An LSTM-based self-excitation cancellation method and a self-excitation cancellation device

Abstract

This invention relates to the field of mobile communication radio frequency signal processing technology, and particularly to a self-oscillation cancellation method and device based on LSTM. The method involves acquiring a mixed signal from the repeater receiver and a self-oscillation interference reference signal from the repeater coupling end, preprocessing both signals to obtain a preprocessed mixed signal and a self-oscillation interference reference signal. These two signals are then input into a pre-set LSTM model, enabling the model to simultaneously learn the correlation between the interference characteristics in the mixed signal and the pure reference signal, thereby generating a cancellation signal whose amplitude, phase, and timing match the actual self-oscillation interference. After amplitude and phase calibration, this cancellation signal is subtracted from the preprocessed mixed signal in the time domain. The residual interference in the resulting clean, useful communication signal is monitored in real time, and the LSTM model is updated online accordingly.

Description

Technical Field

[0001] This invention relates to the field of mobile communication radio frequency signal processing technology, and in particular to a self-oscillation cancellation method and device based on LSTM. Background Technology

[0002] Wireless repeaters are key devices for extending mobile communication network coverage, primarily used to solve coverage problems in mobile communication blind spots and weak signal areas. They are widely used in mountainous areas, basements, tunnels, large stadiums, and other scenarios. During operation, repeaters are prone to self-oscillation due to factors such as insufficient isolation between transmitting and receiving antennas, channel multipath reflections, environmental obstruction changes, and improper equipment gain adjustment. This can lead to communication signal distortion, increased noise, call interruptions, and even damage to the repeater hardware. Therefore, the self-oscillation cancellation (ICS) algorithm is the core technical support for ensuring the stable and efficient operation of repeaters.

[0003] Currently, the self-oscillation cancellation technology commonly used in wireless repeaters is mainly based on classic adaptive filtering algorithms such as LMS and NLMS. These algorithms rely on linear filtering weight iteration to cancel self-oscillation interference. However, they have technical defects in actual complex mobile communication environments: self-oscillation interference signals have strong time-series correlation, nonlinearity, and time-varying characteristics. Traditional linear adaptive filtering algorithms cannot capture the long-term and short-term time-series characteristics of self-oscillation interference signals, nor can they accurately fit the dynamic transmission characteristics of the interference channel. This results in slow model convergence speed, low interference cancellation accuracy, and large residual interference, which cannot meet the high-quality communication requirements of the 5G era. Summary of the Invention

[0004] The technical problem to be solved by this invention is: how to accurately capture the long and short-term time-series characteristics and dynamic transmission characteristics of self-excited interference signals in order to improve the convergence speed and interference cancellation accuracy.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A self-excitation elimination method based on LSTM includes the following steps: S1. Collect the mixed signal from the receiver of the repeater and the self-excited interference reference signal from the coupling end of the repeater. The mixed signal includes the mobile communication useful signal and the self-excited interference signal. S2. Preprocess the mixed signal and the self-excited interference reference signal respectively to obtain the preprocessed mixed signal and the self-excited interference reference signal; S3. Input the preprocessed mixed signal and the self-excited interference reference signal into a preset LSTM model to generate a cancellation signal; S4. After the cancellation signal is calibrated for amplitude and phase, a time-domain subtraction operation is performed with the preprocessed mixed signal to obtain a pure and useful communication signal. S5. Based on the residual interference of the pure useful communication signal, perform online weight updates on the LSTM model, and adjust the weight update frequency of the LSTM model when the residual interference exceeds a preset threshold or the antenna isolation fluctuation exceeds a preset range.

[0006] Another technical solution adopted in this invention is: A self-excitation cancellation device for performing the above-described LSTM-based self-excitation cancellation method includes: The signal acquisition module is used to acquire the mixed signal from the receiver of the repeater and the self-excited interference reference signal from the coupling end of the repeater. The mixed signal includes mobile communication useful signals and self-excited interference signals. A preprocessing module, connected to the signal acquisition module, is used to preprocess the mixed signal and the self-excited interference reference signal, and output the preprocessed mixed signal and the self-excited interference reference signal. An LSTM interference prediction module, connected to the preprocessing module, is used to receive the preprocessed mixed signal and the self-excited interference reference signal, and generate a cancellation signal. An amplitude and phase calibration module, connected to the LSTM interference prediction module, is used to perform amplitude and phase calibration on the cancellation signal and output the calibrated cancellation signal. The time-domain subtraction cancellation module is connected to the preprocessing module and the amplitude and phase calibration module respectively. It is used to perform time-domain subtraction operation on the calibrated cancellation signal and the preprocessed mixed signal to output a clean and useful communication signal. The online update control module is connected to the time-domain subtraction cancellation module and the LSTM interference prediction module, respectively. It is used to update the weights of the LSTM model online according to the interference residue of the clean useful communication signal, and adjust the weight update frequency of the LSTM model when the interference residue exceeds a preset threshold or the antenna isolation fluctuation exceeds a preset range.

