Predictive Maintenance for Ultrafiltration Membranes Using Explainable Similarity-Based Prognostics

TL;DR

The paper tackles UF membrane predictive maintenance by addressing the interpretability gap of traditional ML approaches. It introduces an explainable fuzzy similarity prognostic framework that builds a physics-informed Health Index from metrics like $R_m^*$, TMP$^*$, $J^*$, and Recovery, then fuzzifies these into Gaussian memberships to create a 120-dimensional degradation signature. Similar historical trajectories are retrieved via a set-theoretic fuzzy similarity and combined using a Takagi–Sugeno rule base to produce a RUL estimate: $ ext{RUL ext{_hat}} = ( ext{sum}_i S(q,s_i) r_i)/( ext{sum}_i S(q,s_i))$. The method demonstrates about 4 cycles MAE on industrial UF data, with best performance in the 6–15 cycle horizon, and provides interpretable rules that directly tie predictions to past degradation cases, all while requiring only four hydraulic sensors. This combination of accuracy and transparency supports actionable maintenance planning in desalination pretreatment and can be extended to other membrane processes such as RO and microfiltration.

Abstract

In reverse osmosis desalination, ultrafiltration (UF) membranes degrade due to fouling, leading to performance loss and costly downtime. Most plants rely on scheduled preventive maintenance, since existing predictive maintenance models, often based on opaque machine learning methods, lack interpretability and operator trust. This study proposes an explainable prognostic framework for UF membrane remaining useful life (RUL) estimation using fuzzy similarity reasoning. A physics-informed Health Index, derived from transmembrane pressure, flux, and resistance, captures degradation dynamics, which are then fuzzified via Gaussian membership functions. Using a similarity measure, the model identifies historical degradation trajectories resembling the current state and formulates RUL predictions as Takagi-Sugeno fuzzy rules. Each rule corresponds to a historical exemplar and contributes to a transparent, similarity-weighted RUL estimate. Tested on 12,528 operational cycles from an industrial-scale UF system, the framework achieved a mean absolute error of 4.50 cycles, while generating interpretable rule bases consistent with expert understanding.

Figures (4)

• Figure 1: Conceptual degradation profile illustrating typical UF membrane behavior: progressive TMP rise, flux decline, and HI deterioration, with diminishing chemical cleaning effectiveness (orange) as irreversible fouling accumulates.
• Figure 2: Explainable fuzzy similarity-based prognostic framework for UF membrane RUL estimation.
• Figure 3: (a)Predicted vs. actual RUL for 2,668 test cycles (MAE = 4.08 cycles). Predictions cluster around the ideal diagonal (red dashed line), with optimal accuracy in the medium-term range (6--15 cycles). Scatter increases for imminent failures (lower left) and long-term projections (upper right), reflecting trade-offs in uniform fuzzy partitioning design. (b)Distribution of prediction errors showing near-zero median error. The sharp central peak corresponds to accurate predictions across the dominant 6--15 cycle range, while distribution tails reflect increased uncertainty for extreme RUL values.
• Figure 4: Example degradation trajectory for query cycle 12528 showing Health Index of query cycle in black and top similar trajectories.