SoundCatcher: Acoustic Emission with Machine Learning for Real Time Monitoring of MDI Shot Weights




A novel tool, the SoundCatcher, has been developed for real time monitoring of the quality attribute metered dose inhaler (MDI) shot weights. The application is based on data collection software triggered by the acoustic signal of the MDI actuation and then applying a second step of interpretation both using machine learning algorithms. It was shown that a correlation exists between the MDI actuation sound profile and shot weights, thus forming the scientific basis for the development of the application. Here, a multivariate calibration model is utilized to capture the actuation sound. Next, the relevant section of sound is subjected to another multivariate calibration with the aim to estimate the quality of each shot and display this in real time. The tool was developed using the Python programming language. The software consists of a module with real time sound spectral creation, real time display of sound (updates every 100 ms), model outputs, and indications of detection. Acoustic data spectral interpretation is achieved by machine learning algorithms, e.g., in the form of time lagged PLS models and PLS-batch models.  Prediction times are ≈2 ms for each 100 ms portion of spectra for the lagged detection model. The principles for developing, optimizing and testing the tool are discussed. The capabilities and performance (for instance how well the signal correlates to varying shot weights), when testing real samples are demonstrated.