On the characterization of EM emission of electronic products: Case study for different program modes

The characterization of the EM emissions for electronic products is crucial to ensure that the emissions have met the requirements of the EMC standards. For this, a more comprehensive testing is required to get more meaningful results. While, the emergence of non-stationary emissions is a challenge to obtain valid analysis results. So far, non-stationary EM emissions is not considered and treated properly in the emission analysis. This paper presents a new method for the analysis of EM emissions from electronic devices as a case study by testing three different program modes (scenarios) of Intel Galileo board. These program modes were designed to vary processing intensity in its memory and processor. A comparison was also made between the actual situation (the presence of non-stationary signals) and the hypothetical situation with the assumption that all emissions were stationary. As a result, a significant difference was observed when the analysis considered the real scenario of a non-stationary emission. The ratio between the average autocorrelation using the proposed algorithm and the average correlation by ignoring the non-stationarity of the emission signal was 113.6 times. The study concludes that different program modes produce the different characteristics of EM emissions, making some of them non-stationary. Hence, we strongly suggest the consideration of the non-stationarity of the EM emissions in characterizing complex electronic devices


Introduction
The measurement and analysis of electromagnetic (ЕМ) emissions in electronic products is crucial to ensure that the produced emissions have met the EMC requirements and do not interfere with the surroundings.Electronic products that meet EMC standards and regulations can be marketed anywhere, both at national and international level.
Based on [75], NFS has some more advantages compared to other techniques, where it is more economical and is able to locate the source of the emissions.Far field measurements such as OATS, AC, CATR, and RC are not able to detect the sources of the problematic EM emissions.So, it is more difficult to rectify the issue if far field measurement technique is used.One of the main problems in NFS is the presence of non-stationary behavior in the EM emissions.The emergence of this nonstationary emission takes place in a short period with a specific frequency.A previous study dealing with the problem of nonstationary emissions stated that if a non-stationary series were used as an input for predicting the propagation of the EM emissions, the results would be incorrect [76].
Since processes with high emissions may occur only within a small percentage of the time, the averaging process in calculating the field-field correlation will reduce the effect of this process by assuming stationary emission statistics.
The conventional NFS method uses the frequency domain making the emergence of non-stationary emissions unable to be detected, while the new method using the time domain can detect non-stationary EM emission signals.Most of NFS researchers neglect the existence of non-stationary EM emissions.Some examples of NFS studies that ignore nonstationary EM emissions are [77][78][79][80].
Some facts and phenomena of non-stationary emission have been presented in [81].This non-stationary EM emission is the results of Raspberry Pi 3 and Intel Galileo board measurements [81].Fig. 1, Fig. 2, and Fig. 3 show the examples of nonstationary EM emissions containing a signal of interest, which causes it to be non-stationary in different ways.In Fig. 1 the signal of interest is at the right end and wide, while for EM emissions in Fig. 2 and Fig. 3, the signal of interest is narrower.For Fig. 2, the signal of interest is around 0.18ms, and for Fig. 3, the signal of interest around 0.02ms and 1.05ms.This paper will characterize and analyze the EM emissions from the Galileo board with several program modes that produce stationary and non-stationary EM emissions.The initial hypothesis in this research is that different program modes running in electronic products will produce different emissions.The more complex the program mode being run will produce the different unpredictable EM emission behavior.
The contribution of this research is to improve conventional methods in which in these methods, EM emission analysis is commonly carried out in the frequency domain and does not take into account the non-stationary character of EM emission.In this study, EM emissions analysis was carried out in the time domain and the behavior of EM emissions in certain program modes that were run was studied taking into account nonstationary emissions.The result of the improvement was found more accurate EM emissions and enabled to identify the source of EM emissions in electronic products.This paper consists of 4 sections.Section 1 presents an introduction related to the phenomenon of non-stationary EM emissions and techniques for measuring EM emissions in electronic products.Section 2 presents the research methodology carried out including the research stages, instruments used as well as research scenarios.Furthermore, Section 3 contains the results and discussion.In this section, the results of EM emission measurements with three different scenarios and analysis of non-stationary EM emissions are discussed.The final section of this paper presents the conclusion of the study results.

Materials and Methods
The objective of this study is to characterize EM emissions in electronic products based on different program modes running on the device.To achieve this objective, three research stages were involved: (1) measurement of EM emissions with three program modes (scenarios); (2) stationarity test; and (3) characterization of EM emissions for each scenario.
Fig. 4 shows the process of measuring EM emissions.In this study, the DUT refers to the Intel Galileo board (cf.Table 1).The R50-1 RF magnetic probe is manufactured by Langer EMV (cf.Table 2).Measurements were made by placing the probe tip 2 mm above the Intel Galileo memory.The probe was connected to port 1 of Keysight DSOS804a digital oscilloscope.Table 3 shows the data sheet for this instrument.The sampling rate of this instrument was 20 GSa/s with bandwidth 8 GHz.Based on this data sheet, the instrument can be used to measure EM emissions with the R-50-1 probe, which has a working frequency of 30 MHz -3 GHz.The setup is shown in Fig. 4.After the measurements, the data were further assessed for stationarity.Since there were only three data sequences only three, the stationarity was estimated by the visual inspection of the run-sequence plots (i.e. the oscillograms).If the EM emission was stationary, then its autocorrelation function was calculated.If the EM emission was non-stationary, a signal of interest that caused it to be non-stationary must be detected.The algorithm to detect the signal of interest was by implementing segmentation and automatic detection based on Short time energy (STE).Based on previous research carried out by the author, STE is the technique with the best performance in segmentation and detection of non-stationary components [81].
In the segmentation process a time-limited windows w[m] extracted the signal frames at regular intervals as [81] [ where   1,..., mM  is the local time index, M is the window length, f is the frame index, and h is the hop size.
The STE is defined as the energy of the corresponding signal frame [85]: To analyze the relation of the average signal over the entire signal duration and the maximum amplitude of the signal, we used the autocorrelation function (ACF).The ACF of discrete signal x[n] is expressed as [85] : ( ) where l is the sample lag introduced between the original signal and its sliding copy.Fig. 5 shows the flow of this study.The first stage is to measure EM emissions based upon three program modes run by the Intel Galileo board.After obtaining the EM emissions data, the EM emissions were checked to see if they were stationary or non-stationary.If the EM emission was stationary, then it was characterized directly by the autocorrelation operation.If the EM emission was non-stationary, then the STE algorithm was applied for the segmentation process and detection of the part that causes it to be non-stationary.Furthermore, it was characterized using autocorrelation operations.Three different scenarios were considered while performing the measurement.Each time domain data corresponding to the three scenarios would be checked for stationarity, followed by autocorrelation computation and final characterization.If the signal was estimated as non-stationary, then a change detection parameter was performed, and the signal was segmented into a few stationary portions.In such cases, the autocorrelation was computed for each of them.

