Simultaneous self-interference cancellation method and system based on Conv-BiLSTM and LS for simultaneous self-interference cancellation in digital domain

Abstract

This invention relates to a method and system for simultaneous, full-duplex, digital-domain adaptive self-interference cancellation based on Conv-BiLSTM and LS, and relates to self-interference signal cancellation technology in the field of communication. The purpose of this invention is to propose a method for simultaneous, full-duplex, digital-domain adaptive self-interference cancellation based on Bayesian optimization algorithm to determine the self-interference channel length, combined with Conv-BiLSTM and LS, to solve the problems existing in the background technology. Key technical points: First, the self-interference channel length is estimated using a Bayesian optimization algorithm. Then, based on the determined parameters, LS is used to perform linear self-interference cancellation on the self-interference signal. Next, the constructed Conv-BiLSTM network is trained to fit the nonlinear components of the self-interference. Compared with FFNN, the constructed Conv-BiLSTM network can more effectively handle high-dimensional time-series signals and has a good effect on predicting nonlinear distortion signals with memory. Finally, the reconstructed self-interference signal is subtracted from the received signal at the receiving end to achieve the purpose of self-interference cancellation.

Description

Technical Field

[0001] This invention relates to self-interference signal cancellation technology in the field of communication, and specifically to the field of simultaneous full-duplex digital domain adaptive self-interference cancellation technology, and more specifically to a method and system for simultaneous full-duplex digital domain adaptive self-interference cancellation based on Conv-BiLSTM and LS. Background Technology

[0002] Simultaneous co-frequency co-time full duplex (CCFD) systems transmit and receive signals on the same frequency at the same time, theoretically doubling spectral efficiency and significantly increasing signal throughput. The application of full-duplex technology is inevitable for future communications. However, due to the requirements for miniaturization and high integration of communication equipment, high-intensity self-interference signals inevitably appear in the system, disrupting its operation and causing receiver link blockage. To ensure normal system operation, self-interference signals must be specifically eliminated. Currently, self-interference elimination typically combines antenna propagation domain, analog domain, and digital domain methods.

[0003] Digital domain self-interference cancellation technology has attracted widespread attention from researchers due to its processing flexibility. In recent years, research on digital domain self-interference cancellation has also increased. For example, existing patent document CN106165468B discloses aspects of how adaptive interference cancellation adjusts transmitter configuration for efficient concurrent primary user detection. A wireless transmitting device can schedule transmissions in a shared spectrum. During the transmission, the device can scan at least a portion of the shared spectrum to receive signals. Self-interference cancellation circuitry can be used to eliminate interference caused by the transmission from the received signals. The device can determine whether the received signal indicates primary user use of the shared spectrum. In one aspect, the transmission can be a SISO transmission. In another aspect, carrier aggregation can be used for the transmission, and potential carriers may be subject to primary user detection. The device can determine the self-interference cancellation complexity for a carrier combination including the potential carriers and can select one or more carriers for aggregation based on the self-interference cancellation complexity. The prior art, document CN104716980B, discloses a method for canceling self-interference in simultaneous, full-duplex radio frequency (RF) multi-subband signals at the same frequency. The steps are as follows: First, the transmitted signal is split into two paths: one path is sent to the transmitting antenna to radiate the wireless signal, and the other path is sent to an RF cancellation unit. The signal entering the RF cancellation unit is divided into multiple sub-band signals by several parallel bandpass filters, and RF self-interference signals are reconstructed. Finally, the reconstructed RF self-interference signal is canceled out by the self-interference signal received by the antenna. Simultaneously, an amplitude adjustment algorithm module adjusts the attenuation coefficients of each module to achieve the optimal self-interference cancellation effect. This prior art is applicable to simultaneous, full-duplex RF self-interference cancellation methods.

