Keywords

1 Introduction

According to the prediction of the Global System for Mobile Communications assembly (GSMA), the number of global Internet of things (IoT) devices will reach about 24 billion in 2025. So many IoT devices bring challenges to the communication between different IoT devices. Traditionally, the method to realize the communication between heterogeneous IoT devices is to realize the indirect connection between heterogeneous IoT devices through IoT gateway. This will lead to an increase in cost, requiring Internet of things gateway equipment for transfer, slow data transmission and small traffic [1]. As a new research field, CTC has great application scenarios and good scientific research prospects [2]. According to different implementation schemes, CTC mainly includes packet-based CTC and signal-based CTC [3].

In packet-based CTC, the direct CTC of heterogeneous Internet of things devices is realized by embedding packet length, packet energy, and combined frame. Busybee [4] realized the CTC between WiFi devices and ZigBee devices and designed a scheme to encode channel access parameters. The system can correctly decode WiFi signals and ZigBee signals. Zifi [5] uses the unique interference signature generated by ZigBee radio through WiFi beacon to identify the existence of WiFi network. C-morse [6] It is the first to use traffic to implement CTC. When building recognizable wireless energy mode, c-morse slightly interferes with the WiFi packets. The packet-level CTC avoids hardware modifications, but it reduces the transmission rate and bandwidth.

Compared with packet-based CTC, the signal-based CTC will greatly improve throughput, which is conducive to improving throughput and expanding the application range of CTC [1]. TwinBee [7] realizes CTC by recovering chip errors introduced by imperfect signal simulation. LongBee [8] improves the reception sensitivity through new conversion coding, so as to realize CTC.

In this paper, the coding and decoding problem of the CTC signal is transformed into the classification problem of WiFi CSI. We extract several features of the WiFi CSI sequences, and then classify the CSI signal through machine learning classifiesr and neural network. We mark the CSI signal affected by ZigBee as “1” and the CSI signal not affected by ZigBee as “0”. Specifically, our major contributions are as follows:

  1. (1)

    We propose a CTC technology based on machine learning and neural network, from Zigbee to WiFi, using only WiFi CSI.

  2. (2)

    We use a variety of machine learning methods to classify CSI sequences. We extracted eight CSI sequence features and analyzed the accuracy of machine learning classifier using six machine learning classifiers to improve the classification accuracy of CSI sequences.

  3. (3)

    We use neural networks to classify CSI sequences, and neural network has a high accuracy. The experimental results show that the classification accuracy of CSI sequences by machine learning and neural network has reached a satisfactory level.

This paper consists of five sections, and the overall structure is as follows: The Sect. 2 introduces the preliminary work, the Sect. 3 introduces the system design, the Sect. 4 introduces the result analysis, and the Sect. 5 summarizes this paper.

2 Preliminary

2.1 The Spectrum Usage of ZigBee and WiFi

ZigBee is a new low-cost, low-power, and low-speed technology suitable for short-range wireless communication. It can be embedded in various electronic devices to support geographic positioning functions. This technology is mainly designed for low-speed communication networks. Different transmission speeds. WiFi and ZigBee use the 2.4 GHz wireless frequency band and adopt the direct sequence spread spectrum transmission technology (DSSS). ZigBee, transmission distance 50–300 m, rate 250 kbps, power consumption 5 mA. ZigBee is usually used in smart home. WiFi, fast speed (11Mbps), high power consumption, generally connected to the external power supply.

The spectrum usage of ZigBee and WiFi is shown in Fig. 1. Channel 1 of WiFi and channels 11, 12, 13, and 14 of ZigBee overlap, so we can try to achieve cross-technology communication from Zigbee to WiFi.

Fig. 1.
figure 1

The spectrum distribution

2.2 Channel State Information

In order to realize heterogeneous communication from Zigbee to WiFi, we need to analyze the changes of WiFi signals. Channel state information (CSI) is information used to estimate the channel characteristics of a communication link. Therefore, we use WiFi CSI information to analyze WiFi signals.

