Demodulation method for orthogonal time-frequency air conditioning communication system based on deep learning

Abstract

The invention provides an orthogonal time-frequency air conditioning communication system demodulation method based on deep learning. The method comprises the following steps: constructing an orthogonal time-frequency air conditioning communication system model and an interference model; according to communication system characteristics and anti-interference requirements, constructing a network structure combined by a convolutional neural network and a bidirectional long-short-term memory network; preprocessing the training data, and performing time sequence alignment on the input data and the label data; inputting the preprocessed data into a CNN (Convolutional Neural Network) to extract local spatial features, inputting the data into a combined network, capturing time-dependent features, calculating a mean square error loss function according to a predicted demodulation bit and a real bit, calculating a network parameter gradient by using a back propagation algorithm, and updating network parameters by using an optimizer according to the gradient; and carrying out model training until expected requirements are met. According to the invention, the advanced neural network architecture is innovatively combined, so that the communication reliability and robustness of the communication system in the interference environment are effectively improved.

Description

Technical Field

[0001] The present application relates to the technical field of wireless communication anti-interference, and in particular to a demodulation method for an orthogonal time-frequency-space modulation communication system based on deep learning. Background Art

[0002] With the rapid development of wireless communication technology, Orthogonal Time-Frequency-Space (OTFS) modulation has become a research hotspot due to its outstanding performance in high-speed mobile scenarios and multipath fading channels. However, in real-world communication environments, wireless signals are extremely susceptible to various interferences, such as narrowband interference, wideband interference, and malicious interference. These interferences can severely degrade the performance of communication systems, leading to reduced signal transmission quality, increased bit error rates, and even communication interruptions, compromising the accuracy and reliability of communication data.

[0003] To address the anti-interference problem in OTFS communication systems, patent application number 202410032401.9 discloses an anti-interference method based on traditional signal processing algorithms. This method uses adaptive filtering technology to filter the received signal, suppressing the impact of interference signals and improving signal quality to a certain extent. Separately, patent application number 202210817252.8 proposes an anti-interference solution based on machine learning. This utilizes a machine learning model to identify and classify interference signals, and then adopts corresponding anti-interference strategies, achieving good anti-interference results.

[0004] While these methods have achieved some success in interference mitigation, they still have some shortcomings. Traditional signal processing algorithms typically require relatively accurate prior knowledge of the characteristics of interfering signals. This significantly limits their anti-interference performance when the interference environment is complex and changing. Existing machine learning methods, on the other hand, mostly consider only single signal characteristics, such as those in the time or frequency domain. This makes it difficult to fully capture the complex characteristics of the signal in the time-frequency space, resulting in insufficient anti-interference capabilities in complex interference environments. Furthermore, these methods often fail to adjust their anti-interference strategies in a timely manner when dealing with dynamic interference, and cannot adapt well to the rapid changes in interference signals, thus affecting the stability and reliability of the communication system. Summary of the Invention

[0005] The present application provides a demodulation method for an orthogonal time-frequency-space modulation communication system based on deep learning, which can be used to solve the technical problem in the prior art that machine learning only considers a single feature of the signal.

[0006] The present application provides a deep learning-based orthogonal time-frequency-space modulation communication system demodulation method, the method comprising:

[0007] Step 1: Construct an Orthogonal Time Frequency and Space (OTFS) communication system model and an interference model;

[0008] Step 2: According to the characteristics of the OTFS communication system and the anti-interference requirements, construct a network structure combining a Convolutional Neural Network (CNN) and a Bidirectional Long Short-Term Memory (Bi-LSTM);

[0009] Step 3: Perform preprocessing including normalization and denoising on the training data, and then align the input data and label data in time series;

[0010] Step 4: The preprocessed data is input into the CNN to extract local spatial features, and then input into the Bi-LSTM composed of forward and backward LSTM layers to capture time-dependent features. Calculate the mean square error loss function based on the predicted demodulated bits and the true bits, use the backpropagation algorithm to calculate the gradient of the network parameters, and use the optimizer to update the network parameters according to the gradient;

[0011] Step 5: Conduct model training until the expected requirements are met, and apply the trained model to the signal demodulation of the OTFS communication system to resist interference and improve communication reliability.

