Deep classifier-based clustering of long and incomplete multivariate time series from wastewater treatment plants

Time series clustering is a valuable tool for analyzing water-related process data, but existing methods often struggle with the challenging characteristics of such data, including missing values, long sequences, and temporal heterogeneity. We propose a novel clustering methodology that constructs a distance matrix from the outputs of a classifier model trained to discriminate between windows sampled from different time series. The method leverages partial convolutions to handle missing data and combines convolutional neural networks with attention mechanisms to capture relevant features. We evaluate our approach on two datasets from wastewater treatment plants, containing time series with lengths up to 200,000 data points. Our method outperforms benchmarks under the evaluated settings for the water-related time series data and is competitive on commonly used datasets from the UCR Time Series Classification Archive, while maintaining robustness to large proportions of missing data in our experiments. The approach is scalable to arbitrary time series lengths and provides a flexible framework for incorporating domain knowledge through the classifier model design.