R2 · First author · Scientific Reports, article in press

Which view of the signal does the model trust?

A Vision Transformer that reads six physics-based images of one vibration signal, and shows which it relies on.

Start the walkthrough Read the paper

Best test accuracy: 99.62%
Mean over five seeds: 98.72 ± 0.70%
Unseen conditions, mean: 97.15%

Venue: Scientific Reports · Article in press
Accepted: 20 October 2026
Authors: Pratyush Vishal, Shristi Sahani, Narendiranath Babu T.
DOI: 10.1038/s41598-026-77398-5
Data: Zenodo dataset (opens in a new tab)
Code: Zenodo code (opens in a new tab)

01 / Method

Six views in, one answer out

Nine stages, from a raw window to which view the model trusts.

Stage 01 / 09 · Data

Fig. 1Non-overlapping windows over one recording, and the train, validation and test split.

Stage 02 / 09 · Six views

Fig. 2What each view tends to look like for a faulty bearing. Illustration, not the paper’s images.

Stage 03 / 09 · Fusion

Fig. 3The fused input, dimensioned.

Stage 04 / 09 · Tokens

Fig. 4The patch grid and the token sequence. Table 3 lists 32 × 32 patches; the text and the 196-token count use 16 × 16.

Stage 05 / 09 · Encoder

Fig. 5One encoder layer, repeated six times.

Stage 06 / 09 · Objective

Fig. 6The training objective, and the head-similarity figures the paper reports.

Stage 07 / 09 · Classify

Fig. 7Precision and recall per class on the test set.

Stage 08 / 09 · Attention

Fig. 8Attention energy by view (abstract and conclusion figures), and the per-head shares quoted in the text.

Stage 09 / 09 · Ablation

Fig. 9Test accuracy with each view removed in turn.

01 · Data
Same rig, raw windows
The rig and recordings from the first paper, cut into non-overlapping windows. No augmentation: the natural speed and load variation is the diversity.
Window
8192 samples
Split
70 / 15 / 15 %
02 · Six views
One signal, six images
Each window becomes six images that each stress a different physical side of a fault: harmonics, transients, impulsiveness, energy bursts and the fault frequencies themselves.
Views
STFT, CWT, CQT, SK, TEM, ES
03 · Fusion
Fused before the first layer
The six images are resized and stacked as channels, so every patch the transformer sees already holds all six views.
Input
224 × 224 × 6
04 · Tokens
Patches become tokens
A convolutional patch embedding cuts the input into a grid of patches; a learnable CLS token goes in front.
Patch
16 × 16 (see note)
Tokens
196 + CLS, 512-d
05 · Encoder
One shared transformer
Six encoder layers of multi-head self-attention and feed-forward blocks relate the patches across all six views at once.
Layers, heads
6, 8
MLP, dropout
2048, 0.1
Parameters
19.47 M
06 · Objective
Keep the heads different
Two small extra losses push the attention heads to differ and to specialise. They end up moderately alike, mean similarity 0.409.
Loss
Cross-entropy + λ₁·diversity + λ₂·specialisation
Loss weights
λ₁ = λ₂ = 0.01
Optimizer
AdamW, lr 1e-4, weight decay 0.05
Schedule
Cosine annealing, mixed precision
07 · Classify
Five health states
The CLS token feeds a classifier. Best test accuracy 99.62%; 98.72 ± 0.70% over five seeds. Combined faults are the hardest class.
Best, mean
99.62%, 98.72 ± 0.70%
CF precision
0.964, recall 0.982
08 · Attention
The envelope spectrum leads
Summing the CLS token’s attention by view, the envelope spectrum draws about 42.8% and CQT 20.3%: the view that exposes the bearing’s fault frequencies leads.
Note
Results text gives 34.4% for ES
09 · Ablation
Remove a view, lose accuracy
Dropping the envelope spectrum costs the most, confirming the attention. Spectral kurtosis is redundant: the others cover it.
Without ES
98.50% (−1.12)
Without SK
99.62% (±0)

02 / Results

Consistent, general, and small enough to deploy

Not a lucky initialisation

Five training runs with different random seeds.

Five random seedsTable 4

Seed	Validation	Test
42	100.00%	99.25%
7	98.87%	98.12%
123	99.62%	98.87%
2024	98.87%	97.74%
999	98.87%	99.62%
Mean ± SD	99.25 ± 0.53%	98.72 ± 0.70%

Wilcoxon signed-rank, validation against test: p = 0.3125, no significant gap.

Better than a multi-view CNN

Same views, convolutional model, on this rig and the public CWRU dataset.

Against a Multi-View CNNTables 5, 7

Dataset	MV-ViT	Multi-View CNN
Custom rig	99.62%	96.11%
CWRU (public)	98.91%	96.30%

Over five paired runs: 99.243% against 96.106%; one-sided Wilcoxon p = 0.03125. CWRU models were trained from scratch the same way.

From 77.9 MB to 9.9 MB

Post-training quantization: size, speed and accuracy per format.

Post-training quantizationTables 10–12

Format	Size	Latency	Throughput	Accuracy
FP32Baseline	77.9 MB	25.48 ms	39.24 fps	98.87%
FP16All on the Neural Engine	39 MB	2.45 ms	406 fps	98.50%
W8A16INT8 weights, all on the Neural Engine	19.7 MB	2.43 ms	412.17 fps	98.87%
FP16Q481.67% on the Neural Engine, rest on CPU	9.9 MB	2.74 ms	364.42 fps	98.87%

19.47 M parameters. Accuracy here is on the paper’s quantization run, where FP32 scores 98.87%; INT8 and FP16Q4 match FP32 predictions exactly (Cohen’s κ = 1.0).

Read the full paper

Scientific Reports, in press

Earlier paperHybrid transfer learning

Six views in, one answer out

Same rig, raw windows

One signal, six images

Fused before the first layer

Patches become tokens

One shared transformer

Keep the heads different

Five health states

The envelope spectrum leads

Remove a view, lose accuracy

Consistent, general, and small enough to deploy

Not a lucky initialisation

Speeds and loads it never saw

Better than a multi-view CNN

From 77.9 MB to 9.9 MB