(344c) Application of Deep Neural Networks for Artifact Removal from Sensor Data

Chemical processes inherently contain elements of stochastic noise and disturbances that affect their normal operation. Extensive research efforts have been devoted to filtering noises and attenuating disturbances to improve the process system operation [1]. Artifacts, however, represents a challenge in process operations that necessitates further investigation. Artifacts are considered as disturbances in the measured signal not originating from the process itself. Artifacts may arise due to the nature of the measurement devices or the susceptibility of the measurements to ambient or surrounding conditions. In contrast to the process noise, artifacts can have power spectral densities and characteristics similar to the signal being measured, thus confounding the artifacts and the signal. Biochemical and biomedical systems utilize various sensors and measurement technologies that are prone to artifacts. The effective detection and removal of artifacts from measured signals are important as they can lead to significant interpretive errors [2]. The problem of artifact removal is challenging because artifacts in the signals can arise from several unknown sources. Furthermore, it is not obvious whether fluctuations in the signal are due to underlying variations or artifacts. Models relating artifacts to the observed effects in the output measurements are also unavailable as the origins and the effects of the artifacts vary substantially.

Approaches for artifact removal have focused on passive and active filtering techniques. Passive techniques isolate the measurement devices from external factors, yet it may not be always feasible or practical based on the process design and operation setup. Active filtering relies on the use of a reference signal to adaptively deconstruct the measured signal into components associated either with the artifacts or the underlying signal. Adaptive filtering techniques have been effective in adaptive noise cancellation applications where the on-line estimation of the time-varying correlations between the reference signal and the artifacts is possible. Nevertheless, numerous applications do not have a highly informative and conveniently obtainable reference signal. Recent advances in signal processing and deep learning can further advance the tools available for artifact removal, particularly in highly susceptible and sensitive applications.

One application where artifacts are prevalent and pronounced is the measurement of heart rate using photoplethysmography (PPG), a non-invasive and low-cost optical technique used to measure the heart rate variation. The PPG technique measures the rate of heart beats by detecting changes in the backscattered light corresponding to volumetric changes in blood in peripheral circulation. The PPG measurements are highly susceptible to disruption from artifacts caused by motion and other noise sources such as ambient light interference and skin condition. Motion is a major challenge in obtaining accurate measurements as constant movements result in poor contact between the measurement skin surface and the photo sensor [3]. Motivated by the above considerations, a framework for artifact removal is developed with application to the heart rate estimation problem. The PPG signal is first divided into consecutive overlapped time-windows to facilitate batch-wise processing and a bandpass filter used to reject undesired variations. Wavelet approximation and orthogonal signal reconstruction are then applied to reject the artifacts and noise in the signal. A feedforward Neural Network has been utilized to estimate heart beats for a data set of 5 minutes experiment. The heart rate is estimated every two seconds. The comparison between estimated heart beats versus the actual values proves the fact that regardless of the type of physical activity, especially during running section which PPG signal is corrupted with intense motion artifact, the Deep Neural Network model is capable of the tracking the actual heart rate values.

References:

[1] Downs, J. J., and Vogel, E. F., “A Plant-wide Industrial Process Control Problem, ”Computers & Chemical Engineering, vol.17, pp. 245–255 (1993).

[2] Roy, , Kumar, V. R., Kulkarni, B. D., Sanderson, J., Rhodes, M., and vander Stappen, M., “Simple denoising algorithm using wavelet transform,” AIChE Journal, vol. 45, 1999, pp. 2461–2466.

[3] Zhang, , Pi, Z., and Liu, B., “TROIKA: A general framework for heart rate monitoring using wrist-type photoplethysmographic signals during intensive physical exercise,” IEEE Transactions on Biomedical Engineering, vol. 62, pp. 522–531(2015).