[0004] Currently, self-interference cancellation in the digital domain mainly relies on two methods: multinomial models and feedforward neural networks (FFNNs). The multinomial model-based method first models the self-interference channel, then uses algorithms such as least mean square (LMS) and least squares (LS) to solve the model coefficients, reconstructs the self-interference signal, and finally subtracts it from the received signal to achieve self-interference cancellation. The FFNN-based method leverages its strong fitting ability to train a model on transmitted and received data, and then directly fits the self-interference signal using the trained model. However, both methods have various drawbacks. Adaptive filtering methods rely on prior knowledge of the self-interference signal, which carries the risk of model mismatch leading to a deterioration in self-interference cancellation effectiveness. Existing FFNN-based methods are typically applied directly to self-interference cancellation tasks. However, directly applying FFNNs during model training may overlook the characteristics of some nonlinear interference components, leading to insufficient nonlinear interference suppression. Furthermore, FFNNs only utilize multilayer perceptrons and nonlinear activation functions to extract interference signal features, failing to provide detailed processing for self-interference signals with significant temporal and spatial frequency characteristics. Additionally, before model training, data alignment between transmitted and received data is generally required, and the training data needs to be segmented according to the self-interference channel length to ensure the network can fully extract interference signal features. However, there is no unified solution for determining the channel length of the self-interference signal. Summary of the Invention

[0005] The technical problem to be solved by this invention is:

[0006] To address the limitations of existing technologies in processing self-interference signals with distinct temporal and spatial frequency characteristics, and the need for data alignment between transmitted and received data before model training, as well as segmentation of training data based on the self-interference channel length to ensure the network can fully extract interference signal features, resulting in poor self-interference cancellation performance, this invention proposes a simultaneous, same-frequency, full-duplex digital domain adaptive self-interference cancellation method based on Bayesian optimization algorithm to determine the self-interference channel length and combining Conv-BiLSTM and LS.

[0007] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0008] A simultaneous, in-frequency, full-duplex digital domain adaptive self-interference cancellation method based on Conv-BiLSTM and LS is proposed. This method reduces reliance on prior knowledge of self-interference signals and enhances the adaptability of the overall algorithm model, thereby mitigating the risk of channel model mismatch. The method includes:

[0009] Data preprocessing:

[0010] The channel length of the self-interference signal is determined based on the Bayesian optimization algorithm. The Bayesian optimization algorithm uses a Gaussian process to model the unknown objective function and infers the distribution of the entire function from the distribution of known data points. The transmitted signal is divided into multiple data segments according to the self-interference channel length, and each data segment corresponds to the self-interference signal at a certain moment.

[0011] Establishment of a linear model for self-interference signals:

[0012] After determining the length of the interference channel using the Bayesian optimization algorithm, linear modeling is performed. Once the linear model is built, the LS algorithm is used to solve for the model coefficients, thereby estimating the linear components of the interference.

[0013] Construction of Conv-BiLSTM network:

[0014] By adding convolutional layers to the front end of the network and then using bidirectional LSTM for deep temporal learning, a Conv-BiLSTM network is constructed. The Conv-BiLSTM network is then used to capture and eliminate the features of self-interference nonlinear signals.

[0015] Furthermore, the specific implementation process of the method is as follows:

[0016] Data preprocessing

[0017] Suppose the objective function to be optimized is f(x), where x are the parameters to be optimized. The goal of Bayesian optimization is to find the global optimum x. * The goal is to minimize or maximize the objective function. Generalized assumptions (GPs) are used to make prior assumptions about the objective function. Before sufficient observations are available, GPs model the objective function based on the evaluated points, estimating the distribution at each point. The GP process is represented by the following formula:

[0018] f(x)~GP(m(x),k(x,x'))(1)

[0019] Where m(x) is the mean function and k(x,x') is the kernel function;

[0020] In the early stages of iteration, methods such as random sampling or Latin hypercube sampling can be used to sample several points {x}. i}, evaluate its objective function {f(x)} i After initial data points are available, Gaussian process regression is used to learn kernel function parameters, etc., and based on the observed data points D={(x i ,f(x i ))}, infer the posterior distribution of the objective function at each point x in the search space:

[0021] f(x)|D~p(μ(x),σ 2 (x))(2)

[0022] Where μ(x) is the posterior mean, representing the best guess of the function at x under the current observation; σ 2 (x) is the posterior variance, which represents the uncertainty of the model at x. Through multiple iterations, new observation data are obtained to update the parameters of GP, and the model gradually approaches the true objective function model. When the iteration reaches a certain number of times or the convergence condition is met, the entire optimization process ends.

