WO2022146270A1

WO2022146270A1 - An ann based single calibration impedance measurement system for skin impedance range

Info

Publication number: WO2022146270A1
Application number: PCT/TR2020/051504
Authority: WO
Inventors: Mustafa ISTANBULLU; Mutlu AVCI
Original assignee: Cukurova Universitesi Rektorlugu
Priority date: 2020-12-31
Filing date: 2020-12-31
Publication date: 2022-07-07
Also published as: TR202022620A2

Abstract

This invention relates to an artificial neural network-based signal post-processing algorithm to overcome the calibration requirements of the AD5933 integrated circuit, which requires different impedances for different ranges. In the invention, an application specific artificial neural network topology is developed and trained for high precision impedance measurement using a fixed calibration impedance. The measurement system is designed for operating in the range of nominal skin impedance.

Description

AN ANN BASED SINGLE CALIBRATION IMPEDANCE MEASUREMENT SYSTEM FOR SKIN IMPEDANCE RANGE

TECHNICAL FIELD

BACKGROUND

High precision impedance measurement is a widely utilized method from a biological cell analysis to a fuel cell test. The impedance measurement is a simple, useful, and powerful method for the electrical characterization of cells, tissues, and materials.

Since stable and robust high precision impedance measurement in a wide frequency range is a complicated task, traditional instruments are bulky and expensive. In some cases, such as skin cancer, regular tracking of skin impedance may visualize treatment effectivity. Daily high precision impedance measurement gives physiological information about abnormal tissue. A simple, compact, and low-cost impedance measurement system can be used by patients for regular tracking.

Improvements in semiconductor technology and embedded systems enable to produce smaller, cheaper, and portable single-chip impedance measurement devices. High precision, single-chip network analyzer AD5933 is such an impedance measurement device with direct digital synthesis (DDS) and digital signal processing (DSP) abilities. In the literature, several types of research utilize AD5933 integrated circuit (IC) such as impedance cardiograph, bioimpedance measurement, single-cell measurement, cell culture growth monitoring, biosensors, and technical object diagnostics [1], In the studies of Snajdrova et al. [2] and Shlyakhotka et al. [3], impedance cardiography analyzers were developed based on the AD5933 IC to evaluate cardiac output and stroke volume. Both simple structure devices are suitable for continuous measurement and provide acceptable results. Seoane et al. [4] presented a relatively simple analog front end circuit using the AD5933 IC with a four-electrode topology for biomedical applications. Their impedance measurement results exhibited good performance in the aspects of dynamic load range and accuracy. Chen et al. designed a portable impedance measurement system to monitor the growth of L929 cells. The proposed device comprised of a microprocessor, an AD5933 IC and interdigitating microelectrodes to observe cell growth at low cell concentrations. A low cost and miniaturized eight-channel impedance analyzer was developed with ease of use for handheld impedance spectroscopy applications in [5], The device was a pioneer for standalone, portable biomedical applications. Hoja and Lentka [6] presented the telemetric control of an analyzer family, allowing to measure the impedance between 10 Q and 10 GQ in the frequency range of 10 mHz to100 kHz. In this study, the AD5933 IC was used as an impedance converter. AD5933 IC-based bioimpedance measurement systems are widely available in the literature. Margo et al. [7] presented a new extension circuit for the AD5933. The circuit allows four-electrode usage to reduce the artifacts between electrode and skin. In the study of Hui and Ding [8], low- power and portable bioelectric impedance analyzer was presented. Ferreira et al. [9] used AD5933 with analog front-end for the transthoracic electrical bioimpedance measurement recording. Yang et al. [10] proposed a palm bioimpedance spectroscopy system to distinguish different people. Bioimpedance measurement systems focused on cole parameter estimations [11], or device portability [12] [13] are available in the literature. The control of overall systems was done by commonly used microcontrollers such as ARM7. AD5933 IC based impedance measurement also implemented on fruit aging and ripeness detection [14], [15], As mentioned before, the AD5933 impedance converter chip could be useful in many applications. In this sense, AD5933 IC was included in the pH sensor [16], measurement of cholesterol concentration [17], air-water quality monitoring [18], the realization of digital LCR meter [19], eddy current testing [20] and alternating current zero potential circuit [21] studies.