[0012] Further, in Step 1, constructing an OTFS communication system model and an interference model includes:

[0013] In the delay-Doppler domain, the received signal X[m,n] is transformed to the time-frequency domain through the Inverse Short-Time Fourier Transform (ISFFT):

[0014]

[0015] where, X tf [l,k] is the signal in the time-frequency domain after the ISFFT transformation, l is time, k is the frequency index, and the subscript tf indicates that the signal is in the time-frequency domain, X[m,n] is the symbol matrix in the delay-Doppler domain, m is the Doppler index, n is the delay index, N is the total number of Doppler units, and M is the total number of delay units;

[0016] Then use the Heisenberg transform to generate the continuous-time domain signal s(t):

[0017]

[0018] where \(t\) represents time, \(g\) tx is the transmit pulse shaping function, and the subscript tx indicates that this is the transmit signal, \(T\) is the symbol duration, and \(\Delta f\) is the subcarrier spacing; the received signal \(r(t)\) is subjected to matched filtering and Wigner transform to obtain the time-frequency domain signal \(Y\) tf [l,k]:

[0019]

[0020] where is the conjugate of the receive pulse shaping function; the delayed Doppler domain signal \(Y[m,n]\) is reconstructed by the finite Fourier transform SFFT:

[0021]

[0022] The interference signal \(j(t)\) includes multiple interference components, and each interference component is represented by a complex exponential function:

[0023]

[0024] where \(J\) p is the amplitude of the \(p\)th interference signal, \(f\) j,p is the frequency of the \(p\)th interference, \(\varphi\) p is the phase of the \(p\)th interference, and \(P\) is the total number of interference signal components; the interference frequency \(f\) j,p not only includes the original interference frequency but also the influence of the Doppler shift. The original frequency of the \(p\)th interference signal is expressed as:

[0025]

[0026] where \(v\) r is the relative radial velocity between the interference source and the receiver, and \(c\) is the speed of light; the OTFS signal \(r(t)\) after being transmitted through the wireless channel is collected by the receiving end device, and the expression is:

[0027] \(r(t)=\int h(\tau,\nu)s(t - \tau)e\) j2πνt d\tau dv + j(t)+n(t)

[0028] where \(h(\tau,v)\) is the impulse response of the channel in the delay-Doppler domain, \(\tau\) represents the delay, \(v\) represents the Doppler shift, \(s(t)\) is the transmit signal, \(j(t)\) is the interference signal, and \(n(t)\) is the additive white Gaussian noise.

[0029] Furthermore, the combined network structure includes a combined network consisting of a Convolutional Neural Network (CNN) and a Bidirectional Long Short-Term Memory Network (Bi-LSTM). The CNN is responsible for extracting local spatial features from the input delay-Doppler domain signal, while the Bi-LSTM is used to capture the temporal dependencies of the signal. The two work closely together, with the output of the CNN serving as the input to the Bi-LSTM, enabling the network to comprehensively learn the complex features of OTFS signals in the time-frequency space and effectively tackle the signal demodulation problem in an interference environment.

[0030] The structural method of the combined network includes:

[0031] Build a CNN component with two convolutional layers. Each convolutional layer performs a convolution operation on the input delay-Doppler domain signal. The convolutional kernel slides over the signal to extract local spatial features. The first convolutional operation is:

[0032]

[0033] where, is the value at the (i, j) position of the output feature map of the first convolutional layer, W1 (u,v) is the value at the (u, v) position of the convolutional kernel with size K l ×K w The value of K l is the size of the convolutional kernel in the length direction, and K w is the size of the convolutional kernel in the width direction. Y (i+u,j+v) is the value of the input signal at the (i + u, j + v) position, and b1 is the bias term of the first convolutional layer. After each convolutional layer, a ReLU activation function is connected to introduce non-linear operations:

[0034]

[0035] The value at the (i, j) position after being processed by the ReLU activation function, The value input to the ReLU activation function, which is the output of the first convolutional layer. Then, batch normalization is used. By normalizing the output of the convolutional layer, the internal covariate shift is reduced, the training process is stabilized, and the convergence is accelerated:

[0036]

[0037] where, is the value at the (i, j) position after normalization. μ and σ 2 represent the mean and variance of the feature map respectively, which are used to centralize and scale the feature map. ∈ is a numerically stable constant used to prevent the denominator from being zero during the calculation process. Then, scaling and offset are performed through parameter γ scaling and β offset, that is Through these operations, the internal covariate shift is effectively reduced, accelerating the model convergence. After that, the second convolutional operation is performed:

[0038]

[0039] Among them, is the value of the output feature map of the second convolutional layer at the (i, j) position, and b2 is the bias term of the second convolutional layer;

