Improving Supervision of Cooling Towers with Complementary Machine Learning

Reliable fault detection and diagnostics (FDD) in large cooling systems is crucial to ensure a continuous functional performance and to reduce the energy consumption. The application of FDD methods remains challenging due to the little availability of labeled data, the limited interpretability of models, and the poor transferability of FDD solutions. COMETH is an active-learning framework, implementing a patented method, which combines two machine learning (ML) models with expert-in-the-loop feedback. This interaction addresses the challenge of labeled data and in this paper, we show the results of the application of COMETH to eight cooling towers (CTOs). Our training data set is based on time-series data labeled with predefined expert filters. During the application phase, the COMETH models flagged measurements that could not be classified with confidence, requested expert feedback and underwent iterative retraining on the newly labeled data. After 23 iterations, the mean $F_1$-scores had increased to 74% (DBSCAN) and 99% (decision tree), while the number of manually labeled measurements had decreased on average by 86% per CTO. Since labeling efforts are the limiting factor for the application of ML-based FDD, we then applied a transfer learning method based on a a statistical similarity analysis to identify the most suitable source system among the 8 CTOs. The ML methods trained on this source system could be transferred to the other CTOs with the objective to further reduce the training efforts for the target systems. Our results revealed a 57% reduction in labeling effort for the target systems compared to using untrained ML methods. However, the transfer failed for two CTOs due to data gaps and the remaining training efforts remain high due to the variability of the datasets with respect to the operation conditions of the different CTOs. Future work will explore clustering similar feedback requests to further reduce the labeling workload for experts. The transfer studies also suggest that pretraining the ML models on simulation data representing the most common fault scenarios could provide additional benefits.