shot methods.
  • Architecture selection can be guided by a checkable SIA test before pruning is applied.
  • Inter-stage low-rank recovery improves importance re-estimation enough to support 13.8 times overall reduction.
  • Inference latency drops up to 67.2 percent and peak memory by 62.5 percent on the target industrial hardware once the pruned model is deployed.
  • The same staged procedure scales across model sizes from 88 million to 6.25 billion parameters in the bearing-fault domain.

Where Pith is reading between the lines

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

  • Designers could deliberately choose MHA+GELU blocks when the goal is extreme structured pruning for edge devices.
  • The SIA test itself might be turned into an automated pre-pruning diagnostic for any new transformer variant.
  • Low-rank recovery inserted between pruning stages could be generalized to other iterative compression pipelines.
  • The 74-point collapse on violating architectures indicates that architecture-specific recovery modules may be needed when SIA fails.

Load-bearing premise

The Structural Independence Assumption correctly flags which architectures keep per-component pruning criteria reliable.

What would settle it

A direct measurement showing that GQA+SwiGLU models retain accuracy under the cascaded procedure at the same compression ratios where MHA+GELU models succeed.

Figures

Figures reproduced from arXiv: 2606.26861 by Gaoliang Peng, Jinghan Wang, Tianchen Liu, Wei Zhang, Xiaotong Huang, Yanjun Chen.

Figure 1
Figure 1. Figure 1: Overview of the proposed multi-granularity cascaded pruning framework for IIoT edge deployment. The pipeline proceeds from coarse to fine granularity, including layer pruning, attention head pruning, and FFN channel pruning with LoRA-based staged recovery and importance redistribution between consecutive stages. The three theoretical contributions respectively address how to prune across granularities, why… view at source ↗
Figure 2
Figure 2. Figure 2: Stage-by-stage accuracy trajectory of the cascaded pruning process at three compression levels (3.72×, 5.82×, 13.81×). The sawtooth pattern reflects alternating pruning (accuracy drop) and staged LoRA recovery (accuracy restoration) phases. TABLE III OVERALL PERFORMANCE OF THE PROPOSED FRAMEWORK ACROSS FOUR LLM-BASED FAULT DIAGNOSIS MODELS Case Original Parameters (Million) Pruned Parameters (Million) Comp… view at source ↗
Figure 5
Figure 5. Figure 5: Cascade order ablation across three compression levels. Four orderings are compared: L→H→F (proposed), F→H→L (reverse), H→L→F, and Simultaneous. 2) Per-Granularity Contribution Ablation We ablate individual granularity levels on the Fusion model with the others fixed, plus a Layer-Only baseline; results are in the upper panel of [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: visualizes this difference through head importance heatmaps, where each cell encodes the L1 importance of a head (row) per layer (column). The cascade produces sharper, sparser maps because it evaluates heads after the layer structure has stabilized. The cascade evaluates head importance after the layer structure has stabilized, whereas one-shot methods may preserve heads that appear individually important… view at source ↗
Figure 6
Figure 6. Figure 6: Ablation study results. (a)–(b) Per-granularity contribution ablation: per￾class recall and F1-score under five configurations (Full, No-Layer, No-Head, No-FFN, Layer-Only). (c)–(d) Recovery strategy ablation: per-class recall and F1-score for five recovery strategies (None, LoRA r=4/8/16, Full FT). 3) Recovery Strategy Ablation To validate LoRA-based recovery and determine the optimal rank, we compare fiv… view at source ↗
Figure 7
Figure 7. Figure 7: Layer importance analysis comparing structural (LCR) and task-aware (gradient-based) importance scores across all transformer layers. Blue/red [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Architectural comparison between GPT-2 and ChatGLM-2, including full transformer block diagrams, attention mechanism and FFN contrasts [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Industrial experimental setup and the corresponding edge deployment efficiency profile on NVIDIA DGX Spark 128GB. Five metrics are compared across three pruned models: size reduction, latency reduction, throughput improvement, VRAM reduction, and energy efficiency improvement. REFERENCES [1] I. Kok, O. De irci, and . Ozde ir, “ hen IoT Meet LLMs: A lications and Challenges,” ov. 20, 2024, arXiv: arXiv:2411… view at source ↗
read the original abstract

