Multilayer Grad-CAM: An effective tool towards explainable deep neural networks for intelligent fault diagnosis

As a tool to explain deep neural networks using gradient information, Gradient-weighted Class Activation Map (Grad-CAM) provides a potential way for explainable artificial intelligence. However, for vibration signals in machine fault diagnosis, feature resolution of Grad-CAM decreases with the deepening of network layers, which weakens network explainability. To address this issue, a novel Multilayer Grad-CAM (MLG-CAM) is proposed as an effective tool to explain what networks have learned. Meanwhile, three indicators are defined to quantify explainability of deep neural networks. The MLG-CAM uses gradients flow of multiple convolutional layers to obtain activation maps in various resolutions. A comprehensive activation map is then produced by layer-weighted summation of above activation maps. Experiments indicate MLG-CAM not only highlights cyclo-stationary impulses in time domain but also emphasizes fault characteristic frequency in frequency domain. These results prove MLG-CAM as an effective way to explain deep neural networks and build up trustworthiness of networks.