Smart Network Maintenance in an Edge Cloud Computing Environment: An Adaptive Model Compression Algorithm Based on Model Pruning and Model Clustering

摘要

With the rapid development of communication networks, there are more stringent requirements for maintenance management. To carry out smart network maintenance automatically and effectively, this paper uses wearable devices, robots, and Unmanned Aerial Vehicles (UAVs) to collect on-site video data and detect defects in real-time. However, the deep learning models deployed in these devices should have fewer model parameters and less storage space. Therefore, we propose a model compression algorithm based on model pruning and model clustering in smart network maintenance. First, model pruning combines channel pruning and layer pruning and uses deep reinforcement learning to determine the pruning ratio of each layer automatically. This method can effectively compress the width and depth of the model while maintaining the accuracy of the model. Second, although the pruning operation greatly reduces the redundancy of the number of weight parameters, the number of bits of floating-point weights is still redundant. We propose an adaptive model clustering method to cluster the remaining nonzero parameter weights and compress the model. It combines advanced balanced iterative reducing and clustering using hierarchies (BIRCH) clustering and K-meansII clustering and takes the result k value of BIRCH clustering as the input of K-meansII. It can avoid limitations of prior knowledge and reduce clustering time. Simulation results show that the target detection model can reduce parameter redundancy, save storage space, and simplify calculation using the proposed algorithm. In addition, it can be better applied in the smart network maintenance environment.