Research On Acoustic Detection And Recognition Method Of Small-Rotor Drone

At present,with the continuous maturity of the technology of the small-rotor drone,the market of the small-rotor drone is booming.On the one hand,small-rotor drones have improved production efficiency in agricultural production,firefighting,aerial photography and so on.On the other hand,the small-rotor drone with low flying heights,slow flight speeds and small size is easily used by terrorists for terrorist attacks,which is likely to cause the social panic.Therefore,the regulation of the small-rotor drone is an important guarantee for low-altitude security and this is also a powerful measure to promote the formal development of the drone market.In view of this problem,two points of consensus have been reached.First,the control of drones is divided into three steps—detection,recognition,and regulation,furthermore detection and recognition are precondition for regulation.Second,a single technology cannot effectively solve the problem,and a multi-technology integration method can effectively control the drone.Based on these two points of consensus,after comparing the existing technologies,it is considered that the acoustic detection and recognition method that used the noise of the small-rotor drone does not interfere with other technologies,so the method is an important part of multi-technology fusion for controlling the drone in the future.Due to the ability of the microphone array to enhance the signal,the thesis carried out research on the acoustic detection and recognition of the small-rotor drone based on a microphone array.The details are as follows:(1)The drone noise signal was analyzed by using the frequent features in speech recognition,such as the sound pressure level,the short-term average zero-crossing rate and so on.After analysis,it is considered that there is a difference between the drone noise signal and the background noise in the frequency band of 2000~7000Hz.(2)In view of the frequency of the drone noise signal mainly concentrated in the frequency range of 2000~7000Hz.,Mel-frequency cepstral coefficients(MFCC),widely used as a feature in speech recognition,are improved in order to characterize the drone noise signal better.Since the band-pass filter in the Mel filter banks is densely distributed in the low frequency band and sparsely distributed in the high frequency band,the MFCC is widely used in speech recognition.However,the frequency of the speech signal is significantly different from the frequency of the noise signal generated by the small-rotor drone.Therefore,the MFCC cannot adequately characterize the drone noise signal.So an improved MFCC is proposed in this thesis.Since the transformation relationship between the frequency and frequency of the Mel is changed,the distribution of the Mel filter banks is changed.The band-pass filters in the Mel filter banks are densely distributed in the mid-high frequency band and sparsely distributed in the low and high frequency bands.(3)In practical applications,different microphones on the microphone array will have different recognition results for the same sound source.After analysis,it is found that the distance from the same sound source to each microphone on the microphone array is different,lead to different signals received by each microphone.Therefore,the probability of misjudgment is increased.In order to reduce the occurrence of misjudgment,in this thesis,a weighted recognition detection method based on microphone array and a weighted recognition detection method based on neural network are proposed to determine whether there is a drone.Both of these methods can improve the recognition stability of the microphone array system and reduce the probability of misjudgment.(4)The noise of the small-rotor drone during flying was captured,and the sample dataset containing the training set,test set,and validation set is constructed.In the experiment,the control variable method was used.First,the combined features of the MFCC(including the original waveform,the Log-Mel spectrum and MFCC)and the improved MFCC(including the original waveform,the improved Log-Mel spectrum,and the improved MFCC)are used for feature extraction,respectively.It is found that the accuracy of using the improved MFCC feature extraction method on the test set and the verification set is 2 to 3% higher than that of the MFCC.It is proved that the improved MFCC feature extraction method can improve the recognition accuracy.Second,the improved MFCC and MFCC,the improved Log-Mel spectrum and the Log-Mel spectrum are respectively used as feature extraction methods.Based on the accuracy criteria on the test set,it is found that the accuracy is the highest when using the improved Log-Mel spectrum.Finally,through experiment,it is determined that the improved Log-Mel spectrum is extracted as the final feature extraction scheme.The number of band-pass filters in the Mel filter bank is set to 64 and the frame length is set to 200 ms.