Heterogeneous wireless network cross-technology communication method based on neural network

Abstract

The invention discloses a heterogeneous wireless network cross-technology communication method based on a neural network. The method comprises the following steps: 1, modulating LoRa; 2, performing fast Fourier transform on the signal to obtain a LoRa symbol; 3, constructing a CTC waveform simulation training model based on a neural network based on a LoRa symbol; and 4, introducing physical loss calculation, taking an FFT result in a physical loss function as probability distribution P of KL divergence, taking a to-be-simulated waveform result as probability distribution Q of the KL divergence, and taking a finally calculated KL divergence value as a result of physical loss of the physical information neural network to complete model training. Through introduction of the neural network, the signal simulation and communication process from Wi-Fi to LoRa is optimized, and more efficient long-distance radio communication is realized. The problems in the aspects of transmission rate, power consumption and signal receiving quality are solved, and the experimental result verifies the effectiveness of the method in the actual environment.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communication, and particularly relates to a cross-technology communication method for heterogeneous wireless networks based on neural networks. Background Art

[0002] Cross-Technology Communication (CTC) research has traditionally used waveform simulation strategies to achieve communication from WiFi to LoRa. Among them, simulating LoRa signals based on Complementary Code Keying (CCK) has many advantages. First, the CCK physical layer constraints in the 802.11b standard are relatively loose, making the simulation process simpler and more efficient. Second, compared with the Orthogonal Frequency Division Multiplexing (OFDM) physical layer, CCK does not require a Cyclic Prefix (CP) to avoid Inter-Symbol Interference (ISI), so it will not introduce additional simulation distortion and does not require additional optimization strategies to compensate.

[0003] The traditional waveform simulation process includes the following steps: First, generate a standard LoRa signal, then select the waveform (or phase) closest to the LoRa signal in the CCK codeword, then reverse the 802.11b physical layer modulation process to obtain the WiFi load signal, and finally send it to the LoRa receiver through a standard WiFi device. The advantage of this method is that the simulation process is simple, but its limitation is that it does not consider the demodulation process at the LoRa end and only matches in the time-domain waveform, which may lead to a performance decline in cross-technology communication. At the same time, the CCK codeword selection is based on a static mathematical formula and is difficult to adapt to the waveform simulation requirements in complex environments. Summary of the Invention

[0004] Object of the Invention: The object of the present invention is to provide a cross-technology communication method for heterogeneous wireless networks based on neural networks. The method of introducing neural networks in the traditional CCK simulation process will be optimized to improve the simulation accuracy of the WiFi-to-LoRa transmission and enhance its adaptability to complex scenarios.

[0005] Technical Solution: A cross-technology communication method for heterogeneous wireless networks based on neural networks according to the present invention includes the following steps:

[0006] Step 1: Using the Down-chirp signal as a reference, perform LoRa modulation to obtain the signal r(t);

[0007] Step 2: Perform a Fast Fourier Transform (FFT) on the signal r(t) to obtain the LoRa symbol;

[0008] Step 3: Construct a CTC waveform simulation training model based on neural networks based on the LoRa symbol;

[0009] Step 4: Introduce physical loss calculation. Use the FFT result in the physical loss function as the probability distribution P of the KL divergence, and the FFT result of the to-be-simulated LoRa waveform as the probability distribution Q of the KL divergence. The finally calculated KL divergence value is used as the result of the physical loss of the physical information neural network, and the training of the model is completed.

[0010] Further, step 1 is specifically as follows: In LoRa modulation, the frequency variation of the Chirp signal is used to carry information, expressed in the form of a complex exponential. Let the starting frequency corresponding to the symbol be f0, and the LoRa modulation signal is represented in complex form:

[0011]

[0012] where A is the amplitude of the signal, f0 is the starting frequency corresponding to the symbol, Δf is the frequency variation range, and T is the symbol duration. This formula represents a linearly frequency-modulated Chirp signal, where the frequency varies within time t, and the modulated signal is a complex-form sine wave, and the frequency is determined by f0 and determined;

[0013] In the demodulation process, the Down-chirp signal is used as a reference. The frequency of this signal gradually decreases and is expressed as:

[0014]

[0015] The frequency variation of the Down-chirp signal is opposite to that of the original signal. By performing matched filtering with this signal, the starting frequency is restored. The starting frequencies of different LoRa symbols are different. To correctly demodulate the LoRa symbol, the starting frequency is extracted from the received signal; first, multiply the received signal y(t) by the Down-chirp signal, and this process is carried out in the time domain through the following formula:

[0016]

[0017] By simplification: we get:

[0018] r(t) = A

[0019] At this time, to extract the frequency information, the DC ideal signal A is retained.