[0040] Then, a Bi-LSTM network is constructed. Using its bidirectional processing characteristics, the signals processed by the CNN are processed separately from the forward and backward directions. The forward LSTM layer processes the signal sequence in chronological order. At each time step, the current hidden state and cell state are combined with the current input and the previous hidden state to calculate the current hidden state. The backward LSTM layer processes the signal sequence in reverse order and calculates the hidden state at each time step in the same way. The output of the forward LSTM hidden state is:

[0041]

[0042] Among them, h t is the output of the hidden state of the forward LSTM layer at time step t, which synthesizes the current and previous information and participates in subsequent calculations. LSTM → is the calculation process of the forward LSTM, including input, forget, output gates, and cell state update, integrating the input and previous states to update the hidden state. x t is the input signal at time step t, providing real-time information for the current calculation, is the hidden state of the forward LSTM layer at time step t - 1, carrying the feature information of the previous time step, is the cell state of the forward LSTM layer at time step t - 1;

[0043] The output of the backward LSTM hidden state is:

[0044]

[0045] Among them, is the output of the hidden state of the backward LSTM layer at time step t, capturing the reverse time-dependent features. LSTM ← is the calculation process of the backward LSTM, processing the signal from back to front and updating the hidden state. D t is the data input to the backward LSTM layer at time step t (the data processed by the CNN), which is the input for the current calculation, is the hidden state of the backward LSTM layer at time step t + 1. When processing in reverse, the state of the next time step participates in the current calculation, is the cell state of the backward LSTM layer at time step t + 1, used for reverse long-term dependence information transmission and current hidden state calculation;

[0046] Concatenate the outputs of the processed forward and backward LSTM layers at each time step to obtain features containing past and future context dependencies, effectively capture the temporal correlation of interference signals, and enhance the model's ability to understand signal sequences.

[0047] Further, in step 3, perform preprocessing including denoising and normalization on the training data, and then align the input data and label data in time series, including:

[0048] Filter the obtained signal to remove high-frequency or low-frequency noise and improve the signal-to-noise ratio of the signal; through signal calibration, compensate for the signal amplitude and phase changes caused by channel transmission to ensure the accuracy of the signal; perform normalization processing, and use the formula for the signal x Perform normalization, where μ is the mean of the signal and σ is the standard deviation;

[0049] Use the processed data as the input, and the corresponding original transmitted signal b transmitted As data, align the input data and label data in chronological order to form an ordered training data pair; store the ordered training data pair in a dynamic pool, and the dynamic pool manages the data according to the first-in-first-out (FIFO) principle, providing continuous time series data for the training of the Bi-LSTM to ensure that the model can learn the temporal characteristics of the signal.

[0050] Further, in step 4, the preprocessed data is input into a CNN to extract local spatial features, and then input into a Bi-LSTM composed of forward and backward LSTM layers to capture temporal dependence features. Calculate the mean square error loss function based on the predicted demodulation bits and the true bits, use the backpropagation algorithm to calculate the gradient of the network parameters, and use an optimizer to update the network parameters according to the gradient, including:

[0051] Initialize the convolutional kernel weights W, biases b of the convolutional layer in the CNN component, the weights and gating parameters of the Bi-LSTM component, as well as the learning rate and batch size. Input the training data into the anti-interference demodulator for forward propagation to calculate the output of the model; according to the difference between the output and the label data, use the mean square error loss function:

[0052]

[0053] Calculate the loss value, where Nbatch is the number of samples in the batch, is the predicted demodulation bit, is the true transmitted bit; calculate the gradient of the loss value with respect to the network parameters through the backpropagation algorithm, and use an optimization algorithm (such as stochastic gradient descent) to update the network parameters; the parameter update formula for stochastic gradient descent is:

[0054]

[0055] where θ is the network parameter, α is the learning rate, is the gradient of the loss function at θ t During the training process, the cross-validation method is adopted. The training data is divided into multiple subsets, which are used as the training set and the validation set in turn to evaluate the generalization ability of the model and avoid overfitting. According to the performance of the validation set, the training parameters such as the learning rate and the number of training epochs are adjusted in a timely manner to optimize the model performance.

[0056] Further, in step 5, the model is trained until the expected requirements are met, and the trained model is applied to the signal demodulation of the OTFS communication system, including:

[0057] Deploy the trained demodulator to the receiving end. First, preprocess the received signal with interference, then restore the signal through the demodulator, and finally perform error detection and error correction on the demodulated signal. Evaluate the demodulation performance by calculating the bit error rate; perform error detection on the demodulated signal, and measure the demodulation performance by calculating the bit error rate BER. The calculation formula is:

[0058]

[0059] where N errors is the number of bits with demodulation errors, and N total is the total number of transmitted bits. The demodulation effect is evaluated in real time through the monitoring index.

