Industrial Control Systems (ICS) in water utilities rely on sensors to monitor quality parameters. Sensor drift, a gradual deviation in sensor readings, threatens operational reliability. This paper presents an adaptive Long Short-Term Memory (LSTM) based drift detection method. The proposed approach models normal process behavior with LSTM networks and detects deviations using statistical thresholds. The adaptive method adjusts to the dynamics of ICS without retraining. A case study using the Secure Water Treatment (SWaT) testbed digital twin demonstrates its effectiveness. The proposed method reduces false positives to zero, compared to the 48-per hour found using traditional LSTM methods, and achieves an average detection time of less than 5-s. The adaptive LSTM method enhances drift detection and ensures reliable plant operation, advancing data-driven maintenance strategies in critical infrastructure.