Speed Adaptive Gates for Deep-Learning-based Fault Classification of Rotating Machinery under Varying Speed Conditions: Comparison of Implementations

Deep-learning-based rotating machinery fault classification often suffers from the problem of speed induced fault information imbalance when applied to varying speed conditions. The speed adaptive gate (SAG) is an effective auxiliary branch that assists existing deep learning models addressing that problem. This paper presents four types of implementations for SAGs, i.e., the element-wise, channel-wise, time-wise and sample-wise, and compares their fault classification performances with a baseline Convolutional Neural Network (CNN) over two experimental datasets including a planetary gearbox dataset and a fixed-shaft gearbox dataset. Results show that the element-wise achieves the highest fault classification accuracy while the time-wise returns the least. But element-wise consumes about 20% more training CPU time than that of the baseline CNN. A good trade-off is the sample-wise with average speed as the input to the SAG, which obtains fair accuracy improvement, but spends only about 5% more training time than the baseline CNN.