Improved Method for Analyzing Acoustic Signals of Water Environment Based on Convolutional Neural Network SOP

Abstract

The analysis of acoustic signals in marine environments poses numerous challenges, including handling large volumes of data and adapting to rapidly changing environmental conditions. Passive signal reception is often characterized by a low signal-to-noise ratio (SNR), complicating signal processing. Addressing these challenges requires an efficient and versatile approach, which neural networks can provide. Among the various neural network models for acoustic signal analysis, Convolutional Neural Networks (CNNs) are recognized as some of the most effective due to their ability to handle complex data structures and extract meaningful features. Therefore, it is reasonable to explore an efficient CNN-based method that can be modified to enhance the quality of acoustic signal analysis in aquatic environments. An improved approach has been developed, utilizing the CNN SOP network. CQT features were employed as input characteristics for classification. The original two convolutional layers were replaced with a multi-scale convolution using kernels of different sizes, enabling the extraction of both global and local features of the target object, thereby enhancing the model's ability to process diverse signal properties. The extracted features were processed using second-order pooling and then passed to the second-order pooling (SOP) layer. This layer enhances the identification of temporal correlations. The SOP layer outputs feature vectors, which are normalized element-wise using square root and l2 normalization. The normalized data are then fed into a fully connected layer with batch normalization and the ReLU activation function. Subsequently, the data are passed to another fully connected layer with a Softmax activation function, which performs the final classification. To evaluate the performance of the neural network, three datasets were used: two based on real-world underwater objects and one artificial dataset. Each dataset was further augmented with background noise to produce samples with a low signal-to-noise ratio (SNR). In all cases, the proposed improved method demonstrated higher classification accuracy compared to the original method, showcasing its effectiveness in handling noisy data and improving signal analysis.