Fault detection of electric rope shovels is critical to their performance. Using neural networks for fault diagnosis has become increasingly popular, making the optimization of neural networks highly significant. Using evolutionary algorithms to optimize neural network weights is a popular approach that aims to find the optimal weight configuration by simulating the process of natural selection. Unlike traditional gradient-based optimization methods, such as backpropagation, evolutionary algorithms are gradient-free methods that demonstrate advantages in solving complex or non-convex problems. Recently, a popular evolutionary algorithm known as the large-scale sparse multi-objective evolutionary algorithm (LSMOEA) has attracted significant attention. It not only retains the traditional advantages of evolutionary algorithms but also excels in searching for lightweight neural networks. In this paper, we develop a dynamic variable clustering based LSMOEA and apply it to the fault detection of electric rope shovels. Experimental results show that, compared to traditional evolutionary algorithms, LSMOEAs are capable of finding lightweight neural networks, making them more suitable for real-world industrial applications.