What question did this study set out to answer?

The aim is to evaluate the generalizability of a class-discriminant codebook across robotic and industrial sensors.

June 27, 2026Open Access

A Class-Discriminant Codebook Encoder as a Cross-Domain Sensor Substrate: Robotic Force/Torque, Industrial Vibration, and Articulated-Motion Gesture

RFRandolph James FerlicEP Analytics (United States)KFKimberly Kate FerlicEP Analytics (United States)

Key Points

The aim is to evaluate the generalizability of a class-discriminant codebook across robotic and industrial sensors.
Evaluated the encoder on robotic force/torque, industrial vibration, and articulated-motion gesture data.
Used benchmark datasets to characterize encoder performance across various sensor tasks.
Measured classification accuracy and predictive maintenance capabilities on real public datasets.
Achieved ≈ 0.07 AUC at 240× compression for robotic task classification.
Matched a full classifier at 1,834× compression on industrial bearing fault detection with AUC 0.996.
Demonstrated predictive maintenance capabilities with an anomaly score correlating with remaining useful life (RUL).

Abstract

A class-discriminant codebook compresses each window of a continuous sensor signal to a single decision-oriented token, and the same token stream supports plural downstream capabilities — classification, anomaly detection, and multi-task inference — at near-zero marginal cost. Prior work validated this construction on physiologic (electrocardiography) and human-inertial signals. We ask whether the identical encoder, unchanged, generalizes to genuine robotic and industrial sensor streams, and we characterize the full platform on real public benchmark data with every outcome band frozen before analysis. On robotic six-axis force/torque (UCI Robot Execution Failures), the single token classifies failure type within ≈ 0.07 area-under-curve (AUC) of an uncompressed-feature classifier at ≈ 240× compression; on industrial bearing vibration (Case Western Reserve University) it matches a full classifier (0.996 vs 0.994) on the hard ten-class fault benchmark at ≈ 1,834× compression; on articulated-motion gesture (UEA NATOPS, an open analog of surgical-robotic kinematic gesture) it reaches 0.879 within ≈ 0.10 of the ceiling. Trained on normal data only, the codebook's nearest-centroid distance detects faults and collisions as novelty at AUC 0.987–1.000 — a predictive-maintenance capability on the same token stream with no separate model. One shared codebook serves both fault type and severity within 0.003 AUC of dedicated codebooks. Most strikingly, a codebook trained at one motor load classifies faults flawlessly (1.000) at three other loads without re-fit — the deployability property an industrial buyer requires first. We then broaden each leg: across five robotic force/torque tasks (the full UCI subset family) four of five generalize (family mean single-token AUC 0.805); on a second, independent industrial bearing set (MFPT) the single token again reaches near-ceiling fault classification (0.989 vs 0.998 at ≈ 4,096×); and on run-to-failure turbofan data (NASA C-MAPSS) the codebook's anomaly score becomes a monotone degradation / remaining-useful-life signal (Spearman ≈ −0.64 vs RUL; near-failure-vs-healthy AUC ≈ 0.993), upgrading predictive maintenance from a snapshot to a trajectory. We report the honest negatives verbatim: multi-token residual vector quantization does not improve the decision (the discriminant single token already carries it; the residual tokens encode decision-irrelevant detail — robotic 0.777 → 0.694); the reconstruction-residual input-quality gate generalizes only with a documented bound (strong on richer streams, weaker on the small robotic set and on subtle perturbations); and one of the five robotic subsets (lp1, a small four-class approach-to-grasp task) is not captured by a single token (0.667 vs a 0.985 ceiling). The application is already within the filed claims of the codebook family; the contribution is the demonstration, on real public benchmark data with pre-registered discipline, that one decision-oriented, edge-deployable encoder is a genuine cross-domain sensor substrate spanning physiologic monitoring, robotic-assisted operation, and industrial condition monitoring. Keywords / index terms: class-discriminant codebook; single-token compression; cross-domain generalization; robotics; force/torque; industrial condition monitoring; bearing-fault diagnosis; predictive maintenance; anomaly detection; multi-task learning; domain transfer; residual vector quantization; pre-registration; honest negatives. References: 1. W. A. Smith and R. B. Randall, "Rolling element bearing diagnostics using the Case Western Reserve University data: A benchmark study," Mechanical Systems and Signal Processing, 2015. 2. L. Seabra Lopes and L. M. Camarinha-Matos, "Feature transformation strategies for a robot learning problem" (UCI Robot Execution Failures), Springer, 1998. 3. A. Bagnall et al., "The UEA multivariate time series classification archive" (NATOPS), arXiv:1811.00075, 2018. 4. A. Saxena, K. Goebel, D. Simon, and N. Eklund, "Damage propagation modeling for aircraft engine run-to-failure simulation (C-MAPSS)," PHM, 2008. 5. E. Bechhoefer, "A quick introduction to bearing envelope analysis" (MFPT bearing dataset), 2013. 6. N. Tishby, F. C. Pereira, and W. Bialek, "The information bottleneck method," Allerton, 1999. 7. R. J. Ferlic and K. K. Ferlic, companion deposits (Papers 19–25), Zenodo, 2026. Companion deposits in this Zenodo Community (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

AI से पूछें

Bookmark

View Full Paper