Although there are many successful applications and studies with AD5933, the most significant drawbacks are the nonlinear calibration requirement and measurement error due to calibration effects. Direct digital transfer of the measurement information constitutes an estimation problem related to the analog part of the AD5933 from DDS to the DFT module [22], Thus, a calibration step must be performed before each measurement to acquire the gain factor. The gain factor is used to determine the unknown impedance. The value of the calibration resistor depends on the expected range of the unknown impedance. Theoretically, the closer calibration resistor to test material impedance, the more accurate measurement provided [20], Chen et al. [23] tried to overcome the calibration problem by utilizing a switch array consisting of micro-relays and high precision resistors. Although the measurement system calibration was improved by an adaptive feedback approach, relay and resistor arrays increased the complexity of the system. In the study of Simic [24], a self-calibration system based on employing two analog multiplexers for calibration resistor selection was proposed. The accuracy is observed only in relatively low frequencies due to the limitation of precision resistors used in the calibration of the system. Similarly, Abo Bakr et al. [25] calibrated the measurement unit by using the parallel RC network for different frequency ranges, whereas Breniuc et al. [26] utilized a multiplexer for calibration resistor selection. Measurement errors remain to exist since the approximately valued calibration resistance with the resistance to be measured cannot be used for all measurements in the range.

BRIEF DESCRIPTION OF THE INVENTION

High precision impedance measurements by using AD5933 with a single fixed calibration resistance is obtained with this novel artificial neural (ANN) based measurement system. In the literature, such a study of the AD5933 IC has not been discussed to overcome the calibration problem and increase the measurement accuracy. Another advantage of the proposed system is the high measurement accuracy over a wide frequency range. The invention, a novel signal post-processing algorithm, utilizes real (Re) and imaginary (Im) data obtained from the AD5933 IC. The measurement outputs of the AD5933 in different frequencies are taken as inputs of the neural network. Euclidean distance calculation in exponential form is followed by summation and normalization steps. Finally, the multi-layer perceptron artificial neural network (MLPNN) is executed. After the neural network training, test measurements are implemented. Test results proved the efficiency of the measurement system. LIST OF FIGURES

Figure 1. Flowchart of the training algorithm

Figure 2. Flowchart of data acquisition and impedance measurement of the proposed system

Figure s. Proposed testing platform circuit

Figure 4. The printed circuit board (PCB) of the proposed measurement system

DETAILED DESCRIPTION OF THE INVENTION

The proposed ANN-based solution to estimate high accuracy impedance value includes a unique ANN topology. The topology consists of two main parts. The first part realizes radial basis function-based regression where the second part is a MLPNN with error backpropagation learning. In the first part, the Euclidean distance calculation layer is followed by summation and normalization layers. All real and imaginary measurement data for a single frequency are applied to two particular neurons at the Euclidean distance calculation layer. Also, a single summation and normalization layer neuron is assigned for measurement in each frequency. Euclidean distance calculating neurons with exponential activations also have weight values to relate real and imaginary inputs of each frequency. The least-squares (LS) algorithm is utilized to calculate weight and exponent multiplier terms of this layer. In the second part of the topology, the correlation step takes place in order to realize mapping from inputs to outputs. The backpropagation learning algorithm of the MLPNN is implemented in the second part. Although the stages are connected sequentially, they are trained independently. The target value of each stage is the required impedance output. After the utilization of the first stage, the data are applied to the second stage for better error minimization.

Curve fitting is one of the most popular modeling approaches for a data set. The approach forms a curve or a mathematical function that has the maximum fitting to the data set. The function refers to a mathematical model representing a very close successive change amount with the dataset. Function selection is based on interpreting graphical information of the quantitative data set.

