On-Line Condition Monitoring Of Marine Gearbox Based On CWT-DCNN

Gearbox is widely used in mechanical transmission.As the key transmission component of most mechanical equipment,the failure of the gearbox will undoubtedly pose a threat to the health of the machine,and if the failure is serious,it may also cause damage to other components or even equipment paralysis.Therefore,in practical applications,it is very necessary to monitor the condition of the gearbox,which has a positive significance for production safety and equipment maintenance.Gearbox vibration signals are easy to collect and contain rich gearbox status information,so this paper uses it as the object for monitoring.In this paper,an online condition monitoring method based on continuous wavelet transformation(CWT)and deep convolutional neural network(DCNN)is proposed,which combines the timefrequency analysis advantages of CWT and the image recognition capabilities of DCNN.Firstly,based on dynamics and fault analysis,the generation mechanism and typical frequency domain characteristics of gearbox vibration signals are studied,and several mathematical models of vibration signals in typical states are established,which facilitates the interpretable analysis of non-stationary,nonlinear and impact characteristics of gearbox vibration signals.According to the characteristics of the vibration signal of the gearbox,its timefrequency domain feature is used as the condition classification object.The CWT method based on morlet wavelet is selected to extract the time-frequency domain features of the signal,and the vibration signal of the PHM2009 dataset is extracted by this method,and the wavelet time-frequency map obtained has good time-frequency resolution and categorical resolution.Then,based on the continuous wavelet transform and PHM2009 dataset,a wavelet time-frequency atlas is made for condition classification.Based on the extracted time-frequency plot data,this paper uses the DCNN that performs excellent in the field of image recognition to classify them,and builds a network model with 5 convolutional layers and 3 fully connected layers.The network is trained based on gradient descent,and the trained DCNN model achieves 100%accuracy on the training set,validation set,and test set.The generalization performance and classification reliability of the network are verified by the analysis of test results and t-SNE analysis.On the computer used in this paper,the average time for classification testing based on the trained DCNN is about 0.3s,and the average time for time-frequency feature extraction based on the CWT is about 1s,if based on a better performance computer platform and efficient data acquisition conditions,it can better meet the requirements of the gearbox online condition monitoring.