When Labels Are Scarce: An Oscillatory State Space Model for Vibration Diagnosis

DualRes combines aligned spectral resolutions with normalized selective oscillatory updates; the paper’s evaluation contribution uses recording-level splits and explicit duration accounting to distinguish labelled exposure, recording coverage, and operating-condition shift.

Source: [4], [16]

The study evaluates six bearing datasets—Paderborn, CWRU, KAIST, UORED, HUST, and Ottawa—along with the eight-class MCC5-THU gearbox benchmark. It also reports a separate two-model, 14-way configuration-recognition pilot on PHM2009.

Source: [23]

Repeated development on the evaluated datasets limits claims of independent confirmation. The component ablations are descriptive one-seed effects and do not establish uncertainty across initializations.

Source: [10], [11]

The frontend centers each vibration window and divides it by its root-mean-square amplitude, removing absolute amplitude while retaining within-window variation. It uses a one-sided, unnormalized STFT with a periodic Hann window, no padding and a 128-sample hop. The 256-sample and 1024-sample resolutions provide 4 ms and 16 ms supports, with 250 Hz and 62.5 Hz bin spacing. Cropping aligns their frame centers; oscillatory state rotation and input-dependent damping control the subsequent memory updates. The admitted excerpts do not establish a magnitude-STFT or an additional projection step.

Source: [6], [9], [20]

The encoder has 39,528 parameters; the classification head adds 65C parameters for C classes.

Source: [18]

Macro-F1, the unweighted mean of per-class F1 scores, is the primary metric and accuracy is secondary. Predictions are aggregated at the recording level, and metric values are averaged equally across folds.

Source: [1], [11]

The bearing macro-F1 summary reports the mean and sample standard deviation over three seeds, with each seed averaged over both frozen folds; its labelled exposure is 1.024 unique seconds per class and neural models receive 100 supervised updates.

Source: [14]

The paper asks which compact representation captures diagnostic vibration structure and how data efficiency should be evaluated when labelled recordings are limited.

Source: [5]

Across the bearing tasks, the evaluation distinguishes component identity, operating conditions, geometry and speed profiles: Paderborn and UORED hold out bearing identities; CWRU and KAIST hold out conditions; HUST holds out geometries; Ottawa holds out speed profiles. MCC5 evaluates opposite motion modes, with severity composition also changing. The selected split-label fragment does not securely bind a recording-repeat split and its condition change to the separate PHM2009 pilot.

Source: [21], [23]

For MCC5, label budgets range from 1 to 6 recordings per class with four non-overlapping 0.512-second windows per recording, or 2.048–12.288 labelled seconds per class. The primary budget is 3 × 4 windows, or 6.144 seconds per class. A matched-duration pilot compares 1 × 12, 3 × 4 and 6 × 2 windows to study recording diversity at fixed labelled duration. An extension keeps six recordings and increases windows per recording to eight or twelve, corresponding to 24.576 or 36.864 seconds per class. The bearing protocol uses three overlapping windows covering 1.024 unique seconds per class.

Source: [19]

Evaluation recordings are separated before windows are extracted. Compared methods receive identical supports and vibration channels and are fitted directly on labelled source support without pretraining. Neural MCC5 and bearing runs use 300 and 100 Adam updates, respectively; terminal checkpoints are used rather than target-based early stopping.

Source: [3], [11]