Function estimation for multiple indices trend analysis usingself-organizing mapping

摘要

Since system conditions can be indicated by a group of machinesignal features but not any individual index, multiple indices trendanalysis has foundational importance in system monitoring and diagnosisfor factory automation. The author proposes to employ self-organizingneural network method to perform trend analysis in multi-dimensionalspace as an original exploration. However, experiments show thatKohonen's learning algorithm and constrained topological mappingalgorithm may yield nonfunctional maps in such a prediction analysis. Animprovement on them by unequal scaling the training data can protect thetopological order of netted neurons from being violated. This newapproach achieves more accurate results than the widely usedsingle-variable trend analysis method, and is suitable for interpolationfor a large number of data and extrapolation in few data cases. Theproposed approach is actually a general algorithm which can be widelyused in high-dimensional line function regression