[0023] x is represented as x = hSILen, where hSILen represents the length of the self-interference channel; its objective function is the mean square error between the optimized self-interference signal and the actual self-interference signal.

[0024] Establishment of a linear model for self-interference signals

[0025] After determining the length of the interference channel using the Bayesian optimization algorithm, linear modeling is then performed according to Equation 3.

[0026]

[0027] Where y lin (n) represents the linear part of self-interference, L represents the length of the self-interference channel, and h SI (l) represents the channel response coefficient of the l-th self-interference channel; after the linear model is constructed, the LS algorithm is used to solve the model coefficients, and then the self-interference linear component is estimated, i.e.:

[0028]

[0029] Where x represents the baseband transmitted data; [then] we obtain... Then, the linear component from the disturbance is estimated, i.e.:

[0030]

[0031] Where y represents the received data. As labels for subsequent network training;

[0032] Construction of Conv-BiLSTM Network

[0033] Based on the characteristics of self-interference signals in full-duplex transceiver systems, comprehensive suppression is required at multiple scales, including time and spatial domains. Recurrent neural networks can capture the temporal dependencies of signals, and Conv-BiLSTM network structures are used to capture and eliminate the features of self-interference nonlinear signals.

[0034] First, convolutional layers can effectively extract local temporal features and reduce model parameters through weight sharing, thereby improving training efficiency. Simultaneously, convolutional operations can perform preliminary feature extraction on the signal at the front end, filtering out some environmental noise and weak interference. Second, LSTM, as an improved recurrent neural network, effectively alleviates the gradient vanishing and exploding problems in traditional RNNs through gating mechanisms, preserving contextual information over a longer temporal range. Furthermore, bidirectional LSTM concatenates the hidden states in the forward and backward temporal sequences, enabling the network to comprehensively consider input information from both the forward and backward directions at each time step, establishing a more complete sequence dependency modeling capability for strongly correlated and multipath-intertwined self-interference signals. By adding convolutional layers at the front end of the network and then using bidirectional LSTM for deep temporal learning, multi-scale features of self-interference signals are fully extracted, thus achieving higher-precision self-interference cancellation.

[0035] Furthermore, the Conv-BiLSTM structure separates the real and imaginary parts of the transmitted baseband signal and constructs them into a 3D tensor of dimension (batchsize, hSILen, 2), where the parameters represent the amount of data fed into the network for each training iteration, the length of the self-interference channel, and the real and imaginary parts of the baseband signal, respectively. Then, one-dimensional convolution is performed with 8 kernels of size 3. Each convolution kernel extracts different features from the input data, and each kernel performs a convolution operation on the input at each time step, generating a new feature map. The samples that have undergone feature extraction and dimensionality reduction are used as the input to the BiLSTM layer. After that, a fully connected layer connects to the BiLSTM layer, and the output layer has two neurons, representing the real and imaginary parts of the evaluated self-interference nonlinear components, respectively.

[0036] Furthermore, in the digital cancellation part, a linear channel model is constructed in combination with the LS algorithm to reconstruct the linear self-interference component, and the Conv-BiLSTM method is used to fit the nonlinear self-interference component, thereby reconstructing the complete self-interference signal.

[0037] The system is based on the simultaneous full-duplex digital domain adaptive self-interference cancellation system of Conv-BiLSTM and LS. The system has program modules corresponding to the steps of the technical solution. When running, it executes the steps in the simultaneous full-duplex digital domain adaptive self-interference cancellation method based on Conv-BiLSTM and LS.

[0038] A computer-readable storage medium storing a computer program configured to, when invoked by a processor, implement the steps of the simultaneous full-duplex digital domain adaptive self-interference cancellation method based on Conv-BiLSTM and LS.

[0039] The present invention has the following technical advantages: First, the length of the self-interference channel is estimated using a Bayesian optimization algorithm. Bayesian optimization is a global optimization method based on Bayesian theory, typically used to optimize costly black-box functions. It features high sample efficiency and adaptive learning. It can find the global optimum or near-optimal solution with limited computational resources and is applicable to various complex and irregular objective functions. Then, based on the determined parameters, linear cancellation of the self-interference signal is performed using LS. Next, the constructed Conv-BiLSTM network is trained to fit the nonlinear components of the self-interference. Compared to FFNN, the constructed Conv-BiLSTM network can more effectively handle high-dimensional time-series signals and has a good effect on predicting nonlinear distortion signals with memory. Finally, at the receiving end, the reconstructed self-interference signal is subtracted from the received signal to achieve self-interference cancellation.