[0020] Further, step 2 is specifically as follows:

[0021] Perform FFT transformation on the signal r(t). The Fourier transform converts the time-domain signal into a frequency-domain signal, enabling the observation of the frequency components in the signal. By taking the modulus value of the result after FFT, a signal with strong energy in the frequency domain is obtained. The LoRa coding value corresponding to the strongest frequency-domain component is selected as the final LoRa symbol. Incorporate the demodulation knowledge of the LoRa physical layer into the training model, including waveform quantization, channel model, filtering, multiplying by the LoRa Down-Chirp signal, and the FFT module, and finally obtain the output of the network.

[0022] Further, step 3 specifically includes the following steps:

[0023] Step 3.1, CCK waveform quantization based on a neural network: For the signal s to be simulated, select the best CCK waveform x from the matrix X composed of 256 standard CCK waveforms. The formula is as follows;

[0024]

[0025] In the formula, R(·) and I(·) represent the real and imaginary parts of a complex number, C ij represents the baseband waveform of the i-th LoRa signal to be simulated, represents the baseband waveform of 256 standard CCK codewords, where n ∈ [0, 255], j ∈ [0, 7], and select a one-dimensional convolutional layer network as the core of the network;

[0026] Step 3.2, Use two one-dimensional convolutional layers to calculate the cross-correlation coefficients between the LoRa signal and each CCK waveform on the real and imaginary parts respectively, and then superimpose the values of the real and imaginary parts; To more accurately reflect the degree of coincidence between the LoRa signal and the CCK waveform, normalize the cross-correlation coefficient, that is, divide by the product of the squares of the modulus values of the two waveforms; Pre-store the standard CCK waveforms in the convolutional kernels of the one-dimensional convolutional layer and fix the parameters not to participate in the gradient update during training; To retain degrees of freedom for the neural network, a linear layer and a Relu activation function are placed after the cross-correlation coefficient normalization module, and the correlation is further processed through a linear classifier to output the prediction probability of each CCK symbol; Regard the cross-correlation coefficient values of 256 dimensions as probabilities to generate a differentiable one-hot vector, pass the gradient by manually spanning the one-hot vector, and then perform matrix multiplication between the one-hot vector and the standard CCK symbol matrix X to obtain the quantized CCK waveform.

[0027] Further, step 4 specifically includes the following steps:

[0028] Step 4.1. The loss function of PINNs consists of two parts: data loss and physical loss. The data loss measures the difference between the predicted values of the neural network and the actual observed data, while the physical loss measures whether the network output satisfies the physical information law. The overall loss function is expressed as:

[0029] Loss = αLoss data + βLoss phy

[0030] where α and β are weight coefficients used to balance the relative importance of data loss and physical loss. In order to use the physics-informed neural network to train the already constructed CCK waveform quantization neural network, the loss function Loss is constructed.

[0031] Step 4.2. First, for the data loss Lossdata, the MSE loss function is used to calculate the error value between the quantized CCK waveform and the target signal LoRa waveform. Secondly, for the physical loss Loss phy , a series of operations including the channel model, filtering, multiplying by LoRa, Down-Chirp signal, and FFT are included in the loss function. The Gaussian white noise channel is selected for the channel model according to the environment, and its mathematical model is described as:

[0032] y = x + n, n ~ N(0, σ 2 )

[0033] where n is the additive Gaussian white noise, which follows a complex Gaussian distribution; σ 2 is the variance of the noise, which is related to the signal-to-noise ratio SNR. x is the original signal at the transmitter, and y is the signal received at the receiver.

[0034] Step 4.3. The KL divergence function is selected to calculate the final physical loss. The KL divergence is a measure of the difference between two probability distributions. For two discrete probability distributions P and Q, the KL divergence is defined as:

[0035]

[0036] The FFT result in the physical loss function is used as P in the KL divergence, and the FFT result of the simulated LoRa waveform is used as Q in the KL divergence. The finally calculated KL divergence value is used as the result of the physical loss of the physics-informed neural network.

[0037] Step 4.4. The weight parameters of the data loss α and the physical loss β are set respectively.

