Research On Precise Recognition And Diagnosis Technology Of Planetary Gearbox Fault

The expansion of mechanical system scale and the improvement of data acquisition system performance make it difficult for traditional fault diagnosis technology to meet the diagnostic needs of the era of "big data".With the breakthrough of deep learning technology,the data-driven model opens up a new way of thinking for real-time fault monitoring and diagnosis.In this paper,the fault feature extraction method of planetary gearbox is mainly studied,and an intelligent real-time fault diagnosis method of planetary gearbox is proposed based on Deep Learning technology.The specific research work of this paper is as follows:(1)The characteristics and formation principle of typical faults of planetary gearbox are studied.According to the common working conditions and fault forms of planetary gearbox,the design principle of rotor test-bench is studied.A rotor fault test-bed is designed and built to deal with various kinds of gearbox faults.The modular design idea is added to make the design of the test bed have the advantage of multi-functional modularization.In addition,the inherent characteristics of the transmission shaft are analyzed to ensure the safe and stable operation of the test bench.(2)The non-linear,non-stationary and multi-component characteristics of vibration signals of rolling bearings are studied.The problems in extracting fault features of rolling bearings by Multifractal detrended fluctuation analysis are analyzed.And Multifractal Super Order Analysis method is proposed by introducing the extreme value incremental series.The experimental results show that the new method overcomes the original shortcomings,and can extract the fault feature parameters which are more sensitive to the change of fault state and better discriminate.(3)The Recurrent neural network model of GRU structure is studied.Aiming at the problem of real-time fault diagnosis of planetary gearbox,a new intelligent real-time diagnosis method is proposed by utilizing its ability to process arbitrary constant sequence signals.Dropout technology is introduced in training to reduce the requirement of training data.By dividing the parameters of classification layer and using a small amount of new fault data to fine-tune the learned network parameters,the method can quickly realize the diagnosis of new faults under different working conditions and categories while maintaining the original recognition ability.Finally,the real-time diagnostic ability of the new method is verified by experimental data.