All research

Applied machine learning · 2024

Lightweight deep learning for early gearbox fault detection

GearFaultNet, a compact one-dimensional network that flags a broken gear from four-channel vibration signals, tested on loading conditions it had not seen in training.

Qatar University and Kennesaw State University

Block diagram of GearFaultNet. A four-channel input signal is normalised and passes through five blocks of 1D convolution, batch normalisation and tanh activation; the first block also has max pooling. Feature maps change from 4 by 512 at the input to 32, 16, 32, 24 and 16 channels of length 256. Adaptive average pooling reduces them to 16 by 8, and after flattening a small dense network outputs Healthy or Broken.
Figure 6 from Dutta et al. (2024). GearFaultNet: five small 1D convolutional blocks, adaptive average pooling and a dense classifier that labels each vibration window as healthy or broken.View full-size figureSource articleCC BY 4.0

Approach

Can a small one-dimensional network detect a broken gear from vibration signals reliably enough for real-time monitoring, even under loads it was not trained on?

Gearboxes transmit power in wind and hydro turbines, generators and pumps, and in wind turbines they cause more downtime than any other component. Faults found late lead to unplanned shutdowns and costly repairs. Many detection methods rely on heavy preprocessing or large models that are hard to run in real time. This study built GearFaultNet, a shallow one-dimensional convolutional network that labels short windows of vibration as healthy or broken after simple normalisation.

  1. Segment and normalise vibration signals
  2. Classify each window with GearFaultNet
  3. Test on unseen loading conditions

Preparing four-channel vibration data

IEEE Access · 2024

The study used the public Gear Box Fault Diagnosis dataset, recorded on SpectraQuest’s gearbox fault simulator. Four accelerometers on the gearbox body measured vibration for a healthy gear and a broken gear under ten loading conditions, from 0% to 90% in steps of 10%.

Each channel was cut into windows of 512 samples. Training windows overlapped by half to give the network more examples. Each channel was then standardised to zero mean and unit variance and scaled to a range of −0.2 to 0.2, simple steps that keep preprocessing fast.

The folds were split by load. Each of the ten folds held out one loading condition for testing and trained on the others, so the model was always tested on a load it had not seen. This also avoids the overlap between training and test data that shuffling windows before splitting can create.

Accelerometer channels
4
Samples per window
512
Loading conditions, one per test fold
10
Plot of a vibration signal over 5,000 sample points with amplitudes between about −20 and 20. Rectangles mark 512-sample windows at the start and end of the signal, with each dashed window starting halfway through the previous one.
Figure 3 from Dutta et al. (2024). Vibration signals were cut into 512-sample windows; training windows overlapped by 50%.View full-size figureSource articleCC BY 4.0
Grid of density plots for four channels at 0% and 30% load, for broken and healthy gears. Raw signals spread across roughly −20 to 20 units. After z-score and min-max normalisation, every channel sits within −0.2 to 0.2.
Figure 4 from Dutta et al. (2024). Distributions of the four channels before and after normalisation, for broken and healthy gears at 0% and 30% load.View full-size figureSource articleCC BY 4.0

Designing a lightweight one-dimensional classifier

Model design

GearFaultNet takes each four-channel window as a 4 × 512 input. Five small blocks of 1D convolution, batch normalisation and tanh activation extract features, with 16 to 32 filters per block. Max pooling in the first block halves the length, and adaptive average pooling reduces the final features to 16 × 8 before a small dense classifier labels the window as healthy or broken.

The loss passed the outputs through a softmax and then a mean squared error. Training used the Adam optimiser on Google Colab, and the learning rate was lowered and training stopped early when validation loss stopped improving.

Convolutional blocks
5
Input per window
4 × 512
Features passed to the classifier
16 × 8
Training settings. From Table 1 of the paper.
Value
Batch size4
Maximum epochs200
Learning rate0.0002, reduced by a factor of 0.2 after 10 epochs without improvement
Early stoppingAfter 30 epochs without improvement
OptimiserAdam
Loss functionSoftmax followed by mean squared error

Detecting broken gears under unseen loads

Ten-fold evaluation

Across the ten folds there were 7,488 test windows: 3,697 from the broken gear and 3,791 from the healthy one. GearFaultNet classified 94.06% correctly. It found 3,409 broken-gear windows and missed 288, and it raised 157 false alarms among the healthy windows. Precision was higher for the broken class, at 95.60%, while recall was higher for the healthy class, at 95.86%. Both classes had an area under the ROC curve of 0.98.

Under the same settings, GearFaultNet outperformed a 1D ResNet18 (77.74% accuracy) and Self-ResNet18, a version built with self-organised operational layers (92.74%). Earlier hybrid LSTM methods on the same dataset reported 87.50%. Studies that reported 100% shuffled windows before splitting, whereas the load-based split here keeps each test load out of training.

Accuracy
94.06%
Area under the ROC curve
0.98
Test windows across ten folds
7,488
Comparison with two deeper 1D networks under the same settings (%). From Table 3 of the paper.
AccuracyPrecisionSensitivitySpecificityF1 score
ResNet1877.7478.1077.7477.5877.64
Self-ResNet1892.7493.2092.7492.6092.71
GearFaultNet94.0694.1194.0694.0194.05
Two-by-two confusion matrix. 92% of broken-gear windows (3,409) and 96% of healthy windows (3,634) are classified correctly. 8% of broken-gear windows are labelled healthy and 4% of healthy windows are labelled broken.
Figure 7 from Dutta et al. (2024). Confusion matrix combined over all ten test folds.View full-size figureSource articleCC BY 4.0
ROC curves for the broken class, the healthy class and their micro and macro averages. All four curves rise steeply to a true positive rate above 0.9 at a false positive rate below 0.1, and each has an area of 0.98.
Figure 8 from Dutta et al. (2024). ROC curves for each class and their micro and macro averages, all with an area of 0.98.View full-size figureSource articleCC BY 4.0

My contribution

I contributed to model evaluation, quantitative analysis, interpretation and manuscript review.

Study scope

The study used one public test-rig dataset with a single fault, a broken gear, compared with healthy operation. Other fault types, gearboxes and sensor layouts were not tested, and the model recognised healthy windows more reliably than broken ones.