Intelligent Fault Diagnosis and Fault-Tolerant Control of PV-BLDC Water Pumping Systems Using an STO-Enhanced Hierarchical LSTM–CNN Architecture

Reliable operation of photovoltaic (PV)-powered water pumping systems is essential for sustainable irrigation and water supply, particularly in remote areas where maintenance and continuous monitoring are challenging. This paper proposes a Super-Twisting Observer (STO)-assisted hierarchical LSTM–CNN Fault-Tolerant Control (FTC) framework (LC-FTC) for fault diagnosis and fault-tolerant operation of PV-powered brushless DC (BLDC) water-pumping systems. The study investigates the diagnostic usefulness of residual signals generated by an STO and evaluates their ability to support data-driven fault detection and classification within a simulation framework. The proposed framework integrates model-based state estimation with data-driven fault diagnosis and signal substitution. A dataset is generated using a MATLAB/Simulink model under four operating conditions and multiple single and combined measurement-fault scenarios involving DC-link voltage and current measurements, the speed sensor, and the current sensor. The STO generates residual signals for fault diagnosis and estimated signals for subsequent signal substitution. The diagnostic architecture operates hierarchically: a Long Short-Term Memory (LSTM) network analyzes the temporal evolution of the residual signals to detect and localize fault-related events, after which the corresponding signal segments are classified using a one-dimensional Convolutional Neural Network (1D-CNN). The diagnostic decision is subsequently used by the FTC stage to select either measured or STO-estimated signals. On the continuous blind C1 detection evaluation, the LSTM detector achieved 99.98% accuracy, 99.97% precision, 100% recall, and a 99.99% F1-score. On the fault-containing candidate segments generated by the LSTM-guided procedure, the CNN achieved an Exact Match of 80.10% and a Macro-F1 of 93.36%. However, grouped three-fold cross-validation yielded an Exact Match accuracy of 63.33 ± 16.20%, indicating variability in performance across folds. The LSTM-guided segmentation procedure reduced the number of candidate segments forwarded to the CNN by 67.07% relative to the fixed-window procedure under the evaluated C1 condition. The STO-assisted FTC stage reduced the RMSE relative to the fault-injected measurements by approximately 98.6–100% under the investigated noise-free simulation conditions. Additional simulation-based analyses examined measurement noise, parameter variation, fault severity, and temporal-sequence settings. Overall, the results indicate the feasibility of using STO-generated residual signals for hierarchical data-driven fault diagnosis and subsequent fault-tolerant signal substitution under the investigated simulation conditions. All results were obtained in simulation; experimental and hardware-in-the-loop validation remain future work.