Convolutional feature extraction for process monitoring using ultrasonic sensors

Abstract

Ultrasonic sensors are a low-cost and in-line technique and can be combined with machine learning for industrial process monitoring. However, training accurate machine learning models for process monitoring using sensor data is dependant on the feature selection methodology. This paper compares a convolutional feature extraction method to a traditional, coarse feature engineering approach. The convolutional method uses filter weights pre-trained on an auxiliary task to classify ultrasonic waveform dataset membership using previously obtained sensor data. The filter weights are used to extract features from the ultrasonic waveform. Principal component analysis is then applied to produce five principal components to be input into long short-term memory neural networks. The two approaches are compared on fermentation, mixing and cleaning datasets monitored using ultrasonic sensors. Overall, the convolutional feature method produced more informative waveform features than the coarse feature engineering approach, achieving higher model accuracy for datasets requiring substantial waveform information and for 65% of tasks overall. Multi-task learning also improved feature trajectory learning but led to reduced model accuracy for data points far from the classification decision boundaries. This can be overcome by further optimisation of neural network hyperparameters, though at increased model development time. Once trained, the convolutional feature extraction approach is a fast and convenient way of producing high quality features from ultrasonic waveforms using convolutional neural networks with little training data.