[0038] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored on the memory. The processor executes the computer program to implement the steps of the method of the present invention.

[0039] The present invention also discloses a computer-readable storage medium, on which computer programs / instructions are stored, and when the computer programs / instructions are executed by a processor, the steps of the method of the present invention are implemented.

[0040] The present invention also discloses a computer program product, including computer programs / instructions, and when the computer programs / instructions are executed by a processor, the steps of the method of the present invention are implemented.

[0041] Beneficial effects: Compared with the prior art, the present invention has the following remarkable advantages:

[0042] In the present invention, the traditional CTC simulation process is reconstructed into a neural network, and then the demodulation knowledge of LoRa is incorporated into the loss function of the neural network to construct a physics-informed neural network. Through training, the physics-informed neural network not only approaches the standard LoRa signal in terms of waveform, but also approaches the frequency domain of the standard LoRa signal in the frequency domain. Experimental results show that the physics-informed neural network proposed in this chapter can successfully learn the CTC waveform simulation and can further improve the received signal-to-noise ratio SNR of the simulated waveform.

[0043] In the present invention, the LoRa receiver is simulated in MATLAB software, the WiFi preamble part of the simulated signal is removed, and then filtering, downsampling, multiplying by the Down-Chirp signal are performed in sequence, and then the time-frequency spectrum diagram is drawn. Although the LoRa signal of the traditional waveform simulation has a high correct rate, it is limited by the combination of a limited number of CCK waveforms, and a lot of messy noises are generated in the time-frequency spectrum. This will lead to a lower signal-to-noise ratio (SNR) in the actual transceiver process, thereby reducing the transmission distance. By comparison, it can be found that the LoRa signal simulated based on the physics-informed neural network has lower noise energy, that is, the color on the time-frequency spectrum diagram is lighter. Description of the Drawings

[0044] Figure 1 It is a training model diagram of the CTC waveform simulation based on a neural network.

[0045] Figure 2 It is a structural diagram of the CCK waveform quantization network.

[0046] Figure 3 It is a structural diagram of the physics-informed neural network.

[0047] Figure 4 It is the FFT result when the LoRa symbols '0' and '16' are demodulated. Among them, (a) is the FFT result when '0' is demodulated, and (b) is the FFT result when '16' is demodulated;

[0048] Figure 5It is a simulation result diagram; (a) is the LoRa signal diagram simulated by the model-driven neural network, and (b) is the LoRa signal diagram simulated by the physical neural network;

[0049] Figure 6 It is the received signal-to-noise ratio and received signal strength indication diagram. Specific implementation manners

[0050] The technical solution of the present invention will be further described below in conjunction with the accompanying drawings.

[0051] A heterogeneous wireless network cross-technology communication method based on neural network of the present invention includes the following steps:

[0052] Step 1: Using the Down-chirp signal as a reference, perform LoRa modulation to obtain the signal r(t);

[0053] Step 2: Perform fast Fourier transform FFT on the signal r(t) to obtain LoRa symbols;

[0054] Step 3: Build a CTC waveform simulation training model based on neural network based on LoRa symbols;

[0055] Step 4: Introduce physical loss calculation. Take the FFT result in the physical loss function as the probability distribution P of KL divergence, and the FFT result of the LoRa waveform to be simulated as the probability distribution Q of KL divergence. The finally calculated KL divergence value is used as the result of the physical loss of the physical information neural network to complete the training of the model.

[0056] Further, step 1 is specifically: in LoRa modulation, use the frequency change of the Chirp signal to carry information, expressed in complex exponential form. Let the starting frequency corresponding to the symbol be f0, and represent the LoRa modulation signal in complex form:

[0057]

[0058] where A is the amplitude of the signal, f0 is the starting frequency corresponding to the symbol, Δf is the frequency change range, and T is the symbol duration. This formula represents a linearly frequency-modulated Chirp signal, where the frequency changes within time t, and the modulated signal is a complex form of sine wave, and the frequency is determined by f0 and determined;

[0059] The demodulation process uses the Down-chirp signal as a reference, and the frequency of this signal gradually decreases, expressed as:

[0060]

[0061] The frequency change of the Down-chirp signal is opposite to that of the original signal. By performing matched filtering with this signal, the starting frequency is restored. The starting frequencies of different LoRa symbols are different. To correctly demodulate the LoRa symbol, the starting frequency is extracted from the received signal. First, the received signal y(t) is multiplied by the Down-chirp signal. This process is carried out in the time domain through the following formula:

[0062]

[0063] By simplification: we get:

[0064] r(t) = A

[0065] At this time, to extract the frequency information, the direct-frequency ideal signal A is retained.

