Adaptive pump fault detection for vanadium redox flow batteries based on CNN-BiGRU and transfer learning

Vanadium redox flow batteries (VRBs) are vulnerable to pump malfunctions, which can cause severe problems like overcharging, overdischarging, and irreversible damage to the battery stack. Effective real-time fault detection methods for VRBs are essential but underdeveloped. This study proposes a data-driven approach for VRB pump fault detection using a convolutional neural network combined with bi-directional gated recurrent units to predict voltage based on measurable current, flow rate, and state of charge without relying on complex electrochemical models. To address the impact of battery aging and system variability, which are often overlooked in the literature, an offline transfer learning strategy with model tail fine-tuning is employed to adapt the parameters of the pre-trained model using a small amount of newly collected data. Pump faults are identified by comparing voltage prediction errors against a threshold derived from the 3σ criterion. Experimental results across six fault scenarios encompassing three distinct pump fault modes under two operational conditions demonstrate that the proposed method achieves timely and accurate fault detection, significantly improving the reliability of VRB systems.