Linguistic Ordered Weighted Averaging based deep learning pooling for fault diagnosis in a wastewater treatment plant

TL;DR

This paper introduces a novel deep learning pooling method based on linguistic Ordered Weighted Averaging (OWA) for fault diagnosis in wastewater treatment plants, improving early detection accuracy with fewer samples.

Contribution

It proposes a new linguistic OWA pooling technique integrated into DCNNs, enhancing fault diagnosis performance and interpretability in complex industrial water treatment processes.

Findings

Achieved over 91% accuracy, recall, and F1-score.

Outperformed standard max and average pooling methods.

Enabled earlier fault detection with fewer samples.

Abstract

Nowadays, water reuse is a serious challenge to help address water shortages. Here, the wastewater treatment plants (WWTP) play a key role, and its proper operation is mandatory. So, fault diagnosis is a key activity for these plants. Their high complexity and large-scale require of smart methodologies for that fault diagnosis and safety operation. All these large-scale and complex industrial processes are monitored, allowing the data collection about the plant operation, so data driven approaches for fault diagnosis can be applied. A popular approach to fault diagnosis is deep learning-based methodologies. Here, a fault diagnosis methodology is proposed for a WWTP using a new linguistic Ordered Weighted Averaging (OWA) pooling based Deep Convolutional Neural Network (DCNN) and a sliding and overlapping time window. This window slides over input data based on the monitoring sampling time, then the diagnosis is carried out by the linguistic OWA pooling based DCNN. This alternative linguistic pooling uses well-known linguistic OWA quantifiers, which permit terms such as \textsl{Most, AtLeast, etc.}, supplying new intuitive options for the pooling tasks. This sliding time window and the OWA pooling based network permit a better and earlier fault diagnosis, at each sampling time, using a few monitoring samples and a fewer learning iterations than DCNN standard pooling. Several linguistic OWA operators have been checked with a benchmark for WWTPs. A set of 5 fault types has been used, taking into account 140 variables sampled at 15 minutes time intervals. The performance has been over $91\%$ for $Accuracy$, $Recall$ or $F1-Score$, and better than other competitive methodologies. Moreover, these linguistic OWA operators for DCNN pooling have shown a better performance than the standard \textsl{Max} and \textsl{Average} options.