Development of an artificial neural network software tool for the assessment of the electromagnetic field radiating by electrostatic discharges G.P. Fotis, L. Ekonomou, T.I. Maris and P. Liatsis Abstract: Artificial neural networks (ANNs) are addressed in order to assess the electric and magnetic field radiating by electrostatic discharges (ESDs). Actual input and output data collected from hundreds of measurements carried out in the High Voltage Laboratory of the National Technical University of Athens are used in the training, validation and testing process. The developed ANN method is coded in a comprehensive software tool to be used by laboratories involved in ESD tests, which either face a lack of suitable measuring equipment or want to compare their own measurements. The electromagnetic field produced by radiating ESDs, can be assessed very easily and accurately by simply measuring and providing to the tool the discharge current. The authors strongly believe that the proposed ANN software tool can be extremely useful for laboratories involved in ESD tests according to the current IEC Standard, as the forthcoming revised version of this standard will almost certainly include measurements of the radiating electromagnetic field during the verification of ESD generators.
1
Introduction
Electrostatic discharge (ESD) is defined as the sudden transfer of charge between objects at different electrostatic potentials. Electrostatic charge is created whenever two different materials come into contact and are then separated. Creating electrostatic charge by contact and separation of materials is known as the ‘triboelectric effect’. When the discharge takes place the discharge current may come up to a few Amperes. This makes clear that an ESD may be destructive for electronic or integrated circuits, which are very sensitive to these currents. Although the ESD phenomenon lasts only a few hundred nanoseconds, ESD is a constant threat for electronic or integrated circuits. Therefore the International Electrotechnical Committee (IEC) prescribed the Standard 61000-4-2 [1] in order to define a procedure, which must be followed for immunity tests on electrical or electronic equipment against ESDs. However, during these tests an equipment under test (EUT) may pass with one generator and may fail with another for the same charging voltage and for the same discharge current. Similar results can be observed for the same generator as well, as an EUT may pass the test with one orientation of the ESD generator and fail with another.
# The Institution of Engineering and Technology 2007 doi:10.1049/iet-smt:20060137 Paper first received 28th November 2006 and in revised form 4th March 2007 G.P. Fotis and L. Ekonomou are with the National Technical University of Athens, High Voltage Laboratory, School of Electrical and Computer Engineering, 9 Iroon Politechniou St., Zografou 157 80 Athens, Greece T.I. Maris is with the Department of Electrical Engineering, Technological Educational Institute of Chalkida,, Psachna Evias 334 40, Greece P. Liatsis is with the City University, School of Engineering and Mathematical Sciences, Division of Electrical Electronic and Information Engineering, Information and Biomedical Engineering Centre, Northampton Square, London EC1V 0HB, U.K. E-mail:
[email protected] IET Sci. Meas. Technol., 2007, 1, (5), pp. 261 – 269
Measurements of the electromagnetic fields because of ESD received little attention until the end of the eighties. Wilson and Ma [2] were the first, who simultaneously measured the current and the electric field during ESDs at a distance of 1.5 m, using a broadband, TEM horn antenna. During recent years many researchers have conducted measurements of the electromagnetic fields associated with an ESD event. David Pommerenke [3] measured the electric and the magnetic fields at distances between 0.1 and 1 m, for both air and contact discharges. He found that the magnitude of the magnetic field strongly depends on the 1/R factor (R being the distance from the point where the ESD occurs), while the magnitude of the electric field is decreasing for a time period after which it increases. There have also been studies [4, 5], where the ESD current waveform can be calculated by measuring the electromagnetic field. David Pommerenke and Stefan Frei [6] have also designed an impulse monitoring system, which automatically detects ESDs by their associated magnetic and electric fields. In [7] and [8], measurements of the electromagnetic field produced during ESD at a calibration set-up have been carried out and these demonstrate that the field measurement is a challenging task and its results may vary depending on the field probes and the measurement system. Measurements of the radiating magnetic [9] or electric field [10], when the current transducer (Pellegrini target) is mounted on the centre of a grounded metal plane proved that different generators produce different electromagnetic fields and also that each individual generator produces different electromagnetic fields depending on its orientation to the EUT. The main conclusion of this work for the radiating electromagnetic field produced by the ESD generators is that the electromagnetic field must be also measured during the verification of the ESD generators. It is almost certain that the next revision of the Standard [1] will include measurements of the electromagnetic field produced by the ESD generators during the verification. 261
