WiFi-to-ZigBee cross-technology communication method based on neural network

Abstract

The invention discloses a WiFi-to-ZigBee cross-technology communication method based on a neural network, and relates to the technical field of cross-technology communication. According to the WiFi-to-ZigBee cross-technology communication method based on the neural network, a wireless signal processing assembly in a CTC processing flow is converted into a neural model, a neural network model of a structure from a WiFi transmitter end to a ZigBee receiver end is formed, a data set label is not needed, an optimal CTC load can be automatically deduced, the development complexity is effectively simplified, and the development efficiency is improved. Compared with the prior art, the method has good expandability, the average data packet receiving rate is 92.3%, the average symbol error rate is as low as 1.3%, and the method has relatively excellent error performance for WE-Bee and WIDE.

Description

[0001] A cross-technology communication method from WiFi to ZigBee based on neural network Technical Field

[0002] The present invention relates to the field of cross-technology communication technology, and in particular to a neural network-based WiFi to ZigBee cross-technology communication method. Background Art

[0003] Cross-technology communication (CTC) aims to enable direct communication between different protocols by simulating time-domain waveforms that can be detected and identified by the target device. CTC can alleviate the need for centralized IoT gateways for communication between different wireless technologies. Existing strategies use physical layer simulation to implement CTC communication from WiFi to ZigBee, in which the transmission process of the Wi-Fi OFDM modulator is reverse-engineered to obtain a payload that can generate a similar waveform, which is a signal simulation technology.

[0004] However, most existing CTC research work is static, that is, the parameter optimization process is technology-specific and has high limitations. Modeling the CTC process through a black-box machine learning model cannot meet the lightweight requirements. Summary of the Invention

[0005] (1) Technical problems solved

[0006] In view of the deficiencies in the prior art, the present invention provides a neural network-based WiFi to ZigBee cross-technology communication method to solve the above problems.

[0007] (2) Technical solution

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: A neural network-based WiFi to ZigBee cross-technology communication method, comprising the following steps:

[0009] S1, achieve physical layer cross-technology communication through neural network, perform QAM simulation, and generate payload;

[0010] S2, the OFDM modulator at the WiFi transmitter modulates the payload to generate a modulated signal;

[0011] S3. The modulated signal is received by the ZigBee receiver after O-QPSK modulation, completing the cross-technology communication between the WiFi transmitter and the ZigBee receiver.

[0012] The present invention is further configured as follows: the QAM simulation method in S1 includes:

[0013] Convert the DFT processing block and the IDFT processing block into neural networks, and generate a neural network model based on the DFT processing block and a neural network model based on the IDFT processing block;

[0014] Converting the QAM mapping processing block into a neural network and generating a neural network model based on the QAM mapping processing block;

[0015] Converting quantization processing into a neural network to generate a neural network model based on quantization processing;

[0016] Calculation of ZigBee signal phase shift.

[0017] The present invention is further configured such that the generation of the neural network model based on the DFT processing block and the neural network model based on the IDFT processing block includes:

[0018]

[0019]

[0020] Where X[n] is the frequency domain component, S[n] is the time domain signal, e-j2πniN is the basis function corresponding to the DFT processing block, and ej2πniN is the basis function corresponding to the IDFT;

[0021] Among them, the neural network model based on the DFT processing block is used to set the kernel of the transposed convolution layer to the real and imaginary parts of e-j2πniN;

[0022] The neural network model based on the IDFT processing block is used to set the kernel of the transposed convolutional layer to the real and imaginary parts of ej2πniN.

[0023] The present invention is further configured as follows: the generation of the neural network model based on the QAM mapping processing block includes:

[0024] Build a differentiable network with one-hot vectors for table lookup operations.

[0025] The present invention is further configured as follows: the generation of the neural network model based on quantization processing includes:

[0026] We implement QAM demodulation step-by-step using PyTorch and learn the scaling factor using a neural network layer embedded in PyTorch.