[0060] Compared with the prior art, the present invention has the following remarkable advantages: 1. Significantly improved anti-interference performance: This demodulator combines the advantages of CNN and Bi-LSTM, and can effectively extract the spatial and temporal features of OTFS signals. Under different interference intensities, the anti-interference ability can be significantly improved. Experimental results show that under nearly 30 dB of interference, the same bit error rate as the demodulator using only CNN can be achieved, effectively guaranteeing the reliability of communication; 2. Enhanced convergence stability: The application of batch normalization and soft update strategy effectively enhances the convergence stability of the model. Batch normalization reduces the internal covariate shift and accelerates the training process; the soft update strategy avoids parameter mutations, enabling the model to better adapt to complex dynamic interference scenarios and ensuring reliable and efficient convergence of training; 3. Moderate computational resource requirements: After a detailed analysis of the algorithm complexity, it is found that the demodulator has moderate computational resource requirements while ensuring performance; this makes the present invention suitable for the deployment of various communication devices and solves the application problem under resource constraints. Description of the Drawings

[0061] Figure 1 It is the overall system architecture diagram in the demodulation method of the orthogonal time-frequency space modulation communication system based on deep learning of the present invention.

[0062] Figure 2 This is a schematic diagram of the OTFS modulation and signal transmission process in the demodulation method of the orthogonal time frequency and space modulation communication system based on deep learning of the present invention.

[0063] Figure 3 This is a detailed diagram of the demodulator network architecture in the demodulation method of the orthogonal time frequency and space modulation communication system based on deep learning of the present invention.

[0064] Figure 4 This is a comparison chart of the performance of different demodulators in the demodulation method of the orthogonal time frequency and space modulation communication system based on deep learning of the present invention. Detailed implementation manners

[0065] To make the objectives, technical solutions, and advantages of this application clearer, the following will further describe the implementation manners of this application in detail with reference to the accompanying drawings.

[0066] First, the embodiments of this application will be introduced below with reference to the accompanying drawings.

[0067] This application provides a demodulation method for an orthogonal time frequency and space modulation communication system based on deep learning, and the method includes:

[0068] Step 1: Construct an orthogonal time frequency and space (OTFS) communication system model and an interference model;

[0069] Step 2: Construct a network structure combining a convolutional neural network (CNN) and a bidirectional long short-term memory network (Bi-LSTM) according to the characteristics of the OTFS communication system and the anti-interference requirements;

[0070] Step 3: Perform preprocessing including normalization and denoising on the training data, and then align the input data and label data in time series;

[0071] Step 4: The preprocessed data is input into the CNN to extract local spatial features, and then input into the Bi-LSTM composed of forward and backward LSTM layers to capture time-dependent features. Calculate the mean square error loss function according to the predicted demodulation bits and the true bits, calculate the gradient of the network parameters using the backpropagation algorithm, and update the network parameters according to the gradient using an optimizer;

[0072] Step 5: Perform model training until the expected requirements are met, and apply the trained model to the signal demodulation of the OTFS communication system to resist interference and improve communication reliability.

[0073] Further, in step 1, construct an OTFS communication system model and an interference model, including:

[0074] In the delay-Doppler domain, transform the received signal X[m,n] to the time-frequency domain through the Inverse Short-Time Fourier Transform (ISFFT):

[0075]

[0076] where X tf [l,k] is the signal in the time-frequency domain after ISFFT transformation, l is time, k is the frequency index, and the subscript tf indicates that the signal is in the time-frequency domain. X[m,n] is the symbol matrix in the delay-Doppler domain, m is the Doppler index, n is the delay index, N is the total number of Doppler cells, and M is the total number of delay cells;

[0077] Then, generate the continuous-time domain signal s(t) using the Heisenberg transform:

[0078]

[0079] where t represents time, and g tx is the transmit pulse shaping function, and the subscript tx indicates that this is the transmit signal. T is the symbol duration, and Δf is the subcarrier spacing. Perform matched filtering and Wigner transform on the received signal r(t) to obtain the time-frequency domain signal Y tf [l,k]:

[0080]

[0081] where is the conjugate of the receive pulse shaping function; reconstruct the delay-Doppler domain signal Y[m,n] through the finite Fourier transform SFFT:

[0082]

