Abstract: The operating environment of complex feeding and ramming mechanisms is harsh, and the vibration signals collected by sensors contain violent impact, noise and other components, which are typical non-stationary characteristic signals, and the health state of feeding and ramming mechanisms is difficultly evaluated. To solve these problems, a soft measurement method based on time-frequency data driving is proposed for measuring the health state of feeding and ramming mechanism. The time-frequency graph of vibration acceleration signal, as an input feature, is obtained by Morlet wavelet transform, and a soft measurement model based on deep convolutional network is established. The Dropout regularization term is introduced into the deep convolutional network to relieve overfitting phenomenon, and the uncertainty of soft measurement results is analyzed quantitatively. The bench test of complex feeding and ramming mechanism shows that the proposed soft measurement method can effectively distinguish the health state of feeding and ramming mechanism with the accuracy of 90%. In the stage of performance degradation, the degradation degree of the mechanism performance can be quantitatively analyzed, and the measured error is about 7%. Compared with other data-driven soft measurement methods for health state, the proposed method can effectively improve the identification accuracy of health state and the measuring accuracy of performance degradation of feeding and ramming mechanisms, and reduce the uncertainty of measureed results.

Key words: complexfeedingandrammingmechanism, healthstate, softmeasurement, time-frequencydata