Using mega-trend-diffusion and artificial samples in small data set learning for early flexible manufacturing system scheduling knowledge