Research On Gearbox Status Recognition Based On Multilayer Recurrent Neural Network

Gearbox is an indispensable universal co mponent for connecting and transmitting power in mechanical equipment.Its structure is complex,and it has been in a high load and high speed working environment for a long time.Therefore,the internal components of gearbox are prone to abnormal faults,leading to the failure of the gearbox to work normally,thus affecting the production efficiency.It is of great significance to study the fault diagnosis technology of gearbox,especially gears,for ensuring the safe and stable operation of machinery and equipment.In view of the complex working environment of gearbox for a long time,this paper chooses to collect the acoustic emission signals of gearbox to identify the status of gearbox.Acoustic emission(AE)is an effective monitoring and diagnostic tec hnology,which studies state changes by detecting stress waves released by objects themselves.In addition,AE signal has the following advantages:(1)high frequency,which can avoid low-frequency noise interference;(2)AE technology does not need to be very close to the measured object,and the deformation of the measured object will not easily cause the change of AE signal;(3)involving a wide range of fields,it can be used for both macro-detection and detection of changes in the internal organization of the object.Therefore,acoustic emission technology is very important for real-time monitoring and diagnosis of gearbox faults.To solve the problem of insufficient experimental data,this paper refers to the bearing data acquisition method of Case Western Reserve University in the United States,and takes the power transmission fault diagnosis comprehensive experimental platform as the research object,designs and collects a total of 30 sets of single and compound fault data of gearbox,which provides a powerful data guarantee for the subsequent gear box status recognition.At present,gearbox state recognition is mainly divided into two steps.Firstly,signal processing is used to extract the features of gearbox state signals,and then machine learning is used to identify the status.However,in the process of feature extraction,the uncertainty is strong and the requirement for experimenters is high.It requires not only to learn abundant signal processing methods,but also a lot of practical experience.At the same time,it also requires a lot of time and energy.Therefore,in this paper,a deep learning method is introduced,and a multi-layer cyclic neural network is used to identify the gear box state.The essence of multi-layer cyclic neural network is to stack multiple single-layer cyclic neural networks,that is,the input of the next layer network comes from the output of the upper layer network.Long Short-Term Memory(LSTM)and Gated Recurrent Unit(GRU)are variants of Recurrent Neural Network(RNN).In this paper,LSTM and multi-layer GRU network structures are built,and network samples are produced by using acoustic emission frequency domain signals of gearboxes for training.Finally,state recognition of gearboxes is realized.In the experiment,the layers,learning rate and batch processing parameters of the network are compared and analyzed respectively.At the same time,the recognition effects of multiple models are compared to find the optimal network structure for gear box state recognition.After verification,the multi-layer cyclic neural network method can extract the state features from the acoustic emission signal very well,and achieve more accurate and efficient gear box state recognition.