Abstract: There are so many species of components for complicated products in manufacturing enterprise, which has complicated machining and assembly processes. And the quality, process, equipment and other types of data has multi-dimensional, multi-scale, multi-noise characteristics, which makes it difficult to accurately predict the duration. To address the issues mentioned above, one kind of complicated product duration prediction method based on multi-dimensional nonlinear feature reconstruction and fusion is proposed. First, multi-dimensional nonlinear feature reconstruction and fusion using integrated Stacked Auto Encoder is performed and the corresponding model is constructed, and a complicated correlation coupling connection between features is built to provide a feature pool comprising critical duration factors. Second, complicated product duration prediction model based on multi-dimensional nonlinear feature reconstruction is built to accomplish the complicated product duration prediction. Finally, the proposed model is applied in enterprises to predict the circuit breakers and 3D printed products’ duration. The mean root mean square error of this method reached 1.28, and the mean absolute percentage error reached 3.01%, which is higher than the accuracy of the method without feature reconstruction and fusion, support vector machine, neural network, and other methods. The root mean square error decreased by at least 10.87%, the mean absolute percentage error decreased by at least 7.74%, demonstrating the usefulness and applicability of this method.

Key words: feature reconstruction, feature fusion, duration prediction, deep neural network, integrated stacked auto encoder