However, these measurements require the use of specific magnetic and electric field sensors, which are extremely complicated in construction and in use, and of high cost. In this paper artificial neural networks (ANNs) are addressed in order to assess the electric and the magnetic fields radiating by ESDs. ANNs have seen increased usage in recent years in various fields such as finance [11, 12], medicine [13, 14], industry [15, 16] and engineering [17– 20], because of their computational speed, their ability to handle complex nonlinear functions and their robustness and great efficiency, even in cases where full information for the studied problems is absent. Actual electromagnetic field measurements, radiated by ESDs and carried out in the High Voltage Laboratory of the National Technical University of Athens, are used in order to train, validate and test the proposed ANNs. Two different types of ANNs (multi-layer perceptrons and radialbased function networks), several structures, learning algorithms and transfer functions are tested in order to produce the ANN models with the best generalising ability. The developed ANN method is coded in a comprehensive software tool to be used by laboratories, which are facing either a lack of suitable ESD test equipment or want to compare results with their own measurements. Having in mind that in the forthcoming revision of the IEC 61000-4-2 [1] measurements of the radiating electromagnetic field during the verification of ESD generators will be almost certainly included, the authors strongly believe that the proposed ANN software tool will be broadly useful, as the electromagnetic field produced by ESDs, can be easily and accurately assessed by simply measuring the discharge current. 2
IEC 61000-4-2 ESD model
According to IEC 61000-4-2 [1] an ESD generator must produce a Human Body Model (HBM) pulse as is shown in Fig. 1. The pulse of Fig. 1 is divided into two parts: a first peak called the ‘initial peak’, caused by the discharge of the hand, where there is the maximum current and a second peak, which is caused by the discharge of the body. The rise time (tr) of the initial peak is between 0.7 and 1 ns and its amplitude depends on the charging voltage of the ESD simulator. According to the specifications of the standard for the verification of ESD generators there are four parameters whose values must be confined by certain limits. These parameters are: the rise time (tr), the maximum discharge
current (Imax), the current at 30 ns (I30) and the current at 60 ns (I60). As is shown in Fig. 1, these two current values are calculated for a time period of 30 and 60 ns, respectively, starting from the time point, when the current equals 10% of the maximum current. The limits of these parameters, shown in Table 1, are valid for contact discharges only. 3 3.1
Measurement system Experimental set-up
The experimental set-ups for the measurement of the magnetic or the electric field are shown in Fig. 2. The current and field (electric or magnetic) produced by contact discharges for various charging voltage levels were measured simultaneously, by the four-channel Tektronix oscilloscope model TDS 7254B, whose bandwidth ranges from dc to 2.5 GHz. The ESD generators used were the NSG-433 and the NSG-438 of Schaffner. The positioning of the highvoltage cable was kept constant during the experimental procedure. In order for the current to be measured, a resistive load was used, as the IEC [1] defines and it was placed at the centre of a 1.5 m 1.5 m horizontal metal plane, which was placed 70 cm above the ground. The sensors that were used for the experiment were ground-based field sensors with active integration developed by Pommerenke [21]. They were placed at various distances (20, 35, 50 and 65 cm) from the discharge point on the metal plane and in three perpendicular directions (directions A, C and D) as it can be seen in Fig. 3. Measurements in direction B were not conducted because of the interference of the ground strap of the ESD generator. It is known that the position of the ground strap affects the falling edge of the current’s waveform. In order to minimise the uncertainty of this fact on the measurement of the magnetic field, the ground strap was at a distance of 1 m from the target as defined in [1] and the loop was as large as possible. 3.2
Experimental results
Fig. 4a and b depict representative H-field and E-field waveforms in relation to the discharge current for the first 200 ns, respectively, when the field sensors are placed at a distance of 20 cm from the discharge point in direction A for the NSG-438 ESD generator. Fig. 5a and b depict the peak H-field and E-field, respectively, for charging voltage of þ2 kV in direction D for the NSG-433 and NSG-438 ESD generators. It is obvious that each generator produces different electromagnetic field. Fig. 6a and b show the comparison between the peaks of the absolute value of the magnetic field and of the electric field for all three directions for the NSG-438 ESD generator, proving that the produced electromagnetic field differs depending on the orientation of the ESD generator.
