R1 · First author · Scientific Reports, 2026

Reading machine health from vibration.

Which deep-learning models can tell bearing faults apart while load and speed keep changing, and how they hold up in noise.

CQ-NSGT · one window as a surfaceIllustration

Loading 3D surface…

Impacts rise as ridges across every bin; fault harmonics run as lines through time. Drag to turn.

Start the walkthrough Read the paper

Five-fold CV accuracy: 99.83%
Accuracy at 1 dB SNR: 98.8%
Friedman test: p = 0.0005

Venue: Scientific Reports · Nature Portfolio
Published: 11 April 2026 · Vol. 16, Art. 16892
Authors: Pratyush Vishal, Rohith Nair, Jonathan Jacob, Narendiranath Babu T., M. Pandiyan
DOI: 10.1038/s41598-026-42592-y

01 / Method

From vibration to fault class

Nine stages, from a seeded defect to a fault class.

Stage 01 / 09 · Test rig

Fig. 1Rig in side elevation: motor, support bearing, rotor load and the instrumented test bearing.

Stage 02 / 09 · Raw signal

Fig. 2Illustrative outer-race signal at 1500 rpm. The dimension marks one impact period, 1/BPFO.

Stage 03 / 09 · Windowing

Fig. 3One window shown in detail, then the two augmentations applied to it.

Stage 04 / 09 · CQ-NSGT

Fig. 4Illustrative spectrogram (left) and the measured intra-class variance of each transform, Table 5 (right).

Stage 05 / 09 · Input

Fig. 5The model input, dimensioned.

Stage 06 / 09 · Backbone

Fig. 6Schematic feature-map stack (not the real layer count) and each backbone’s cost, Tables 15–16.

Stage 07 / 09 · Residual blocks

Fig. 7The three blocks in line, with Detail A redrawn from the paper’s Fig. 11.

Stage 08 / 09 · Head

Fig. 8Head from the paper’s Fig. 12. Dense layers drawn with a sample of their units; hollow units show dropout.

Stage 09 / 09 · Output

Fig. 9Illustrative softmax for one outer-race window, with the published accuracy figures.

01 · Test rig
A seeded fault, under load
Bearings with seeded defects on the inner race, outer race, a ball, or all three, run at three speeds and three loads on a Spectra Quest MFS-PKG5 rig.
Health states
HB · IRF · ORF · BF · CF
Conditions
5 × 3 speeds × 3 loads = 45
Sampling
12,800 Hz
02 · Raw signal
Raw vibration, unfiltered
A fault shows up as a train of decaying impacts, one per rolling-element pass. The model gets the signal raw, unfiltered.
Files
45
Raw points
18.9 million
03 · Windowing
Cut into windows
Fixed-length windows with no overlap, so nothing leaks between folds. Flips and time rotations augment training.
Overlap
None
Augmentation
H-flip, time rotation
04 · CQ-NSGT
Signal to image
Each window becomes an image with geometric frequency bins. Of four transforms tested, CQ-NSGT kept classes tightest: variance 70.1 against 205–283.
05 · Input
224 × 224 × 3
Resized to the shape the pretrained networks expect.
06 · Backbone
A frozen, pretrained backbone
A pretrained network extracts features, weights frozen. Unfreezing a quarter of it cost 30–100% more training time for no accuracy gain.
Options
EfficientNetB0, MobileNetV2, InceptionV3
Weights
Frozen
07 · Residual blocks
Three trainable residual blocks
Three trainable blocks adapt the general features to bearing spectrograms; the skip path keeps gradients healthy.
Main path
Conv 3×3/s · BN · ReLU · Conv 3×3 · BN
Shortcut
Conv 1×1/s · BN
Merge
Add · ReLU
08 · Head
A small, shared head
Pooling, two small dense layers and a five-way softmax, shared by all three backbones for a fair comparison.
Optimizer
Adam
Epochs
Up to 100
Early stopping
val_loss, patience 3
LR schedule
×0.5 after 5 flat epochs, min 1e-5
09 · Output
A fault class
99.83% cross-validated accuracy, and 98.8% with noise at 1 dB SNR, where it carries almost as much power as the signal.
Evaluation
Stratified fivefold CV
Friedman test
p = 0.0005

02 / Model

The whole network on one sheet

General arrangement

Input

224 × 224 × 3

Frozen

Backbone

pretrained

Trainable · residual

Residual block 1

3×3, 3×3, 1×1

Residual block 2

3×3, 3×3, 1×1

Residual block 3

3×3, 3×3, 1×1

Trainable · head

Global avg pool

Dense 128

ReLU · drop .05

Dense 64

ReLU · drop .05

Dense 5

softmax

Fault class

HB IRF ORF BF CF

Frozen weights Trained on bearing data

After the paper’s Figs 7, 11 and 12.

Layer schedule

No.	Layer	Configuration	Trains
01	Input224 × 224 × 3 CQ-NSGT image	224 × 224 × 3 CQ-NSGT image	—
02	BackboneEfficientNetB0 / MobileNetV2 / InceptionV3, pretrained	EfficientNetB0 / MobileNetV2 / InceptionV3, pretrained	—
03	Residual block 1Conv 3×3/s → BN → ReLU → Conv 3×3 → BN, + Conv 1×1/s → BN shortcut, Add → ReLU	Conv 3×3/s → BN → ReLU → Conv 3×3 → BN, + Conv 1×1/s → BN shortcut, Add → ReLU	Yes
04	Residual block 2Conv 3×3/s → BN → ReLU → Conv 3×3 → BN, + Conv 1×1/s → BN shortcut, Add → ReLU	Conv 3×3/s → BN → ReLU → Conv 3×3 → BN, + Conv 1×1/s → BN shortcut, Add → ReLU	Yes
05	Residual block 3Conv 3×3/s → BN → ReLU → Conv 3×3 → BN, + Conv 1×1/s → BN shortcut, Add → ReLU	Conv 3×3/s → BN → ReLU → Conv 3×3 → BN, + Conv 1×1/s → BN shortcut, Add → ReLU	Yes
06	Global avg poolingFeature maps → one vector	Feature maps → one vector	—
07	Dense128 units · ReLU, L2	128 units · ReLU, L2	Yes
08	DropoutRate 0.05	Rate 0.05	—
09	Dense64 units · ReLU, L2	64 units · ReLU, L2	Yes
10	DropoutRate 0.05	Rate 0.05	—
11	Dense5 units · Softmax	5 units · Softmax	Yes

s = stride, not stated in the paper.