[0040] This invention uses transmitted and received signal data (the publicly available dataset fdTestbedData20MHz10dBm.mat from the National Instruments Institute) for training the aforementioned model. To enhance comparability, the performance of the proposed LS+ConvBiLSTM digital domain self-interference cancellation method is compared with that of linear models, multinomial models, LS+FFNN, and FFNN and ConBiLSTM used alone. Verification shows that the Bayesian optimization algorithm proposed in this invention is more effective and has a more significant cancellation effect. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the Conv-BiLSTM structure. Figure 2 This is a schematic diagram of the digital domain self-interference cancellation scheme described in this invention. Figure 3 The graph shows the cancellation effect of LS+ConvBiLSTM. Figure 4 The graph shows the training results of ConvBiLSTM nonlinear cancellation. Figure 5 The graph shows the training results of the FFNN nonlinear cancellation. Detailed Implementation

[0042] Combined with appendix Figure 1-5 The implementation of the simultaneous, same-frequency, full-duplex digital domain adaptive self-interference cancellation method based on Conv-BiLSTM and LS described in this invention is explained as follows:

[0043] The method proposed in this invention includes three parts: data preprocessing, establishment of a linear model of self-interference signals, and construction of a Conv-BiLSTM network.

[0044] Data preprocessing

[0045] To enable the model to fully learn the mapping relationship between the received and transmitted signals, the channel length of the self-interference signal must first be determined. Then, based on the self-interference channel length, the transmitted signal is divided into multiple data segments. Theoretically, each data segment corresponds to the self-interference signal at a specific moment. The preprocessed data is then linearly canceled before the data segments are fed into the constructed model for more detailed nonlinear self-interference signal fitting. Currently, systematic research on self-interference cancellation is lacking in the absence of prior knowledge about the self-interference signal.

[0046] This invention proposes a scheme based on Bayesian optimization to determine the length of self-interference channels. Bayesian optimization (BO) is a probabilistic optimization method mainly applied to black-box problems. In the absence of prior knowledge, it can quickly find the global optimum by modeling the objective function through efficient global search.

[0047] The core idea of ​​Bayesian optimization algorithm is to use a Gaussian process (GP) to model an unknown objective function and infer the distribution of the entire function from the distribution of known data points.

[0048] Suppose the objective function to be optimized is f(x), where x are the parameters to be optimized. The goal of Bayesian optimization is to find the global optimum x. * The goal is to minimize or maximize the objective function. It typically uses Generalized Probability (GP) to make prior assumptions about the objective function. Before sufficient observations are available, GP models the objective function based on the evaluated points, estimating the distribution at each point. The GP process can be represented by the following formula:

[0049] f(x)~GP(m(x),k(x,x'))(1)

[0050] Where m(x) is the mean function and k(x,x') is the kernel function.

[0051] In the early stages of iteration, methods such as random sampling or Latin hypercube sampling can be used to sample several points {x}. i}, evaluate its objective function {f(x)} i After obtaining initial data points, Gaussian process regression is used to learn kernel function parameters, etc., and based on the observed data points D={(x i ,f(x i ))}, infer the posterior distribution of the objective function at each point x in the search space:

[0052] f(x)|D~p(μ(x),σ 2 (x))(2)

[0053] Where μ(x) is the posterior mean, representing the best guess of the function at x under the current observations. σ 2 (x) represents the posterior variance, indicating the uncertainty of the model at x. Through multiple iterations, new observation data is obtained to update the parameters of the generalized objective function (GP), eventually bringing the model closer to the true objective function. The optimization process ends when a certain number of iterations are completed or the convergence condition is met.

[0054] For the target scenario in this paper, x can be represented as x = hSILen, where hSILen represents the length of the self-interference channel. The objective function is the mean square error between the optimized self-interference signal and the actual self-interference signal.

[0055] Establishment of a linear model for self-interference signals

[0056] After determining the length of the interference channel using the Bayesian optimization algorithm, linear modeling is then performed according to Equation 3.