Table 1: Waveform parameters Voltage (kV) +2
Fig. 1 Typical waveform of the output current of the electrostatic discharge generator [1] 262
Imax (A) 6.7528.25
tr (ns) 0.721
I30 (A) 2.825.2
I60(A) 1.422.6
+4
13.50216.50
0.721
5.6210.4
2.825.2
+6
20.25224.75
0.721
8.4215.6
4.227.8
+8
27.00233.00
0.721
11.2220.8
5.6210.4
IET Sci. Meas. Technol., Vol. 1, No. 5, September 2007
Fig. 2 Experimental set-ups for the measurement of a Magnetic field b Electric field
dynamic state response to external inputs. ANNs are capable of learning patterns by being trained with a number of known patterns. The learning process automatically adjusts the weights and thresholds of the processing elements. Once adjusted with minimal differences between the ANN output and the targeted output, the neural network is said to be trained. There are plenty of learning rules in the literature [22 – 24]. Gradient descent and Levenberg-Marquardt are the most popular and effective. ANNs also have many structures and types. The simplest and most commonly used are the multi-layer perceptron (MLP) [23, 24] and the radial basis function network (RBF) [25, 26]. 4.1 Fig. 3 Measurement points where the field sensors were placed
4
Artificial neural networks
ANNs are biologically inspired. They are generally made of a number of simple and highly interconnected processing elements organised in layers as shown in Fig. 7. These processing elements or neurons process information by their
Neural network architecture
The goal of this work is to develop two ANN models capable of accurately assessing the electric and magnetic fields radiated by each of the two ESD generators (NSG-433 and NSG-438) mentioned before. Seven parameters that play an important role in the assessment of the electric and magnetic fields radiating by ESDs are considered as the inputs to the neural networks, while as outputs the peak value of the electric and the peak value of the
Fig. 4 Representative H-field and E-field waveforms a Electrostatic discharge (ESD) current and H-field for the NSG-438 ESD generator at 20 cm from the discharge point, in direction A b ESD current and E-field for the NSG-438 ESD generator at 20 cm from the discharge point, in direction A (charging voltage ¼ þ2 kV) IET Sci. Meas. Technol., Vol. 1, No. 5, September 2007
263
Fig. 5
H-field peak and E-field peak versus distance
a Peak of H-field for various distances from the discharge point in direction D, using the NSG-433 and NSG-438 electrostatic discharge (ESD) generators b Peak of E-field for various distances from the discharge point in direction D, using the NSG-433 and NSG-438 ESD generators
Fig. 6
Comparison between the peaks of the three perpendicular directions
a Peak of H-field for the NSG-438 electrostatic discharge (ESD) generator for three perpendicular directions on the horizontal plane b Peak of E-field for the NSG-438 ESD generator for three perpendicular directions on the horizontal plane (charging voltage ¼ þ2 kV)
magnetic field are considered. These data, which are presented in Table 2, constitute actual data recorded during measurements carried out in the High Voltage Laboratory of the National Technical University of Athens [9, 10], using the method and the measurement system presented in Section 3.1. It must be mentioned that besides these input and output parameters more parameters (such as the E and H field rise times) could be used in the training, validation and testing processes.
