Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

FISHER: A Foundation Model for Multi-Modal Industrial Signal Comprehensive Representation

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper argues that one small encoder, pre-trained only on audio and music, can represent sound, vibration, voltage, and current industrial signals at their native sampling rates and beat 24 larger baselines on a new 19-dataset benchmark.

desk verdict A genuinely new sub-band mechanism for variable-rate industrial signals, backed by broad experiments, but the headline SOTA margin needs held-out validation and error bars before I'd trust the 1.56-point claim. read the letter →

arxiv 2507.16696 v3 pith:QDMIQTJB submitted 2025-07-22 cs.LG cs.AIcs.MMcs.SD

Pingyi Fan , Anbai Jiang , Shuwei Zhang , Xinhu Zheng , Zhiqiang Lv , Bing Han , Wenrui Liang , Junjie Li
show 4 more authors
Wei-Qiang Zhang Yanmin Qian Xie Chen Jia Liu
This is my paper
classification cs.LGcs.AIcs.MMcs.SD
keywords foundationmodelindustrialsignalanalysismulti-sampling-ratesub-bandmodelingself-distillationanomalydetectionfaultdiagnosisrepresentationlearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

FISHER is a proposed foundation model meant to serve as a single feature extractor for heterogeneous industrial signals, covering sound, vibration, voltage, and current without resampling or task-specific fine-tuning. The paper reduces the heterogeneity to an 'M5 problem' — multi-modal, multi-sampling-rate, multi-scale, multitask, minim fault — and argues that the multi-sampling-rate issue is the first obstacle to remove. Its mechanism is to split a short-time Fourier transform spectrogram into fixed-width sub-bands, encode each sub-band independently with a shared Vision Transformer, and concatenate the resulting [CLS] embeddings, so a higher sampling rate simply adds more sub-bands. Pre-trained by teacher-student self-distillation on 17k hours of audio and music, FISHER reports an RMIS score of 62.23%, ahead of 24 baselines by at least 1.56% with model sizes of 5.5M-22M, and the paper attributes the gain to adapting to variable sampling rates rather than to domain-specific training data.

What carries the argument

The load-bearing mechanism is STFT sub-band modeling with fixed time duration. Using a window of fixed time duration (25 ms) makes the frequency resolution independent of the sampling rate, and each sub-band covers a fixed bandwidth measured in STFT bins; the spectrogram is split into these sub-bands, each is encoded separately by the shared ViT, and their [CLS] embeddings are concatenated. Pre-training couples this with EMA teacher-student self-distillation, in which masked student views are pushed toward the teacher's full-band representations at both patch and sub-band levels, and mask cloning multiplies views per clip; at inference only the student encoder is used.

What would settle it

Construct a fault whose signature spans two adjacent sub-bands, such as a modulation sideband crossing the boundary, and compare FISHER's zero-shot diagnosis accuracy against a model that processes the full spectrogram jointly at the native sampling rate; if the joint model wins while the per-band embeddings stay unchanged, the concatenation hypothesis is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that sampling-rate heterogeneity, the largest obstacle to a universal industrial signal model, can be neutralized by treating the spectrum as a stack of independent fixed-width sub-bands. Because sensors anti-alias before sampling, a recording made at a higher rate is assumed to contain exactly the lower-rate information plus additional high-frequency sub-bands; FISHER therefore processes each sub-band with one shared encoder and concatenates the sub-band [CLS] embeddings, giving representations whose length scales with the native sampling rate. With this design, a 5.5M-22M encoder pre-trained only on audio and music reaches 62.23% on the RMIS benchmark, at least 1.56% above all 24 baselines under zero-shot KNN evaluation, and the paper concludes that full-bandwidth processing at native rates, not resampling or pretraining on industrial data, is what drives the result.

Load-bearing premise

The design assumes that a higher sampling rate adds only new, self-contained high-frequency sub-bands, so each fixed-width sub-band can be encoded alone and concatenated without losing information that crosses band boundaries.

Editorial extensions

If this is right

  • A fixed FISHER checkpoint can be deployed across sensors with different native sampling rates without resampling or per-rate fine-tuning.
  • High-frequency fault content, often diluted by mel-scale processing or discarded during resampling, remains available up to the native Nyquist frequency.
  • RMIS gives the community a zero-shot protocol with sealed train-test splits so future encoders can be compared fairly on anomaly detection and fault diagnosis across four modalities.
  • Pre-training for industrial signal foundation models can draw on open audio and music corpora, sidestepping the scarcity and stationarity of labeled industrial recordings.
  • Because the representation is built from independent sub-bands, removing high sub-bands degrades performance smoothly, directly linking the model's generalization to the bandwidth it is allowed to see.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension is to check whether the integer-multiple sub-band assumption matters on the 44.1 kHz, 50 kHz, and 51.2 kHz RMIS datasets, since those rates do not divide cleanly into 2 kHz or 4 kHz bands.
  • The results imply that temporal variability, not domain identity, is the main driver of representation quality; pre-training on synthetic signals with controlled variability would test this directly.
  • If sub-band independence holds, the same fixed-resolution STFT band concatenation could transfer to other variable-rate series problems, such as biomedical or heterogeneous sensor monitoring, although the paper does not evaluate those.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces the M5 problem for industrial signals, proposes FISHER, an STFT sub-band-based ViT encoder pre-trained by teacher-student self-distillation on audio and music data, and presents the RMIS benchmark of 19 datasets across sound, vibration, current, and voltage. All models are evaluated by KNN without fine-tuning; FISHER-small is reported to reach SRMIS=62.23%, outperforming 24 baselines by at least 1.56 points with far fewer parameters. The paper also claims that native high-sampling-rate adaptation and audio/music pretraining are the keys to this performance.

