Design Of Fast Gunshot Type Recognition System Based On Knowledge Distillation

In regional conflicts,timely and accurate acquisition of enemy firearm types is one of the important contents of battlefield situational awareness and a key basis for commander’s tactical deployment.The current Russian-Ukrainian conflict shows that the most intense fighting often occurs in narrow urban streets.In urban warfare,optical and electromagnetic sensing devices cannot accurately identify enemy firepower composition due to special reasons such as obstruction and shielding.However,sound waves,due to their independent propagation characteristics,are not affected by conventional interference and camouflage,making acoustic feature analysis one of the most effective means of obtaining enemy information in narrow street fighting.However,the unique factors of gun sound waves,such as mixing and background noise,reduce the signal-to-noise ratio of the signal,confuse acoustic features,and reduce the accuracy of firearm identification.To address these issues,this dissertation proposes a gun sound type rapid identification system based on knowledge distillation.Through the identification algorithm and corresponding hardware deployment platform,the system achieves end-to-end output of firearm types from gun sound signals.First,an adaptive time-frequency feature extraction method is used to extract low-dimensional features of gun sound signals under strong reverberation,reducing the interference of mixed noise in the environment.Secondly,a deep neural network model based on transfer learning is constructed to extract high-dimensional acoustic features under strong interference conditions,ensuring the accuracy of identification under conditions where gun sound features are not obvious.On this basis,the model’s feature extraction ability is transferred to a lightweight network model through knowledge distillation,reducing the delay of the identification method and improving the model’s inference time and response speed.Thirdly,a hardware platform with a deep neural network model is built to achieve rapid and accurate identification of firearm types in portable devices.Finally,field experiments were conducted indoors and outdoors.Through field experiments,the system can achieve gun sound acquisition and identification functions within 3.32 seconds,with a recognition accuracy of 83.2% in indoor signal mixing conditions,and a recognition accuracy of over 81.7% within 80 m of the shooting point in open outdoor areas.The recognition accuracy outside 80 m is 68.2%.Even in the case of losing important Mach wave features,the system can still achieve an accuracy of over 70%.The experiments prove that this design can quickly and accurately identify gun sound signals in close-range urban and street areas,and has certain theoretical significance and engineering application value.