Research on noise prediction of multiple sound sources in vehicles based on FIM-LSTM

Abstract

This study addresses the challenge of predicting in-vehicle noise under multi-source excitation conditions by proposing a novel prediction model—FIM-LSTM—which integrates an improved Fisher Information Matrix (FIM) with a Long Short-Term Memory (LSTM) network. The model uses vehicle body vibration signals as reference inputs. To overcome the limitations of traditional approaches—such as arbitrary selection of reference signals and high computational cost—we introduce an enhanced FIM-based signal selection method. By incorporating regularization parameters and a frequency-band focusing strategy, the method efficiently identifies the optimal set of reference signals within the 20–600 Hz target band, mitigating the ill-conditioning issue of the FIM and significantly reducing computational complexity. The FIM-LSTM prediction model, constructed using the optimized signal subset, demonstrates strong performance in predicting background noise. Experimental results show that, compared to the baseline LSTM model using all 28 reference signals, the FIM-LSTM model achieves comparable accuracy with only 20 selected signals, while reducing the number of model parameters by 8.49% and computational load by 7.99%. Moreover, when compared to a FIM-Convolutional Neural Network (CNN) model using the same selected signals, the FIM-LSTM model exhibits notably superior prediction accuracy in both the time and frequency domains. Overall, the proposed method offers a computationally efficient and accurate solution for in-cabin noise prediction, providing valuable support for the development of active noise control (ANC) systems and holding strong potential for practical engineering applications.