Significance. If the claims hold, the paper is a useful practical contribution: a small encoder that handles multiple industrial modalities and native sampling rates, a broad baseline comparison, a curated benchmark with sealed splits, and open-sourced models and data. The strengths are the breadth of the comparison (24 baselines, Table II), the sealed-split protocol for fault diagnosis (Section IV-B), and a conceptually clean sub-band mechanism supported by a systematic ablation suite. However, the headline SOTA margin is not fully certified because the final benchmark was also used for hyperparameter selection, no error bars are reported, and a stated modeling assumption (integer-aligned sub-bands at common rates) is not satisfied by several datasets in Table I.

major comments (3)
  1. [V-F, Table II, IV-B] FISHER's hyperparameters—sub-band width w (Fig. 8b), STFT window and hop (Fig. 8d), loss weight λ (Fig. 8e), mask-cloning budget m_b (Fig. 8f), SSL framework (V-F.7), pretraining sampling-rate schedule (Fig. 8c), and data-source mixture (Fig. 5b)—are all selected by their RMIS scores. The baseline checkpoints, by contrast, are fixed official releases with no equivalent selection on RMIS. The reported 62.23% is therefore the maximum of a search over the same benchmark used for the final comparison, and the margin over the best baseline is only 1.56 points. Section IV-B states that fault-diagnosis results are averaged over 10 sealed splits, but Table II reports no standard deviations, so it is unclear whether the margin is above run-to-run noise. I ask the authors to report per-split results and standard deviations, to evaluate the final configuration on a held-out subset of RMIS (e.g., leave out one or more datasets from all selection), and to report performance for all searched configurations rather than only the selected one.
  2. [III-B, Eq. (1), Table I/III] The sub-band construction assumes N=t_win·fs is an integer and that fs is an integer multiple of the physical sub-band bandwidth. With the shared t_win=25ms, N=1102.5 for the 44.1kHz IIEE dataset, and 44.1, 50, 51.2, and 25.6kHz are not integer multiples of the stated f_base values (2kHz or 4kHz) used for the tiny/mini and small variants. The text also calls w a 'bandwidth' although the shape (B×n,T,w) and Table III imply w is a number of STFT bins (w=f_base·t_win; the physical bandwidth is f_base). The manuscript does not explain how STFT and sub-band splitting are computed for these rates, and n=floor(F/w) discards any residual high-frequency bins. This directly affects 10 of the 19 RMIS datasets and the claim of 'full signal bandwidth without resampling.' Please specify the exact n_fft, frequency-bin alignment, and residual-band handling for every sampling rate in Table I.
  3. [V-E.1, Fig. 6] The conclusion that FISHER's gain 'primarily stems from seamless adaptation to variable sampling rates' is not fully controlled. In Fig. 6, increasing the number of sub-bands also increases the representation dimension by d per sub-band, and KNN accuracy can change simply because the feature dimension changes. The comparison with models that happen to accept ≥32kHz inputs (PaSST, ECHO) is confounded by architecture, pretraining data, and context length. A cleaner test would fix the total representation dimension (e.g., pool sub-band embeddings to a constant size) or compare the same model trained with and without native-rate inputs while holding inference dimension constant; the current evidence is suggestive but not conclusive.
minor comments (5)
  1. [I, III-A] The term 'Minim fault' appears to be a typo for 'minimal fault' or 'minor fault'; it should be corrected in the M5 list and the roadmap.
  2. [Table III] The table header formatting is broken ('mb1 ml Param Head Size 1'); please align the column names with the entries and define m_b and m_l in the caption.
  3. [Appendix II] The text says 'select the top performing open-sourced checkpoints' for baselines but does not enumerate the exact checkpoint used for each model; a reproducible checklist of model versions is needed.
  4. [V-C] Because ECHO is described as a follow-up work with 'the identical sub-band modeling scheme,' it is not an independent validation; the sentence 'the superiority of FISHER has also been verified in other works' should be softened or supported by third-party evaluation.
  5. [III-B] The statement that common sampling rates '16 kHz, 32 kHz, 44.1 kHz, and 48 kHz' are integer multiples of f_base (2kHz or 4kHz) is incorrect for 44.1kHz; this should be reworded and the sub-band alignment rule stated precisely.

Circularity Check

1 steps flagged · score 4.0 of 10

FISHER's headline RMIS score is generated by choosing core hyperparameters on the same RMIS benchmark that is then reported as the SOTA comparison, so the 1.56-point margin is an optimistic-selection estimate rather than an independent prediction.

  1. fitted input called prediction [Section V-C (Results on the RMIS Benchmark), Section V-E.1 (High Frequency Gain), Section V-F (Ablation Study)]
    "Figure 8b experiments with different bandwidth w ... Figure 8d experiments with different t_win and t_hop on FISHER-small, where the optimal setting is (25 ms, 10 ms) ... Figure 8e experiments with different λ on FISHER-small. The best score is achieved at λ=0.5 ... Figure 8f experiments with different m_b on FISHER-small, where S_RMIS grows monotonically with m_b. Thus, we set m_b to the maximum value allowed by video random access memory (VRAM). ... FISHER achieves the highest RMIS score of 62.23% and surpassing all baselines by at least 1.56%."