[0007] The beneficial effects of this invention are as follows: This scheme acquires a mixed signal from the repeater receiver and a self-excited interference reference signal from the repeater coupling end. Both signals are preprocessed to obtain a preprocessed mixed signal and a self-excited interference reference signal. These two signals are then input into a pre-defined LSTM model, allowing the model to simultaneously learn the correlation between the interference characteristics in the mixed signal and the pure reference signal. This generates a cancellation signal whose amplitude, phase, and timing match the actual self-excited interference. After amplitude and phase calibration, the cancellation signal is subtracted from the preprocessed mixed signal in the time domain. The residual interference in the resulting clean, useful communication signal is monitored in real time, and the LSTM model is updated online accordingly. When the residual interference exceeds a preset threshold or the antenna isolation fluctuation exceeds a preset range, the weight update frequency is further adjusted. In this method, the generation effect of the LSTM model directly affects the accuracy of the cancellation calibration, while the residual interference after cancellation controls the update intensity of the LSTM model in a closed loop. This enables the entire method to adaptively track channel changes, significantly improving convergence speed and interference cancellation accuracy. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating the steps of the self-excitation elimination method based on LSTM of the present invention. Figure 2 This is a connection block diagram of the LSTM-based self-excitation elimination device of the present invention; Label Explanation: 1. Signal acquisition module; 2. Preprocessing module; 3. LSTM interference prediction module; 4. Amplitude and phase calibration module; 5. Time-domain subtraction cancellation module; 6. Online update control module. Detailed Implementation

[0009] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0010] Please refer to Figure 1 A self-excitation elimination method based on LSTM includes the following steps: S1. Collect the mixed signal from the receiver of the repeater and the self-excited interference reference signal from the coupling end of the repeater. The mixed signal includes the mobile communication useful signal and the self-excited interference signal. S2. Preprocess the mixed signal and the self-excited interference reference signal respectively to obtain the preprocessed mixed signal and the self-excited interference reference signal; S3. Input the preprocessed mixed signal and the self-excited interference reference signal into a preset LSTM model to generate a cancellation signal; S4. After the cancellation signal is calibrated for amplitude and phase, a time-domain subtraction operation is performed with the preprocessed mixed signal to obtain a pure and useful communication signal. S5. Based on the residual interference of the pure useful communication signal, perform online weight updates on the LSTM model, and adjust the weight update frequency of the LSTM model when the residual interference exceeds a preset threshold or the antenna isolation fluctuation exceeds a preset range.

[0011] As can be seen from the above description, the beneficial effects of the present invention are as follows: This scheme acquires a mixed signal from the repeater receiver and a self-excited interference reference signal from the repeater coupling end. Both signals are preprocessed to obtain a preprocessed mixed signal and a self-excited interference reference signal. These two signals are then input into a pre-defined LSTM model, allowing the model to simultaneously learn the correlation between the interference characteristics in the mixed signal and the pure reference signal. This generates a cancellation signal whose amplitude, phase, and timing match the actual self-excited interference. After amplitude and phase calibration, the cancellation signal is subtracted from the preprocessed mixed signal in the time domain. The residual interference in the resulting clean, useful communication signal is monitored in real time, and the LSTM model is updated online accordingly. When the residual interference exceeds a preset threshold or the antenna isolation fluctuation exceeds a preset range, the weight update frequency is further adjusted. In this method, the generation effect of the LSTM model directly affects the accuracy of the cancellation calibration, while the residual interference after cancellation controls the update intensity of the LSTM model in a closed loop. This enables the entire method to adaptively track channel changes, significantly improving convergence speed and interference cancellation accuracy.

[0012] Furthermore, the preprocessing in step S2 includes baseband digitization processing, adaptive Kalman filter noise reduction processing, and amplitude normalization processing; The core formula for adaptive Kalman filter noise reduction is: State prediction equation: ; Covariance prediction equation: ; Kalman gain equation: ; State update equation: ; Covariance update equation: ; in, for The predicted state value at time 10:00. Here is the state transition matrix. for State estimate at time 10:00 For the control matrix, for Control input value at any time, Predict the covariance at time k. For process noise covariance, for Kalman gain at time step For the observation matrix, To observe the noise covariance, for The observed value at time, It is the identity matrix; The amplitude normalization process uses the min-max normalization algorithm, the core formula of which is: ; in, The normalized signal amplitude mapped to the interval [-1, 1] The original signal amplitude, The maximum amplitude of the original signal. This represents the minimum amplitude of the original signal.

