Convolutional feature extraction for process monitoring using ultrasonic sensors

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