Research on Graphene Thermoacoustic Speakers Based on a Sparse Autoencoder-Back Propagation Neural Network Model

To mitigate the complexities of traditional theoretical models for graphene thermoacoustic speakers, which are hampered by multifarious influencing factors that are challenging to unify and constrained by discrete data, leading to imprecise characterizations, this study delves into a deep learning model for graphene thermoacoustic speakers. Initially, nine parameters pertinent to graphene thermoacoustic speakers are identified as initial input features: the density, thermal conductivity, and specific heat capacity of both the film and substrate, the root mean square (RMS) value of the input voltage, frequency, and the measured distance. A deep learning model is then established for predicting the sound pressure level (SPL), serving as the output feature. Subsequently, a sparse autoencoder-back propagation (SA-BP) neural network model is constructed by integrating the encoding network of a sparse autoencoder with a fully connected BP neural network. Optimization is achieved when the encoder network of the SA-BP model comprises three hidden layers with 11 neurons each, reducing the nonlinear feature dimensionality to five, and the BP neural network comprises seven hidden layers with 18 neurons. Finally, the predicted SPL is computed using the SA-BP neural network model on a test set. Comparison with a BP neural network reveals that while the latter exhibits a maximum relative error of 6.5%, the SA-BP model achieves a notably lower maximum relative error of 1.5%. The SA-BP model outperforms the BP neural network in both feature extraction capability and SPL prediction accuracy, particularly demonstrating enhanced adaptability at lower voltage levels. This research offers a highly accurate solution for precise feature extraction and output prediction in a wide range of sensor applications.