A novel feature extraction method is developed for impact localisation based on Artificial Neural Networks (ANNs) in sensorized composite structures subjected to environmental and operational conditions. Impact induced lamb waves are investigated for the first time for different impact scenarios (angle, mass and energy) on flat and curved plates under environmental (temperature range) and operational (vibration) conditions. The Time of Arrival (TOA) is significantly influenced by these conditions hence complicating the impact localisation. To overcome this complication, a novel and robust TOA extraction method is proposed. It is based on Normalised Smoothed Envelope Threshold (NSET) coupled with a high pass filter to remove vibration noise prior to TOA extraction. Localisation ANNs were trained with data from a single baseline impact condition and were tested under impacts with varying conditions. It was shown that by using the proposed method for TOA extraction, the trained ANN is able to better predict the location of impacts compared to an ANN trained with data from common TOA extraction methods (detection area 0.99–56.08% of sensing region versus 0.28–1.55% for NSET). The developed method gives consistent accuracy and significantly reduces the required training data, making ANN based impact localisation more feasible for real life application.