How to Turn Vibrations into Images to Spot Cracks in Turbines?

Wind turbines are like giant airplane wings: a cracked blade reduces the whole machine's efficiency. Researchers have devised a way to "see" cracks using only turbine vibrations. First, they recorded signals from a three‑axis accelerometer, then turned them into images with the Gramian Angular Field (GAF). Imagine converting a song into a picture: the rhythm becomes a color pattern. These images preserve the signal's time relationships, revealing textures invisible to the human eye. Next, a convolutional neural network (CNN) – the same algorithm that recognizes faces – analyzes the GAF textures and learns to differentiate normal vibrations from those indicating a crack. The outcome? A classification accuracy above 99.9 %, meaning almost no false alarms. This approach is faster and less noise‑sensitive than traditional methods that need experts and manual inspections. In practice, turbines could self‑diagnose, cutting maintenance costs and boosting wind‑farm safety.

The Gramian Angular Field (GAF) is an encoding technique that maps a time series x(t) into a symmetric matrix G, where G_{i,j}=cos(\phi_i+\phi_j) with \phi_i=\arccos(x_i). This transformation preserves temporal structure and converts the sequence into a two‑dimensional representation suitable for CNNs. The CNN employed consists of three convolutional blocks (3×3 kernels, ReLU, 2×2 max‑pool) followed by two fully‑connected layers and a softmax classifier. The dataset includes vibration recordings from low‑power turbines, labeled as no crack, minor crack, and severe crack. Training used the Adam optimizer (learning rate 0.001) for 50 epochs, with data augmentation via Gaussian noise to enhance robustness. The ROC curve shows an AUC of 0.9995, confirming the model's high discriminative power. Saliency map analysis reveals that high‑intensity regions in the GAF correspond to vibration transients typical of fractures, providing model interpretability. This approach can be extended to other rotating systems, such as generators or motors, where predictive diagnostics are critical.