Deploying large language models (LLMs) on Industrial Internet of Things (IIoT) edge devices demands extreme compression, yet existing structured pruning methods collapse at high compression ratios due to one-shot importance estimation, and their cross-architecture behavior remains unpredictable. This article presents a cascaded multi-granularity pruning framework that removes layers, attention heads, and feed-forward channels in coarse-to-fine order, with lightweight low-rank recovery between stages to re-estimate component importance. An information-theoretic analysis motivates this ordering, and the Structural Independence Assumption (SIA) is formalized as a checkable condition predicting whether per-component pruning criteria are reliable for a given architecture: Multi-Head Attention (MHA)+GELU designs satisfy the SIA, whereas Grouped Query Attention (GQA)+SwiGLU designs violate it. On bearing fault diagnosis spanning 88M to 6.25B-parameter models, the framework extends achievable compression to 13.8 times on MHA+GELU architectures with 83.82% accuracy (+3.70 percentage points (pp) over the strongest baseline), while exposing a ~74pp accuracy collapse on GQA+SwiGLU architectures that violate the SIA. Deployed on an industrial slewing bearing fault diagnosis platform with NVIDIA DGX Spark, compressed models reduce inference latency by up to 67.2% and peak memory by 62.5%, demonstrating viability for IIoT edge inference.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 0 minor

Summary. The paper introduces a cascaded multi-granularity pruning framework for extreme compression of LLMs on IIoT edge devices. It performs layer-, head-, and channel-level pruning in coarse-to-fine order with lightweight low-rank recovery between stages. An information-theoretic analysis motivates the ordering, and the Structural Independence Assumption (SIA) is defined as an architecture-dependent checkable condition: MHA+GELU models satisfy SIA while GQA+SwiGLU models violate it. Experiments on bearing fault diagnosis (88M–6.25B parameter models) report 13.8× compression at 83.82% accuracy (+3.70 pp over baseline) for SIA-satisfying architectures and a ~74 pp accuracy collapse for violating ones; hardware deployment shows up to 67.2% latency and 62.5% memory reduction.

Significance. If reproducible, the empirical demonstration of architecture-specific pruning reliability and the SIA concept would be useful for guiding structured pruning choices in resource-constrained industrial settings. The reported compression ratios and hardware gains on real IIoT hardware are practically relevant, but the absence of error bars, dataset details, and the full SIA derivation limits the strength of the contribution.

major comments (3)
  1. [Abstract] Abstract: The central accuracy (83.82%) and compression (13.8×) claims are reported without error bars, number of runs, or dataset split details for the bearing fault diagnosis task; this directly affects the reliability of the +3.70 pp improvement and the architecture-dependent contrast.
  2. [Abstract] Abstract: The SIA is presented as an independent, checkable condition derived from information-theoretic analysis that predicts pruning reliability, yet the manuscript provides neither the derivation nor the explicit checkable criterion; without this it is impossible to confirm that SIA is not circular with the reported outcomes.
  3. [Abstract] Abstract: Full experimental methods, model architectures, and baseline implementations are stated to be unavailable for inspection, preventing verification of the low-rank recovery step and the cross-architecture comparison that underpins the SIA claim.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their detailed review and constructive comments. We will revise the manuscript to improve the clarity and completeness of the reported results, the SIA derivation, and the experimental details as outlined below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central accuracy (83.82%) and compression (13.8×) claims are reported without error bars, number of runs, or dataset split details for the bearing fault diagnosis task; this directly affects the reliability of the +3.70 pp improvement and the architecture-dependent contrast.

    Authors: We agree that the absence of error bars and run details limits the assessment of result reliability. In the revised version, we will report the mean and standard deviation over 5 independent runs for all accuracy figures, including the 83.82% and the +3.70 pp improvement. We will also specify the dataset splits (e.g., 70/15/15 for train/val/test) used in the bearing fault diagnosis experiments to allow full reproducibility of the claims. revision: yes

  2. Referee: [Abstract] Abstract: The SIA is presented as an independent, checkable condition derived from information-theoretic analysis that predicts pruning reliability, yet the manuscript provides neither the derivation nor the explicit checkable criterion; without this it is impossible to confirm that SIA is not circular with the reported outcomes.

    Authors: The full manuscript includes a formal definition of the SIA as a checkable condition based on whether component importance scores remain stable under pruning (derived from mutual information considerations between layers and components). However, to address this concern, we will expand the main text with the complete information-theoretic derivation, including the mathematical steps showing why MHA+GELU satisfies SIA while GQA+SwiGLU violates it, ensuring it is not circular but predictive. revision: yes

  3. Referee: [Abstract] Abstract: Full experimental methods, model architectures, and baseline implementations are stated to be unavailable for inspection, preventing verification of the low-rank recovery step and the cross-architecture comparison that underpins the SIA claim.