[0066] Furthermore, step 2 is specifically as follows:

[0067] Perform FFT transformation on the signal r(t). The Fourier transform converts the time-domain signal into a frequency-domain signal, enabling the observation of the frequency components in the signal. By taking the modulus value of the result after FFT, the signal with strong energy in the frequency domain is obtained, and the LoRa coding value corresponding to the strongest frequency-domain component is selected as the final LoRa symbol. As Figure 4 shown, for the FFT results of LoRa symbol '0' and symbol '16', there are FFT energy peaks in their specific frequency domains, while the energy values in other frequency-domain components are 0, indicating no noise signal during demodulation.

[0068] After completing the training, the trained quantization model can be used to implement the CTC waveform simulation based on the neural network. As Figure 1 shown, the demodulation knowledge of the LoRa physical layer is incorporated into the training model, including waveform quantization, channel model, filtering, multiplying by the LoRa Down-Chirp signal, and FFT module, and finally the output of the network is obtained.

[0069] Furthermore, step 3 specifically includes the following steps:

[0070] Step 3.1, Neural-network-based CCK waveform quantization: For the signal s to be simulated, the best CCK waveform x is selected from the matrix X composed of 256 standard CCK waveforms. The formula is as follows;

[0071]

[0072] In the formula, R(·) and I(·) represent the real and imaginary parts of the complex number, and C ij represents the baseband waveform of the i-th LoRa signal to be simulated, Represents the baseband waveform of 256 standard CCK codewords, where n ∈ [0, 255], j ∈ [0, 7], and a one-dimensional convolutional layer network is selected as the core of the network;

[0073] And a one-dimensional convolutional layer network is selected as the core of the network;

[0074] Step 3.2, as Figure 2 shown, use two one-dimensional convolutional layers to calculate the cross-correlation coefficients between the LoRa signal and each CCK waveform on the real and imaginary parts respectively, and then superimpose the values of the real and imaginary parts; in order to more accurately reflect the degree of coincidence between the LoRa signal and the CCK waveform, normalize the correlation coefficients, that is, divide by the product of the sum of the squares of the moduli of the two waveforms; pre-store the standard CCK waveform into the convolutional kernel of the one-dimensional convolutional layer, and fix the parameters not to participate in the gradient update during training; in order to retain degrees of freedom for the neural network, a linear layer and a Relu activation function are placed after the correlation coefficient normalization module, and a linear classifier is used to further process the correlation, and the predicted probability of each CCK symbol is output; regard the cross-correlation coefficient values of 256 dimensions as probabilities to generate a differentiable one-hot vector, pass the gradient by manually striding across the one-hot vector, and then perform matrix multiplication on the one-hot vector and the standard CCK symbol matrix X to obtain the quantized CCK waveform.

[0075] Furthermore, step 4 specifically includes the following steps:

[0076] Step 4.1, the loss function of PINNs consists of two parts: including data loss and physical loss; where the data loss measures the difference between the predicted value of the neural network and the actual observed data; and the physical loss measures whether the network output satisfies the physical information law, and the overall loss function is expressed as:

[0077] Loss = αLoss data + βLoss phy

[0078] where α and β are weight coefficients used to balance the relative importance of data loss and physical loss. In order to use the physics-informed neural network to train the already constructed CCK waveform quantization neural network, the loss function Loss is constructed;

[0079] Step 4.2, first for the data loss Lossdata, use the MSE loss function to calculate the error value between the quantized CCK waveform and the target signal LoRa waveform. Secondly, for the physical loss Loss phy , include a series of operations such as the channel model, filtering, multiplying by LoRa, Down-Chirp signal, and FFT into the loss function. For the channel model, it is selected according to the actual environment. For example, for the Gaussian white noise channel, its mathematical model is described as:

[0080] y = x + n, n ~ N(0, σ 2 )

[0081] where n is additive white Gaussian noise, following a complex Gaussian distribution; σ 2 is the variance of the noise, related to the signal-to-noise ratio SNR, x is the original signal at the transmitter, and y is the signal received at the receiver;

[0082] Step 4.3: Select the KL divergence function to calculate the final physical loss. The KL divergence is a measure of the difference between two probability distributions; for two discrete probability distributions P and Q, the KL divergence is defined as:

[0083]

[0084] Take the FFT result in the physical loss function as P in the KL divergence, and the FFT result of the simulated LoRa waveform to be simulated as Q in the KL divergence. The finally calculated KL divergence value is used as the result of the physical loss of the physics-informed neural network; Figure 3 shows the basic structure of the physics-informed neural network of the present invention, mainly including the calculation processes of data loss and physical loss.