More specifically several hundreds of measurements have been performed using each one of the two ESD generators, that is, the NSG-433 and the NSG-438 produced by Schaffner. This extensive number of measurements is because of the many parameters. These parameters were: the generator’s charging voltage (+2, +4, +6 and +8 kV), the distances from the discharge point on the metal plane (20, 35, 50 and 65 cm), the three perpendicular directions (directions A, C and D) (Fig. 3) and the current waveform parameters (rise time, maximum discharge current, current at 30 ns and current at 60 ns) presented in Table 1. These several hundreds of measurements, for
Table 2: ANN architectures Input variables
Output variables
Charging voltage (U)
Peak value of the electric field (Emax)
Rise time (tr)
Peak value of the magnetic field (Hmax)
Maximum discharge current (Imax) Current at 30 ns (I30) Current at 60 ns (I60) Distance (d )
Fig. 7 264
Direction (D)
Structure of a three-layer artificial neural network IET Sci. Meas. Technol., Vol. 1, No. 5, September 2007
each one of the two generators, constitute combinations of all the earlier parameters. 4.2
Training, validation and testing
In order to address the assessment of the electric and magnetic fields radiated by ESDs two ANN types were considered, that is, the MLP and the RBF network. Each ANN model is determined according to its structure, the transfer function and the learning rule, which are used in an effort by the network to learn the fundamental characteristics of the problem examined. The structure of the networks, that is, the number of hidden layers and the number of nodes in each hidden layer, is generally decided by trying various combinations for selecting the structure with the best generalising ability among the combinations tried, considering that one hidden layer is adequate to distinguish input data that are linearly separable, whereas extra layers can accomplish nonlinear separations [26, 27]. This approach was followed, as the selection of an optimal number of hidden layers and nodes for an MLP network is still an open issue, although some papers have been published in these areas [28, 29]. The designed and tested MLP ANN models were combinations of two learning algorithms (the gradient descent and the LevenbergMarquardt), two transfer functions (the tan-sigmoid and the log-sigmoid) and several different structures (1-4 hidden layers and 2-40 neurons in each hidden layer). RBF neural networks are three layer networks. Each node of the hidden layer of RBF corresponded to one basis function centre. In this study, the Gaussian function was used as the kernel function. Bias, which determined the size of the receptive field, was a free parameter. The weights in the output layer were derived using the least square error learning algorithm. At each epoch (iteration), centres were added dynamically until the desired minimum square error was achieved. However, performance of the RBF neural network model critically depended on the chosen centres, which may require an unnecessarily large RBF network to obtain a given level of accuracy and cause numerical ill conditioning [30].
The proposed ANN models were trained using the MATLAB Neural Network Toolbox [31]. Two sets of 1536 values of each input and output data, were used to train and validate each one of the two ANNs. These data refer to measurements conducted with each one of the two ESD generators in every possible combination of charging voltage, distance from the discharge point on the metal plane, perpendicular direction and current waveform parameter. In each training iteration, 20% of random data were removed from the training set and a validation error was calculated for these data. The training processes were repeated until a root mean square error between the actual outputs (peak value of electric field and peak value of magnetic field) and the desired outputs reached the goal of 0.5% or a maximum number of epochs (it was set to 15 000), were accomplished. Finally, the estimated values of the electric and magnetic fields were checked with the values obtained from situations encountered in the training, that is, each one of the two sets of 1536 values for each ESD generator. After extensive simulations with all possible combinations and ANN types (MLP and RBF), it was found that the two ANN models for the NSG-433 and the NSG-438 ESD generators that exhibited the best generalising ability, had a compact structure, had a fast training process and consumed lower memory than all the other tried combinations were: MLP ANN – two hidden layers with 17 and 25 neurons in each hidden layer – Levenberg-Marquardt learning rule-tan-sigmoid transfer function for the NSG-433 generator and MLP ANN – two hidden layers with 21 and 20 neurons in each hidden layer-gradient descent learning rule-tan-sigmoid transfer function for the NSG-438 generator. The mean square error was minimised to the final value of 0.005 within 13 463 epochs for the first ANN model and to the final value of 0.005 within 9176 epochs for the second ANN model. 5
Developed ANN software tool
The two ANN models are coded in the software tool (written also in MATLAB) presented in the simple flow chart of Fig. 8. The software tool which is simple in use,
Fig. 8 Artificial neural network software tool’s simplified flow chart IET Sci. Meas. Technol., Vol. 1, No. 5, September 2007
265
(f) Assessment of the electric and magnetic fields. The peak values of the electric and magnetic fields are assessed and produced. (g) Option of another assessment. The user has an option to perform a new assessment or to exit the software tool.
flexible and user-friendly can be used by laboratories that lack suitable equipment for electromagnetic field measurements or by laboratories which want to compare simulation results with their own measurements. The developed ANN software tool is organised as follows:
It must be mentioned that the developed software tool can produce reliable results only for similar parameters, that is, generator’s charging voltage from +2 to +8 kV, distance from the discharge point on the metal plane from 20 to 65 cm, three perpendicular directions (Fig. 3) and current waveform parameters in the range of those presented in Table 1. For situations with totally different parameters, new ANN models must be trained, validated and tested [32], and a new software tool which will include these new ANN models must be developed.