As shown in Fig. 2, the left figure shows the WiFi CSI signal when there is ZigBee, and the right figure shows the WiFi CSI signal when there is no ZigBee. It can be seen from the figure that ZigBee will affect the WiFi CSI signal. We can judge whether there is ZigBee by analyzing the WiFi CSI signal. Therefore, cross-technology communication from Zigbee to WiFi can be realized.

Fig. 2.
figure 2

The impact of ZigBee on WiFi CSI signal

2.3 The Support Vector Machines (SVM) Classifier

In this paper, we use machine learning classifiers to classify WiFi CSI signals. The experimental demonstrates that SVM classifier is the best classifier in our CSI sequence. Next, we introduce the SVM classifier.

Support vector machine (SVM) is a two class machine learning classifier. It is a supervised model, which is usually used for data classification of small samples. Support vector machine is the segmentation surface used to segment data points. Its position is determined by the support vector (if the support vector changes, the position of the segmentation surface will change). Therefore, this surface is a classifier determined by the support vector, that is, the support vector machine.

3 System Design

Figure 3 illustrates our system design, we first collect CSI data, then process the collected data, through the feature selection module and classification module, and finally analyze the classification results.

3.1 Hardware Setting

We conduct data acquisition on WiFi and ZigBee devices. We use the Intel 5300 network card as the WiFi device and the TelosB node as the ZigBee device. The transmission interval of WiFi packets is 0.5 ms and the length is 145 bytes. ZigBee packets are sent at an interval of 0.192 ms and 28 bytes in length. The experiment was conducted in a real environment. We extract some features of the WiFi CSI signal, and then classify the CSI signal through a machine learning classifier and neural network. We mark the CSI signal affected by ZigBee as “1” and the CSI signal not affected by ZigBee as “0”.

Fig. 3.
figure 3

System design

3.2 Feature Extraction

The length of the classifier window is 16, which can obtain the optimal classification accuracy and transmission rate. In each window, we extract 8 features of CSI sequence: variance, peak to peak, kurtosis, bias, standard deviation, mean, mode and median. We classify the extracted features of CSI sequences with machine learning classifiers, and the classification results will be analyzed in Sect. 4.

3.3 Machine Learning Classification Selection and Neural Network Design

We use machine learning classifiers such as complex tree, quadratic discriminator, cubic SVM, fine KNN, medium tree, bagged trees and logistic regression. The classification results will be analyzed in Sect. 4.

Long short term memory network (LSTM) is a kind of time recurrent neural network (RNN), LSTM avoids long-term dependence through deliberate design. LSTM neural network is more suitable for dealing with timing problems. Our CSI sequences are timing problems, so we can use LSTM to classify them. Figure 4 illustrates the LSTM network structure we use.

Fig. 4.
figure 4

The LSTM structure

4 Result Evaluation

4.1 Hardware

We experimented with off-the-shelf hardware. Figure 5 shows the placement of our transmitting antenna and receiving antenna. We used one WiFi transmitter and three WiFi receivers for the experiment. The distance between the transmitter and the receiver is about 100 cm, which can obtain better classification accuracy. ZigBee transmitter is between transmitting antenna and receiving antenna. The distance between the ZigBee transmitter and WiFi transmitting antenna and receiving antenna is about 50 cm.

Fig. 5.
figure 5

Experimental setup diagram

4.2 Evaluation of Experiment Results

We extract 8 features of CSI sequence and train them with 10 machine learning classifiers. The classification results are shown in Table 1. Different machine learning classifiers have different classification accuracy, among which SVM classifier has the highest accuracy. The accuracy of Cubic SVM is 93.8%.This is the highest accuracy of machine learning classifier, reaching a high level.

Then we use the LSTM network introduced in Sect. 3 for training. The accuracy of LSTM is 94.2%, which is higher than that of SVM in machine learning classifier. LSTM is more suitable for training time series. Our CSI sequence is time series, which improves the accuracy of CSI sequence classification.

Table 1. Classification results our dataset.

5 Conclusion and Next Work

We realize the cross-technology communication from Zigbee to WiFi through CSI classification. In future work, we will explore how to realize cross-technology communication from WiFi to ZigBee, and use other neural networks to classify CSI sequences. CTC technology is an important technology in the Internet of things, which can realize the communication between different Internet of things devices. There is still a lot of work to be done in the future.