Intelligent Model Compression And Migration Algorithm For Smart Maintenance Of Communication Network

With the communication network structure and communication services more and more complex and diverse,the problems of large on-site maintenance workload,various abnormal phenomena are prominent.Intelligent wearable devices,inspection robots,and inspection Unmanned Aerial Vehicles have been applied in communication network maintenance.These intelligent terminal devices will realize the automation and intellectualization of communication network maintenance.However,these intelligent terminal devices generally have the problems of limited resources.And the models trained by different edge nodes are relatively independent and difficult to be directly migrated to other edge nodes.Therefore,this paper proposes an intelligent model compression and model migration algorithm for smart network maintenance.At present,there are some researches on model compression and model migration algorithms in smart network maintenance,but there are still some limitations.For model compression,the existing compression methods use a single method and rely on a prior knowledge,which can not meet the needs of devices in smart network maintenance.In terms of model migration,the existing migration algorithms do not consider whether the edge nodes have similar service requirements and do not consider the migration task and maintenance task simultaneously,which can not improve the efficiency of maintenance.According to the shortcomings,we optimize the communication network maintenance mechanism,mainly from model compression and model migration.(1)An adaptive model compression algorithm based on pruning and quantization is proposed.Firstly,the YOLOv3 model is the primary target detection model to identify the network entity target in the smart network maintenance.Secondly,according to the maintenance requirements of the communication network,this paper designs the model pruning algorithm based on reinforcement learning.Using this algorithm to obtain the compression ratio of each layer of the model and cut off some unimportant channels and convolution layers according to the compression ratio to obtain the sparse model and realize the compression of model size and processing time.Based on the above,an adaptive clustering method is designed to perform weight clustering on the pruned model,which can constrain the bit number of effective weights stored during hardware deployment by allowing multiple connections to share the same weight.Finally,simulation experiments show that this model compression algorithm effectively reduces the size and processing time of the model and is more suitable for intelligent terminal devices in smart network maintenance.(2)A model migration strategy is proposed based on a service demand similarity model and delay optimization.Firstly,to reduce the blindness of edge model migration,a service demand similarity model is constructed based on the type similarity,name similarity,and interpretation similarity.Then,according to the similarity of service demand,the model migration strategies under different model similarities are designed,which ensures that the training model is reused in similar target task areas.Moreover,a large number of edge resources will be consumed in model migration,which may affect the normal execution of maintenance tasks.So,this paper jointly considers the migration tasks and maintenance tasks in the communication network maintenance,fully considers the delay constraints of different tasks,and establishes an unloading scheme based on the improved discrete particle swarm optimization algorithm to minimize the unloading delay and realize the reasonable allocation of resources.Moreover,it also ensures that migration tasks and maintenance tasks are executed efficiently in the meantime.Finally,simulation experiments show that the model migration strategy effectively reduces the task delay,ensures the task completion rate,and performs better.