A Single-Token Sensor Substrate for Industrial Condition Monitoring: Fault Classification at Classifier Parity, Free Anomaly and Remaining-Useful-Life Signals, and Multi-Task Decoding at 2,000–4,000× Compression Randolph James Ferlic, M.D., and Kimberly Kate Ferlic · Fieldstone Analytics, LLC · randolphf@fieldstoneanalyticsllc.com Community: spiral-domain-encoder-campaign · Version 2 (the concept DOI resolves to this latest version). Version 2 note This version reframes and substantially strengthens the paper around industrial condition monitoring and characterizes the single-token class-discriminant codebook as a multi-function sensing substrate. It adds a modern strong-baseline comparison (1D convolutional network, MiniROCKET, random forest, histogram gradient boosting, multilayer perceptron) on identical data splits; a supervised-versus-unsupervised remaining-useful-life comparison; an edge memory-footprint and per-window latency characterization; and an honest, verbatim boundary. It adopts a shrinkage-regularized discriminant, which recovers most of the small-sample boundary tax at no cost in bits, and reports that a supervised / mutual-information codebook does not improve on K-means placement. No new subject matter is disclosed relative to the prior version. Summary Modern sensor deployments must classify faults, flag anomalies, forecast degradation, and adapt across operating conditions — typically with a separate model for each and full-bandwidth data. We ask how much of that stack an extreme-compression front end can deliver from a single discrete token per window. A class-discriminant codebook reduces each window of a multivariate sensor stream to one 8-bit token (K = 256) chosen to preserve the decision; the same token stream then drives classification, anomaly detection, remaining-useful-life (RUL) estimation, and multi-task inference at near-zero marginal cost. Results • Industrial fault classification at classifier parity. On the standard bearing-fault benchmarks — Case Western Reserve University (hard 10-class) and MFPT — the single token matches strong modern classifiers (1D-CNN, MiniROCKET, random forest; all ≥ 0.99 macro-AUC) to within 0.011 AUC at 2,000–4,000× compression (CWRU-10 0.997; MFPT 0.989). • Free anomaly detection. An unsupervised codebook fit on normal windows only separates fault from normal at macro-AUC 0.99–1.00 — no separate detector. • RUL as a free byproduct. On NASA C-MAPSS run-to-failure turbofan data, the same unsupervised distance flags near-failure at AUC 0.97 with no labels (a supervised 1D-CNN regressor reaches 1.00); the token's signal is a zero-cost degradation early-warning. • Multi-task from one stream. One shared codebook serves fault type and severity within 0.003 AUC of dedicated models. • Cross-condition transfer. Trained at one motor load, the token classifies faults at macro-AUC 1.000 across three other loads without re-fit. • Edge cost. The deployed encoder occupies 12–18 KB and encodes a window in under 2 ms on a single CPU thread; the dominant benefit is bandwidth — one token per window in place of the full waveform. Honest boundary (reported verbatim) On small multivariate force/torque tasks (UCI Robot Execution Failures lp1–lp5) and articulated-motion gesture (UEA NATOPS), strong classifiers outperform the single token by 0.04–0.10 AUC. Much of the extreme force/torque tax was a small-sample regularization artifact: a Ledoit-Wolf shrinkage-regularized discriminant recovers most of it (the smallest task rises from 0.68 to 0.96) at no cost in bits, with industrial parity preserved. A supervised, mutual-information-maximizing codebook does not reliably beat K-means on the discriminant embedding — the projection, not the quantizer, is the accuracy lever. Reproducibility and pre-registration PYTHONHASHSEED=0; train-only fitting; five seeds; identical shared splits; paired-bootstrap confidence intervals; real public data only. Every phase was pre-registered with frozen outcome bands; honest negatives are reported verbatim. Datasets are public: CWRU and MFPT bearing, NASA C-MAPSS FD001 turbofan, UCI Robot Execution Failures, and UEA NATOPS. The reproducibility directory contains the self-contained runners and per-run result summaries. Conflicts of interest The authors have filed U.S. provisional patent applications related to the methods described herein and are the principals of Fieldstone Analytics, LLC. The applications reported here are within the filed claims; no new subject matter is disclosed. Keywords condition monitoring; bearing fault diagnosis; remaining useful life; single-token compression; class-discriminant codebook; edge sensing; anomaly detection; multi-task learning; predictive maintenance; time-series classification; pre-registration; honest negatives. References 1 N. Zeghidour et al., "SoundStream: An end-to-end neural audio codec," IEEE/ACM Trans. Audio, Speech, Language Process., 2021. 