Wind turbine signal fault diagnosis using deep neural networks-inspired model

This work presents a deep neural network-inspired solution to intelligent signal fault diagnosis for wind turbine gearbox systems. A 1D convolution deep neural network architecture is proposed, constructed and validated. The proposed model was constructed of 1D signal for the input layer, ten different learned kernels as signal features, convolution layer, activation layer using rectified linear unit function, max-pooling layer, flatten layer and traditional multi-perceptron neural network for classification with soft-max class assignment. The data was acquired from real-world experiments conducted on real wind turbine gearboxes and archived by the National Renewable Energy Labs of the US Department of Energy. Ten independent experiments were conducted on 2,400,000 data points and the proposed model produced a mean classification accuracy of 96.14% for normal signals with a standard deviation of 0.0027 and a mean classification accuracy of 99.87% for faulty signals with a standard deviation of 0.0016.