It has been widely recognized that the practical application of deep learning to condition monitoring of industrial equipment relies heavily on having comprehensive coverage of essential features in the training dataset, i.e., the data collected should include all potential feature modes under various operating conditions. Given this limitation, this paper proposes a multi-level branch neural network (MLBNN) with self-evolution capability. In particular, the MLBNN structure can be updated online when emerging fault modes or operating conditions are encountered during equipment operation. Two case studies employing an industrial robot arm joint bearing and a motor bearing fault diagnosis dataset demonstrate the efficacy of MLBNN. We then demonstrate that MLBNN is suitable for deployment under cloud-edge computing architecture, which can further improve the speed of model update operations during self-evolution process and model inference to meet online condition monitoring requirements. Finally, the performance of MLBNN is compared to classical deep learning and transfer learning approaches under an incomplete training dataset scenario, and results show that the accuracy of MLBNN is superior to others.