The LS algorithm is one of the most effective methods of curve fitting. It is based on the sum of squares minimization of the error function. The LS algorithm is applicable to any differentiable function. The most widely used functions are polynomial, exponential, Gaussian, and Fourier, etc. In this study, the two-term exponential function model is used in the LS algorithm to fit data as given in equation (1).

where a_i, b_i , c_i and d_i are coefficients of the model whereas _magi is the magnitude of each Re and im data as expressed in equation (2).

Thus, the error function is defined as in equation (3).

The smallest error can be obtained by solving the partial derivative of the error function with respect to each variable. The obtained a,b,c,d coefficients are used for the preprocessing step of the artificial neural network, as shown in Figure 1.

Table 1. Preprocessing schematic of input data

MLPNN, one of the most popular types of ANN, contains an input layer, one or more hidden layers, and an output layer. There is not an exact specific method to specify optimum neuron and hidden layer number of the network. Generally, the number of hidden layers and corresponding neuron numbers are determined by heuristically.

MLPNN, with the error backpropagation learning algorithm, provides the approximation of input-output mappings of multivariate, nonlinear functions. In the backpropagation learning phase, the learning originates from the output neurons. The instant error between the desired output and estimated output is continuously calculated for each iteration and backpropagated through the network. The backpropagation process modifies the weight and bias parameters with a particular learning rate and/or momentum term. The proposed solution for impedance measurement by using the LS and MLPNN is given below, gradually. In this study, two-term exponential function approximation and a modified MLPNN with additional range arrangement neurons are hybridized. The two-term exponential function approximation part used in the LS algorithm to generate the MLPNN input data is given in equation (1). The proposed ANN structure is shown in Figure 2. The structure contains six layers. The first two layers are assigned for the realization of the two-term exponential function These two layers are in the first stage. The last four layers, generating the second stage, are assigned for modified MLPNN realization.

Table 2. Proposed ANN structure with preprocessing unit

In the first two layers, the magnitude of each real and imaginary pair is applied to exponential function after multiplication with z> and d_t constants. Outputs of exponential function units are multiplied by a_i, and c_i constants before summation. The summation is divided by a normalization parameter, N . The operations of the first and the second layer are given in equations (3) and (2), respectively. After these operations, inputs of the MLPNN part are generated by merging the first layer and the second layer outputs, as shown in equations (4) and (5).

for i = 1,3,5 .

for i = 2,4,6 . The index i represents the neuron number of the third layer, where they are enumerated from one to six in ascending order. The net calculations of the third layer are given in equation (9).

where j = 1,...,9 and b_j is bias value. At the next step, all ner calculations are applied to the logarithmic sigmoid activation function, as given in equation (7).

where _Uj is the output of the third layer. After the activation function, the inputs of the fourth layer are obtained. In the next layer, the outputs of the fourth layer are separated into three groups. Each group accepts three successive outputs of previous layer neurons for the net calculations, as given in equations (8), (9) and (10).

where k = 10.

where k = 11.

where k = 12 , b_k is bias value and d_C1,d_C2 and dc₃ are distance coefficients calculated as in equations (11), (12) and (13), respectively.

where hardlim represents the hard-limit transfer function. At the next step, all net_k calculations are applied to the logarithmic sigmoid activation function as given in equation (14) to obtain the fifth layer outputs.

where k = 10,...,12 . Finally, output y , referring to the MLPNN estimation, is achieved by completing the sixth layer operations as in equation (15).

According to the backpropagation algorithm, the error is calculated as in equation (16). e = t - y (16) where _e is the instantaneous error expressed as the difference between the target (t) and estimated output. In the standard backpropagation learning algorithm with gradient descent, weight updates are done as in equation (17).

where and are defined as updated weight and existing weight of iteration,

respectively, whereas 7 is the learning rate,

is the momentum factor, and E expresses the mean square error.

The error function can be obtained by summing over a training set of N examples as given in equation (18).

where the y_k is determined by using equation (15) with k^th input vector. The derivative of E with respect to hidden layer to output layer weights w_kj can be written by using the chain rule of differentiation as given in equation (19).

