Zotero Chunk RAG

Error between the custom impedance analyser (CIA) and the theoretical impedance of
the control DUT over the range of 1 Hz–100 kHz (magnitude) and 1 Hz–10 kHz (phase),
quantified by the root-mean-square error normalised using the mean of the theoretical
impedance of the control DUT.
ZA
ZB
Data set
|𝑍| (%)
𝜙(%)
|𝑍| (%)
𝜙(%)
1
4.7
2.7
5.7
5.3
2
5.3
1.6
5.6
1.5
3
3.8
3.4
4.6
4.2
Mean (SD)
4.6 (0.7)
2.6 (0.9)
5.3 (0.6)
3.6 (2.0)
impedance compensation. As the residual error was a function of both
magnitude and phase, the frequency range was limited to that of the
phase data (1 Hz–9.4 kHz). The NLLS regression was implemented using
the interior point algorithm in MATLAB (R2019b, MathWorks). The
bounds of the NLLS regression was set to physical limits desired for the
compensation network. The resistive components were bound between
1 Ω and 10 MΩ, and the capacitive elements were bound between 1 nF
and 10 μ𝐹. To reduce the possibility of a solution being based on a
local minima, the NLLS regression was run 100 times. The first 50
initial value vectors were formed using the lower and upper bounds
in a logarithmically ascending order. The same method was used to
produce the remaining 50 initial value sets; however, 𝑅′
𝐵and 𝐶′
𝐵had
a logarithmically descending order. Initially the NLLS regression was
run using coarse exit conditions to decrease the processing time of the
solver. The initial conditions that resulted in the least error, were run
again with finer exit criteria. A RMSE between the two compensated
impedance data sets was calculated over the phase data frequency span,
permitting a quantitative simulated impedance imbalance reduction.
4. Results
4.1. Proxy Device Under Test (DUT)
The accuracy of the CIA using the proxy DUT is outlined in Table 2,
resulting in a mean error and standard deviation of (4.6 ± 0.7) % for
the magnitude and (2.6 ± 0.9) % for the phase of electrode A; and (5.3
Error between the custom impedance analyser (CIA) and the Keysight impedance
analyser (KIA) measured combined impedance of electrodes A and B on human subjects
over the range of 1 Hz–100 kHz (magnitude) and 1 Hz–9.4 kHz (phase), quantified by
the root-mean-square error normalised using the mean of the KIA.
Ag/AgClNSP
Ag/AgClSP
Subject
|𝑍| (%)
𝜙(%)
|𝑍| (%)
𝜙(%)
1
3.3
0.9
1.7
14.9
2
7.6
3.7
6.4
11.3
3
14.4
2.7
3.4
4.4
4
1.8
0.8
1.2
24.0
5
16.9
1.7
1.6
19.0
6
9.4
0.6
3.2
4.2
7
4.4
1.3
17.4
5.5
8
64.8
12.8
11.3
9.8
9
1.5
0.9
8.6
4.2
10
0.8
1.1
2.6
8.9
Mean (SD)
12.5 (19.2)
2.7 (3.7)
5.7 (5.3)
10.6 (6.8)
electrode configuration), and the most balanced data set (Ag/AgClSP
electrode configuration); Subject 7 had the largest recording error
(Ag/AgClNSP electrode configuration, but as the recording error is
below 20 Hz, the validation error is still low), and the least balanced
data set at 50 Hz (Ag-SP electrode configuration); Subject 10 had the
least balanced phase data set (Ag/AgClNSP electrode configuration).
4.3. Electrode–skin impedance compensation
The mean and standard deviation of the electrode–skin impedance
imbalance and the reduction in impedance imbalance after applying
the compensation network are outlined in Table 5. The mean and
standard deviation of the compensation network component values are
outlined in Table 6. Compensation data for Subjects 4, 7 and 10 for
the Ag/AgClNSP electrode configuration are shown in Fig. 7. This elec-
trode configuration was chosen as it was the most imbalanced. These
results highlight the potential for a compensation network to reduce
the impedance imbalance of the electrode–skin interface, therefore,
reducing differential-mode interference due to the potential divider
effect.