    The reported 62.23% is not an independent estimate of FISHER's generalization: the same RMIS benchmark was used as the selection criterion for the model's core design choices (bandwidth w, STFT t_win/t_hop, loss weight λ, mask-cloning budget m_b, SSL framework, pre-training sampling-rate schedule, and pre-training data mixture, Figs. 5b/8b-f and Section V-F.7). The final configuration is the locally optimal one on RMIS, and the same RMIS score is then presented as the head-to-head SOTA result in Table II, with no error bars for the 10 seeded splits.

full rationale

The derivation of the sub-band representation itself is not circular: Eq. (1) is an identity defining frequency resolution from fixed time duration, and the concatenated [CLS] architecture, EMA self-distillation, and external audio/music pretraining are design choices tested against 24 fixed official baseline checkpoints. AnoPatch's KNN pipeline is applied uniformly to every model, so it does not by itself force FISHER's ranking. The ECHO passage is a self-citation with overlapping authorship and is not independent support, but it is not load-bearing for the central result. The genuine circularity is evaluative rather than definitional: RMIS is used both to select FISHER's crucial hyperparameters and to certify the SOTA claim, and the small 1.56-point margin over ECHO has no reported variance. The central claim still has independent content (external pretraining data, diverse public datasets, uniform KNN evaluation), so the score is 4 rather than 6+.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claims rely on the sub-band decomposition hypothesis (listed as axiom 3), the anti-aliasing premise (axiom 1), and the alignment of sampling rates to fixed sub-band grids (axiom 2). The model's design also depends on hand or ablation-chosen hyperparameters (sub-band width, STFT window and hop, loss weight). No new physical entities are introduced.

free parameters (3)
  • Sub-band bandwidth w = 100 Hz (tiny/mini), 50 Hz (small)
    Chosen by ablation on RMIS (Fig 8b); directly sets the number of sub-bands and is central to the sampling-rate-adaptation mechanism.
  • STFT window and hop durations (t_win, t_hop) = 25 ms / 10 ms
    Selected by ablation on FISHER-small (Fig 8d); determines time-frequency resolution and interacts with sub-band width.
  • Loss combination coefficient lambda = 0.5
    Chosen by ablation (Fig 8e); balances band-level and patch-level distillation losses.
assumptions (3)
  • domain assumption All sensors apply anti-aliasing filtering, so spectrograms contain no information above half the sampling rate.
    Invoked in Section III-B to argue that higher sampling rates only add new sub-bands; if a sensor lacks proper anti-aliasing, the sub-band decomposition premise fails.
  • ad hoc to paper Common sampling rates are integer multiples of a fundamental frequency f_base, making sub-bands a natural unit.
    Section III-B uses this to justify fixed sub-band width; however RMIS includes 44.1, 50, and 51.2 kHz datasets where N = t_win * fs is non-integer or f_base alignment is unclear, and no handling is described.
  • ad hoc to paper The signal representation is the concatenation of per-sub-band representations, and independent per-band processing preserves task-relevant information.
    Core modeling assumption (Section III-B and Fig 3); if diagnostic features span band boundaries, independent ViT passes lose cross-band interactions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FISHER: A Foundation Model for Multi-Modal Industrial Signal Comprehensive Representation." pith.science (2026). https://pith.science/paper/QDMIQTJB

@misc{pith2026250716696,
  author       = {Pith},
  title        = {Pith review of: FISHER: A Foundation Model for Multi-Modal Industrial Signal Comprehensive Representation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QDMIQTJB}},
  note         = {Machine review of arXiv:2507.16696}
}
read the original abstract

Industrial signal analysis is hindered by severe data heterogeneity, which we characterize as the M5 problem. Existing solutions rely on specialized models that lack robustness and scalability, while large-scale pre-training has rarely been investigated in this area. In this work, we derive a prioritized roadmap for the M5 problem and propose FISHER, a Foundation model for multi-modal Industrial Signal compreHEnsive Representation. To address the foremost multi-sampling-rate problem, FISHER utilizes a novel sub-band modeling approach that treats sampling rate increments as concatenated sub-band information, enabling the adaptive usage of full signal bandwidth without resampling. FISHER is pre-trained by teacher-student self-distillation over external audio and music data. We also establish the RMIS benchmark, comprising 19 datasets across four modalities. In the experiment, FISHER outperforms 24 state-of-the-art series encoders (up to 2B) with much smaller sizes (up to 16x), showcasing groundbreaking diagnostic accuracy and remarkable versatility. We further demonstrate that 1) seamless adaptation to variable sampling rates is the key to generalization 2) audio and music data provide better temporal variability, which is essential for pre-training. Both FISHER and RMIS are open-sourced.

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. ECHOv2: Two-Level Band-Splitting Representation Learning for Anomalous Sound Detection

    eess.AS 2026-07 conditional novelty 5.5 of 10

    Band-splitting pre-training with two-level self-distillation and multi-summary-token aggregation yields modest but consistent ASD gains over ECHO and strong audio baselines on a unified DCASE 2020–2025 benchmark.

  2. Mind the Gap: Detecting Cluster Exits for Robust Local Density-Based Score Normalization in Anomalous Sound Detection

    eess.AS 2026-02 conditional novelty 5.0 of 10

    Adaptively choosing the neighborhood size by detecting distance-ratio jumps (cluster exits) makes local-density score normalization more robust and improves anomalous sound detection by about 0.2 percentage points on average.

