Abstract:To learn partbased representation of data and enhance sparseness, this study demonstrates the embedding of nonnegativity constraints in the deep network. A state recognition method for rolling bearing is proposed based on the deep autoencoder neural network with nonnegative constrains. Multiple autoencoders and a classification layer are stacked to formulate an integrated model for feature selflearning and state recognition. The bearing vibration timefrequency spectrogram is taken as input, and the model is optimized by combining unsupervised layerwise pretraining and supervised finetuning. Both of them are with the nonnegativity constraints embedding. The deep network encodes and extracts the intrinsic feature of data layer by layer. The nonnegative constrains and denoising encoding improve the partbased representation ability of deep network. And the influence of condition variation and noise interference is decreased. The proposed method is applied to the vibration data analysis of two kinds of rolling bearings. The average recognition accuracy of four different state bearings under variable conditions and eight different state bearings under constant conditions are 9799% and 9732%, respectively. The average recognition accuracy of bearings with different retainer wear levels is 9564%. Meanwhile, the proposed method shows good antinoise capability under different levels of noise.