[0027] Specifically, if the input of the neural network quantizer is a three-dimensional vector, given as: Input = [B, S, 2], where B is the batch size, S is the number of symbols in each coding block, and 2 is a specific value representing the real and imaginary parts of the complex signal, the corresponding output shape is: Output = [B, S, 1], where B is the batch size, S is the number of symbols in each coding block, and 1 is the standard constellation point to be solved.

[0028] The present invention is further configured as follows: the calculation of the ZigBee signal phase shift includes:

[0029] Use torch.angle to obtain the phase of the time domain signal and calculate the phase difference between adjacent sampling points.

[0030] The present invention is further configured as follows: the QAM simulation includes analog simulation and digital simulation;

[0031] The simulation includes: stacking a neural network model based on a DFT processing block, a neural network model based on an IDFT processing block, a neural network model based on a QAM mapping processing block, and a neural network model based on quantization processing into an autoencoder simulation model, inputting a signal waveform into the autoencoder simulation model, outputting a reconstructed signal waveform, and performing simulation;

[0032] The digital simulation includes: constructing a simulation from the WiFi transmitter end to the ZigBee receiver end, calculating the ZigBee signal phase shift, and performing digital simulation.

[0033] The present invention is further configured as follows: after performing QAM simulation in S1, QAM post-simulation processing and channel coding simulation are performed in sequence to derive a binary payload as the payload.

[0034] The present invention is further configured as follows: the QAM post-emulation processing includes:

[0035] The ZigBee symbol is divided into four segments, each of which has the same duration as a Wi-Fi frame. A selective boundary flipping strategy is introduced, and an additional linear layer is added to select data subcarriers.

[0036] The present invention is further configured such that: the channel coding simulation includes: additional linear layer selection data.

[0037] (3) Beneficial effects

[0038] The present invention provides a neural network-based WiFi to ZigBee cross-technology communication method. It has the following beneficial effects:

[0039] The present invention converts the wireless signal processing components in the CTC processing flow into a neural model to construct a neural network model of the WiFi transmitter-to-ZigBee receiver structure. Without the need for dataset labels, the optimal CTC payload can be automatically derived, effectively simplifying development complexity and having good scalability. In addition, the average packet reception rate is 92.3%, and the average symbol error rate is as low as 1.3%, which has relatively excellent error performance compared to WE-Bee and WIDE. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a workflow diagram for NNCTC and conventional CTC in an embodiment of the present invention;

[0041] Figure 2 This is a flowchart of the WiFi and ZigBee workflow in an embodiment of the present invention;

[0042] Figure 3 A diagram of a conventional simulation process in an embodiment of the present invention;

[0043] Figure 4 A schematic diagram of the configuration from mathematical concepts to neural networks in an embodiment of the present invention;

[0044] Figure 5 Schematic diagram of the framework of an automatic encoder for QAM simulation in an embodiment of the present invention;

[0045] Figure 6 The constellation diagrams of analog simulation and digital simulation in the embodiment of the present invention;

[0046] Figure 7 Schematic diagram of traditional Post-QAM simulation in an embodiment of the present invention;

[0047] Figure 8 Schematic diagram of subcarriers selected for use in a fully connected layer for the present invention;

[0048] Figure 9 Schematic diagram of the layout of the experimental transmitter device in an embodiment of the present invention;

[0049] Figure 10 This is a diagram of an experimental scene in an embodiment of the present invention;

[0050] Figure 11 1 is a table showing the overall performance comparison in the embodiments of the present invention. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention.

