1. Introduction
With the proposal of the industrial 4.0 strategy, advanced technologies such as the Internet of Things, edge computing, and Big Data have emerged to provide technical support for the development of intelligent manufacturing [
1,
2,
3,
4]. The Industrial Internet of Things connects machines, objects, and people in factories [
5,
6]. Edge computing transfers data processing, application running, and even the realization of some functional services from the network center to the nodes on the edge of the network. It solves the disadvantage that cloud computing cannot support computing-intensive and delay-sensitive tasks and enables data processing and analysis in intelligent factories. In order to support the intelligent interconnection and application of all kinds of equipment in the industrial field more effectively, edge calculation should be made close to or embedded in all kinds of sensor, robot, instrument, and other equipment nodes of industrial wireless networks [
7]. In order to better access or embed edge computing on various device nodes of industrial wireless networks, the characteristics of the Industrial Internet of Things should be studied.
As a typical case of the fifth generation mobile communication (5G) mMTC (massive machine type of communication) scenario, the Industrial Internet of Things (IIoT) has become the key competitiveness of the country and has attracted extensive attention in academic and industrial fields [
8]. Wireless technology, as the information transmission channel in the IIoT system, plays a significant role in the industrial field. At present, improving production efficiency and realizing energy saving and emission reduction are the main strategic tasks of manufacturing industry. With the development of the industrial internet, intelligent manufacturing will run through all aspects of enterprise production and operation, bringing profound changes to the development of manufacturing industry. IIoT connects mechanical equipment, people, and items together and enables them to exchange data. It will be widely used in automobile manufacturing and other industrial scenarios. However, the electromagnetic noise in the industrial environment is significantly different from that in other typical environments such as office areas [
9]. When the machine is working, different types of electromagnetic noises will be emitted due to the increase of temperature, mechanical vibration, spark discharge, and other physical phenomena. Moreover, there are many kinds of electromagnetic noise distributed at any random position in the factory, and the frequency distribution range is very large [
10]. Electromagnetic noise causes electromagnetic interference to the industrial environment, which reduces the signal-to-noise ratio of the communication system and affect the communication quality and the capacity of the communication system.
Noise is usually considered as additive white Gaussian noise in general scenes. However, the noise characteristics in IIoT scenes are quite different from the additive white Gaussian noise (AWGN) hypothesis [
11]. First, the noise generated by machines, motors, and other equipment may have a non-flat power spectrum, so the noise in the IIoT scene does not always meet the characteristics of white noise [
12]. Second, the frequency domain impulse noise generated by welding equipment in infrared environments does not obey the time domain Gaussian distribution, so the noise in IIoT scenes does not always obey the Gaussian distribution. Third, the noise process in IIoT scenes is not always stationary, and there often exist sudden noises. In addition, in order to prolong the service life, equipment in industrial applications is in a low power transmission mode, which makes them vulnerable to industrial noise generated by machines, motors, and other equipment [
13]. Therefore, it is meaningful to study the wireless network design of the Industrial Internet of Things.
At present, previous scholars had carried out research on electromagnetic noise in the factory environment. In [
14], the first-order Markov process was used to model noise in the factory environment, and noise in the factory was compared with additive white Gaussian noise. The analysis results showed that factory noise was not a conventional additive white Gaussian noise, and the properties between them were significantly different. However, the difference between noise in factories and additive white Gaussian noise was only discussed theoretically in this paper, and the actual measurement was not considered. In [
15], the time domain measurement of noise was carried out in paper mills and iron and steel mills by using a spectrum analyzer, and the amplitude probability distribution (APD) information of noise at 439 MHz, 440 MHz, 570 MHz, and 2450 MHz was provided. However, the time-varying characteristics of noise were not discussed in this paper. In [
16], the researchers measured the impulse noise generated by an automotive ignition system and an automotive electronic parts manufacturing system at 2.4 GHz frequency point, but the typical frequency bands of the IIoT system such as 315 MHz, 779 MHz, and other possible frequencies were not considered.
In this paper, log-period antennas and spectrum analyzers were used as measurement and analysis equipment. Based on the measurement results in the frequency domain, the frequency occupation in the factory environment was analyzed. According to the time domain measurement results at 315 MHZ, 779 MHz, and 916 MHZ, the amplitude probability distribution (APD), noise amplitude distribution (NAD), pulse duration distribution (PDD), and pulse separation distribution (PSD) of noise were extracted, and the characteristics of electromagnetic noise were investigated. Finally, the continuous hidden Markov models (CHMM) were used to model the time-varying characteristics of electromagnetic noise.
The rest of this article is organized as follows. The descriptions of the measurement environment and methods are provided in
Section 2. In
Section 3, we use four types of parameter extraction methods, APD, NAD, PDD, and PSD, to analyze the statistical properties of the gathered noises. The results of noise measurement in the frequency domain and time domain are given in
Section 4 and
Section 5, respectively. In
Section 6, a brief conclusion is given.
3. Statistical Parameters
As seen in
Figure 3, a simple noise baseband model was plotted [
17]. Three important characteristics of the electromagnetic noise baseband model were considered: the amplitude of the pulse, the duration of the pulse, and the interval between the two adjacent pulses. Four statistical parameters were employed to analyze the three characteristics of the noise spectrum [
17]: amplitude probability distribution (APD), pulse duration distribution (PDD), pulse separation distribution (PSD), and noise amplitude distribution (NAD). The APD and NAD were used to study the amplitude characteristics of pulses, PDD was used to analyze the duration of pulses, and PSD could well characterize the arrival interval of adjacent pulses.