[0083] The interference signal j(t) includes multiple interference components, and each interference component is represented by a complex exponential function:

[0084]

[0085] where J p is the amplitude of the p-th interference signal, f j,p is the frequency of the p-th interference, φ p is the phase of the p-th interference, and P is the total number of interference signal components; the interference frequency f j,pIt not only contains the original interference frequency but also the influence of Doppler frequency shift. The original frequency of the p-th interference signal is expressed as:

[0086]

[0087] where v r is the relative radial velocity between the interference source and the receiver, and c is the speed of light; the OTFS signal r(t) after being transmitted through the wireless channel is collected by the receiving-end device, and the expression is:

[0088] r(t) = ∫h(τ,v)s(t - τ)e j2πvt dτdν + j(t) + n(t)

[0089] where h(τ,ν) is the impulse response of the channel in the delay-Doppler domain, τ represents the delay, ν represents the Doppler frequency shift, s(t) is the transmitted signal, j(t) is the interference signal, and n(t) is the additive white Gaussian noise.

[0090] Furthermore, the combined network structure includes a combined network consisting of a convolutional neural network CNN and a bidirectional long short-term memory network Bi-LSTM; the CNN is responsible for extracting local spatial features from the input delay-Doppler domain signal, and the Bi-LSTM is used to capture the temporal dependencies of the signal; the two cooperate closely, and the output of the CNN is used as the input of the Bi-LSTM, enabling the network to comprehensively learn the complex features of the OTFS signal in the time-frequency space and effectively cope with the signal demodulation problem in the interference environment;

[0091] The structural method of the combined network includes:

[0092] Build a CNN component containing two convolutional layers. Each convolutional layer performs a convolution operation on the input delay-Doppler domain signal. The convolutional kernel slides on the signal to extract local spatial features; the first-layer convolution operation is:

[0093]

[0094] where, is the value of the output feature map of the first convolutional layer at the (i,j) position, W1 (u,v) is the value of the convolutional kernel of size K l ×K w at the (u,v) position, K l is the size of the convolutional kernel in the length direction, K w is the size of the convolutional kernel in the width direction, Y (i+u,j+v) is the value of the input signal at the (i + u,j + v) position, and b1 is the bias term of the first convolutional layer; after each convolutional layer, a ReLU activation function is connected to introduce non-linear operations:

[0095]

[0096] The value at the (i, j) position after being processed by the ReLU activation function The value input to the ReLU activation function, which is the output of the first convolutional layer; then batch normalization is used to reduce internal covariate shift by normalizing the output of the convolutional layer, stabilizing the training process and accelerating convergence:

[0097]

[0098] Among them, is the value at the (i, j) position after normalization processing, and μ and σ 2 respectively represent the mean and variance of the feature map, which are used to centralize and scale the feature map. ∈ is a numerically stable constant used to prevent the denominator from being zero during the calculation process; then scaling and offset are performed through parameter γ scaling and β offset, that is Through these operations, internal covariate shift is effectively reduced and the model convergence is accelerated. After that, the second convolutional operation is performed:

[0099]

[0100] Among them, is the value at the (i, j) position of the output feature map of the second convolutional layer, and b2 is the bias term of the second convolutional layer;

[0101] Then a Bi-LSTM network is constructed. Using its bidirectional processing characteristics, the signals processed by the CNN are processed separately from the forward and backward directions; the forward LSTM layer processes the signal sequence in chronological order. At each time step, the current hidden state and cell state are combined with the current input and the previous hidden state to calculate the current hidden state; the backward LSTM layer processes the signal sequence in reverse order and also calculates the hidden state at each time step; the output of the forward LSTM hidden state is:

[0102]

[0103] where h t is the output of the hidden state of the forward LSTM layer at time step t, which combines the current and previous information and participates in subsequent calculations. LSTM → is the calculation process of the forward LSTM, including input, forget, output gates and cell state update, integrating the input and previous states to update the hidden state. x t is the input signal at time step t, providing real-time information for the current calculation, is the hidden state of the forward LSTM layer at time step t - 1, carrying the feature information of the previous time step, is the cell state of the forward LSTM layer at time step t - 1;

[0104] The output of the reverse LSTM hidden state is as follows:

[0105]

[0106] Among them, is the hidden state output of the backward LSTM layer at time step t, capturing reverse time-dependent features, and LSTM ← is the backward LSTM calculation process, processing signals from back to front and updating the hidden state, D t is the data input to the backward LSTM layer at time step t (data processed by CNN), which is the input for the current calculation, is the hidden state of the backward LSTM layer at time step t + 1. During reverse processing, the state of the next time step participates in the current calculation, is the cell state of the backward LSTM layer at time step t + 1, used for reverse long-term dependence information transmission and current hidden state calculation;

[0107] Concatenate the outputs of the processed forward and backward LSTM layers at each time step to obtain features containing past and future context dependencies, effectively capturing the temporal correlation of interference signals and enhancing the model's ability to understand signal sequences.