[0085] The loss function is written in the form of pseudocode in Table 1. It can be intuitively seen that before calculating the physical loss, we need to perform a series of pre-operations. Among them, Targets represents the standard LoRa waveform input to the network, and the Re and Im functions are used to find the real and imaginary parts of the complex signal respectively.

[0086] Table 1 Physics-informed neural network loss function

[0087]

[0088] Step 4.4: Set the weight parameters of the data loss α and the physical loss β respectively. In the simplest case, the respective weights can be set to 0.5, that is, the waveform loss and the frequency domain loss are equally important.

[0089] In the indoor environment of the present invention, the transmission distance between the fixed transmitter and receiver is 3 meters. Then, let the USRP regularly send three types of LoRa frames, namely the standard LoRa signal, the LoRa signal simulated by the traditional waveform, and the LoRa signal simulated by the physics-informed neural network. The LoRa payload lengths of these three signals are all set to 3 bytes, and the coding rate is 4 / 5. Using the USRP at the transmitter can more conveniently unify the power and transmission gain values of the transmitted signal. At the receiver, the present invention derives the RSSI value and SNR value from a commercial LoRa device.

[0090] By Figure 6It can be seen that under different bandwidth and SF configurations, the LoRa signals simulated by the physics-informed neural network are slightly better than or equal to the traditional waveform simulation in terms of SNR and RSSI metrics. Moreover, the performance of both the traditional waveform simulation and the physics-informed neural network waveform simulation is weaker than that of the standard LoRa because errors inevitably occur during the simulation process, resulting in a decrease in signal quality. Through experiments, it can be concluded that the waveform simulation based on the physics-informed neural network improves the signal-to-noise ratio and signal strength indication value at the LoRa receiver to a certain extent.

Claims

1. A cross-technology communication method for heterogeneous wireless networks based on neural networks, characterized in that It includes the following steps: Step 1: Using the Down-chirp signal as a reference, perform LoRa modulation to obtain the signal r(t); Step 2: Perform a fast Fourier transform (FFT) on the signal r(t) to obtain LoRa symbols; Step 3: Based on the LoRa symbols, construct a CTC waveform simulation training model based on a neural network; Step 4: Introduce physical loss calculation. Take the FFT result in the physical loss function as the probability distribution P of the KL divergence, and the FFT result of the LoRa waveform to be simulated as the probability distribution Q of the KL divergence. The finally calculated KL divergence value is used as the result of the physical loss of the physical information neural network to complete the training of the model.

2. The cross-technology communication method for heterogeneous wireless networks based on neural network according to claim 1, wherein Specifically, Step 1 is as follows: In LoRa modulation, the frequency change of the Chirp signal is used to carry information, which is expressed in complex exponential form. Let the starting frequency corresponding to the symbol be f0, and the LoRa modulation signal is represented in complex form: where A is the amplitude of the signal, f0 is the starting frequency corresponding to the symbol, Δf is the frequency change range, and T is the symbol duration. This formula represents a chirp signal with linear frequency modulation, where the frequency changes over time t, and the modulated signal is a complex-form sine wave with the frequency determined by f0 and determined; The demodulation process uses the Down-chirp signal as a reference. The frequency of this signal gradually decreases and is expressed as: The frequency change of the Down-chirp signal is opposite to that of the original signal. By performing matched filtering with this signal, the starting frequency is restored. The starting frequencies of different LoRa symbols are different. To correctly demodulate LoRa symbols, extract the starting frequency from the received signal; first, multiply the received signal y(t) by the Down-chirp signal. This process is carried out in the time domain through the following formula: By simplification: Obtain: r(t) = A At this time, to extract frequency information, retain the direct frequency ideal signal A.