[0057]

[0058] Where y lin (n) represents the linear part of self-interference, L represents the length of the self-interference channel, and h SI (l) represents the channel response coefficient of the l-th self-interference channel. After the linear model is constructed, the LS algorithm is used to solve for the model coefficients, and then the self-interference linear component is estimated, i.e.:

[0059]

[0060] Where x represents the baseband transmitted data. (The result is...) Then, the linear component from the disturbance is estimated, i.e.:

[0061]

[0062] Where y represents the received data. As labels for subsequent network training.

[0063] Construction of Conv-BiLSTM Network

[0064] In full-duplex transceiver systems, self-interference signals typically exhibit the following characteristics: First, the intensity of the self-interference signal is much greater than the target useful signal, and nonlinear distortion and multipath effects may accumulate in the interference path. Second, the self-interference signal and the useful signal highly overlap in the time and frequency domains, making traditional linear filtering difficult to separate them effectively. Third, the self-interference signal has strong temporal correlation and potential non-stationary statistical characteristics, requiring comprehensive suppression across multiple scales, including time and space. To address this issue, recurrent neural networks (RNNs) have attracted widespread attention due to their ability to model time-series data, capturing the temporal dependencies of signals. However, unidirectional RNNs may not fully utilize the forward and reverse temporal information of the self-interference signal, and their learning efficiency and generalization performance are limited in complex multipath and strong self-interference scenarios.

[0065] To better capture and eliminate the features of self-interference nonlinear signals, this invention proposes a Conv-BiLSTM network structure to address this problem. First, convolutional layers can effectively extract local temporal features and reduce model parameters through a weight-sharing mechanism, improving training efficiency. Simultaneously, convolutional operations can perform preliminary feature extraction at the front end, filtering out some environmental noise and weak interference. Second, LSTM, as an improved recurrent neural network, effectively alleviates the gradient vanishing and exploding problems in traditional RNNs through a gating mechanism, preserving contextual information over a longer temporal range. Furthermore, bidirectional LSTM concatenates the hidden states in the forward and backward time sequences, enabling the network to comprehensively consider input information from both directions at each time step, establishing a more complete sequence dependency modeling capability for strongly correlated and multipath-intertwined self-interference signals. By adding convolutional layers at the network front end and then using bidirectional LSTM for deep temporal learning, this scheme can more fully extract the multi-scale features of self-interference signals, thereby achieving higher-precision self-interference cancellation.

[0066] The network structure designed in this invention is as follows: Figure 1 As shown, the real and imaginary parts of the transmitted baseband signal are separated and constructed into a 3D tensor of dimension (batchsize, hSILen, 2), where the parameters represent the amount of data fed into the network for each training iteration, the length of the self-interference channel, and the real and imaginary parts of the baseband signal, respectively. Then, a one-dimensional convolution is performed with 8 kernels of size 3. Each convolution kernel extracts different features from the input data. Each kernel performs a convolution operation on the input at each time step, generating a new feature map. The samples after feature extraction and dimensionality reduction are used as input to the BiLSTM layer. A fully connected layer then connects to the BiLSTM layer, and the output layer has two neurons, representing the real and imaginary parts of the evaluated self-interference nonlinear components, respectively.

[0067] In summary, the overall technical approach for digital domain self-interference cancellation proposed in this invention is as follows: Figure 2 As shown.

[0068] In the digital cancellation part of this invention, a linear channel model is constructed combined with the LS algorithm to reconstruct the linear self-interference component, and the Conv-BiLSTM method is used to fit the nonlinear self-interference component. This reconstructs the complete self-interference signal. This method reduces reliance on prior knowledge of the self-interference signal, enhances the adaptability of the overall algorithm model, and reduces the risk of channel model mismatch.

[0069] The technical effects of the present invention are verified as follows:

[0070] This invention uses the transmitted and received signal data (the publicly available dataset fdTestbedData20MHz10dBm.mat from the National Instruments Institute) for training the aforementioned model. To enhance comparability, the performance of the proposed LS+ConvBiLSTM digital domain self-interference cancellation method is compared with that of linear models, multinomial models, LS+FFNN, and FFNN and ConBiLSTM alone. Furthermore, to verify the effectiveness of the proposed Bayesian optimization algorithm, the parameters obtained by the optimization algorithm are compared with the actual parameters of the self-interference signals in this dataset and randomly determined parameters.

