Research On Time-frequency Texture Characterization And Recognition Of Acoustic Signals

Acoustic signals are important information carriers.Perceiving environmental information through acoustic signals is one of the important research contents in the field of machine hearing.The acoustic signal recognition technology has the advantages of small equipment size,low hardware cost,and is not restricted by terrain,angle and light during the working process.Therefore,acoustic signal recognition technology has broad development prospects in the fields of security supervision,medical monitoring,ecosystem investigation,and anti-terrorism and anti-riot.This paper focuses on two aspects of acoustic signal time-frequency texture characterization and recognition technology.The main contents of the paper are as follows:This paper studies the principle of audio recognition technology and the current research status at home and abroad,and establishes the overall scheme of acoustic signal recognition technology based on the characteristics of the research object.Firstly,the time-frequency texture features of the acoustic signal are extracted by a filter bank that mimics the auditory characteristics of the human ear through a Mel filter bank and a Gammatone filter bank.For a single time-frequency texture feature cannot fully characterize the evolution of the acoustic signal in the time-frequency domain.Further extracting the first-order difference features of the time-frequency texture features of the acoustic signal in the time domain and the frequency domain,to obtain the change information of the acoustic signal energy in the time-frequency domain.The three features are combined to form a multi-dimensional time-frequency texture feature of the acoustic signal.This multi-dimensional feature can more effectively provide the recognition model with different time-frequency characteristics of different acoustic signals.Secondly,a convolutional neural network model was designed to extract the high-level features of the acoustic signal.According to the different characteristics of the time-frequency texture feature dimension information,a separate convolution method was used to extract the high-level features of different dimensions of the time-frequency texture feature.A high-level feature extraction model applied to the research object in this paper,and a classification model is designed.At the same time,DS evidence theory is used to fuse the identification information of two different time-frequency texture features,which further improves the performance of the acoustic signal recognition model.The acoustic signal recognition model established in this paper achieved 97.2% and 87.1% recognition rates in the ESC-10 and ESC-50 data sets,respectively.Finally,based on the acoustic signal recognition model proposed in this paper,a real-time acoustic signal recognition system in the real environment is established.The system mainly includes two parts: hardware system and software system.The software system can be divided into three operating interfaces: simulation mode,learning mode and online mode.According to the difficulty of acoustic signal identification in the real environment,active calibration technology is introduced to effectively avoid some problemsbrought by traditional noise reduction algorithms.The acoustic signal recognition system constructed in this paper realizes the real-time recognition of footsteps,guns,helicopters,and human voices in a real environment.When the signal-to-noise ratio is above 0dB,the recognition rate can reach 91.8%,and the average recognition time is 2.4 seconds.