Reference graph

Works this paper leans on

76 extracted references · 61 canonical work pages · cited by 2 Pith papers

  1. [1]

    Qwen2.5-omni technical report,

    J. Xu, Z. Guo, J. He, H. Hu, T. He, S. Bai, K. Chen, J. Wang, Y . Fan, K. Dang, B. Zhang, X. Wang, Y . Chu, and J. Lin, “Qwen2.5-omni technical report,”arXiv preprint arXiv:2503.20215, 2025

    arXiv 2025

  2. [2]

    An image is worth 16x16 words: Trans- formers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Trans- formers for image recognition at scale,” inInternational Conference on Learning Representations, 2021

    2021

  3. [3]

    Masked au- toencoders are scalable vision learners,

    K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick, “Masked au- toencoders are scalable vision learners,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 16 000–16 009

    2022

  4. [4]

    Sim ´eoni, H

    O. Sim ´eoni, H. V . V o, M. Seitzer, F. Baldassarre, M. Oquab, C. Jose, V . Khalidov, M. Szafraniec, S. Yi, M. Ramamonjisoaet al., “Dinov3,” arXiv preprint arXiv:2508.10104, 2025

    arXiv 2025

  5. [5]

    wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” Advances in neural information processing systems, vol. 33, pp. 12 449– 12 460, 2020. 10 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. XX, NO. XX, XXXX

    work page 2020

  6. [6]

    Robust speech recognition via large-scale weak supervi- sion,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervi- sion,” inProceedings of the 40th International Conference on Machine Learning, ser. ICML’23. JMLR.org, 2023

    2023

  7. [7]

    Masked autoencoders that listen,

    P.-Y . Huang, H. Xu, J. Li, A. Baevski, M. Auli, W. Galuba, F. Metze, and C. Feichtenhofer, “Masked autoencoders that listen,”Advances in Neural Information Processing Systems, vol. 35, pp. 28 708–28 720, 2022

    2022

  8. [8]

    BEATs: Audio pre-training with acoustic tokenizers,

    S. Chen, Y . Wu, C. Wang, S. Liu, D. Tompkins, Z. Chen, W. Che, X. Yu, and F. Wei, “BEATs: Audio pre-training with acoustic tokenizers,” in Proceedings of the 40th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 202. PMLR, 23– 29 Jul 2023, pp. 5178–5193

    work page 2023

Show all 76 references
  1. [9]

    Eat: self-supervised pre-training with efficient audio transformer,

    W. Chen, Y . Liang, Z. Ma, Z. Zheng, and X. Chen, “Eat: self-supervised pre-training with efficient audio transformer,” in Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, ser. IJCAI ’24, 2024. [Online]. Available: https://doi.org/10.2...

    2024 doi

  2. [10]

    Unsupervised anomaly detection and localization of machine audio: A gan-based approach,

    A. Jiang, W.-Q. Zhang, Y . Deng, P. Fan, and J. Liu, “Unsupervised anomaly detection and localization of machine audio: A gan-based approach,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  3. [11]

    Anopatch: Towards better consistency in machine anomalous sound detection,

    A. Jiang, B. Han, Z. Lv, Y . Deng, W.-Q. Zhang, X. Chen, Y . Qian, J. Liu, and P. Fan, “Anopatch: Towards better consistency in machine anomalous sound detection,” inInterspeech 2024, 2024, pp. 107–111

    2024

  4. [12]

    Self-supervised learning for anomalous sound detec- tion,

    K. Wilkinghoff, “Self-supervised learning for anomalous sound detec- tion,” inICASSP 2024-2024 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 276–280

    2024

  5. [13]

    Exploring self-supervised audio models for generalized anomalous sound detection,

    B. Han, A. Jiang, X. Zheng, W.-Q. Zhang, J. Liu, P. Fan, and Y . Qian, “Exploring self-supervised audio models for generalized anomalous sound detection,”IEEE Transactions on Audio, Speech and Language Processing, 2025

    2025

  6. [14]

    Auto-embedding transformer for inter- pretable few-shot fault diagnosis of rolling bearings,

    G. Wang, D. Liu, and L. Cui, “Auto-embedding transformer for inter- pretable few-shot fault diagnosis of rolling bearings,”IEEE Transactions on Reliability, vol. 73, no. 2, pp. 1270–1279, 2023

    2023

  7. [15]

    A rolling bearing fault diagnosis method based on multimodal knowledge graph,

    C. Peng, Y . Sheng, W. Gui, Z. Tang, and C. Li, “A rolling bearing fault diagnosis method based on multimodal knowledge graph,”IEEE Transactions on Industrial Informatics, vol. 20, no. 11, pp. 13 047– 13 057, 2024

    2024

  8. [16]

    Bearllm: A prior knowledge-enhanced bearing health management framework with uni- fied vibration signal representation,

    H. Peng, J. Liu, J. Du, J. Gao, and W. Wang, “Bearllm: A prior knowledge-enhanced bearing health management framework with uni- fied vibration signal representation,” inProceedings of the AAAI Confer- ence on Artificial Intelligence, vol. 39, no. 19, 2025, pp. 19 866–19 874

    2025

  9. [17]

    Cows: Self-supervised representation pre-training for cross-machine fault diagnosis,

    X. Chen, R. Yang, Y . Xue, and Z. Wang, “Cows: Self-supervised representation pre-training for cross-machine fault diagnosis,”IEEE Transactions on Emerging Topics in Computational Intelligence, 2026

    2026

  10. [18]

    Gearbox fault diagnosis using a deep learning model with limited data sample,

    S. R. Saufi, Z. A. B. Ahmad, M. S. Leong, and M. H. Lim, “Gearbox fault diagnosis using a deep learning model with limited data sample,” IEEE Transactions on Industrial Informatics, vol. 16, no. 10, pp. 6263– 6271, 2020

