Recurrent normalizing flow-based monitoring framework for cutter seal temperature sensors in earth pressure balance shield tunneling: A sensor validation approach

• Data-driven sensor validation promotes EPB safety by detecting faulty temperature sensors. • Hybrid LSTM-AE and RAF model captures temporal patterns of EPB operational data. • This method achieved a 100 % fault detection rate and 96 % overall accuracy. • This method achieved low errors in reconstruction, RMSE: 0.35–0.45 °C, MAPE: 0.81–1.13 %. • This method enables proactive and accurate maintenance through accurate sensor correction. Reliable monitoring of cutter seal temperature is essential for the safe and efficient operation of earth pressure balance (EPB) machines. Faulty temperature sensors can mislead operators, potentially leading to seal degradation, slurry or water ingress, and costly downtime. This study introduces a data-driven monitoring framework that automatically detects and reconstructs faulty temperature sensor readings to ensure robust information during tunneling. The proposed method combines a long short-term memory-based autoencoder with a recurrent autoregressive flow, namely RAF-LSTM, to learn normal temperature patterns over time and model the probability of those patterns to distinguish normal from abnormal behavior. The proposed method was assessed using operational data from an EPB project, where synthetic faults representing realistic sensor failures have been introduced. Results show a fault detection accuracy of 96 % with 100 % recall and reconstruction errors below 0.6 °C. These outcomes demonstrate that the proposed framework can reliably identify and correct faulty sensor signals, reducing false alarms and supporting proactive maintenance strategies for EPB machine operations.