(a) Selection of NSG-433 or NSG-438 ESD generator. The user selects the generator according to which the assessment of the electric and magnetic fields will be performed. (b) Introduction of current waveform parameters (tr , Imax , I30 and I60). The current waveform parameters are entered to the software tool. (c) Introduction of direction (D). The perpendicular direction at which the sensors are placed is entered to the software tool. (d) Introduction of distance from the charging point (d ). The distance at which the sensors are placed from the discharge point on the metal plane is entered to the software tool. (e) Introduction of generator’s charging voltage (U). The generator’s charging voltage is entered to the software tool.
6
Results and discussion
The developed software tool for the assessment of the electric and magnetic fields radiating by ESDs has been applied to 60 different test examples which were not part of the training validation and testing processes. Thirty different
Table 3: Measured electric and magnetic field versus ANN results for ESD generator NSG-433 Case I – Schaffner ESD generator NSG 2 433 Varying parameters No.
U (kV)
tr (ns)
1
þ2
2
þ2
3 4
Measured Imax(A)
I30 (A)
I60 (A)
0.77
6.95
3.78
0.75
7.17
3.69
þ2
0.76
7.13
þ2
0.73
7.28
5
þ2
0.73
6
þ2
0.76
E (kV/m)
ANN
d (cm)
D
H (A/m)
E (kV/m)
H (A/m)
2.44
35
A
4.04
2.03
4.49
2.31
2.45
65
A
2.36
1.38
2.68
1.57
3.95
2.50
35
C
3.42
1.64
3.69
1.89
3.93
2.39
65
C
2.55
0.95
2.77
0.79
7.34
3.67
2.43
35
D
4.08
1.68
4.26
1.88
7.16
3.62
22.40
65
D
3.52
1.12
3.89
1.33
7
22
0.75
27.19
23.99
22.41
20
A
29.75
22.98
29.25
22.78
8
22
0.78
27.12
23.80
22.42
35
A
24.11
21.88
23.85
21.57
9
22
0.76
27.13
23.87
22.43
50
A
22.65
21.75
22.95
21.95
10
22
0.74
27.18
23.99
22.41
65
A
22.26
21.61
22.62
21.88
11
22
0.73
27.19
23.92
22.37
20
C
26.55
22.93
26.02
23.24
12
22
0.73
27.18
23.95
22.41
35
C
23.35
21.67
23.04
21.93
13
22
0.77
27.18
23.90
22.42
50
C
22.69
20.94
22.29
21.13
14
22
0.72
27.12
23.99
22.44
65
C
22.52
20.81
22.85
21.18
15
þ4
0.71
14.75
7.19
5.08
20
A
3.68
12.28
3.95
12.92
16
þ4
0.75
14.75
7.05
5.04
35
A
2.96
7.18
3.22
7.52
17
þ4
0.78
14.87
6.97
5.01
50
A
2.80
5.14
2.99
5.42
18
þ4
0.79
14.98
6.87
5.09
65
A
2.55
4.51
2.73
3.71
19
þ4
0.72
14.75
7.09
5.12
35
C
6.56
2.89
6.83
2.99
20
þ4
0.72
14.88
6.94
5.13
65
C
5.21
1.64
5.45
1.95
21
þ4
0.74
14.93
6.84
5.17
35
D
7.45
3.18
7.95
3.58
22
þ4
0.73
14.75
6.86
5.11
65
D
6.19
2.10
6.03
1.92
23
24
0.74
215.46
27.36
25.10
35
A
28.31
22.99
28.69
22.77
24
24
0.76
214.78
27.23
25.19
65
A
24.12
22.84
24.35
22.99
25
24
0.79
214.59
27.24
25.16
35
C
26.17
22.72
26.75
22.49
26
24
0.72
214.79
27.13
25.18
65
C
24.77
21.59
24.25
21.86
27
24
0.74
214.42
27.14
24.94
20
D
212.87
23.18
212.06
23.55
28
24
0.76
214.81
27.12
25.01
35
D
27.63
22.92
27.05
22.60
29
24
0.79
215.38
27.25
25.14
50
D
26.17
22.65
26.52
22.89
30
24
0.76
214.83
27.42
24.99
65
D
27.41
22.20
27.65
22.45
ANN, artificial neural network; ESD, electrostatic discharge 266
IET Sci. Meas. Technol., Vol. 1, No. 5, September 2007
Table 4: Measured electric and magnetic field versus ANN results for ESD generator NSG-438 Case II-Schaffner ESD generator NSG 2 438 Varying parameters No.
