Feature Visualization and Classification using Adaptive VMD and Grad-CAM

As a key component in industrial system, the accurate detection of the wear condition of the ball screw is crucial for ensuring the operational accuracy and product quality. This paper proposes a feature visualization and classification approach for ball screw wear detection based on parameter-adaptive variational mode decomposition (VMD) and gradient-weighted class activation mapping (Grad-CAM). A VMD decomposition parameter adaptive optimization model based on the particle swarm optimization algorithm is constructed. The adaptive matching of the parameter pair is achieved through the mode mixing suppression criterion and approximate entropy minimization, resulting in multiple intrinsic mode function (IMF) components. Furthermore, Grad-CAM is utilized to quantify the contribution weights of the features. By computing the input channel weights and combining them with the IMF components, the input feature's influence on the convolutional neural network classification model's decision-making process is visualized. Finally, the method is employed to classify the wear states of ball screws. Experimental results show that the proposed method not only reveals the model's sensitivity to input features but also provides a basis for feature selection and model optimization. It effectively enhances the accuracy and robustness of wear detection, offering a novel technical approach for the condition monitoring of ball screws.