Research On Underwater Acoustic Signal Recognition Technology Based On Machine Learning

With the development of military technology,the underwater detection are getting higher and stronger,Because the identification technology is the key technology of detection means.Therefore,it is necessary to combine the characteristics of the original underwater acoustic pulse signal to study and have better effect.Universal recognition technology to adapt to the development of the times.Firstly,the conventional pulse modulation signals and spread spectrum communication signals in the field of underwater acoustic engineering are studied.A total of 8 types of underwater acoustic modulation pulse signals are simulated.The time-frequency analysis method is used as a non-stationary signal analysis tool to transform the one-dimensional time domain signal into time.Combined with the frequency domain,the paper studies the short-time Fourier transform and adaptive optimal kernel time-frequency distribution(AOK)method,and compares the performance of other classical time-frequency analysis methods.The time-frequency distribution of different signals has very obvious shape differences,and the shape recognition can be used to perform signal recognition classification very quickly.The basic theory of pattern recognition is studied,and the effective feature information of time-frequency images of different signals is extracted.The invariant moment features and HOG image features of time-frequency images of common underwater acoustic pulse signals are extracted.In order to extract the invariant moment features of images more effectively,the paper adopts common image preprocessing methods,when the image features of HOG high-dimensional are used.In order to solve the computational problems caused by high-dimensional features,the PCA dimensionality reduction technique was studied,and then a variety of traditional machine learning models based on statistical theory were studied.Then the training was verified to include naive Bayes,support vector machines,Four models of random forest and BP neural network were used to obtain the average recognition accuracy rate by cross-validation method.In this thesis,using the deep learning theory,the deep learning model is designed to automatically extract the time-frequency image of the signal from the original signal time-frequency image sample set,which eliminates the steps of artificial design feature extraction in the machine learning recognition method.Three types of deep learning models,including deep neural networks,convolutional neural networks,and circulatory neural networks,were compared using the time-frequency distribution gray maps of different typesof signals,and the classification and recognition results of the models were compared.Combining the self-learning ability of convolutional neural network and the excellent classification performance of support vector machine,a CNN model is trained as the feature extractor to obtain the feature data of the time-frequency distribution map,and then the feature data is performed by using the support vector machine model.Learn.At the same time,the model with simple input and network structure is studied.A one-dimensional frequency domain sequence using signals is proposed for the signal recognition classification problem,and then a one-dimensional convolutional neural network is used for signal classification and recognition.The simulation and measured data show that the recognition performance of multi-class underwater acoustic pulse signals based on time-frequency analysis signal processing technology and simple frequency domain sequence processing technology combined with deep learning theory is higher than traditional pattern recognition methods.Therefore,the deep learning method is efficient and feasible in the classification and recognition of underwater acoustic pulse signals.