Convolutional neural networks for accurate and robust noninvasive pressure measurements in sealed systems

Abstract

Accurate noninvasive pressure measurement in sealed systems is essential for many applications in process control, chemical transport, and chemical safety assessment. This paper introduces a pioneering approach to enhance pressure estimation using Acoustic Resonance Spectroscopy (ARS), a promising technique that analyzes the system's exterior vibrations to infer internal pressure. Despite its potential, extracting precise measurements from noisy ARS data remains challenging. We present a novel method employing a convolutional neural network (CNN) to improve the accuracy and generalizability of the pressure prediction. The CNN performance is compared against benchmark ARS processing methods such as k-nearest neighbors (kNN) and a peak tracking method. The models were trained and tested using labeled data from laboratory-controlled tests on vacuum vessels subjected to varying internal pressures. Evaluation under various signal processing and field measurement scenarios was conducted to assess the accuracy and generalizability of the proposed models. The CNN outperformed other models by significant margins, achieving a mean absolute error (MAE) of approximately 15 Torr, compared to approximately 30 Torr for kNN and 60 Torr for peak tracking. Furthermore, we tested the generalizability of the models by introducing synthetic data augmentations such as spectrum shift and dropout, which approximate real-world measurement errors in ARS measurements. We found that the CNN maintained high accuracy under shifting and dropout data scenarios, showcasing its robustness, while the other models showed larger increases in error. This suggests CNN as a strong candidate for noninvasive pressure measurements, as well as a wide range of other applications detectible via spectral measurements, such as corrosion detection, structural integrity monitoring, or reaction tracking in closed vessels, offering high accuracy and resilience to environmental fluctuations.