[0071] Table 1. Self-interference cancellation performance (dB) of each method under different hSILen conditions.

[0072]

[0073] In Table 1, hSILen:12 represents the parameters determined by Bayesian optimization, hSILen:13 represents the prior knowledge of the interference signal in the dataset, and hSILen:8 represents the randomly determined parameters. The data in Table 1 shows that the performance of various methods with parameters obtained through Bayesian optimization is comparable to that with hSILen=13, and also shows improvement compared to hSILen=8. This confirms the effectiveness of the Bayesian optimization method. In fact, when the self-interference channel length is sufficiently large, theoretically, various methods can reach their theoretical extreme values. However, excessively large self-interference channel lengths lead to increased computational costs, which may result in a decrease in real-time performance. Furthermore, it can be seen that the performance of using FFNN and ConvBiLSTM alone is generally worse than other methods. For FFNN alone, its performance only reaches the level of linear cancellation. This is because the linear component in the self-interference signal has a large weight, causing the FFNN fitting process to ignore the nonlinear characteristics of the self-interference signal. For ConvBiLSTM used alone, its performance doesn't even reach the level of FFNN. This is because, while ConvBiLSTM has stronger feature extraction and nonlinear modeling capabilities, when the dominant information in the training data is linear, it may distract the model, reducing its efficiency in learning the linear components. Furthermore, due to its large number of network parameters, it is prone to overfitting or unstable gradient propagation when primarily describing simple linear relationships, resulting in a less stable training process compared to FFNN. Therefore, when the dominant component in the self-interference signal is predominantly linear, but the nonlinear component cannot be ignored, combining LS with a neural network is an effective approach. Compared to multinomial models, it reduces reliance on prior knowledge of the self-interference signal, lowering the risk of model mismatch. Compared to using a neural network alone, LS can first eliminate the dominant linear component, allowing the neural network to focus on canceling the nonlinear components, reducing the difficulty of network training.

[0074] exist Figure 3 Clearly, the method proposed in this invention achieves a cancellation performance of 45.5 dB, and the residual self-interference signal level is very close to the receiver's noise floor. Furthermore, comparing Table 1 shows that LS+ConvBiLSTM improves performance by 18.8 dB compared to using ConvBiLSTM alone. Based on linear cancellation, by... Figure 4 , Figure 5It can be seen that the nonlinear cancellation performance of ConvBiLSTM reaches 7dB after 13 training epochs, and eventually reaches approximately 8dB, which is unattainable by FFNN. Furthermore, no obvious overfitting or underfitting phenomena were observed during training, and the performance on the test set was relatively stable. Combined with the specific data in Table 1, it can be concluded that the nonlinear cancellation amount of ConvBiLSTM reaches 7.6dB. In terms of nonlinear cancellation performance, LS+ConvBiLSTM improves by 10% compared to the multinomial model and by 17% compared to LS+FFNN.

[0075] In summary, this invention studies the self-interference cancellation problem in the digital domain for simultaneous, same-frequency, full-duplex systems. Addressing the lack of prior knowledge about the self-interference signal, a Bayesian optimization algorithm is proposed for adaptive estimation of the self-interference channel length. To address the issues of traditional polynomial models relying on prior knowledge of the self-interference signal and the risk of model mismatch, and the inability of FFNNs to effectively handle the high-dimensional temporal characteristics of the self-interference signal, an LS+ConvBiLSTM processing scheme is proposed, which better addresses these problems and further improves the final self-interference cancellation effect. Experimental results show that the proposed self-interference cancellation method based on Conv-BiLSTM and LS achieves a suppression capability of 45.5 dB. In terms of nonlinear cancellation performance, the proposed method improves by 10% compared to the polynomial model and by 17% compared to FFNN. In terms of overall cancellation performance, it shows an 18.8 dB improvement compared to using the Conv-BiLSTM network alone, demonstrating a more significant cancellation effect. This invention has potential application value for self-interference cancellation in simultaneous, same-frequency, full-duplex systems.

[0076] Verification has shown that the method proposed in this invention solves the technical problem raised in this invention, and the method described in this invention has been verified to achieve the technical effects claimed in this invention.

