Residual convolutional neural network for continuous identification of aircraft noise

To fully utilize the possibilities given by the aircraft noise monitoring system, it should quickly, automatically and accurately identify whether the limit-exceeding noise event is caused by the aircraft operation. Due to the often delayed access to airport operation logs, the system should operate with minimal or no non-acoustic data. The paper proposes the architecture of an aircraft noise detection method, meeting the above requirements and attempts to assess its effectiveness. Proposed approach involves using the residual convolutional neural network for solving the task. The network operates on 1/3 octave noise input data and determines the similarity of the input sound to the aircraft noise. The accuracy of the proposed method for a single data frame using real-life measurements exceeds 95% for a frame length of at least 30 seconds. Further work is progressing, focusing mainly on improving the quality of training data and refining the hyperparameters of the network.