[0108] Furthermore, in step 3, perform preprocessing including denoising and normalization on the training data, and then align the input data and label data in time series, including:

[0109] Filter the obtained signal to remove high-frequency or low-frequency noise and improve the signal-to-noise ratio of the signal; through signal calibration, compensate for signal amplitude and phase changes caused by channel transmission to ensure signal accuracy; perform normalization processing, and normalize the signal x using the formula for normalization, where μ is the mean of the signal and σ is the standard deviation;

[0110] Use the processed data as the input, and the corresponding original transmitted signal b transmitted as the data, align the input data and label data in chronological order to form an ordered training data pair; store the ordered training data pair in a dynamic pool, and the dynamic pool manages the data according to the first-in-first-out (FIFO) principle, providing continuous time series data for the training of Bi-LSTM to ensure that the model can learn the temporal features of the signal.

[0111] Further, in step 4, the preprocessed data is input into a CNN to extract local spatial features, and then input into a Bi-LSTM composed of forward and backward LSTM layers to capture temporal dependence features. The mean square error loss function is calculated based on the predicted demodulated bits and the true bits. The gradient of the network parameters is calculated using the backpropagation algorithm, and the network parameters are updated according to the gradient using an optimizer, including:

[0112] Initialize the convolutional kernel weights W, biases b of the convolutional layer in the CNN component, the weights and gating parameters of the Bi-LSTM component, as well as the learning rate and batch size. Input the training data into the anti-interference demodulator for forward propagation to calculate the output of the model; according to the difference between the output and the label data, use the mean square error loss function:

[0113]

[0114] Calculate the loss value, where N batch is the number of samples in the batch, is the predicted demodulated bit, is the true transmitted bit; calculate the gradient of the loss value with respect to the network parameters using the backpropagation algorithm, and update the network parameters using an optimization algorithm (such as stochastic gradient descent); the parameter update formula for stochastic gradient descent is:

[0115]

[0116] where θ is the network parameter, α is the learning rate, is the gradient of the loss function at θ t ; during the training process, use the cross-validation method to divide the training data into multiple subsets, take turns as the training set and the validation set, evaluate the generalization ability of the model, and avoid overfitting; according to the performance of the validation set, adjust the training parameters, such as the learning rate and the number of training epochs, in a timely manner to optimize the model performance.

[0117] Further, in step 5, perform model training until the expected requirements are met, and apply the trained model to the signal demodulation of the OTFS communication system, including:

[0118] Deploy the trained demodulator to the receiving end. First preprocess the received signal with interference, then restore the signal through the demodulator, and finally perform error detection and correction on the demodulated signal. Evaluate the demodulation performance by calculating the bit error rate; perform error detection on the demodulated signal, and measure the demodulation performance by calculating the bit error rate BER. The calculation formula is:

[0119]

[0120] where N errors is the number of bits with demodulation errors, and N totalis the total number of transmitted bits, and the demodulation effect is evaluated in real time through monitoring indicators.

[0121] To verify the effectiveness of the proposed solution of the present invention, the following experimental design is carried out. The system simulation uses the Python language and is based on the Pytorch neural network. The parameter settings do not affect generality. In this embodiment, several common neural networks are used for prediction: Convolutional Neural Network (CNN), Attention Convolutional Neural Network, Transformer Network, Performer Network, Resnet Network, and CNN-Bi-LSTM Hybrid Network. Figure 4 is the bit error rate of communication under different neural networks, and the results show that the present invention can reduce the bit error rate.

[0122] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0123] The above-described embodiments only represent several implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A demodulation method for an orthogonal time-frequency space modulation communication system based on deep learning, characterized in that The method includes: Step 1: Construct an Orthogonal Time Frequency and Space (OTFS) communication system model and an interference model; Step 2: Construct a network structure combining a Convolutional Neural Network (CNN) and a Bidirectional Long Short-Term Memory (Bi-LSTM) according to the characteristics of the OTFS communication system and anti-interference requirements; Step 3: Perform preprocessing including normalization and denoising on the training data, and then align the input data and label data in time series; Step 4: The preprocessed data is input into the CNN to extract local spatial features, and then input into the Bi-LSTM composed of forward and backward LSTM layers to capture time-dependent features. Calculate the mean square error loss function according to the predicted demodulated bits and the true bits, use the backpropagation algorithm to calculate the gradient of the network parameters, and use the optimizer to update the network parameters according to the gradient; Step 5: Perform model training until the expected requirements are met, and apply the trained model to the signal demodulation of the OTFS communication system.