U (kV)
tr (ns)
1
þ2
2 3
Measured Imax(A)
I30 (A)
I60 (A)
0.72
7.12
3.89
þ2
0.72
6.97
þ2
0.73
7.16
4
þ2
0.74
5
þ2
6 7
E (kV/m)
ANN
d (cm)
D
H (A/m)
E (kV/m)
H (A/m)
2.42
20
A
5.13
2.65
5.51
2.93
3.67
2.48
50
A
2.42
1.34
2.70
1.78
3.98
2.49
65
A
5.81
1.09
6.29
1.51
7.35
3.82
2.54
20
C
5.19
2.91
5.55
2.69
0.75
7.17
3.95
2.43
35
C
2.34
1.87
2.09
1.52
þ2
0.79
7.29
3.75
2.49
50
C
2.36
1.39
2.71
1.13
þ2
0.72
7.25
3.89
2.46
65
C
5.84
1.09
6.14
1.37
8
þ2
0.74
7.39
3.79
2.39
20
D
5.63
2.66
5.89
2.98
9
þ2
0.77
7.16
3.93
2.30
35
D
2.98
2.21
3.29
2.64
10
þ2
0.79
7.22
3.67
2.45
50
D
3.02
1.73
3.28
1.54
11
þ2
0.77
6.97
3.95
2.39
65
D
2.73
1.34
2.41
1.03
12
22
0.72
27.07
3.82
22.32
20
A
25.94
22.95
26.22
23.28
13
22
0.71
27.06
23.99
22.41
50
A
22.60
21.31
22.33
21.03
14
22
0.75
27.09
23.84
22.52
20
C
22.83
25.87
22.95
26.12
15
22
0.78
27.09
23.65
22.36
50
C
22.40
21.21
22.22
21.48
16
22
0.74
27.19
24.12
22.44
20
D
24.13
22.91
24.33
23.22
17
22
0.72
27.12
23.93
22.24
35
D
23.93
21.82
23.67
21.69 21.20
18
22
0.72
27.29
23.98
22.32
50
D
23.42
21.42
23.15
19
þ4
0.72
14.69
6.98
5.03
20
A
11.61
3.39
12.21
3.62
20
þ4
0.73
14.72
6.89
5.04
35
A
7.22
2.88
7.44
2.99
21
þ4
0.78
14.79
6.72
5.06
50
A
4.83
2.34
4.99
2.01
22
þ4
0.74
14.59
6.58
4.54
65
A
3.58
1.92
3.87
1.68
23
þ4
0.72
14.67
6.84
4.09
20
C
9.69
3.47
9.97
3.72
24
þ4
0.78
14.72
7.18
4.65
50
C
4.79
2.45
5.19
2.69
25
þ4
0.75
14.89
6.39
4.66
20
D
10.63
3.24
11.35
3.66
26
þ4
0.73
14.70
6.81
4.78
50
D
5.49
2.69
5.75
2.25
27
24
0.76
214.89
27.32
24.81
20
A
212.39
23.36
211.75
23.55
28
24
0.77
214.63
27.31
24.66
50
A
25.19
22.26
24.91
22.56
29
24
0.74
214.79
27.21
24.05
20
C
212.27
23.33
212.57
23.56
30
24
0.78
214.92
27.29
25.01
50
C
24.47
22.36
24.09
22.02
ANN, artificial neural network; ESD, electrostatic discharge
test examples were tested for the Schaffner ESD generator NSG-433 and another 30 test examples for the Schaffner ESD generator NSG-438. Tables 3 and 4 present the results produced using the ANN software tool. The same tables also present actual measurements of the electric and the magnetic fields measured during experiments performed in the High Voltage Laboratory of the National Technical University of Athens for exactly the same parameters. The results obtained using the proposed ANN software tool are very close to the actual measured ones, something which clearly implies that the proposed ANN method and the software tool are well working and have an acceptable accuracy. Figs. 9 and 10 present the good correlation between measured and predicted results for both case studies (NSG-433 and NSG-438 ESD generators) for both peak values of electric and magnetic fields. The ANN software tool will be a useful tool for laboratories involved in ESD tests, as the verification of the ESD generators will be conducted by simply measuring the ESD current. The produced electromagnetic field, which will be included in the next revision of the IET Sci. Meas. Technol., Vol. 1, No. 5, September 2007
Standard [1], can be easily calculated by the proposed ANN software tool. 7
Conclusions
