Research On The Construction Of Capsule Network Model For Intelligent Diagnosis Of Mechanical Faults

Intelligent mechanical fault diagnosis technology is the key to ensure the safe operation of equipment.Intelligent diagnosis can realize the transformation of mechanical equipment from "scheduled maintenance" to "predictive maintenance" by early detection of faults and their causes.However,there are still some challenges in the current intelligent diagnosis technology.Firstly,mechanical fault signals are a group of nonlinear structured time series,and the fault signals are diversified and complex,which leads to difficulty in fault feature extraction.Secondly,the vibration signal sensor is exposed to multiple vibration sources,complex transmission path and environmental interference for a long time,and the final accuracy of the diagnosis work is insufficient when the fault signal is annihilated by noise.Thirdly,most mechanical fault diagnosis models require large computing resources and storage capacity.Therefore,this thesis studies the solutions to the above problems in the mechanical fault diagnosis scenario.Compared with convolutional neural networks and other models,capsule networks can capture the feature hierarchy more effectively,perform more robust to the geometric changes of one-dimensional sequences,and cope with the sequence being disturbed by the outside world.In this thesis,capsule network as the basic framework,carried out a number of experimental platforms and engineering application case studies,step by step,the construction of three mechanical fault intelligent diagnosis methods with strong feature extraction ability,good anti-noise ability,lightweight model,the main content of this thesis is as follows.(1)In order to enhance the feature extraction capability of the original capsule network for fault signals,a temporal convolutional capsule network method is proposed in this thesis.Firstly,based on the traditional model’s difficulty in grasping the long time sequence of signals,a multi-scale temporal convolutional network with a larger receptive field was proposed to adaptively extract the long time sequence features of different levels of fault samples.Secondly,in order to improve the ability of the routing mechanism in the original capsule network to obtain effective information,the Sim attention mechanism is added to the dynamic routing mechanism to suppress the interference of irrelevant features on the model.The ablation experiment proves the effectiveness of each module of the temporal convolutional capsule network,and the comparison results with other methods prove that the proposed method has better feature extraction ability.(2)In order to apply the capsules network in(1)to a variety of complex scenes and improve the anti-noise ability,this thesis puts forward the comparison capsules network method.This method applies the contrast learning theory to construct a contrastive capsule network integrating encoder module,projection module and classification module.The network uses supervised contrastive loss to measure the distribution of sample features in the projection module.The results show that the signal-to-noise ratio ranges from-10 d B to-1 d B,and the comparison capsule network has better noise adaptability than the capsule network in(1)and other comparison methods.(3)In order to simplify the time and space complexity of capsule network in(2)and construct a mechanical fault diagnosis model requiring smaller computing and storage resources,a lightweight method of capsule network is proposed in this thesis.This method starts with the network model structure and parameter number in(2),constructs a multiscale temporal convolutional network with fewer residual modules from the perspective of knowledge distillation to reduce the parameter number,and uses a shared parameter fault capsule layer to reduce the calculation amount from the perspective of lightweight structure design.The experimental results show that the compression ratio of model parameters is equal to 51.56%,and the accuracy of the method is 87.89% and 83.60% in two experimental cases when the signal-to-noise ratio is-10 d B.