    2020

  11. [19]

    A review on deep learning in planetary gearbox health state recognition: methods, applications, and dataset publication,

    D. Liu, L. Cui, and W. Cheng, “A review on deep learning in planetary gearbox health state recognition: methods, applications, and dataset publication,”Measurement Science and Technology, vol. 35, no. 1, p. 012002, 2023

    2023

  12. [20]

    A comprehensive gear eccentricity dataset with multiple fault severity levels: Description, characteristics analysis, and fault diagnosis applications,

    J. Li, H. Chen, X.-B. Wang, and Z.-X. Yang, “A comprehensive gear eccentricity dataset with multiple fault severity levels: Description, characteristics analysis, and fault diagnosis applications,”Mechanical Systems and Signal Processing, vol. 224, p. 112068, 2025

    2025

  13. [21]

    Bearingfm: Towards a foundation model for bearing fault diagnosis by domain knowledge and contrastive learning,

    Z. Lai, C. Yang, S. Lan, L. Wang, W. Shen, and L. Zhu, “Bearingfm: Towards a foundation model for bearing fault diagnosis by domain knowledge and contrastive learning,”International Journal of Produc- tion Economics, vol. 275, p. 109319, 2024

    2024

  14. [22]

    Rmgpt: A foundation model with generative pre-trained transformer for fault diagnosis and prognosis in rotating machinery,

    Y . Wang, Y . Yu, K. Sun, P. Lei, Y . Zhang, E. Zio, A. Xia, and Y . Li, “Rmgpt: A foundation model with generative pre-trained transformer for fault diagnosis and prognosis in rotating machinery,”IEEE Internet of Things Journal, 2025

    2025

  15. [23]

    Hse: A plug-and- play module for unified fault diagnosis foundation models,

    Q. Li, B. Chen, Q. Chen, X. Li, Z. Qin, and F. Chu, “Hse: A plug-and- play module for unified fault diagnosis foundation models,”Information Fusion, p. 103277, 2025

    2025

  16. [24]

    Scaling up masked audio encoder learning for general audio classification,

    H. Dinkel, Z. Yan, Y . Wang, J. Zhang, Y . Wang, and B. Wang, “Scaling up masked audio encoder learning for general audio classification,” in Proc. Interspeech 2024, 2024, pp. 547–551

    2024

  17. [25]

    Ced: Consistent ensemble distillation for audio tagging,

    H. Dinkel, Y . Wang, Z. Yan, J. Zhang, and Y . Wang, “Ced: Consistent ensemble distillation for audio tagging,” inICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024

    2024

  18. [26]

    Openbeats: A fully open-source general- purpose audio encoder,

    S. Bharadwaj, S. Cornell, K. Choi, S. Fukayama, H.-J. Shim, S. Desh- mukh, and S. Watanabe, “Openbeats: A fully open-source general- purpose audio encoder,” in2025 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), 2025, pp. 1–5

    2025

  19. [27]

    Qwen2-audio technical report,

    Y . Chu, J. Xu, Q. Yang, H. Wei, X. Wei, Z. Guo, Y . Leng, Y . Lv, J. He, J. Linet al., “Qwen2-audio technical report,”arXiv preprint arXiv:2407.10759, 2024

    2024 arXiv

  20. [28]

    Audio flamingo 3: Advancing audio intelligence with fully open large audio language models,

    S. Ghosh, A. Goel, J. Kim, S. Kumar, Z. Kong, S.-g. Lee, C.-H. Yang, R. Duraiswami, D. Manocha, R. Valleet al., “Audio flamingo 3: Advancing audio intelligence with fully open large audio language models,”Advances in Neural Information Processing Systems, vol. 38, pp. 41 819–4...

    2026

  21. [29]

    Midashenglm: Efficient audio understanding with general audio captions,

    H. Dinkel, G. Li, J. Liu, J. Luan, Y . Niu, X. Sun, T. Wang, Q. Xiao, J. Zhang, and J. Zhou, “Midashenglm: Efficient audio understanding with general audio captions,”arXiv preprint arXiv:2508.03983, 2025

    2025

  22. [30]

    Time-moe: Billion-scale time series foundation models with mixture of experts,

    X. Shi, S. Wang, Y . Nie, D. Li, Z. Ye, Q. Wen, and M. Jin, “Time-moe: Billion-scale time series foundation models with mixture of experts,” in International conference on learning representations, vol. 2025, 2025, pp. 34 635–34 667

    2025

  23. [31]

    Sundial: A family of highly capable time series foundation models,

    Y . Liu, G. Qin, Z. Shi, Z. Chen, C. Yang, X. Huang, J. Wang, and M. Long, “Sundial: A family of highly capable time series foundation models,” inInternational Conference on Machine Learning. PMLR, 2025, pp. 39 295–39 317

    2025

  24. [32]

    Bearing fault diagnosis based on an enhanced image representation method of vibration signal and conditional super token transformer,

    J. Li, H. Liu, J. Liang, J. Dong, B. Pang, Z. Hao, and X. Zhao, “Bearing fault diagnosis based on an enhanced image representation method of vibration signal and conditional super token transformer,” Entropy, vol. 24, no. 8, 2022

    2022

  25. [33]

    Parinfogpt: An llm-based two- stage framework for reliability assessment of rotating machine under partial information,

    Z. Pang, Y . Luan, J. Chen, and T. Li, “Parinfogpt: An llm-based two- stage framework for reliability assessment of rotating machine under partial information,”Reliability Engineering & System Safety, vol. 250, p. 110312, 2024

    2024

  26. [34]

    Data-efficient motor condition monitoring with time series foundation models,

    D. Li, X. Liao, S. Chen, and S. Zhao, “Data-efficient motor condition monitoring with time series foundation models,” in2026 IEEE Applied Power Electronics Conference and Exposition (APEC). IEEE, 2026, pp. 3650–3656

    2026

  27. [35]

    Adaptive prototype learning for anomalous sound detection with partially known attributes,

    A. Jiang, X. Zheng, B. Han, Y . Qiu, P. Fan, W.-Q. Zhang, C. Lu, and J. Liu, “Adaptive prototype learning for anomalous sound detection with partially known attributes,” inICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE,...