    Authors: We apologize for any difficulty in locating the details; the full manuscript and supplementary material describe the model architectures (e.g., specific MHA and GQA configurations from 88M to 6.25B parameters), the low-rank recovery implementation using SVD-based approximation, and the baseline pruning methods. To facilitate verification, in the revision we will add a dedicated section or appendix with pseudocode for the cascaded pruning and low-rank recovery, and clarify that code will be released upon acceptance. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper motivates the cascaded ordering via an information-theoretic analysis and formalizes SIA as an independent checkable condition that predicts architecture-dependent reliability of pruning criteria. Central results are empirical (13.8× compression on MHA+GELU, ~74pp collapse on GQA+SwiGLU). No quoted equations or self-citations reduce the SIA definition, ordering choice, or reported outcomes to fitted inputs or prior self-work by construction. The derivation chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim depends on the SIA holding for the tested architectures and on the low-rank recovery step correctly re-estimating component importance after each pruning stage; no free parameters or invented entities are stated in the abstract.

axioms (1)
  • domain assumption Structural Independence Assumption (SIA) is a valid checkable condition that determines reliability of per-component pruning criteria for a given architecture.
    Abstract states that MHA+GELU satisfies SIA while GQA+SwiGLU violates it, and uses this to explain the observed accuracy collapse.

pith-pipeline@v0.9.1-grok · 5812 in / 1240 out tokens · 30843 ms · 2026-06-26T04:55:44.448186+00:00 · methodology

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 · 19 canonical work pages · 3 internal anchors

  1. [1]

    hen IoT Meet LLMs: A lications and Challenges,

    I. Kok, O. De irci, and . Ozde ir, “ hen IoT Meet LLMs: A lications and Challenges,” ov. 20, 2024, arXiv: arXiv:2411.17722. doi: 10.48550/arXiv.2411.17722

    work page doi:10.48550/arxiv.2411.17722 2024

  2. [2]

    dge hard: fficient LLM Inference via Colla orative dge Co uting,

    M. Zhang, X. hen, J. Cao, Z. Cui, and . Jiang, “ dge hard: fficient LLM Inference via Colla orative dge Co uting,” IEEE Internet of Things Journal , vol. 12, no. 10, pp. 13119 –13131, May 2025, doi: 10.1109/JIOT.2024.3524255

    work page doi:10.1109/jiot.2024.3524255 2025

  3. [3]

    LLM-based fra ework for earing fault diagnosis,

    L. Tao, . Liu, G. ing, . Cao, . uang, and C. Lu, “LLM-based fra ework for earing fault diagnosis,” Mechanical Systems and Signal Processing, vol. 224, p. 112127, Feb. 2025, doi: 10.1016/j.ymssp.2024.112127

    work page doi:10.1016/j.ymssp.2024.112127 2025

  4. [4]

    FD-MVLLM: Fault diagnosis based on multimodal vi ration data and large language odel for earing syste ,

    D. Li et al. , “FD-MVLLM: Fault diagnosis based on multimodal vi ration data and large language odel for earing syste ,” Mechanical Systems and Signal Processing , vol. 239, p. 113226, Oct. 2025, doi: 10.1016/j.ymssp.2025.113226

    work page doi:10.1016/j.ymssp.2025.113226 2025

  5. [5]

    A Simple and Effective Pruning Approach for Large Language Models

    M. un, Z. Liu, A. air, and J. Z. Kolter, “A i le and ffective Pruning A roach for Large Language Models,” May 06, 2024, arXiv: arXiv:2306.11695. doi: 10.48550/arXiv.2306.11695

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2306.11695 2024

  6. [6]

    arseGPT: Massive Language Models Can be Accurately Pruned in One - hot,

    Frantar and D. Alistarh, “ arseGPT: Massive Language Models Can be Accurately Pruned in One - hot,” in Proceedings of the 40th International Conference on Machine Learning , PMLR, Jul. 2023, pp. 10323–10337. Accessed: Jun. 12, 2026. [Online]. Available: https://proceedings.mlr.press/v202/frantar23a.html

    2023

  7. [7]

    LLM-Pruner: On the Structural Pruning of Large Language Models,

    X. Ma, G. Fang, and X. ang, “LLM-Pruner: On the Structural Pruning of Large Language Models,” Advances in Neural Information Processing Systems, vol. 36, pp. 21702–21720, Dec. 2023

