Research On Underwater Acoustic Target Recognition Technology Based On Machine Learning

With the continuous promotion and deepening of the strategy of marine power,hydroacoustics has become more and more popular.Hydroacoustic target recognition is the focus of hydroacoustics research,but also the technical difficulties.Analyze the noise information contained in the water signal and use it to judge its category,mainly relying on manual listening discrimination or automatic machine recognition.There are various limitations of manual listening discrimination.So use machine for automatic recognition of water sounds can play a greater role in real life.Therefore,this thesis proposes a machine learning based hydroacoustics target recognition technology to replace the manual recognition.It based on the analysis of ship noise dataset,the water sound feature extraction method and classifier design are studied.Aim to improve the recognition accuracy of the classifier,the overall recognition result accuracy reached 95.36%.The specific research contents can be concluded as follows.(1)After analyzing the real water sound data,we found that there are high-frequency attenuation and loudness change in the process of sound propagation.So we adopt preemphasis and amplitude regularization to process the original noise signal,and increase the sample data by adding windows in frames to reduce the possibility of inaccurate model target identification due to insufficient data.Inspired by the data enhancement method,this thesis proposes a data enhancement method based on the time-frequency signal information,which expands data size and adds noise data to make the classifier model more robust.(2)A new feature extraction method is proposed for the pre-processed hydroacoustic data by simulating the reception of sound by the human ear auditory system.The hydroacoustic signal sequences are stitched together with the log-Meier spectral parameters before and after differencing for feature fusion.The features contain the original static information in the time and frequency domains of the hydroacoustic signal,and the dynamic features containing the whole frame information from the differencing operation.The features contain the original static information in the time and frequency domains of the hydroacoustic signal,and the dynamic features containing the before and after frame information from the differencing operation.The combination of dynamic and static features is used to improve the Mel frequency cepstral coefficients feature,which has performed well.This feature is used to express the hydroacoustic signal more comprehensively and improve the performance of the hydroacoustic target recognition model.It is proved that this feature extraction method can effectively improve the recognition accuracy,which is 20 percentage points higher than the original feature accuracy and 14 percentage points higher than the traditional Meier sign before the improvement.(3)In this thesis,hydroacoustic target recognition model based on machine learning is proposed.An attentional convolution module and a memory module based on transfer learning are incorporated into the model.The model can reasonably allocate the resources of each convolutional channel so that there is a dependency relationship between channels.Recognition accuracy and model robustness are improved by this method.The model accuracy is improved by 1 percentage point by using the transfer learning approach.Memory module allows the model to fully learn the information contained in the temporal dimension,improving the recognition accuracy by 3 percentage points.The final accuracy exceeds 92% in all categories in the actual hydroacoustic target recognition task.