    2025

  28. [36]

    ToyAD- MOS: A dataset of miniature-machine operating sounds for anomalous sound detection,

    Y . Koizumi, S. Saito, H. Uematsu, N. Harada, and K. Imoto, “ToyAD- MOS: A dataset of miniature-machine operating sounds for anomalous sound detection,” inProceedings of IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), November 2019, pp. 308–312

    2019

  29. [37]

    MIMII Dataset: Sound dataset for malfunctioning industrial machine investigation and inspection,

    H. Purohit, R. Tanabe, T. Ichige, T. Endo, Y . Nikaido, K. Suefusa, and Y . Kawaguchi, “MIMII Dataset: Sound dataset for malfunctioning industrial machine investigation and inspection,” inProceedings of the Detection and Classification of Acoustic Scenes and Events 2019 Worksh...

    2019

  30. [38]

    Description and discussion on DCASE2020 challenge task2: Unsuper- vised anomalous sound detection for machine condition monitoring,

    Y . Koizumi, Y . Kawaguchi, K. Imoto, T. Nakamura, Y . Nikaido, R. Tan- abe, H. Purohit, K. Suefusa, T. Endo, M. Yasuda, and N. Harada, “Description and discussion on DCASE2020 challenge task2: Unsuper- vised anomalous sound detection for machine condition monitoring,” in Proc...

    2020

  31. [39]

    MIMII DUE: Sound dataset for malfunctioning industrial machine investigation and inspection with domain shifts due to changes in operational and environmental conditions,

    R. Tanabe, H. Purohit, K. Dohi, T. Endo, Y . Nikaido, T. Nakamura, and Y . Kawaguchi, “MIMII DUE: Sound dataset for malfunctioning industrial machine investigation and inspection with domain shifts due to changes in operational and environmental conditions,”IEEE Workshop on Ap...

    2021

  32. [40]

    Description and discussion on dcase 2021 challenge task 2: Unsupervised anomalous detection for machine condition monitoring under domain shifted conditions,

    Y . Kawaguchi, K. Imoto, Y . Koizumi, N. Harada, D. Niizumi, K. Dohi, R. Tanabe, H. Purohit, and T. Endo, “Description and discussion on dcase 2021 challenge task 2: Unsupervised anomalous detection for machine condition monitoring under domain shifted conditions,” in Proceedi...

    2021

  33. [41]

    ToyADMOS2: Another dataset of miniature-machine operating sounds for anomalous sound detection under domain shift conditions,

    N. Harada, D. Niizumi, D. Takeuchi, Y . Ohishi, M. Yasuda, and S. Saito, “ToyADMOS2: Another dataset of miniature-machine operating sounds for anomalous sound detection under domain shift conditions,” in Proceedings of the 6th Detection and Classification of Acoustic Scenes an...

    2021

  34. [42]

    K. Dohi, T. Nishida, H. Purohit, R. Tanabe, T. Endo, M. Yamamoto, Y . Nikaido, and Y . Kawaguchi, “MIMII DG: Sound dataset for mal- AUTHORet al.: FISHER: A FOUNDATION MODEL FOR MULTI-MODAL INDUSTRIAL SIGNAL COMPREHENSIVE REPRESENTATION 11 functioning industrial machine investi...

    2022

  35. [43]

    Description and discussion on DCASE 2022 challenge task 2: Unsu- pervised anomalous sound detection for machine condition monitoring applying domain generalization techniques,

    K. Dohi, K. Imoto, N. Harada, D. Niizumi, Y . Koizumi, T. Nishida, H. Purohit, R. Tanabe, T. Endo, M. Yamamoto, and Y . Kawaguchi, “Description and discussion on DCASE 2022 challenge task 2: Unsu- pervised anomalous sound detection for machine condition monitoring applying dom...

    2022

  36. [44]

    Description and discussion on DCASE 2023 challenge task 2: First-shot unsupervised anomalous sound detection for machine condition monitoring,

    K. Dohi, K. Imoto, N. Harada, D. Niizumi, Y . Koizumi, T. Nishida, H. Purohit, R. Tanabe, T. Endo, and Y . Kawaguchi, “Description and discussion on DCASE 2023 challenge task 2: First-shot unsupervised anomalous sound detection for machine condition monitoring,”In arXiv e-prin...

    2023 arXiv

  37. [45]

    First- shot anomaly detection for machine condition monitoring: A domain generalization baseline,

    N. Harada, D. Niizumi, D. Takeuchi, Y . Ohishi, and M. Yasuda, “First- shot anomaly detection for machine condition monitoring: A domain generalization baseline,”In arXiv e-prints: 2303.00455, 2023

    2023 arXiv

  38. [46]

    Description and discussion on DCASE 2024 challenge task 2: First-shot unsupervised anomalous sound detection for machine condition monitoring,

    T. Nishida, N. Harada, D. Niizumi, D. Albertini, R. Sannino, S. Pradolini, F. Augusti, K. Imoto, K. Dohi, H. Purohit, T. Endo, and Y . Kawaguchi, “Description and discussion on DCASE 2024 challenge task 2: First-shot unsupervised anomalous sound detection for machine condition...