This paper describes an ANN software tool that assesses the electric and the magnetic fields radiated by ESDs. Two different types of ANNs and actual input and output data collected from measurements carried out in the High Voltage Laboratory of the National Technical University of Athens were used in order to select the ANN models which produced the best generalising ability, had a compact structure, had a fast training process and consumed lower memory. These models were coded in a comprehensive software tool to be used by laboratories involved in ESD tests. The developed software tool assesses easily and accurately the electromagnetic field radiating by ESDs by simply measuring the discharge current. The software tool will be extremely useful for the laboratories involved in the ESD tests, as the forthcoming revision of the current Standard [1] will almost certainly include 267
measurements of the radiating electromagnetic field during the verification of ESD generators. 8
Acknowledgments
The authors would like to thank Professor David Pommerenke of the Missouri-Rolla University in the United States for the field sensors he lent to them in order to measure the electric and the magnetic fields. 9
Fig. 9 Correlation between measured and predicted artificial neural network results for electrostatic discharge generator NSG-433
Fig. 10 Correlation between measured and predicted artificial neural network results for electrostatic discharge generator NSG-438 268
References
1 IEC 61000-4-2: Electromagnetic Compatibility (EMC), Part 4: Testing and measurement techniques, Section 2: Electrostatic discharge immunity test – Basic EMC Publication, 1995 2 Wilson, P.F., and Ma, M.T.: ‘Field radiated by electrostatic discharges’, IEEE Trans. Electromagn. Compat., 1991, 33, (1), pp. 10–18 3 Pommerenke, D.: ‘ESD: transient fields, arc simulation and rise time limit’, J. Electrostat., 1995, 36, pp. 31– 54 4 Ki-Chai Kim, Kwang-Sik Lee, and Dong-In Lee.: ‘Estimation of ESD current waveshapes by radiated electromagnetic fields’, IEICE Trans. Commun., 2000, E83-B, (3), pp. 608– 612 5 Ishigami, S., Gokita, R., Nishiyama, Y., and Yokoshima, I.: ‘Measurements of fast transient fields in the vicinity of short gap discharges’, IEICE Trans. Commun., 1995, E78-B, (2), pp. 199–206 6 Frei, S., and Pommerenke, D.: ‘A transient field measurement system to analyze the severity and occurrence rate of electrostatic discharge (ESD)’, J. Electrostat., 1998, 44, pp. 191– 203 7 Leuchtmann, P., and Sroka, J.: ‘Transient field simulation of electrostatic discharge (ESD) in the calibration setup (ace. IEC 61000-4-2)’. Int. Symp. on EMC, August 2000, pp. 443–448 8 Leuchtmann, P., and Sroka, J.: ‘Enhanced field simulations and measurements of the ESD calibration setup’. Int. Symp. on EMC, August 2001, pp. 1273– 1278 9 Fotis, G.P., Rapanakis, A.G., Gonos, I.F., and Stathopulos, I.A.: ‘Measurement of the magnetic field radiating by electrostatic discharge generators along various directions’. 12th Biennial IEEE Conference on Electromagnetic Field Computation CEFC 2006, Poster Session PE 1: Static and Quasi – Static Fields VII 10 Fotis, G.P., Gonos, I.F., and Stathopulos, I.A.: ‘Measurement of the electric field radiating by electrostatic discharges’, Instit. Phys. (IOP), Proc., Meas., Sci. Technol., 2006, 17, pp. 1292–1298 11 Bodyanskiy, Y., and Popov, S.: ‘Neural network approach to forecasting of quasiperiodic financial time series’, Eur. J. Operation. Res., 2006, 175, (3), pp. 1357–1366 12 An-Sing Chen, and