[0052] See also Figure 1-11 The embodiment of the present invention provides the following technical solution: a neural network-based WiFi to ZigBee cross-technology communication method, comprising the following steps:

[0053] S1. Implementing cross-technology communication at the physical layer through a neural network, performing QAM simulation, then sequentially performing QAM post-simulation processing and channel coding simulation, and deriving a binary payload as the payload;

[0054] QAM simulation methods include:

[0055] As attached Figure 3 As shown, the DFT processing block and the IDFT processing block are converted into neural networks to generate a neural network model based on the DFT processing block and a neural network model based on the IDFT processing block:

[0056]

[0057]

[0058] Where X[n] is the frequency domain component, S[n] is the time domain signal, e-j2πniN is the basis function corresponding to the DFT processing block, and ej2πniN is the basis function corresponding to the IDFT;

[0059] Among them, the neural network model based on the DFT processing block is used to set the kernel of the transposed convolution layer to the real and imaginary parts of e-j2πniN;

[0060] The neural network model based on the IDFT processing block is used to set the kernel of the transposed convolution layer to the real and imaginary parts of ej2πniN;

[0061] Convert the QAM mapping processing block into a neural network and generate a neural network model based on the QAM mapping processing block:

[0062] The QAM mapper works like a lookup table, mapping bits to complex symbols according to the constellation diagram in the protocol. In order to convert the QAM mapping process into a neural network, it is necessary to process and convert non-differentiable operations, that is, to build a differentiable network with one-hot vectors for table lookup operations.

[0063] Convert the quantization process into a neural network and generate a neural network model based on the quantization process:

[0064] Implement QAM demodulation step by step using PyTorch and embed neural network layers in PyTorch to learn the scaling factors;

[0065] The calculation of ZigBee signal phase shift is implemented through the built-in functions in the neural network development framework:

[0066] Use torch.angle to obtain the phase of the time domain signal and calculate the phase difference between adjacent sampling points;

[0067] QAM simulation includes analog simulation and digital simulation;

[0068] The simulation includes: stacking a neural network model based on a DFT processing block, a neural network model based on an IDFT processing block, a neural network model based on a QAM mapping processing block, and a neural network model based on quantization processing into an autoencoder simulation model, inputting a signal waveform into the autoencoder simulation model, outputting a reconstructed signal waveform, and performing simulation;

[0069] The digital simulation includes: constructing a simulation from the WiFi transmitter end to the ZigBee receiver end, and calculating the ZigBee signal phase shift to perform digital simulation;

[0070] Post-QAM simulation processing includes:

[0071] The ZigBee symbol is divided into four segments, each of which has the same duration as a Wi-Fi frame. A selective boundary flipping strategy is introduced, and a linear layer is added to select data subcarriers.

[0072] Channel coding simulation includes: additional linear layer to select data subcarriers;

[0073] S2, the OFDM modulator at the WiFi transmitter modulates the payload to generate a modulated signal;

[0074] S3. The modulated signal is received by the ZigBee receiver after O-QPSK modulation, completing the cross-technology communication between the WiFi transmitter and the ZigBee receiver.

[0075] As a detailed description,

[0076] The basic process of OFDM-based Wi-Fi transmission and O-QPSK-based ZigBee reception used in traditional CTC simulation. The schematic diagram of the traditional CTC link is shown in the attached figure. Figure 2As shown, at the Wi-Fi transmitter end, the artificially crafted payload generates a bit sequence through channel coding. These bits are mapped to complex symbols according to the QAM constellation. The frequency component of the OFDM signal consists of complex symbols and multiple pilot symbols. The frequency component is converted into a time domain component by applying the inverse discrete Fourier transform IDFT. In addition, a repeated segment, namely the cyclic prefix CP, is appended to the end of each frame of the Wi-Fi signal. By designing the payload, the Wi-Fi transmitter can generate a signal, one of which can be detected and received by the ZigBee device. At the Bee receiver end, the RF front-end first captures the signal in the air. The analog-to-digital converter (ADC) samples the signal and outputs discrete-time samples s[n]. The O-QPSK demodulator in the ZigBee receiver calculates the phase shift from the conjugate product of s[n] and s[n-1]. If the phase shift is greater than 0°, the demodulator outputs '1', otherwise it outputs '0'. According to the DSSS process in IEEE802.15.4, after collecting 32 binary phase offsets, the ZigBee receiver maps this binary phase offset sequence to a 4-bit symbol.