The APD and NAD were used to characterize the amplitude of the pulse. The APD represents the probability that the pulse amplitude exceeds a certain threshold value. It can be calculated by the ratio of the total time when the pulse amplitude exceeds the threshold value to the total time measured as [
18].
where
is the noise amplitude and
is the threshold level; we refer to the method proposed in patent [
19] to determine it in the actual measurement and analysis process.
is the total measurement time,
tk is the duration of a pulse whose amplitude exceeds the fixed value. For instance, the APD of the signal in
Figure 3 is
. The influence of electromagnetic noise on bit error rate (BER) can be mapped directly to the APD value, which can directly reflect the size of the BER.
The NAD is defined as the number of pulses whose value is higher than the threshold divided by the total measurement time [
20,
21], and is expressed as
where
is the total measurement time and
is the number of pulses whose value is higher than the threshold. For instance, the NAD of the signal in
Figure 3 is
.
The PDD is used to analyze pulse duration. It means the probability that the duration of the pulse exceeds a threshold value
, which can be expressed by the ratio of the number of pulses lasting time beyond
to the total number of pulses as [
17]:
where
is the number of pulses lasting time beyond
and
is the total number of pulses.
PSD can well characterize the arrival interval of adjacent pulses. It means the probability that the interval of adjacent pulses exceeds a fixed value
[
16]. It can be expressed as
where
is the probability that the duration of the pulse exceeds the fixed value.
According to the above theory, the time domain electromagnetic noise signals at 315 MHz, 779 MHz, and 916 MHz frequencies were measured, and the APD, PDD, PSD, and NAD statistical parameters of electromagnetic noises at different frequencies could be extracted.
4. Characterization Using the Frequency Domain Measurement Data
The measurement results of the welder robot are shown in
Figure 4. As shown in
Figure 4a, the received signals contained electromagnetic noise, background noise, and operating mobile communication signals, such as 2G (Global System for Mobile Communications, GSM) and 4G (Long Term Evolution, LTE) signals.
Table 3 shows the spectrum occupancy rate of existing communication signals, in which the operating mobile communication signals were mainly between 900 MHz and 1800 MHz. In
Figure 4a, we can see that mobile communication signals had wideband spectrums and high power levels. However, the power of electromagnetic noise was small, mainly below 1 GHz.
Figure 4b shows that the results measured by horizontal and vertical polarization of the antenna were similar.
Low power wireless sensors are often used in industrial scenarios. The noise power has a great impact on the wireless links of sensors. Low power wireless sensors usually work in five frequency bands, i.e., 315 MHz, 433 MHz, 779 MHz, 868 MHz, and 916 MHz, and the bandwidth is usually from 10 kHz to 200 kHz. Therefore, we focused our measurements on the spectral bands near the above five bands.
Figure 5 shows the spectrum near the above frequency bands. Note that there were mobile communication signals in the frequency bands above 870 MHz and below 915 MHz. We extracted the average power levels of five bands of noise, as shown in
Table 4. According to the extracted results, we could see that the power of the electromagnetic noise was decreased with the increase of the frequency.
However, according to the principle of the transmission theory, the path loss increases with the increase of frequency. The frequency selection of the IIoT needs to consider not only noise but also path loss. Therefore, we needed to calculate the path loss in different frequency bands. Thus, we could judge which frequency band is most suitable for IIoT.
The expression of path loss in free space is:
where
is the frequency in MHz and
is the distance between the transmitter and the receiver. The unit is kilometers. We assumed that
was equal to 0.1 km. We calculated the path loss according to Equation (5) and the results are shown in
Table 5.
Taking 315 MHz as a reference, let its gain be 0. We defined the gain equal to the reduction of the average power of the noise minus the increase of the transmission loss. Taking 868 MHz as an example, the average noise power was reduced by 8.3 dB ((−111 dBm) − (−119.3 dBm)) compared to 315 MHz, and the path loss was increased by 8.8 dB (71.2 dB − 62.4 dB). Therefore, the gain of 916 MHz was −0.5 dB (8.3 dB − 8.8 dB). Similarly, we calculated the gain of other frequency bands. The specific results are listed in
Table 6. According to the results in
Table 6, the best choice of frequency of IIoT was 916 MHz, and the others were 315 MHz, 433 MHz, 868 MHz, and 779 MHz.
6. Conclusions
In recent years, edge computing and the Internet of Things (IOT) have provided a technological basis for the development of intelligent manufacturing. In order to support the intelligent interconnection and application of all kinds of equipment in the industrial field, edge computing should be close to or embedded in all kinds of equipment nodes in the industrial wireless network. In order to better access or embed edge computing on various device nodes of industrial wireless networks, the channel characteristics of the Industrial Internet of Things (IIoT) should be studied first. Therefore, it is very meaningful to study the wireless network design of the IIoT.
In this paper, the measurement results of electromagnetic noise at three different test positions were given in an automobile factory. The spectrum occupancy of a factory wireless environment in the 300 MHz–3 GHz band was obtained by frequency domain measurement. By calculating the average power of path loss and electromagnetic noise in different frequency bands, the optimal frequency band for the IIoT is 916 MHz. In time domain measurements, we analyzed the distribution of electromagnetic noise in different plant areas and in different frequency bands by measuring four statistical parameters of 315 MHz, 433 MHz, and 916 MHz. According to the measurement results, we concluded that the time-varying characteristics of electromagnetic noise can be characterized by CHMM.
According to the research results, we suggest that under an IIOT scenario, 916 MHz is the best frequency band for the equipment. The 779 MHz band with the most serious noise interference should be avoided as much as possible. The focus of this paper is the noise characteristics in an IIOT scene and its modeling analysis. Therefore, the research in this paper has some limitations. In the following research, the authors will further explore the channel and noise characteristics in the IIOT scene, understand the IIOT scene more deeply, and try to find effective methods to reduce electromagnetic noise interference in the IIOT scene.