[0013] As described above, after baseband digitization of the mixed signal and the self-excited interference reference signal, adaptive Kalman filtering is used for noise reduction, which can eliminate radio frequency noise and abnormal pulse data, thereby improving the signal-to-noise ratio. Then, amplitude normalization is performed to map the signal amplitude to the [-1,1] interval, avoiding excessive amplitude differences from affecting the training and computation accuracy of the LSTM model. The quality of the preprocessed signal directly determines the reliability of the input data of the LSTM model, thus laying the foundation for the subsequent generation of high-precision cancellation signals and further improving the convergence speed and interference cancellation accuracy.

[0014] Furthermore, the LSTM model employs a single-layer or multi-layer long short-term memory network structure, and the gating structure formula of the LSTM model includes: Forget gate operation formula: ; Input gate operation formula: ; Cell state candidate value formula: ; Cell state update formula:

[0015] Output gate operation formula: ; Hidden state update formula: ; in, The output value of the forget gate. The output value of the input gate. The output value of the output gate. It is the sigmoid activation function. The hyperbolic tangent activation function is used. Here is the weight matrix for the forget gate. Here is the weight matrix of the input gate. This is the weight matrix for candidate cell states. This is the weight matrix of the output gate. For the bias term of the forget gate, For the bias term of the input gate, This is a bias term for candidate cell state values. This is the bias term for the output gate. for The hidden state at all times for The hidden state at all times for Input signal at time, for Candidate values ​​of cell state at time t. for Cellular state at any given moment for Cellular state at any given moment This is an element-wise multiplication operation.

[0016] As described above, the LSTM model employs gating structures such as forget gate, input gate, and output gate. The forget gate determines the degree to which the cell state from the previous time step is retained, the input gate controls the addition of new information, and the output gate determines the output of the current hidden state. Through the coordinated operation of these gating formulas, the model can selectively remember or forget the long-term and short-term temporal characteristics of self-excited interference signals, effectively solving the gradient vanishing and gradient explosion problems of traditional RNNs, enabling the model to stably fit the dynamic transmission characteristics of interference under complex time-varying channels.

[0017] Furthermore, the output formula for the canceled signal in step S3 is: ; in, for The cancellation signal output at any time, This is the output layer weight matrix. For output layer bias terms, To suppress model overfitting.

[0018] As described above, the hidden state of the LSTM model is mapped to the cancellation signal by the output formula of the cancellation signal. This formula directly utilizes the time-series features extracted by the model to ensure that the output cancellation signal is highly matched with the actual self-excited interference signal in terms of amplitude, phase and time sequence, providing an accurate reference signal for subsequent subtraction cancellation operation, thereby effectively reducing the amount of interference residue.

[0019] Furthermore, the LSTM model also includes a dropout regularization layer, the formula for which the dropout regularization layer is calculated is: ; in, This is the hidden state after dropout processing. For dropout probability, For a mask matrix that follows a Bernoulli distribution, To suppress model overfitting.

[0020] As described above, adding a dropout regularization layer to the LSTM model suppresses the model's excessive reliance on training data by randomly discarding some neuron outputs. This mechanism can prevent the model from overfitting in complex time-varying channel scenarios, improve the model's generalization ability and adaptability to different channel environments, and enable the same model to work stably in various repeater application scenarios.

[0021] Furthermore, in step S4, the amplitude and phase calibration employs the minimum mean square error calibration algorithm. The core formula of the minimum mean square error calibration algorithm includes: Calibration error objective function: ; Amplitude calibration formula: ; Phase calibration formula: ; in, For calibration error, To calibrate the data length, For ideal cancellation signal, The original cancellation signal output by the LSTM model. The amplitude of the calibrated cancellation signal. The amplitude of the cancellation signal before calibration. For amplitude calibration coefficient, The phase of the calibrated cancellation signal. The phase of the cancellation signal before calibration. This is for phase calibration deviation.

[0022] As described above, the minimum mean square error calibration algorithm is used to calibrate the amplitude and phase of the cancellation signal output by the LSTM model. The calibration coefficients are optimized by the calibration error objective function to make the amplitude deviation ≤0.5dB and the phase deviation ≤5°. The consistency between the calibrated cancellation signal and the real self-excited interference is significantly improved, providing a higher precision input for time-domain subtraction operation and further improving the final effect of self-excited interference elimination.