    2023

  8. [8]

    Fluctuation -Based Adaptive tructured Pruning for Large Language Models,

    An, X. Zhao, T. u, M. Tang, and J. ang, “Fluctuation -Based Adaptive tructured Pruning for Large Language Models,” AAAI, vol. 38, no. 10, pp. 10865–10873, Mar. 2024, doi: 10.1609/aaai.v38i10.28960

    work page doi:10.1609/aaai.v38i10.28960 2024

  9. [9]

    L., Nascimento, M

    S. Ashkboos, M. L. Croci, M. G. do Nascimento, T. Hoefler, and J. ens an, “ liceGPT: Co ress Large Language Models y Deleting ows and Colu ns,” resented at the The Twelfth International Conference on Learning Representations, Jan. 2024. doi: 10.48550/arXiv.2401.15024

    work page doi:10.48550/arxiv.2401.15024 2024

  10. [10]

    hortGPT: Layers in Large Language Models are More edundant Than ou x ect,

    X. Men et al., “ hortGPT: Layers in Large Language Models are More edundant Than ou x ect,” in Findings of the Association for Computational Linguistics: ACL 2025 , Association for Computational Linguistics, 115 2024, pp. 20192 –20204. doi: 10.48550/arXiv.2403.03853

    work page doi:10.48550/arxiv.2403.03853 2025

  11. [11]

    LaCo: Large Language Model Pruning via Layer Colla se,

    ang, Z. Cao, and . Zhao, “LaCo: Large Language Model Pruning via Layer Colla se,” in Findings of the Association for Computational Linguistics: EMNLP 2024 , Oct. 2024. doi: 10.18653/v1/2024.findings - emnlp.372

    work page doi:10.18653/v1/2024.findings 2024

  12. [12]

    A study on quantum reservoir recurrent models for time-constrained volatile sequence forecasting,

    L. Mugnaini et al., Efficient LLMs with AMP: Attention Heads and MLP Pruning. 2025, p. 8. doi: 10.1109/IJCNN64981.2025.11227985

    work page doi:10.1109/ijcnn64981.2025.11227985 2025

  13. [13]

    GPTQ: Accurate Post-Training Quantization for Generative Pre -trained Transfor ers,

    Frantar, . Ashk oos, T. oefler, and D. Alistarh, “GPTQ: Accurate Post-Training Quantization for Generative Pre -trained Transfor ers,” ArXiv, Oct. 202 2, Accessed: Jun. 12, 2026. [Online]. Available: https://www.semanticscholar.org/paper/GPTQ%3A-Accurate-Post- Training-Quantization-for-Frantar- Ashkboos/7da0f2501034522e3d50af7e9b8fa7ec9d7b65b6

    2026

  14. [14]

    A Q: Activation -aware Weight Quantization for On - Device LLM Co ression and Acceleration,

    J. Lin et al., “A Q: Activation -aware Weight Quantization for On - Device LLM Co ression and Acceleration,” Proceedings of Machine Learning and Systems, vol. 6, pp. 87–100, May 2024

    2024

  15. [15]

    Learning both Weights and Connections for Efficient Neural Networks

    an, J. Pool, J. Tran, and . J. Dally, “Learning oth eights and Connections for fficient eural etworks,” in Advances i n Neural Information Processing Systems (NeurIPS) , Oct. 2015, pp. 1135 –1143. doi: 10.48550/arXiv.1506.02626

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1506.02626 2015

  16. [16]

    Pruning Filters for fficient Conv ets,

    Li, A. Kadav, I. Durdanovic, . a et, and . P. Graf, “Pruning Filters for fficient Conv ets,” resented at the International Conference on Learning Representations, Feb. 2017. Accessed: Jun. 12, 2026. [Online]. Available: https://openreview.net/forum?id=rJqFGTslg

    2017

  17. [17]

    Lo A: Low- ank Ada tation of Large Language Models,

    E. J. Hu et al., “Lo A: Low- ank Ada tation of Large Language Models,” presented at the International Conference on Learnin g Representations, Oct. 2022. Accessed: Jun. 12, 2026. [Online]. Available: https://openreview.net/forum?id=nZeVKeeFYf9

    2022

  18. [18]

    Lo APrune: tructured Pruning Meets Low-Rank Parameter-Efficient Fine-Tuning,

    M. Zhang et al., “Lo APrune: tructured Pruning Meets Low-Rank Parameter-Efficient Fine-Tuning,” Oct. 2023, Accessed: Jun. 12, 2026. [Online]. Available: https://openreview.net/forum?id=9KVT1e1qf7

