Research On Edge Condition Monitoring Based On Model Compression Via Deep Reinforcement Learning

Modern machinery and equipment are becoming increasingly automated,complex and large-scale,and when such machinery and equipment fail,it can cause significant economic losses and safety accidents.In order to avoid accidents,it is necessary to monitor and maintain the equipment to ensure the safety and stable operation of the equipment.Current state monitoring based on cloud computing has been widely attention,but with the rapid development of Internet of things industry,using industrial equipment of all kinds of sensors to collect the amount of data is more and more big,if the real time to upload the data synchronization carries on the analysis to the clouds,can bring huge to condition monitoring based on cloud storage,network and calculation of the pressure.The mode of combining state monitoring and edge computing emerged as The Times require.By carrying out state monitoring on devices at the edge and utilizing the computing storage capacity carried by devices at the edge,data processing and analysis of industrial equipment can be completed at the edge,which can share the computing pressure on the cloud and reduce the deployment cost of enterprises.Deep learning is the hotspot in various fields of research and application,in order to improve the performance of the neural network model for structural layers,and quantity calculation,the model complexity,again and again leading to deep memory overhead of network model and with the increase of computational complexity and depth of network model on resource-constrained edge side equipment deployment more difficult.Model compression can reduce the computation and storage resource occupation demand of the deep network model,which makes it possible to deploy the edge side of the deep network model.The main work of this thesis is as follows:1.For the analysis and processing of equipment monitoring data and state recognition,a state recognition method based on wavelet packet reconstruction and convolutional neural network is proposed.Wavelet packet transform is used to obtain the low frequency subband and high frequency subband signals of the state monitoring data,and recombine them according to the sequence to retain all the signal characteristics of the data.Then the convolutional neural network is introduced to further extract the characteristics of the input state monitoring data with its strong learning ability and adaptability.2.For model compression of convolutional neural network,a model compression method based on maximum entropy reinforcement learning algorithm is proposed.By using the reward function obtained by the maximum entropy reinforcement learning algorithm,the compression ratio of continuous motion mapping is controlled to achieve a compression selection method for model autonomous learning.3.Built the edge-side state monitoring system compressed by the deep reinforcement learning optimization model,analyzed the edge-side state signals,deployed the pruned model to the edge nodes,and carried out edge-side data collection,edge-side monitoring,and cloud backup of key historical information.