We found the same issue with non-normal distributions in our time series data sets when using SAX (developed by this paper's authors - assumes normality for dim

We found the same issue with non-normal distributions in our time series data sets when using SAX (developed by this paper's authors - assumes normality for dimensionality reduction) and addressed it by using quantiles in the piecewise aggregate approximation step. The quantile breakpoints behaved much better than the "normal" breakpoints.