    2024 arXiv

  39. [47]

    First-shot anomaly detection for machine condition monitoring: A domain generalization baseline,

    N. Harada, D. Niizumi, D. Takeuchi, Y . Ohishi, and M. Yasuda, “First-shot anomaly detection for machine condition monitoring: A domain generalization baseline,”Proceedings of 31st European Signal Processing Conference (EUSIPCO), pp. 191–195, 2023

    2023

  40. [48]

    Description and discussion on DCASE 2025 challenge task 2: First-shot unsupervised anomalous sound detection for machine condition monitoring,

    T. Nishida, N. Harada, D. Niizumi, D. Albertini, R. Sannino, S. Pradolini, F. Augusti, K. Imoto, K. Dohi, H. Purohit, T. Endo, and Y . Kawaguchi, “Description and discussion on DCASE 2025 challenge task 2: First-shot unsupervised anomalous sound detection for machine condition...

    2025

  41. [49]

    Compressed air leakage detection using acoustic emissions with neural networks,

    D. Johnson, J. Kirner, S. Grollmisch, and J. Liebetrau, “Compressed air leakage detection using acoustic emissions with neural networks,” in INTER-NOISE and NOISE-CON Congress and Conference Proceedings, vol. 261, no. 1. Institute of Noise Control Engineering, 2020, pp. 5662– 5673

    2020

  42. [50]

    Sounding industry: Challenges and datasets for industrial sound analysis,

    S. Grollmisch, J. Abeßer, J. Liebetrau, and H. Lukashevich, “Sounding industry: Challenges and datasets for industrial sound analysis,” in2019 27th European signal processing conference (EUSIPCO). IEEE, 2019, pp. 1–5

    2019

  43. [51]

    Mafaulda-machinery fault database,

    F. M. L. Ribeiro, “Mafaulda-machinery fault database,” 2016. [Online]. Available: https://www02.smt.ufrj.br/ ∼offshore/mfs/page 01. html#SEC2

    2016

  44. [52]

    Attention guided multi-wavelet adversarial network for cross domain fault diagnosis,

    J. Wang, X. Zhang, Z. Zhang, B. Han, X. Jiang, H. Bao, and X. Jiang, “Attention guided multi-wavelet adversarial network for cross domain fault diagnosis,”Knowledge-Based Systems, vol. 284, p. 111285, 2024

    2024

  45. [53]

    Integrated decision-making with adaptive feature weighting adversarial network for multi-target domain compound fault diagnosis of machinery,

    X. Zhang, J. Wang, Z. Zhang, B. Han, H. Bao, and X. Jiang, “Integrated decision-making with adaptive feature weighting adversarial network for multi-target domain compound fault diagnosis of machinery,”Advanced Engineering Informatics, vol. 62, p. 102730, 2024

    2024

  46. [54]

    A novel domain adaptive fault diagnosis method for bearings based on unbalance data generation,

    B. Han, X. Jiang, J. Wang, and Z. Zhang, “A novel domain adaptive fault diagnosis method for bearings based on unbalance data generation,” IEEE Transactions on Instrumentation and Measurement, vol. 72, pp. 1–11, 2023

    2023

  47. [55]

    A novel rolling bearing fault diagnosis method based on generalized nonlinear spectral sparsity,

    B. Han, Z. Yang, Z. Zhang, H. Bao:, J. Wang, Z. Liu, and S. Li, “A novel rolling bearing fault diagnosis method based on generalized nonlinear spectral sparsity,”Measurement, vol. 198, p. 111131, 2022

    2022

  48. [56]

    Condition monitoring of bearing damage in electromechanical drive systems by using motor current signals of electric motors: A benchmark data set for data-driven classification,

    C. Lessmeier, J. K. Kimotho, D. Zimmer, and W. Sextro, “Condition monitoring of bearing damage in electromechanical drive systems by using motor current signals of electric motors: A benchmark data set for data-driven classification,” inPHM society European conference, vol. 3,...

    2016

  49. [57]

    Intelligent fault diagnosis of type and severity in low-frequency, low bit-depth signals,

    T. Spadini, K. Nose-Filho, and R. Suyama, “Intelligent fault diagnosis of type and severity in low-frequency, low bit-depth signals,”arXiv preprint arXiv:2411.06299, 2024

    2024 arXiv

  50. [58]

    Audio set: An ontology and human- labeled dataset for audio events,

    J. F. Gemmeke, D. P. Ellis, D. Freedman, A. Jansen, W. Lawrence, R. C. Moore, M. Plakal, and M. Ritter, “Audio set: An ontology and human- labeled dataset for audio events,” in2017 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2017, p...