3. A cross-technology communication method for heterogeneous wireless networks based on neural networks according to claim 1, characterized in that, Specifically, Step 2 is as follows: Perform an FFT transformation on the signal r(t). The Fourier transform converts the time-domain signal into a frequency-domain signal, enabling the observation of the frequency components in the signal. By taking the modulus value of the result after FFT, obtain the signal with strong energy in the frequency domain, and select the LoRa coding value corresponding to the strongest frequency-domain component as the final LoRa symbol. Incorporate the demodulation knowledge of the LoRa physical layer into the training model, including waveform quantization, channel model, filtering, multiplying by the LoRa Down-Chirp signal, and the FFT module, and finally obtain the output of the network.

4. A cross-technology communication method for heterogeneous wireless networks based on neural networks according to claim 1, characterized in that Specifically, Step 3 includes the following steps: Step 3.1: CCK waveform quantization based on a neural network: For the signal s to be simulated, select the best CCK waveform x from the matrix X composed of 256 standard CCK waveforms. The formula is as follows: In the formula, R(·) and I(·) represent the real and imaginary parts of a complex number, and C ij represents the baseband waveform of the i-th LoRa signal to be simulated, represents the baseband waveform of the 256 standard CCK codewords, where n ∈ [0, 255], j ∈ [0, 7], and a one-dimensional convolutional layer network is selected as the core of the network; Step 3.2: Use two one-dimensional convolutional layers to calculate the cross-correlation coefficients between the LoRa signal and each CCK waveform on the real and imaginary parts respectively, and then superimpose the values of the real and imaginary parts. To more accurately reflect the degree of coincidence between the LoRa signal and the CCK waveform, normalize the correlation coefficients, that is, divide by the product of the sum of the squares of the moduli of the two waveforms. Pre-store the standard CCK waveform into the convolutional kernel of the one-dimensional convolutional layer and fix the parameters not to participate in the gradient update during training. To retain degrees of freedom for the neural network, a linear layer and a Relu activation function are placed after the correlation coefficient normalization module, and the correlation is further processed through a linear classifier to output the prediction probability of each CCK symbol. Treat the cross-correlation coefficient values of 256 dimensions as probabilities to generate a differentiable one-hot vector, pass the gradient through manually spanning the one-hot vector, and then perform matrix multiplication between the one-hot vector and the standard CCK symbol matrix X to obtain the quantized CCK waveform.

5. A cross-technology communication method for heterogeneous wireless networks based on neural networks according to claim 1, characterized in that Step 4 specifically includes the following steps: Step 4.1: The loss function of PINNs consists of two parts: including data loss and physical loss. Among them, the data loss measures the difference between the predicted value of the neural network and the actual observed data. And the physical loss measures whether the network output satisfies the physical information law, and the overall loss function is expressed as: Loss = αLoss data + βLoss phy where α and β are weight coefficients used to balance the relative importance of data loss and physical loss. To use the physics-informed neural network to train the already constructed CCK waveform quantization neural network, construct the loss function Loss; Step 4.2: First, for the data loss Lossdata, use the MSE loss function to calculate the error value between the quantized CCK waveform and the target signal LoRa waveform. Secondly, for the physical loss Lossphy, include a series of operations such as the channel model, filtering, multiplying by LoRa, Down-Chirp signal, and FFT into the loss function. Select the Gaussian white noise channel for the channel model according to the environment, and its mathematical model is described as: y = x + n, n ~ N(0, σ 2 ) where n is additive white Gaussian noise, which follows a complex Gaussian distribution; σ 2 is the variance of the noise, related to the signal-to-noise ratio SNR, x is the original signal at the transmitter, and y is the signal received at the receiver; Step 4.3: Select the KL divergence function to calculate the final physical loss. KL divergence is a measure of the difference between two probability distributions. For two discrete probability distributions P and Q, the KL divergence is defined as: Take the FFT result in the physical loss function as P in the KL divergence, and the FFT result of the to-be-simulated LoRa waveform as Q in the KL divergence. The finally calculated KL divergence value is used as the result of the physical loss of the physics-informed neural network. Step 4.4: Set the weight parameters of the data loss α and the physical loss β respectively.

6. A computer device, comprising a memory, a processor, and a computer program stored on the memory, characterized in that, The processor executes the computer program to implement the steps of the method described in claim 1.

7. A computer-readable storage medium having computer programs / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, the steps of the method described in claim 1 are implemented.

8. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, the steps of the method described in claim 1 are implemented.

Images