Abstract:This work introduces a wavelet neural network to learn a filter-bank specialized to fit non-stationary signals and improve interpretability and performance for digital signal processing. The network uses a wavelet transform as the first layer of a neural network where the convolution is a parameterized function of the complex Morlet wavelet. Experimental results, on both simplified data and atmospheric gravity waves, show the network is quick to converge, generalizes well on noisy data, and outperforms standard network architectures.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to effectively process non - stationary signals through neural networks and improve interpretability and performance in digital signal processing. Specifically, the author proposes a neural network based on wavelet transform (WaveNet), aiming to learn a filter bank specifically for non - stationary signals, thereby enhancing the understanding of signal characteristics while maintaining high accuracy. ### Main contributions: 1. **Propose WaveNet**: Combine the advantages of wavelet transform and neural networks. By using the parameterized complex Morlet wavelet as part of the convolutional layer, the network can adaptively learn the frequency characteristics of the signal. 2. **Optimize the training process**: Optimize the learning of wavelet parameters by setting different learning rates to ensure the stability and performance of training. 3. **Experimental verification**: Experiments were carried out on the simplified data set and the atmospheric gravity wave data set, demonstrating the advantages of WaveNet in terms of fast convergence, generalization ability and classification accuracy. ### Specific problem - solving methods: - **Wavelet transform**: Use wavelet transform as the first layer of the neural network, where the convolution operation is a parameterized complex Morlet wavelet function. This can capture the time - frequency information of the signal. - **Parameterized learning**: The frequency \( f \) and width \( w \) of the wavelet are trainable parameters and are optimized through back - propagation. - **Batch normalization**: Apply batch normalization after the transformation layer to regularize the data and stabilize the training process. - **Multilayer structure**: The subsequent fully - connected layers use the tanh non - linear activation function, and finally classification is carried out through the linear output layer. ### Experimental results: - **Simplified data set**: WaveNet can converge quickly, accurately identify events with different frequencies, and finally achieve a 100% test accuracy rate. - **Atmospheric gravity wave data set**: WaveNet performs excellently in the task of automatically detecting atmospheric gravity waves in satellite images, reaching a maximum test accuracy rate of 88.32%. ### Conclusion: WaveNet provides an effective method for processing non - stationary signals. By learning a specific filter bank, it not only improves the classification performance but also enhances the interpretability of the model. Future work can further explore different wavelet functions and other network architectures to better adapt to different types of data.

Trainable Wavelet Neural Network for Non-Stationary Signals

Learning Sparse Wavelet Representations

Complex-Valued Wavelet Network

Robust Time Series Denoising with Learnable Wavelet Packet Transform

AdaWaveNet: Adaptive Wavelet Network for Time Series Analysis

Deep Learning in the Wavelet Domain

Wavelet Networks: Scale-Translation Equivariant Learning From Raw Time-Series

Wavelet-Driven Spatiotemporal Predictive Learning: Bridging Frequency and Time Variations

Orthogonal Wavelet Networks for Nonlinear Dynamic System Identification

Neural network compression via learnable wavelet transforms

Distributed wavelet neural networks

Wavelet J-Net: A Frequency Perspective on Convolutional Neural Networks

T-WaveNet: Tree-Structured Wavelet Neural Network for Sensor-Based Time Series Analysis

Detection of Weak Targets with Wavelet and Neural Network

Wavelet neural operator: a neural operator for parametric partial differential equations

A Learnable and Explainable Wavelet Neural Network for EEG Artifacts Detection and Classification

Wavelet decomposition and neural networks: a potent combination for short term wind speed and power forecasting

Deep Adaptive Wavelet Network

Fully learnable deep wavelet transform for unsupervised monitoring of high-frequency time series

Learnable wavelet neural networks for cosmological inference

Wavelet Dynamic Selection Network for Inertial Sensor Signal Enhancement