Results and Discussion
Fig. 6, Fig. 7, and Fig. 8 present the results of measurements in the time domain for Scenario 1, Scenario 2 and Scenario 3, respectively.It is interesting to note that the program mode variation from Intel Galileo produced different EM emissions.Scenario 1 refers to Galileo performing mathematical calculations, which showed that the signal appeared stable without any spikes in EM emissions.Meanwhile, Scenario 2 refers to a process with random numbers filled into a large array of sequential addresses.In this scenario, the EM emission produced had a spike on the edge (red box in Fig. 7).Meanwhile, Scenario 3 showed that Galileo performed one process with random numbers filled into multiple random addresses.Fig. 6 to Fig. 8 show almost the same amplitude.The amplitude of Fig. 6 and Fig. 7 was 0.04 volts, while the amplitude in Fig. 8 was 0.05 Volts.In Fig. 7 there was a visible segment of the emission that had more intensity than other parts (in the red line box making EM emissions non-stationary).The next step was to determine the stationarity of the EM emissions for the three scenarios.The stationarity could visually be estimated (by Eye-ball test) directly in the time domain (cf.Figs. 6 to 8).In addition, one may use the spectrograms of the signals of interest for the same goal.The signal is considered stationary if its time or time-frequency domain pattern does not change [86].The process of stationarity estimation could also be done using the test given in [87].
Fig. 9, Fig. 10 and Fig. 11 show the EM emissions in the time-frequency domain for Scenario 1, Scenario 2 and Scenario 3 respectively.From these three Figures, it was found that the EM emissions of Scenario 1 and Scenario 3 were stationary, while the EM emissions of Scenario 2 were non-stationary.After identifying stationarity, the next step was the autocorrelation operation for stationary emission.This correlation operation was based on equation 3. Fig. 12 and Fig. 14 show the autocorrelation results for Scenario 1 and Scenario 3. As shown in Fig. 12 and Fig. 14, the ensemble average of the autocorrelation function for Scenario 1 and Scenario 3 was acceptable since the ACF itself was stationary and its average was representative.The maximum peak to average magnitude ratio was found at 1.82 and 2.76 for scenario 1 and scenario 3 respectively.For EM emissions resulting from Scenario 2, a signal of interest detection operation was required because EM emissions were non-stationary.The results would be biased if we applied autocorrelation directly to EM emission signals from Scenario 2, like Scenario 1 and Scenario 3. In Fig. 14, it can be observed that the average was far below the maximum peak.In other words, there was a high gap between the average value and the maximum peak.The maximum peak to average magnitude ratio was 23.22 for scenario 2. This is in line with previous studies showing that results would be misleading if a non-stationary series was applied as input and considered stationary.The solution is to detect the signal's change points, extract the piecewise stationary segments, and thread them in an ordinary way.
The change detection was performed using the STE property after which the signal was segmented, the results of which are shown in Fig. 15.We can see that the algorithm successfully detected the signals of interest in non-stationary emissions.As shown in Fig. 16, there was a significant difference between the average autocorrelation using the proposed algorithm and the average correlation by ignoring the non-stationarity of the emission signal.In this case, the difference was 113.6 times.Analysis that considers the nonstationary behavior of EM emissions will provide more accurate results.Also, by using autocorrelation, we could determine the source or location of problematic EM emissions precisely.This is very helpful in redesigning electronic products and makes it easier for company to produce electronic products that comply with EMC.The limitations of this method are that it consumes a long-time during measurement, and requires large data and an oscilloscope with a high working frequency (GHz Oscilloscope).

Conclusion
In this paper, EM emissions from an electronic product have been characterized in a new manner in a study organized as a case report.EM emissions produced were found different for each of three program scenarios.The nature of EM emissions was also different; some were stationary, and other were not.It is suggested that to overcome the problem of characterizing the nonstationary EM emissions is by presenting it based upon a change detection via STE feature.Prior to characterization with autocorrelation, nonstationary emissions were segmented, and the signal of interest was detected.The results showed a significant difference between the ACF in the case of stationarity and nonstationarity.Any neglect of this fact leads to a dramatic underestimation of the EM emissions quantitatively and qualitatively.The findings of this research are very useful for obtaining more accurate emission measurements and analysis.This makes it easier for manufacturers to comply with EMC standards.In addition, the manufacturers will be able to determine which components contribute to the problems related to EM emissions, thus making it easier for them to improve the design.This work is a basis for further studies with a larger sample of measurement data.Future research will also more deeply investigate the factors determining the occurrence of nonstationary emissions.

Fig. 15 .
Fig. 15.Detection of the signal of interest using STE