2. The demodulation method of the orthogonal time-frequency space modulation communication system based on deep learning according to claim 1, wherein Step 1, constructing an OTFS communication system model and an interference model, includes: In the delay-Doppler domain, the received signal X[m,n] is transformed to the time-frequency domain through the Inverse Short-Time Fourier Transform (ISFFT): Among them, X tf [l,k] is the signal in the time-frequency domain after ISFFT transformation, where l is time and k is the frequency index, and the subscript tf indicates that the signal is in the time-frequency domain. The symbol matrix of X[m,n] in the delay-Doppler domain, where m is the Doppler index, n is the delay index, N is the total number of Doppler cells, and M is the total number of delay cells; Then use the Heisenberg transform to generate the continuous-time domain signal s(t): where t represents time, and g tx is the transmit pulse shaping function, and the subscript tx indicates that this is the transmit signal, T is the symbol duration, and Δf is the subcarrier spacing; the received signal r(t) is subjected to matched filtering and Wigner transformation to obtain the time-frequency domain signal Y tf [l,k]: wherein, is the conjugate of the received pulse shaping function; the delayed Doppler domain signal Y[m,n] is reconstructed by the finite Fourier transform SFFT: The interference signal j(t) includes multiple interference components, and each interference component is represented by a complex exponential function: where J p is the amplitude of the p-th interference signal, f j,p is the frequency of the p-th interference, φ p is the phase of the p-th interference, and P is the total number of interference signal components; the interference frequency f j,p includes not only the original interference frequency but also the influence of the Doppler shift. The original frequency of the p-th interference signal is expressed as: where v r is the relative radial velocity between the interference source and the receiver, and c is the speed of light; the receiving-end device is used to collect the OTFS signal r(t) after transmission through the wireless channel, and the expression is: r(t) = ∫h(τ, v)s(t - τ)e j2πvt dτdv + j(t) + n(t) Where h(τ,ν) is the impulse response of the channel in the delay-Doppler domain, τ represents the delay, v represents the Doppler frequency shift, s(t) is the transmitted signal, j(t) is the interference signal, and n(t) is the additive Gaussian white noise.

3. The demodulation method of the orthogonal time-frequency space modulation communication system based on deep learning according to claim 1, wherein The combined network structure includes a combined network consisting of a Convolutional Neural Network (CNN) and a Bidirectional Long Short-Term Memory (Bi-LSTM); the CNN is responsible for extracting local spatial features from the input delay-Doppler domain signal, and the Bi-LSTM is used to capture the time-dependent relationship of the signal; the output of the CNN is used as the input of the Bi-LSTM, so that the network can comprehensively learn the complex features of the OTFS signal in the time-frequency space; The structural method of the combined network includes: Build a CNN component containing two convolutional layers. Each convolutional layer performs a convolution operation on the input delay-Doppler domain signal. The convolutional kernel slides on the signal to extract local spatial features; the first convolutional operation is: Among them, is the value of the output feature map of the first convolutional layer at the position (i, j), is the value of the convolutional kernel of size K l ×K w at the position (u, v), and K l is the size of the convolutional kernel in the length direction, and K w is the size of the convolutional kernel in the width direction. Y (i +u,j+v) is the value of the input signal at the position (i + u, j + v), and b1 is the bias term of the first convolutional layer; a ReLU activation function is connected after each convolutional layer to introduce non-linear operations: The value at the (i, j) position after being processed by the ReLU activation function The value input to the ReLU activation function, which is the output of the first convolutional layer; then batch normalization is used to normalize the output of the convolutional layer: Among them, is the value at the (i, j) position after normalization. μ and σ 2 respectively represent the mean and variance of the feature map, which are used to centralize and scale the feature map. ∈ is a numerically stable constant used to prevent the denominator from being zero during the calculation process. Then, scaling and offsetting are performed through parameter γ scaling and β offset, that is After that, the second convolutional operation is performed: wherein, is the value of the output feature map of the second convolutional layer at the position (i, j), and b2 is the bias term of the second convolutional layer; Then, a Bi-LSTM network is constructed. Using its bidirectional processing feature, the signals processed by the CNN are processed separately from the forward and backward directions. The forward LSTM layer processes the signal sequence in chronological order. At each time step, the current hidden state is calculated by combining the current input, the hidden state, and the cell state at the previous moment. The backward LSTM layer processes the signal sequence in reverse order and also calculates the hidden state at each time step. The output of the forward LSTM hidden state is: where h t is the hidden state output of the forward LSTM layer at time step t, and LSTM → is the forward LSTM calculation process, including input, forget, output gates and cell state update, integrating the input and the previous state to update the hidden state, and x t is the input signal at time step t, providing real-time information for the current calculation, is the hidden state of the forward LSTM layer at time step t-1, carrying the feature information of the previous time step, is the cell state of the forward LSTM layer at time step t-1; The output of the backward LSTM hidden state is: Among them, is the hidden state output of the backward LSTM layer at time step t, capturing reverse time-dependent features, and LSTM ← is the backward LSTM calculation process, processing the signal from back to front and updating the hidden state, D t is the data input to the backward LSTM layer at time step t (data processed by CNN), which is the input for the current calculation, is the hidden state of the backward LSTM layer at time step t + 1. During backward processing, the state of the next time step participates in the current calculation, is the cell state of the backward LSTM layer at time step t + 1, which is used for reverse long-term dependence information transmission and current hidden state calculation; The outputs of the processed forward and backward LSTM layers at each time step are concatenated to obtain features containing past and future context dependencies.