Training setup

Optimizer: Adam
Epochs: Up to 100
Early stopping: val_loss, patience 3
LR schedule: ×0.5 after 5 flat epochs, min 1e-5
Evaluation: Stratified fivefold CV
Noise test: AWGN at 5, 3 and 1 dB SNR

Why the backbone stays frozen

Unfreezing the last 25% made training slower and validation no better.

Backbone	Frozen training	25% unfrozen	Val. accuracy
MobileNetV2	121.4 s	+30–40%	No gain
EfficientNetB0	177 s	+50%	No gain
InceptionV3	230.5 s	≈ 2×	No gain

03 / Results

Accurate, robust, and cheap enough to ship

Is it real, or a lucky split?

Five folds per model. Consistency matters as much as the average.

Fivefold cross-validation

Accuracy per fold, no noise

Fold 1: train on four fifths, test on the fifth

testtraintraintraintrain

Stratified: each fold has a balanced share of all five classes and every speed and load. Windows do not overlap, so near-identical fragments can’t leak across folds.

fold accuracy mean (value at right)

Friedman test

χ²(4) = 20.00, p = 0.0005

Effect size

η² = 1.00

The ranking (EfficientNetB0, MobileNetV2, InceptionV3, CNN, ViT) was identical in every fold of the grouped evaluation, so the differences are systematic rather than luck of the split.

Show data table

Per-fold accuracy by model (Tables 7–11)
Model	F1	F2	F3	F4	F5	Mean
EfficientNetB0	99.80%	99.50%	100.00%	100.00%	100.00%	99.9%
MobileNetV2	97.70%	99.10%	99.30%	99.10%	99.50%	99.0%
InceptionV3	97.70%	98.40%	98.20%	97.70%	99.10%	98.2%
Base CNN	97.00%	93.00%	86.00%	94.00%	94.00%	92.8%
ViT	75.34%	85.11%	82.89%	81.34%	90.22%	83.0%

Per-fold values from Tables 7–11 · statistics from Table 17

What 98.8% at 1 dB actually means

White noise at 5, 3 and 1 dB SNR, every model re-tested.

Noise stress test

Additive white Gaussian noise

Synthetic signal · 60 msSNR 1 dB

Illustration, not paper data: outer-race impacts at 76.25 Hz (1500 rpm) plus white Gaussian noise scaled to the selected SNR. At 1 dB, the noise carries almost as much power as the signal.

EfficientNetB0: 98.8%
MobileNetV2: 95.2%
InceptionV3: 90.0%
Base CNN: 83.8%
ViT: 73.3%

Accuracy as noise increases (Table 12)

EfficientNetB0 loses just 1.1 points at 1 dB. The CNN and ViT baselines drop sharply and then fluctuate, rather than degrading smoothly.

With simulated speed fluctuation

EfficientNetB099.90% → 99.83%
MobileNetV299.00% → 98.23%
InceptionV398.20% → 95.07%

Grey dot: steady speed · filled dot: with fluctuation · axis 94–100%

Show data table

Accuracy by model and noise level (Table 12)
Model	No noise	5 dB	3 dB	1 dB
EfficientNetB0	99.9%	99.6%	99.2%	98.8%
MobileNetV2	99.0%	98.4%	97.4%	95.2%
InceptionV3	98.2%	95.2%	92.2%	90.0%
Base CNN	92.8%	83.6%	81.6%	83.8%
ViT	83.0%	69.0%	74.4%	73.3%

Accuracy from Table 12 · speed fluctuation from Table 13

A lower-bound robustness check; speed fluctuation was tested separately.

The best model isn’t always the one you ship

On edge hardware, compute, latency and memory matter too.

Accuracy vs cost

Up and to the left is better

MobileNetV2 and EfficientNetB0 need 0.50 and 0.59 GFLOPS per image; InceptionV3, ViT and the baseline CNN need roughly five to eight times as much, for lower accuracy.

Show data table

Accuracy and cost by model (Tables 6, 14–16)
Model	Accuracy	GFLOPS	Training (s)	Latency (ms)	Memory (MB)
EfficientNetB0	99.9%	0.59	177	12.214	24.623
MobileNetV2	99.0%	0.5	121.4	8.167	17.788
InceptionV3	98.2%	3.47	230.5	20.735	96.096
Base CNN	92.8%	4.05	195.5	–	–
ViT	83.0%	3.26	213.6	–	–

Accuracy from Table 6 · compute from Table 15

Read the full paper

Scientific Reports, 2026

Next paperMulti-view Vision Transformer

From vibration to fault class

A seeded fault, under load

Raw vibration, unfiltered

Cut into windows

Signal to image

224 × 224 × 3

A frozen, pretrained backbone

Three trainable residual blocks

A small, shared head

A fault class

The whole network on one sheet

Layer schedule

Training setup

Why the backbone stays frozen

Accurate, robust, and cheap enough to ship

Is it real, or a lucky split?

What 98.8% at 1 dB actually means

The best model isn’t always the one you ship