[0077] The algorithm (method) proposed in this invention is the underlying technical core of this invention, and various products can be derived based on the algorithm. Based on the algorithm (method) proposed in this invention, a system for adaptive self-interference cancellation in the digital domain based on Conv-BiLSTM and LS at the same frequency and full duplex is developed using a programming language. This system has program modules corresponding to the steps of the above technical solution, and executes the steps in the aforementioned adaptive self-interference cancellation method in the digital domain based on Conv-BiLSTM and LS at the same frequency and full duplex.

[0078] The developed system (software) computer program is stored on a computer-readable storage medium. This computer program is configured to implement the steps of the aforementioned simultaneous, same-frequency, full-duplex digital domain adaptive self-interference cancellation method based on Conv-BiLSTM and LS when called by a processor. In other words, the invention is materialized on a carrier, becoming a computer program product.

[0079] A simultaneous full-duplex digital domain adaptive self-interference cancellation device based on Conv-BiLSTM and LS is disclosed. The device includes at least one processor and a memory communicatively connected to the at least one processor. The memory stores instructions executable by the at least one processor. The instructions are executed by the at least one processor to enable the at least one processor to execute the aforementioned simultaneous full-duplex digital domain adaptive self-interference cancellation method based on Conv-BiLSTM and LS, thereby achieving feature capture and cancellation of self-interference nonlinear signals.

[0080] Various implementations of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, application-specific integrated circuits (ASICs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include: implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transferring data and instructions to the storage system, the at least one input device, and the at least one output device.

[0081] The computational programs (also referred to as programs, software, software applications, or code) of this invention include machine instructions of a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or apparatus (e.g., disk, optical disk, memory, programmable logic device PLD) for providing machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0082] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this application can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this application can be achieved, they are all within the protection scope of this invention.

Claims

1. A simultaneous, in-frequency, full-duplex digital domain adaptive self-interference cancellation method based on Conv-BiLSTM and LS, wherein the method reduces reliance on prior knowledge of self-interference signals and enhances the adaptability of the overall algorithm model, thereby reducing the risk of channel model mismatch, characterized in that... The method includes: Data preprocessing: The channel length of the self-interference signal is determined based on the Bayesian optimization algorithm. The Bayesian optimization algorithm uses a Gaussian process to model the unknown objective function and infers the distribution of the entire function from the distribution of known data points. The transmitted signal is divided into multiple data segments according to the channel length of the self-interference signal, and each data segment corresponds to the self-interference signal at a certain moment. Establishment of a linear model for self-interference signals: After determining the channel length of the interference signal using the Bayesian optimization algorithm, linear modeling is performed. Once the linear model is constructed, the LS algorithm is used to solve for the model coefficients, thereby estimating the linear components of the interference. Construction of Conv-BiLSTM network: By adding convolutional layers to the front end of the network and then using bidirectional LSTM for deep temporal learning, a Conv-BiLSTM network is constructed. The Conv-BiLSTM network is then used to capture and eliminate the features of self-interference nonlinear signals.

