Research On Fault Diagnosis Of Commercial Vehicle Main Reducer Based On Capsule Network

After the commercial vehicle main reducer design and development is completed,it is necessary to conduct fatigue endurance test on the main reducer to verify the design life of the main reducer.In the fatigue endurance test,as the main reducer cannot be disassembled at will to observe the internal damage,it is necessary to diagnose the fault of the main reducer in real time through the collected information.With the continuous development of fatigue endurance test of main deceleration,enterprises need to stop at a specific fault(pitting corrosion,tooth breaking,etc.).At present,manual fault diagnosis is inefficient,requires high professional skills of personnel,and the fault diagnosis accuracy of expert system can’t meet the needs of enterprises.Therefore,it is imperative to find a high precision fault diagnosis method.In fatigue endurance test,the amount of failure data produced is small,and each failure data means the damage of a main reducer.Capsule network adopts vector calculation,which contains more valid information and requires less data.In this paper,a fault diagnosis and classification method for commercial vehicle main reducer based on capsule network is proposed.By optimizing and improving the network structure,algorithm and super parameters,efficient and accurate fault classification and recognition can be achieved for the main decelerators during fatigue endurance.In order to reduce intermediate data processing links and improve classification efficiency,this paper adopts the "end-to-end" recognition process(vibration original time domain signal as input,directly output diagnosis results).In this paper,transfer learning is introduced to improve the accuracy of fatigue fault diagnosis of the newly designed main reducer.The research contents of this paper are as follows:(1)The dynamic routing formula of capsule network is deduced,and the encoding structure and decoding structure are introduced to build the capsule network model.The vibration data set of the main reducer is converted into two-dimensional gray scale map and input into the capsule network,and the algorithm of the capsule network is introduced in detail.The research shows that the capsule network has a good classification ability when facing the vibration data of the main reducer.(2)In this paper,the main reducer structure,the fatigue damage theory and fatigue durability test failure characteristics are summarized,based on the principle of the gear fault signal modulation,analyzed to summarize the characteristic of fatigue failure of gear,for the first time put forward the relative edge band value the monitoring indicators,and Peak-to-peak,RMS,kurtosis,such as monitoring indicators for comparative analysis,It provides reference for determining the specific fault of the main reducer and assists engineers to judge the specific fault of the main reducer.According to this monitoring index and observation by disassembling the main reducer housing,the vibration data set of the main reducer is labeled accurately.The fatigue and durable vibration signal data set of commercial vehicle main reducer was established,which included five kinds of data:normal,gear pitting,gear crack,gear tooth breaking and bearing fault.(3)The classification ability of capsule network is still insufficient,and the test accuracy of the previously established data set is only 84.4%,which is difficult to meet the requirements of fault diagnosis accuracy in fatigue durability.Therefore,the capsule network needs to be further optimized.Firstly,the processing process of data set is reduced,and one-dimensional signal in time domain is directly input into the capsule network,and the two-dimensional capsule network is improved into one-dimensional capsule network.Then,the one-dimensional capsule network is optimized from three aspects of structure,algorithm and hyperparameter.The multi-channel parallel structure is used to extract more signal features,and the cavity convolution is used to expand the receptive field to improve the feature extraction ability of the network model.The accuracy of one-dimensional capsule network improved to 96.4% compared with two-dimensional capsule network.(4)In fatigue endurance test,it is often encountered that the number of vibration data sets of the newly designed main reducer is insufficient.The training network of neural network needs a lot of data,so the study of fault diagnosis of main reducer based on transfer learning and capsule network is proposed.After "freezing" the parameters of the front feature extraction layer,only train the digital capsule layer through a small amount of data,and transfer the capsule network trained with vibration data of other models of the main reducer to the new data set.In the case of only a small amount of labeled data,the accuracy was improved from 79.8% to 94.8%.