loading page

A novel neural network-based data acquisition system targeting high-speed electrical tomography systems
  • +2
  • Varun Kumar Tiwari ,
  • Mahmoud Meribout ,
  • Khalid Alhammadi ,
  • Naji Al Sayari ,
  • Lyes Khezzar
Varun Kumar Tiwari
Khalifa University, Khalifa University

Corresponding Author:[email protected]

Author Profile
Mahmoud Meribout
Author Profile
Khalid Alhammadi
Author Profile
Naji Al Sayari
Author Profile
Lyes Khezzar
Author Profile

Abstract

In electrical impedance tomography (EIT), high signal-to-noise ratio (SNR) and fast data acquisition are simultaneously required in several applications. However, these two targets are usually compromising and require careful hardware design for the data acquisition (DAQ) system. In this paper, the design of a quick DAQ for an EIT system is suggested to accommodate a large number of electrodes (i.e., 32 electrodes or more). The solution consists of using a very high-speed analog-to-digital converter (ADC) to digitize the output analog signals corresponding to all possible pairs of electrodes within the same cycle in a time-multiplexed manner. Feature extraction is then performed for each channel in real-time within the actual electric current excitation cycle to proceed with the image reconstruction. Various artificial neural network (ANN) models with customized loss functions were assessed, and the optimal model selection approach using the grid search technique is presented. The feature considered in this paper was the peak-to-peak signal amplitude for each channel. Contrary to other previous hardware accelerators, the suggested apparatus does not need to use a multi-frequency current source. Thus, the data acquisition system is simpler since it does not require the use of high-quality, narrow-band pass-band filters for different frequencies. This gives the opportunity for electrical tomography systems to operate at a very high throughput that can easily exceed 2,800 frames per second for a 50 kHz excitation signal using 32 electrodes, for instance. An extensive experimental testing work using the suggested approach reveals that peak estimation can be achieved with an accuracy of more than 98% even for signals with 40 dB SNRs.