2. The method for simultaneous, same-frequency, full-duplex digital domain adaptive self-interference cancellation based on Conv-BiLSTM and LS according to claim 1, characterized in that, The implementation process of the method is as follows: Data preprocessing Suppose the objective function to be optimized is ,That Given the parameters to be optimized, the goal of Bayesian optimization is to find the global optimum. The goal is to minimize or maximize the objective function. Generalized assumptions (GPs) are used to make prior assumptions about the objective function. Before sufficient observations are available, GPs model the objective function based on the evaluated points, estimating the distribution at each point. The GP process is represented by the following formula: （1） in It is the mean function. For kernel functions; In the early stages of the iteration, a number of points are sampled using random or Latin hypercube sampling methods. Evaluate its objective function. ; With initial data points available, the kernel function parameters are learned through Gaussian process regression, and then based on the observed data points... Inferring the objective function at every point in the search space The posterior distribution: in Let be the posterior mean, representing the mean of the function under the current observations. The optimal guess value at that location; Here, denoted as posterior variance, represents the model's performance on... The uncertainty of the objective function is measured; through multiple iterations, new observation data is obtained to update the parameters of the GP, and the model gradually approaches the true objective function model. When the iteration reaches a certain number of times or the convergence condition is met, the entire optimization process ends. Represented as , Indicates the channel length of the self-interference signal; Its objective function is the mean square error between the optimized self-interference signal and the actual self-interference signal; Establishment of a linear model for self-interference signals After determining the channel length of the interference signal using the Bayesian optimization algorithm, linear modeling is then performed according to Equation 3. （3） in The linear component of the self-interference signal is represented by L, which represents the channel length of the self-interference signal. Representing the Channel response coefficients of each self-interference channel; Once the linear model is constructed, the LS algorithm is used to solve for the model coefficients, thereby estimating the self-disturbed linear components, i.e.: （4） in Send data to baseband; obtain Then, the linear component from the disturbance is estimated, i.e.: （5） in In order to receive data, As labels for subsequent network training; Construction of Conv-BiLSTM Network Based on the characteristics of self-interference signals in full-duplex transceiver systems, it is necessary to comprehensively suppress them at multiple scales in time and space. Recurrent neural networks can capture the dependence of signals in the temporal dimension, and Conv-BiLSTM network structure is used to capture and eliminate the features of self-interference nonlinear signals. First, convolutional layers are used to extract local temporal features and reduce model parameters through weight sharing, thereby improving training efficiency. Simultaneously, convolutional operations can perform preliminary feature extraction on the signal at the front end, filtering out some environmental noise and weak interference. Second, LSTM, as an improved recurrent neural network, effectively alleviates the gradient vanishing and exploding problems in traditional RNNs through gating mechanisms, preserving contextual information over a longer temporal range. Bidirectional LSTM concatenates the hidden states in the forward and backward temporal sequences, enabling the network to comprehensively consider input information from both the forward and backward directions at each time step, establishing a more complete sequence dependency modeling capability for strongly correlated and multipath-intertwined self-interference signals. By adding convolutional layers at the front end of the network and then using bidirectional LSTM for deep temporal learning, multi-scale features of self-interference signals are extracted, thereby achieving higher-precision self-interference cancellation.

3. The method for simultaneous, same-frequency, full-duplex digital domain adaptive self-interference cancellation based on Conv-BiLSTM and LS according to claim 2, characterized in that, The Conv-BiLSTM structure separates the real and imaginary parts of the transmitted baseband signal and constructs a 3D tensor of dimension (batchsize, hSILen, 2), where the parameters represent the amount of data fed into the network for each training iteration, the channel length of the self-interference signal, and the real and imaginary parts of the baseband signal, respectively. Then, a one-dimensional convolution is performed with 8 kernels of size 3. Each kernel extracts different features from the input data, performing a convolution operation on the input at each time step and generating a new feature map. The samples after feature extraction and dimensionality reduction are used as input to the BiLSTM layer. A fully connected layer then connects to the BiLSTM layer, and the output layer has two neurons, representing the real and imaginary parts of the evaluated self-interference nonlinear components, respectively.

4. The method for simultaneous, same-frequency, full-duplex digital domain adaptive self-interference cancellation based on Conv-BiLSTM and LS according to claim 3, characterized in that, In the digital cancellation part, a linear channel model is constructed and the LS algorithm is used to reconstruct the linear self-interference component, and the Conv-BiLSTM method is used to fit the nonlinear self-interference component, thereby reconstructing the complete self-interference signal.

5. The simultaneous, same-frequency, full-duplex digital domain adaptive self-interference cancellation method based on Conv-BiLSTM and LS according to claim 4, characterized in that: The Bayesian optimization algorithm uses a Gaussian process (GP) to model an unknown objective function and infers the distribution of the entire function from the distribution of known data points.

6. A simultaneous, same-frequency, full-duplex digital domain adaptive self-interference cancellation system based on Conv-BiLSTM and LS, characterized in that: The system includes a module for performing the steps corresponding to any one of the claims 1-5. When the module is running, it performs the steps in the simultaneous full-duplex digital domain adaptive self-interference cancellation method based on Conv-BiLSTM and LS.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program configured to, when invoked by a processor, implement the steps of the simultaneous full-duplex digital domain adaptive self-interference cancellation method based on Conv-BiLSTM and LS as described in any one of claims 1-5.

Images