Multi-residual unit fusion and Wasserstein distance-based deep transfer learning for mill load recognition

This paper proposes the ball mill load recognition algorithm (MRUF-WD) based on multi-residual unit fusion (MRUF) and Wasserstein distance transfer learning to address the problem of accurately judging the working state parameters during ball mill grinding operations. The proposed algorithm can significantly enhance the processing capability of under-labeled or unlabeled data, while ensuring excellent transfer capability even when there is a large data gap between the target and source domains. The MRUF-WD transfer algorithm first inputs collected ball mill vibration data under different working conditions into the MRUF feature extraction network based on ResNet-18, where MRUF can extract feature information from the upper, middle, and lower layers of the network for learning. This approach equips the algorithm with enhanced feature extraction capability, leading to a significant improvement in the processing capability for under-labeled or unlabeled data. This algorithm has a better ability to deal with under-labeled or unlabeled data through its improved feature extraction capability. It then inputs a classifier that uses Wasserstein distance as a loss function to measure distance between source and target domain data, as well as a domain alignment network. The Wasserstein distance loss function, also known as EDM (Earth Mover’s Distance), measures the distance between two probability distributions and exhibits broader applicability, better robustness, and improved continuity when the probability distribution of the data changes, compared to other common distance metrics. After the parameters are updated by back propagation, the algorithm uses the invariant features learned in the source domain to identify the load in the target domain.