2 A. Défossez et al., "High fidelity neural audio compression (EnCodec)," arXiv:2210.13438, 2022. 3 N. Tishby, F. C. Pereira, and W. Bialek, "The information bottleneck method," Proc. Allerton Conf., 1999. 4 W. A. Smith and R. B. Randall, "Rolling element bearing diagnostics using the CWRU data: A benchmark study," Mech. Syst. Signal Process., vol. 64–65, 2015. 5 L. Seabra Lopes and L. M. Camarinha-Matos, "Feature transformation strategies for a robot learning problem," Springer, 1998 (UCI Robot Execution Failures). 6 A. Bagnall et al., "The UEA multivariate time series classification archive," arXiv:1811.00075, 2018 (NATOPS). 7 E. Bechhoefer, "A quick introduction to bearing envelope analysis," MFPT bearing fault dataset, 2013. 8 A. Saxena, K. Goebel, D. Simon, and N. Eklund, "Damage propagation modeling for aircraft engine run-to-failure simulation (C-MAPSS)," PHM, 2008. 9 A. Dempster, D. F. Schmidt, and G. I. Webb, "MiniRocket: A very fast (almost) deterministic transform for time series classification," Proc. ACM SIGKDD, 2021. 10 L. Wen, X. Li, L. Gao, and Y. Zhang, "A new convolutional neural network-based data-driven fault diagnosis method," IEEE Trans. Ind. Electron., vol. 65, no. 7, 2018. 11 X. Li, Q. Ding, and J.-Q. Sun, "Remaining useful life estimation in prognostics using deep convolution neural networks," Reliab. Eng. Syst. Saf., vol. 172, 2018. 12 L. Breiman, "Random forests," Machine Learning, vol. 45, 2001. 13 W. Shi, J. Cao, Q. Zhang, Y. Li, and L. Xu, "Edge computing: Vision and challenges," IEEE Internet Things J., vol. 3, no. 5, 2016. 14 A. Bagnall, J. Lines, A. Bostrom, J. Large, and E. Keogh, "The great time series classification bake off," Data Min. Knowl. Discov., vol. 31, 2017. 15 Y. Lei et al., "Machinery health prognostics: A systematic review from data acquisition to RUL prediction," Mech. Syst. Signal Process., vol. 104, 2018. 16 A. Dempster, F. Petitjean, and G. I. Webb, "ROCKET: Exceptionally fast and accurate time series classification using random convolutional kernels," Data Min. Knowl. Discov., vol. 34, 2020. 17 S. Zhang, S. Zhang, B. Wang, and T. G. Habetler, "Deep learning algorithms for bearing fault diagnostics — A comprehensive review," IEEE Access, vol. 8, 2020. 18 A. A. Alemi, I. Fischer, J. V. Dillon, and K. Murphy, "Deep variational information bottleneck," Proc. ICLR, 2017. 19 S. Zheng, K. Ristovski, A. Farahat, and C. Gupta, "Long short-term memory network for remaining useful life estimation," Proc. IEEE ICPHM, 2017. 20 S. Talukder, Y. Yue, and G. Gkioxari, "TOTEM: Tokenized time series embeddings for general time series analysis," Trans. Mach. Learn. Res. (TMLR), 2024. 21 D. Gündüz et al., "Beyond transmitting bits: Context, semantics, and task-oriented communications," IEEE J. Sel. Areas Commun., vol. 41, no. 1, 2023. 22 L.-Y. Duan, J. Liu, W. Yang, T. Huang, and W. Gao, "Video coding for machines: A paradigm of collaborative compression and intelligent analytics," IEEE Trans. Image Process., vol. 29, 2020. 23 C. R. Banbury et al., "Benchmarking TinyML systems: Challenges and direction," arXiv:2003.04821, 2020. 24 R. David et al., "TensorFlow Lite Micro: Embedded machine learning for TinyML systems," Proc. MLSys, 2021. 25 R. Zhao, R. Yan, Z. Chen, K. Mao, P. Wang, and R. X. Gao, "Deep learning and its applications to machine health monitoring," Mech. Syst. Signal Process., vol. 115, 2019. 26 T. Zonta et al., "Predictive maintenance in the Industry 4.0: A systematic literature review," Comput. Ind. Eng., vol. 150, 2020. Companion deposits (spiral-domain-encoder-campaign) Paper 19 — 10.5281/zenodo.20788187 · Paper 20 — 10.5281/zenodo.20802759 · Paper 21 — 10.5281/zenodo.20802826 · Paper 22 — 10.5281/zenodo.20805321 · Paper 23 — 10.5281/zenodo.20821668 · Paper 24 — 10.5281/zenodo.20821779 · Paper 25 — 10.5281/zenodo.20821903. License Released under Creative Commons Attribution 4.0 International (CC-BY 4.0).
Ferlic et al. (Thu,) studied this question.