Research On The Intelligent Fault Diagnosis System For The Rolling Bearing On The Edge Computing Platform

As a key component in the aero engine,high-speed rail hub and vehicle radar slewing support system,bearing components are usually subjected to alternating loads in mechanical systems.The health status of the bearing directly affects the performance of the precision electromechanical systems.When working,the bearing component is very easy to suffer the performance degradatuion or damages due to the action of alternating stress.This will bring safety hazards to the equipment and human.Thus,it is necessary to monitor the health status of the bearing component online.When applied in the field,the intelligent bearing fault diagnosis method is facing the problems of the low fault diagnosis accuracy under the unstable working conditions.The bearing fault diagnosis networks often suffer the problems of the structural and parameter redundancy and insufficient performance of lightweight fault diagnosis network.In view of the above problems,this thesis conducts the research on the intelligent fault diagnosis method for the rolling bearing on the edge computing platform.The main research contents the design method of the intelligent fault diagnosis neural network under the unstable working conditions,the autonomous compression method of the fault diagnosis network,the performance optimization method of the compressed lightweight fault diagnosis networks and the deployment method of the fault diagnosis networks on the FPGA chip.With the above researches,the practical process of the intelligent fault diagnosis algorithms represented by the deep neural network is expected to be promoted.The main works are as follows:Aiming at the problem of the low accuracy under the variable speeds,an intelligent fault diagnosis algorithm based on the order tracking(OT)and the onedimensional convolutional neural networks(1-DCNN)for the bearing under the variable speeds is proposed.This work carries out the research on the intelligent classification method for the non-stationary vibration signals.The OT method is adopted to resample the original signals.This process can overcome the problem of modal aliasing caused by the variable speeds.A convolutional neural network is builded for the fault feature extraction and the fault classification.In order to improve the robustness of the fault diagnosis network,the wide convolutional kernels are adopted in the first layer.This step is expected to improve the anti-interference ability of the network.With the above steps,the proposed algorithm can realize the autonomous classification task for the non-stationary fault signals under the variable speed conditions.To tackle the issue of redundant structure and parameters faced by the current large-scale fault diagnosis neural network and the difficulty of practical deployment on small edge computing platform,an autonomous compression method of bearing fault diagnosis network structure based on network pruning and reinforcement learning technology is proposed.The network model compression and structural optimization of the initially designed large-scale fault diagnosis neural network are carried out.The network pruning process is viewed as a Markov process and statistical mathematical models are developed for floating point operations and network parameters.Pruning criteria for various types of neural network layers are constructed.An agent model is devised to observe the environment status during the network pruning process.Simultaneously,the agent provides the pruning strategy for each step of pruning.Utilizing the reinforcement learning method,the parameters of the networks in the agent are updated to enable the output of the optimal pruning strategy for real-time fault diagnosis.Lastly,experimental studies using two datasets are conducted to verify the efficacy of the proposed network compression method.To address the issue of the degradation performance of lightweight fault diagnosis neural network classification obtained after compression,a lightweight fault diagnosis network performance optimization method based on knowledge distillation and parameter quantization method is proposed,and the performance of the lightweight fault diagnosis neural network obtained after compression is optimized.A general framework for optimizing the performance of lightweight fault diagnosis networks is constructed based on knowledge distillation and parameter quantization.A mathematical model of the fault diagnosis network performance optimization process based on knowledge distillation is established to explore the mathematical principles of knowledge transfer between teacher and student networks during knowledge distillation.Based on the basic principles of parameter lowprecision quantization,a mathematical model is established for the integer parameter calculation process of quantized convolution layers and fully connected layers,further compressing the parameter storage space of the fault diagnosis neural network and ultimately achieving the goal of improving the classification accuracy and robustness of lightweight fault diagnosis networks.In order to further verify the effectiveness of the bearing fault intelligent diagnosis network design,network compression and network optimization algorithm proposed for edge computing platform scenarios,a bearing fault experimental platform is built,and the intelligent diagnosis network design,network compression and network optimization research are carried out in turn,and the bearing fault under unstable working conditions is intelligently diagnosed.The optimized network is successfully deployed on small FPGA chips.The intelligent diagnosis task of bearing fault data under unstable working conditions is realized on the small embedded computing platform.