[0077] Analog simulation and digital simulation follow the attached Figure 3 The simulation process is similar to the one shown, but the difference lies in the desired signal. The desired signal is transformed into the frequency domain through DFT, and then the frequency components are quantized to the nearest constellation symbols and standardized in the Wi-Fi protocol. The data bits corresponding to these QAM symbols are then derived. After completing these steps, the derived data bits are sent to the normal transmission path of Wi-Fi, including QAM mapping and IDFT, to generate a simulated signal close to the expected signal. The neural network-based cross-technology communication proposed in this application, namely NNCTC, covers these two simulation ideas by implementing waveform simulation and phase offset simulation through a neural network model.

[0078] To provide a detailed explanation of the mathematical basis for the transition from the physical layer to the neural network:

[0079] In wireless communications, signal space analysis is used to model modulation schemes including single-carrier amplitude / phase modulation and orthogonal frequency division multiplexing (OFDM). The symbol-to-signal modulation process is considered as a linear combination of the symbol si and the corresponding basis function, expressed as:

[0080]

[0081] Where sij represents the j-th dimension of si, Si(t) represents the time domain modulated signal, and φj(t) represents the j-th function in the basis function set {φ(t)};

[0082] Then convert the continuous form in the above formula into discrete time form, which is:

[0083]

[0084] Where Si[n] and φj[n] represent the sample of the original signal Si(t) and the basis function φj(t), respectively. After sequential modulation, the output signal sampling is:

[0085]

[0086] Where L is the number of signal samples per symbol. Since the symbols and basis functions are complex numbers, the complex-valued signal samples are decomposed into real and imaginary parts:

[0087]

[0088] Apply transposed convolutional layers and linear layers, as shown in the attached Figure 4 As shown in the figure, the real and imaginary parts are fed into the transposed convolution layer. The kernel parameters of the transposed convolution layer are set according to the basis function. The output of the transposed convolution layer is the component of the signal sample, and these components are combined together through a fully connected layer to generate the modulated signal.

[0089] In order to realize the detection of the simulation performance of neural network, as shown in the attached Figure 5 As shown in Figure 1, the QAM simulation is designed to be similar to the autoencoder structure. The required time domain signals u(t) and v(t) are the simulated signals generated by the autoencoder for QAM simulation. The discrete forms of the u(t) and v(t) signals are u[n] and v[n] respectively, where n is the sampling point. The loss value is calculated using the MSELoss function. The average loss value of u[n] and v[n] of N sampling points is expressed as:

[0090]

[0091] Where N is the total number of sampling points;

[0092] U[n] and V[n] are the DFT calculation results corresponding to u[n] and v[n] respectively. According to Parseval's theorem, the total energy of the time domain signal is equal to the total energy of the frequency domain signal, that is:

[0093]

[0094] Get the relationship between the loss value and the frequency domain signal:

[0095]

[0096] It can be seen that the gradient descent is performed according to the loss value, reducing the absolute value of the difference between the expected frequency domain component U[n] and the simulated frequency domain component V[n], and the nearest constellation point will be selected on the constellation diagram, that is, the simulation can be realized through the neural network.

[0097] In order to realize the detection of the digital simulation performance of neural network, as shown in the attached Figure 6 As shown, it can be seen that there are differences in the selection of constellation points between analog simulation and digital simulation:

[0098] ZigBee uses phase difference for demodulation, so some errors may occur during simulation. Figure 6 As shown in (b), according to the simulation strategy, selecting the nearest constellation point will lead to a larger phase error, which may cause the ZigBee demodulation accuracy to decrease. Due to this problem, the phase of the simulated signal is calculated and the loss value is calculated:

[0099] u[n] and v[n] represent the expected time domain signal and the simulated time domain signal respectively. The phase corresponding to u[n] is h[n], and the phase corresponding to v[n] is q[n]. The mean square error loss value MSELoss is:

[0100]

[0101] Where N is the number of signal sampling points;

[0102] Under the guidance of the loss function, the neural network will use the gradient descent method to minimize the gap between h[n] and q[n], and the phase of the simulated signal is ultimately closer to the phase of the desired signal.