    2023

  19. [19]

    Phyelds: A pythonic framework for aggregate computing,

    T. Chen, T. Ding, . adav, I. Zharkov, and L. Liang, “Lo A hear: Efficient Large Language Model Structured Pruning and Knowledge ecovery,” 2023, doi: 10.48550/A XIV.2310.18356

    work page doi:10.48550/a 2023

  20. [20]

    Dutta, R

    O. Dutta, R. Gupta, and S. Agarwal, VTrans: Accelerating Transformer Compression with Variational Information Bottleneck based Pruning

  21. [21]

    doi: 10.48550/arXiv.2406.05276

    work page doi:10.48550/arxiv.2406.05276

  22. [22]

    The Lottery Ticket y othesis: Finding Sparse, Trai na le eural etworks,

    J. Frankle and M. Car in, “The Lottery Ticket y othesis: Finding Sparse, Trai na le eural etworks,” resented at the International Conference on Learning Representations, Sep. 2018. Accessed: Jun. 12,

    2018

  23. [23]

    Available: https://openreview.net/forum?id=rJl-b3RcF7

    [Online]. Available: https://openreview.net/forum?id=rJl-b3RcF7

  24. [24]

    Deep learning and the information bottleneck principle

    Tish y and . Zaslavsky, “Dee learning and the infor mation ottleneck rinci le,” in 2015 IEEE Information Theory Workshop (ITW), Apr. 2015, pp. 1–5. doi: 10.1109/ITW.2015.7133169

    work page doi:10.1109/itw.2015.7133169 2015

  25. [25]

    T. M. Cover and J. A. Thomas, Elements of Information Theory, 2nd ed. John Wiley & Sons, Ltd, 2006. doi: 10.1002/047174882X.ch17

    work page doi:10.1002/047174882x.ch17 2006

  26. [26]

    GQA: Training Generalized Multi-Query Transformer Models from Multi- ead Check oints,

    J. Ainslie, J. Lee-Thorp, M. de Jong, Y. Zemlyanskiy, F. Lebron, and S. anghai, “GQA: Training Generalized Multi-Query Transformer Models from Multi- ead Check oints,” resented at the The 2023 Conference on Empirical Methods in Natural Langu age Processing, Dec. 2023. Accessed: Jun. 12, 2026. [Online]. Available: https://openreview.net/forum?id=hmOwOZWzYE

    2023

  27. [27]

    GLU Variants Improve Transformer

    N. Shazeer, GLU Variants Improve Transformer . 2020. doi: 10.48550/arXiv.2002.05202

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2002.05202 2020

  28. [28]

    Progressive Knowledge-Guided Large Language Model Framework for Bearing Fault Diagnosis,

    J. Wang, G. Peng, Y. Chen, W. Zhang, W. Wu, and T. Liu, “Progressive Knowledge-Guided Large Language Model Framework for Bearing Fault Diagnosis,” Jun. 15, 2026, arXiv: arXiv:2606.16684. doi: 10.48550/arXiv.2606.16684

    work page doi:10.48550/arxiv.2606.16684 2026

  29. [29]

    Polyanskiy and Y

    Y. Polyanskiy and Y. Wu, Information Theory: From Coding to Learning. Cambridge University Press, 2025. doi: 10.1017/9781108966351

    work page doi:10.1017/9781108966351 2025

  30. [30]

    Download a Data File | Case chool of ngineering

    “Download a Data File | Case chool of ngineering.” Accessed: Jun. 15,

  31. [31]

    Available: https://engineering.case.edu/bearingdatacenter/download-data-file

    [Online]. Available: https://engineering.case.edu/bearingdatacenter/download-data-file

  32. [32]

    equential Fuzzy Diagnosis Method for Motor Roller Bearing in Variable Operating Conditions ased on Vi ration Analysis,

    K. Li, X. Ping, H. Wang, P. Chen, and . Cao, “ equential Fuzzy Diagnosis Method for Motor Roller Bearing in Variable Operating Conditions ased on Vi ration Analysis,” Sensors, vol. 13, no. 6, pp. 8013–8041, Jun. 2013, doi: 10.3390/s130608013. In ut Tokens esidual Add esidual Add Out ut idden tates In ut Tokens esidual Add esidual Add Out ut idden tates Q1...

    work page doi:10.3390/s130608013 2013