Research on evaluation model for vehicle interior sound quality based on an optimized BiLSTM using genetic algorithm.

• An optimized bidirectional long short-term memory (BiLSTM) using genetic algorithm (GA), called GA-BiLSTM model, is proposed to evaluate vehicle interior sound quality. • Raw noise is converted to log-mel spectrum containing time–frequency information in line with human auditory perception. It is more suitable as the input feature representation than waveform. • GA-BiLSTM model perform better than BiLSTM model without GA. At the training, validation and testing stages, the former has higher evaluation accuracy and lower loss than the latter. • The evaluation obtained by the proposed model is in good agreement with human subjective evaluation. Interior sound quality strongly affects passengers' physiological and psychological perceptions. Therefore, it is important to evaluate vehicle interior sound quality. Compared with assessing by humans, the artificial intelligent-based evaluation model can acquire an evaluation efficiently. However, this type of models determines initial learnable parameters at random before training, which is easy to cause final trained model to trap in local optima. This paper proposes an evaluation model based on an optimized bidirectional long short-term memory using genetic algorithm. Firstly, interior noise measurement and subjective evaluation are completed. Secondly, to obtain the time–frequency information in line with human auditory perception, log-mel spectrum is used to preprocess the noise. Thirdly, the evaluation model is constructed, which consists of two bidirectional long short-term memory layers, two fully connected layers and one Softmax output unit. Next, to avoid model trapping in local optima, initial learnable parameters are optimized using genetic algorithm. After optimization, average fitness and best fitness decreased by 6.5136% and 1.4415%, respectively. The training accuracy is 95.79%. The validation accuracy is 93.15%. The testing accuracy is 93.33%. Only two samples are misclassified in the confusion matrix of testing stage. These suggest that genetic algorithm can greatly enhance the model's performance by optimizing initial learnable parameters. The evaluation obtained by the optimized model is very close to human subjective evaluation. [ABSTRACT FROM AUTHOR]