Function estimation for multiple indices trend analysis using self-organizing mapping

Since system conditions can be indicated by a group of machine signal features but not any individual index, multiple indices trend analysis has foundational importance in system monitoring and diagnosis for factory automation. The author proposes to employ self-organizing neural network method to perform trend analysis in multi-dimensional space as an original exploration. However, experiments show that Kohonen´s learning algorithm and constrained topological mapping algorithm may yield nonfunctional maps in such a prediction analysis. An improvement on them by unequal scaling the training data can protect the topological order of netted neurons from being violated. This new approach achieves more accurate results than the widely used single-variable trend analysis method, and is suitable for interpolation for a large number of data and extrapolation in few data cases. The proposed approach is actually a general algorithm which can be widely used in high-dimensional line function regression.<>