[0103] Methods for implementing QAMDemapper / Mapper based on neural network NN include:

[0104] Use the linear layer or torch.masked_select() or Embedding layer in PyTorch to pre-load the constellation parameters predefined in IEEE802.11 into the Embedding dictionary or the weights of the linear layer to implement table lookup operations. However, structures involving index operations are often non-differentiable, and there is no way to transfer gradients, which will cause problems for our end-to-end structure. In order to achieve differentiable operations, a floating-point One-hotVertor structure that supports gradient transmission is constructed, and then operations are performed on the linear layer with standard constellation point weight parameters to implement differentiable Mapper or DeMapper operations.

[0105] Necessity research on QAM post-simulation processing and channel coding simulation:

[0106] The cyclic prefix introduced in the OFDM modulator is prone to distortion. The cyclic prefix CP is a signal fragment copied from the end of each frame and appended to the beginning. For OFDM in a Wi-Fi transmitter running in a 20MHz bandwidth, there are 64 subcarriers in total, so a 64-point IDFT is used. The duration of each frame signal is 3.2us, and a 0.8us CP is appended. Each WiFi frame lasts a total of 4us. In the receiving workflow, the OFDM demodulator first discards the first 0.8us long CP, and the remaining 3.2us long signal is passed to the DFT to derive the QAM symbol. In the simulation process, a ZigBee symbol (16us) is divided into 4 fragments, and the duration of each fragment is the same as that of a Wi-Fi frame, which is 4us, as shown in the attached figure. Figure 7 As shown in (a), due to the CP removal process, some ZigBee segments will not be processed in the QAM simulation, as shown in the attached figure. Figure 7 As shown in (b), this leads to inherent post-QAM simulation distortion. A selective boundary flipping strategy is adopted. By selective boundary flipping, the error area can be made more concentrated, thereby reducing the error in the simulation signal. Due to the DSSS scheme in the ZigBee receiver, the remaining errors will be corrected. However, the selective boundary flipping strategy still has severe distortion and becomes more complicated on the transmitter side.

[0107] In addition to CP, post-QAM simulation also needs to deal with the impact of pilot subcarriers and carefully select data subcarriers. The symbols on the pilot subcarriers are fixed, so we cannot manipulate these pilot subcarriers to transmit payload. Therefore, we must carefully select data subcarriers. In NNCTC, signal simulation distortion is reduced in a manner similar to traditional manual parameter adjustment. At the same time, a neural network for CP processing and subcarrier selection is constructed. The CP addition and removal process is modeled through linear layers, and other linear layers are attached to select data subcarriers.

[0108] The linear layers for CP processing and subcarrier selection are configured with fixed weights and are not trained. The main purpose of NN-based post-QAM simulation is to integrate these processes into the end-to-end process of CTC simulation. Therefore, the autoencoder can be trained to reduce the distortion introduced by CP processing and find the correct QAM constellation that meets the requirements. In order to construct the binary payload for CTC transmission, the bits are derived from the symbols. Figure 1 As shown in Figure 2, in the WiFi transmission path, the binary payload is encoded by the channel encoder and the encoded bits are mapped to QAM symbols. Therefore, after we infer the symbols from the QAM simulation model, we can derive the binary payload based on the QAM mapping and channel coding scheme.

[0109] Experimental verification

[0110] According to the above method, NNCTC is designed using PyTorch and run on an x86 architecture laptop. At the same time, an NVIDIA RTX 3060 graphics card with 12G memory is used to train the designed NNCTC. At the CTC evaluation level, USRPB210 with IEEE802.11b / gPHY and commercial WiFi network card RTL8812AU are used as the transmitter of NNCTC, and USRPB210 with IEEE802.15.4PHY and TICC2650 are used as receivers. USRPB210 is only used for evaluation purposes, which can help us obtain low-level physical layer information, such as symbol error rate.