Leung, M.T.: ‘Regression neural network for error correction in foreign exchange forecasting and trading’, Comput. Operation Res., 2004, 31, (7), pp. 1049–1068 13 Paetz, J.: ‘Knowledge-based approach to septic shock patient data using a neural network with trapezoidal activation functions’, Artif. Intell. Med., 2003, 28, (2), pp. 207– 230 14 Frize, M., Ennett, C.M., Stevenson, M., and Trigg, H.C.E.: ‘Clinical decision support systems for intensive care units: using artificial neural networks’, Med. Eng. Phys., 2001, 23, (3), pp. 217–225 15 Sloleimani-Mohseni, M., Thomas, B., and Per, Fahlen.: ‘Estimation of operative temperature in buildings using artificial neural networks’, J. Energy Buildings, 2006, 38, pp. 635 –640 16 Miti, G.K., and Moses, A.J.: ‘Neural network-based software tool for predicting magnetic performance of strip-wound magnetic cores at medium to high frequency’, IEE Proc., Sci. Meas. Technol., 2004, 151, (3), pp. 181–187 17 Chen, Y.J., Chen, Y.M., Wang, C.B., Chu, H.C., and Tsai, T.N.: ‘Developing a multi-layer reference design retrieval technology for knowledge management in engineering design’, Expert Syst. Appl., 2005, 29, (4), pp. 839–866 18 Rajasekaran, S., and Amalraj, R.: ‘Predictions of design parameters in civil engineering problems using SLNN with a single hidden RBF neuron’, Comput. Struct., 2002, 80, (31), pp. 2495– 2505 19 Mahanty, R.N., and Gupta, P.B.D.: ‘Application of RBF neural network to fault classification and location in transmission lines’, IEE Proc., Generat. Trans. Distrib., 2004, 151, (2), pp. 201–212 20 Patnaik, A., and Mishra, R.K.: ‘ANN techniques in microwave engineering’, IEEE Microwave Mag., 2000, 1, (1), pp. 55–60 21 Wang, K., Pommerenke, D., Chundru, R., Doren, T.V., Drewniak, J.L., and Shashindranath, A.: ‘Numerical modeling of electrostatic discharge generators’, IEEE Trans. Electromagn. Compat., 2003, 45, (2), pp. 258–270 22 Abe, S.: ‘Neural networks and fuzzy systems’ (Kluwer Academic Publishers, Boston, 1997) IET Sci. Meas. Technol., Vol. 1, No. 5, September 2007
23 Cichocki, A., and Unbehauen, R.: ‘Neural networks for optimization and signal processing’ (Wiley, 1993) 24 Lippmann, R.: ‘An introduction to computing with neural nets’, IEEE ASSP Mag., 1987, 4, (2), pp. 4 –22 25 Gurney, K.: ‘An introduction to neural networks’ (UCL Press, London, 1999) 26 Haykin, S.: ‘Neural networks: a comprehensive foundation’ (PrenticeHall, Englewood Cliffs, NJ, 1999) 27 Maghami, P.G., and Sparks, D.W.: ‘Design of neural networks for fast convergence and accuracy: dynamics and control’, IEEE Trans. Neural Netw., 2000, 11, (1), pp. 113–123
IET Sci. Meas. Technol., Vol. 1, No. 5, September 2007
28 Tamura, S.I., and Tateishi, M.: ‘Capabilities of a four-layered feedforward neural network: four layers versus three’, IEEE Trans. Neural Nets, 1997, 8, (2), pp. 251– 255 29 Hornik, K.: ‘Some new results on neural network approximation’, Neural Netw., 1993, 6, (8), pp. 1069– 1072 30 Nolles, O.: ‘Nonlinear system identification: from classical approaches to neural networks and fuzzy models’ (Spring-Verlag, Berlin, 2001) 31 Demuth, H., and Beale, M.: ‘Neural network toolbox user’s guide for use with MATLAB’, 2002 32 Hagan, M.T., Demuth, H.P., and Beale, M.: ‘Neural network design’ (Boston, PWS Publishing, 1996)
269
