Based on the Cao, mutual information, and small-data methods, the embedding dimension, delay time, and maximum Lyapunov exponent are calculated, respectively, to analyze the chaos characteristics of rolling bearings. The curve attractor of x(t) − x(t + (m − 1)τ)x(t) is constructed to parse the dynamics features on friction torque in the phase space of time series. According to the five prediction methods utilized—one-rank local-region, adding-weight one-rank local-region, improved adding-weight one-rank local-region, dial basis function (RBF) neural network, and Volterra series—the time series of bearings A, B, and C are forecasted with the first 400 experiment data as training values and the latter 57 data as test values to verify the prediction models’ feasibility. Finally, the bootstrap-maximum-entropy method is proposed to effectively fuse the results of these five prediction methods, and obtain the estimation interval and true value of friction torque. Experimental investigation shows that the friction torque phase trajectory has a linear increasing trend. These five forecasting models are effective for friction torque time series prediction with small error and high precision. The range of the fused estimation interval is relatively small, and the maximum error between the estimated true value and the experiment value is only 5.183 %, so the fluctuation information and the change trends of friction torque are accurately described. Moreover, the proposed models do not consider the probability distribution and trend information of the research system, breaking from the features of traditional statistical models.

No abstract