[0111] The overall performance of NNCTC was evaluated, and each experiment was repeated ten times to calculate the average value. The experiments included different transmission distances, payload lengths, and transmit powers in indoor and outdoor scenarios. NNCTC was compared with WEBee and WIDE, and their symbol error type SER, packet reception rate PRR, and effective throughput Goodput were evaluated respectively. The comparison results are shown in the attached figure. Figure 11 As shown in the results, it was found that NNCTC was effective in optimizing traditional CTC.

Claims

1. A neural network-based WiFi to ZigBee cross-technology communication method, characterized by: The steps include: S1, achieve physical layer cross-technology communication through neural network, perform QAM simulation, and generate payload; S2, the OFDM modulator at the WiFi transmitter modulates the payload to generate a modulated signal; S3. The modulated signal is received by the ZigBee receiver after O-QPSK modulation, completing the cross-technology communication between the WiFi transmitter and the ZigBee receiver.

2. The neural network-based WiFi to ZigBee cross-technology communication method according to claim 1, characterized in that: The QAM simulation method in S1 includes: Convert the DFT processing block and the IDFT processing block into neural networks, and generate a neural network model based on the DFT processing block and a neural network model based on the IDFT processing block; Converting the QAM mapping processing block into a neural network to generate a neural network model based on the QAM mapping processing block; Converting quantization processing into a neural network to generate a neural network model based on quantization processing; Calculation of ZigBee signal phase shift.

3. The neural network-based WiFi to ZigBee cross-technology communication method according to claim 2, characterized in that: The generation of the neural network model based on the DFT processing block and the neural network model based on the IDFT processing block includes: Where X[n] is the frequency domain component, S[n] is the time domain signal, e-j2πniN is the basis function corresponding to the DFT processing block, and ej2πniN is the basis function corresponding to the IDFT; Among them, the neural network model based on the DFT processing block is used to set the kernel of the transposed convolution layer to the real and imaginary parts of e-j2πniN; The neural network model based on the IDFT processing block is used to set the kernel of the transposed convolutional layer to the real and imaginary parts of ej2πniN.

4. The neural network-based WiFi to ZigBee cross-technology communication method according to claim 3, characterized in that: The generation of the neural network model based on the QAM mapping processing block includes: Build a differentiable network with one-hot vectors for table lookup operations.

5. The neural network-based WiFi to ZigBee cross-technology communication method according to claim 4, characterized in that: The generation of the neural network model based on quantization processing includes: We implement QAM demodulation step by step using PyTorch and learn the scaling factor using a neural network layer embedded in PyTorch.

6. The neural network-based WiFi to ZigBee cross-technology communication method according to claim 5, characterized in that: The calculation of the ZigBee signal phase shift includes: Use torch.angle to obtain the phase of the time domain signal and calculate the phase difference between adjacent sampling points.

7. The neural network-based WiFi to ZigBee cross-technology communication method according to claim 6, characterized in that: The QAM simulation includes analog simulation and digital simulation; The simulation includes: stacking a neural network model based on a DFT processing block, a neural network model based on an IDFT processing block, a neural network model based on a QAM mapping processing block, and a neural network model based on quantization processing into an autoencoder simulation model, inputting a signal waveform into the autoencoder simulation model, outputting a reconstructed signal waveform, and performing simulation; The digital simulation includes: constructing a simulation from the WiFi transmitter end to the ZigBee receiver end, calculating the ZigBee signal phase shift, and performing digital simulation.

8. The neural network-based WiFi to ZigBee cross-technology communication method according to claim 1, characterized in that: After the QAM simulation is performed in S1, QAM post-simulation processing and channel coding simulation are performed in sequence to derive a binary payload as a payload.

9. The neural network-based WiFi to ZigBee cross-technology communication method according to claim 8, characterized in that: The QAM post-emulation processing includes: The ZigBee symbol is divided into four segments, each of which has the same duration as a Wi-Fi frame. A selective boundary flipping strategy is introduced, and an additional linear layer is added to select data subcarriers.

10. The neural network-based WiFi to ZigBee cross-technology communication method according to claim 8, characterized in that: The channel coding simulation includes: adding a linear layer to select data subcarriers.

Images