    2017

  51. [59]

    The mtg-jamendo dataset for automatic music tagging,

    D. Bogdanov, M. Won, P. Tovstogan, A. Porter, and X. Serra, “The mtg-jamendo dataset for automatic music tagging,” inMachine learning for music discovery workshop, international conference on machine learning (ICML 2019). Long Beach, CA, United States, 2019, pp. 1–3

    2019

  52. [60]

    Music4all: A new music database and its applications,

    I. A. P. Santana, F. Pinhelli, J. Donini, L. Catharin, R. B. Mangolin, V . D. Feltrim, M. A. Domingueset al., “Music4all: A new music database and its applications,” in2020 International Conference on Systems, Signals and Image Processing (IWSSIP). IEEE, 2020, pp. 399–404

    2020

  53. [61]

    Efficient training of audio transformers with patchout,

    K. Koutini, J. Schl ¨uter, H. Eghbal-zadeh, and G. Widmer, “Efficient training of audio transformers with patchout,” inProc. Interspeech 2022, 2022, pp. 2753–2757

    2022

  54. [62]

    Mert: Acoustic music understanding model with large-scale self-supervised training,

    Y . Li, R. Yuan, G. Zhang, Y . Ma, X. Chen, H. Yin, C. Xiao, C. Lin, A. Ragni, E. Benetoset al., “Mert: Acoustic music understanding model with large-scale self-supervised training,” inInternational Conference on Learning Representations, vol. 2024, 2024, pp. 12 181–12 204

    2024

  55. [63]

    Muq: Self-supervised music representation learning with mel residual vector quantization,

    H. Zhu, Y . Zhou, H. Chen, J. Yu, Z. Ma, R. Gu, Y . Luo, W. Tan, and X. Chen, “Muq: Self-supervised music representation learning with mel residual vector quantization,”IEEE Transactions on Audio, Speech and Language Processing, 2025

    2025

  56. [64]

    Ensemble empirical mode decomposition: a noise-assisted data analysis method,

    Z. Wu and N. E. Huang, “Ensemble empirical mode decomposition: a noise-assisted data analysis method,”Advances in adaptive data analysis, vol. 1, no. 01, pp. 1–41, 2009

    2009

  57. [65]

    Permutation entropy: a natural complexity measure for time series,

    C. Bandt and B. Pompe, “Permutation entropy: a natural complexity measure for time series,”Physical review letters, vol. 88, no. 17, p. 174102, 2002

    2002

  58. [66]

    Tfpred: Learning discriminative representations from unlabeled data for few-label rotating machinery fault diagnosis,

    X. Chen, R. Yang, Y . Xue, B. Song, and Z. Wang, “Tfpred: Learning discriminative representations from unlabeled data for few-label rotating machinery fault diagnosis,”Control Engineering Practice, vol. 146, p. 105900, 2024

    2024

  59. [67]

    Liconvformer: A lightweight fault diagnosis framework using separable multiscale con- volution and broadcast self-attention,

    S. Yan, H. Shao, J. Wang, X. Zheng, and B. Liu, “Liconvformer: A lightweight fault diagnosis framework using separable multiscale con- volution and broadcast self-attention,”Expert Systems with Applications, vol. 237, p. 121338, 2024

    2024

  60. [68]

    A unified rotating machinery health management framework leveraging large language models for diverse components, conditions, and tasks,

    H. Peng, J. Gao, J. Liu, J. Du, and W. Wang, “A unified rotating machinery health management framework leveraging large language models for diverse components, conditions, and tasks,”Engineering Applications of Artificial Intelligence, vol. 162, p. 112544, 2025

    2025

  61. [69]

    Echo: Frequency-aware hierarchical encoding for variable-length signals,

    Y . Zhang, J. Liu, and M. Li, “Echo: Frequency-aware hierarchical encoding for variable-length signals,” inICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2026, pp. 4301–4305

    2026

  62. [70]

    Sim- pleshot: Revisiting nearest-neighbor classification for few-shot learning,

    Y . Wang, W.-L. Chao, K. Q. Weinberger, and L. Van Der Maaten, “Sim- pleshot: Revisiting nearest-neighbor classification for few-shot learning,” arXiv preprint arXiv:1911.04623, 2019. APPENDIXI DETAILS OF THERMIS BENCHMARK For ASD tasks, all models adopt the identical KNN-base...

    1911 arXiv

  63. [71]

    For Whisper, we evaluate the encoders of five official pre-trained checkpoints

    Speech Encoders:For Wav2Vec 2.0, we employ the of- ficial wav2vec2-base-960h, wav2vec2-xls-r-300m, wav2vec2- xls-r-1b, wav2vec2-x1s-r-2b checkpoints. For Whisper, we evaluate the encoders of five official pre-trained checkpoints

  64. [72]

    For Open- BEATs, we employ the base and large checkpoints of iter3

    Audio Encoders:For PaSST, we employ the offi- cial passt-s-f128-p16-s10-ap.476-swa checkpoint. For Open- BEATs, we employ the base and large checkpoints of iter3

  65. [73]

    LALMs:We evaluate the audio encoder of each LALM

  66. [74]

    Thus, only two models are compared

    Time Series Encoders:Time series models are not de- signed to process kHz-level signals, since 1) they model raw time points, which is highly susceptible to noise 2) their input windows (thousands of points) are much too short to capture second-level dependency, and their comp...

  67. [75]

    For LiConvFormer, we utilize the weights trained on CWRU

    Signal Encoders:For PEEMD, we follow the works of [64], [65], which combine IMFs extraction, instability detection and ensemble averaging. For LiConvFormer, we utilize the weights trained on CWRU. For BearLLM and RotLLM, we employ their feature encoders. Pingyi Fan(Senior Memb...

    2022

  68. [2025]

    degree with the Department of Electronic Engi- neering, Tsinghua University, Beijing, China

    He is currently working toward the Ph.D. degree with the Department of Electronic Engi- neering, Tsinghua University, Beijing, China. His research interests include signal representation, analysis and generation. Xinhu Zhengreceived the B.Eng. degree in Electronic Engineering ...

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.