Loudspeaker parameterization using physics-informed transfer learning

A loudspeaker’s electrical impedance can be modeled as an equivalent electrical circuit that is characterized by its Thiele–Small parameters. The Thiele–Small parameters are related to the electrical, mechanical, and acoustical properties of the loudspeaker. Monitoring changes in these parameters throughout a loudspeaker’s life is expected to provide health-significantinsights to the cause of a given failure. To test this, a series of loudspeakers were run to failure while continually measuring the total electrical impedance across its electric terminals. The Thiele–Small parameters for each loudspeaker were measured prior to each run-to-failure experiment for reference to predicted values. A physics-informed neural network was pretrained as a source model to predict the Thiele–Small parameters from a loudspeaker equivalent circuit model, which was used to generate the synthetic training data. Transfer learning was accomplished by training the source model with a small subset of the data obtained from the run-to-failure experiments to then monitor the Thiele–Small parameters throughout the life of a loudspeaker. While not exact predictions, the trend in the monitored parameters provides physical insight into the cause of each failure, making it easier to predict when failure may occur, which is applicable to other health monitoring applications.