4. The demodulation method of the orthogonal time-frequency space modulation communication system based on deep learning according to claim 1, wherein Step 3: Perform preprocessing on the training data, including denoising and normalization. Then, align the input data and label data in time series, including: Filter the obtained signal to remove high-frequency or low-frequency noise and improve the signal-to-noise ratio; through signal calibration, compensate for the signal amplitude and phase changes caused by channel transmission; perform normalization processing, and use the formula for signal x to perform normalization, where μ is the mean of the signal and σ is the standard deviation; Using the processed data as input, along with the corresponding original transmitted signal b transmitted As data, align the input data and the label data in chronological order to form an ordered pair of training data; store the ordered pair of training data in a dynamic pool, and the dynamic pool manages the data according to the first-in-first-out (FIFO) principle.

5. The demodulation method of the orthogonal time-frequency space modulation communication system based on deep learning according to claim 1, wherein Step 4: The preprocessed data is input into the CNN to extract local spatial features, and then input into the Bi-LSTM composed of forward and backward LSTM layers to capture time-dependent features. Calculate the mean square error loss function based on the predicted demodulated bits and the true bits. Use the backpropagation algorithm to calculate the gradient of the network parameters, and use the optimizer to update the network parameters according to the gradient, including: Initialize the convolution kernel weights W, biases b of the convolutional layer in the CNN component, the weights and gating parameters of the Bi-LSTM component, as well as the learning rate and batch size. Input the training data into the anti-interference demodulator for forward propagation to calculate the output of the model. According to the difference between the output and the label data, use the mean square error loss function: Calculate the loss value, where N batch is the number of samples in the batch, are the predicted demodulated bits, are the true transmitted bits; calculate the gradient of the loss value with respect to the network parameters by the backpropagation algorithm, and update the network parameters using an optimization algorithm; the parameter update formula for stochastic gradient descent is: where θ is the network parameter and α is the learning rate, is the gradient of the loss function at θ t During the training process, the cross-validation method is adopted to divide the training data into multiple subsets, which are used as the training set and the validation set in turn to evaluate the generalization ability of the model and avoid overfitting. According to the performance of the validation set, the training parameters are adjusted timely to optimize the model performance.

6. The demodulation method of the orthogonal time-frequency space modulation communication system based on deep learning according to claim 1, characterized in that Step 5: Train the model until the expected requirements are met, and apply the trained model to the signal demodulation of the OTFS communication system, including: Deploy the trained demodulator to the receiving end. First, preprocess the received signal with interference, then restore the signal through the demodulator, and finally perform error detection and correction on the demodulated signal. Evaluate the demodulation performance by calculating the bit error rate. Perform error detection on the demodulated signal, and measure the demodulation performance by calculating the bit error rate BER. The calculation formula is: where N errors is the number of bits with demodulation errors, and N total is the total number of transmitted bits, and the demodulation effect is evaluated in real time by monitoring indicators.

Images