Finally, the error is backpropagated to the previous layer neurons, as given in equation (20).

where δ_k is the backpropagated error, /' is the first derivative of the activation function and dc_m is the distance coefficient. The subscript m is 1 ,2 or 3. Similarly, the derivative of

E with respect to the weights between input and the first hidden layers, w_kj , must be calculated by using equation (21).

Since net_k depends on the _Uj , after applying the chain rule of differentiation, the result in equation (22) is obtained.

Thus, substituting equation (22) into (21), the backpropagated error at the i^th neuron is found as in equation (23).

The derivative of the error at the output layer with respect to weights between input and the first hidden layer is given in equation (24).

This approach is applied until the third layer of the proposed ANN. The weight vectors and bias parameters of all the neurons are updated until the mean square error goal, or the maximum number of iterations are achieved. The proposed ANN training algorithm flowchart is given in Figure 1 .

Human skin is a resistive medium as the other bodily tissues [2], Measuring accurate skin impedance is crucial in many biomedical applications [3], According to the tape stripping experiment, the impedance between outermost and underlying skin layers has real and imaginary parts between 0.1-100 kQ and 0-200 kQ ranges, respectively [4], The frequency range of the applied signal in the experiment is altered from 1.22kHz to 1 MHz before and after 30, 60, and 90 tape stripping [4],

In this study, similar to the tape stripping experiment boundaries, impedance data acquisition is performed between 1 kΩ - 200 kΩ interval at the frequency range of 10 kHz - 100 kHz. Totally, 1930 impedances are measured in the specified impedance and frequency ranges for training the neural network. After ten repetitive measurements of each impedance, the average value is taken to form the data set. Additionally, 1900 impedance measurements, apart from the training data, are collected for testing the proposed system. Figure 2 shows the flowchart of the data collection procedure.

Impedance measurement circuit with the AD5933 impedance analyzer is shown in Figure 3. The measurement system is supplied with 3.3V DC voltage. An internal clock with 16.776 MHz frequency is defined as the system clock for the AD5933. The control of the AD5933 is done by the ATMEGA328P microcontroller on the serial I2C communication. 16 MHz HC49S crystal oscillator is used for the microcontroller. Two 10 kQ pull-up resistors are wired to the AD5933 to keep the serial data and serial clock pins from floating.

The voltage gain of the programmable gain amplifier (PGA) and output excitation voltage is set to x1 and 2.2v_pp , respectively. A 10 kQ resistor is chosen as a fixed calibration impedance. After connecting the unknown impedance to the VI N and VOUT pins, the measurement result is displayed on the LCD screen as soon as the completion of execution. The printed circuit board (PCB) of the proposed measurement system is given in Figure 4.

Claims

CLAIMS Artificial Neural Networks (ANN) based single calibration impedance measurement system for skin impedance range system characterized by; Artificial Neural Networks (ANN) based topology, consisting of the first part performing radial- based-function (RBF) based regression and the second part, which is a multilayer perceptron neural network (MLPNN) with error-backpropagation learning. It is the first section that performs the radial-basis-function (RBF) based regression mentioned in Claim 1 characterized in that;

- monitoring the Euclidean distance calculation layer, the summation and normalization layers,

- all real and virtual measurement data for a single frequency are applied to two specific neurons in the Euclidean distance calculation layer,

- assigning a summing and normalizing layer neuron for measurement at a frequency,

- exponentially activated euclidean distance calculation neurons, each frequency also has weight values to correlate real and virtual inputs,

- the least squares (LS) algorithm is characterized by the use of this layer to calculate weight and exponential factor terms. It is the second section, which is the error-backpropagation learning multi-layer perceptron neural network (MLPNN) mentioned in Claim 1 characterized in that;

- involvement of the correlation stage to perform the matching from inputs to outputs,

- implementation of MLPNN's back propagation learning algorithm,

- although the stages here are connected one after another, they are trained independently of each other,

- target value of each step, required impedance output,

- after using the first stage, the data is characterized by applying it to the second stage for better error minimization.