[0023] Furthermore, the core formula for the time-domain subtraction operation in step S4 is: ; in, For pure and useful communication signals, This is the preprocessed mixed signal. This is to cancel out the signal.

[0024] As can be seen from the above description, the operation structure is simple and has low processing latency. It can remove the self-excited interference component in the mixed signal in real time and output a clean and useful communication signal, which meets the latency requirements of the wireless repeater for real-time signal processing, while ensuring the accuracy of interference cancellation.

[0025] Furthermore, the formula for calculating the residual amount of interference in step S5 is as follows: ; in, The residual interference level (unit: dB). To compensate for the residual interference signal after cancellation, The original mixed signal, This represents the length of the signal data.

[0026] As can be seen from the above description, the residual interference level after cancellation is quantified in real time through the interference residual amount calculation formula. This residual amount directly reflects the cancellation effect of the current LSTM model and can be used as the optimization target for online weight update, so that the model can automatically adjust the update intensity according to the size of the residual amount, ensuring the long-term stability of the self-excitation elimination effect.

[0027] Furthermore, in step S5, the LSTM model is updated online using the Adam optimization algorithm. The core formula of the Adam optimization algorithm is: First-order moment estimation: ; Second-order moment estimation: ; First-order moment deviation correction: ; Second-order moment deviation correction: ; Weight update formula: ; in, This is a first-moment estimate of the gradient. For the second moment estimation of the gradient, For momentum parameters, =0.9, For momentum parameters, =0.999, for The gradient of the loss function at time step 1. The corrected first moment, The corrected second moment, for Model weights at time step for Model weights at time step For learning rate, The initial value is 0.001~0.01. To prevent tiny constants with a denominator of 0, =10 -8 .

[0028] As described above, the Adam optimization algorithm is used to update the weights of the LSTM model online. Its first-order moment estimation, second-order moment estimation, and bias correction formula can adaptively adjust the update step size of each weight. The initial learning rate is set to 0.001~0.01 and dynamically adjusted according to the channel interference intensity and interference residue. No offline pre-training is required throughout the process. When the interference residue exceeds the threshold or the antenna isolation fluctuation exceeds the preset range, the model can automatically speed up or slow down the weight update frequency to quickly adapt to channel changes and ensure that the signal processing delay is ≤1ms.

[0029] Please refer to Figure 2 A self-excitation cancellation device for performing the above-described LSTM-based self-excitation cancellation method includes: Signal acquisition module 1 is used to acquire the mixed signal at the receiver end of the repeater and the self-excited interference reference signal at the coupling end of the repeater. The mixed signal includes mobile communication useful signal and self-excited interference signal. Preprocessing module 2, connected to signal acquisition module 1, is used to preprocess the mixed signal and self-excited interference reference signal, and output the preprocessed mixed signal and self-excited interference reference signal; The LSTM interference prediction module 3 is connected to the preprocessing module 2 and is used to receive the preprocessed mixed signal and the self-excited interference reference signal to generate a cancellation signal. The amplitude and phase calibration module 4 is connected to the LSTM interference prediction module 3 and is used to perform amplitude and phase calibration on the cancellation signal and output the calibrated cancellation signal. The time-domain subtraction cancellation module 5 is connected to the preprocessing module 2 and the amplitude and phase calibration module 4 respectively. It is used to perform time-domain subtraction operation on the calibrated cancellation signal and the preprocessed mixed signal to output a clean and useful communication signal. The online update control module 6 is connected to the time-domain subtraction cancellation module 5 and the LSTM interference prediction module 3, respectively. It is used to update the weights of the LSTM model online according to the interference residue of the clean useful communication signal, and adjust the weight update frequency of the LSTM model when the interference residue exceeds a preset threshold or the antenna isolation fluctuation exceeds a preset range.

[0030] As can be seen from the above description, the beneficial effects of the present invention are as follows: This scheme comprises a signal acquisition module 1, a preprocessing module 2, an LSTM interference prediction module 3, an amplitude and phase calibration module 4, a time-domain subtraction cancellation module 5, and an online update control module 6. The signal acquisition module 1 simultaneously acquires the mixed signal from the repeater receiver and the self-excited interference reference signal from the coupling end. These two signals are fed into the preprocessing module 2 for preprocessing, and the output high-quality signal is directly used as the input to the LSTM interference prediction module 3. The LSTM interference prediction module 3 uses its internal control structure to extract long- and short-term time-series features and generate a cancellation signal. After the residual deviation is corrected by the amplitude and phase calibration module 4, this cancellation signal is subtracted from the preprocessed mixed signal in the time-domain subtraction cancellation module 5 to obtain a clean and useful communication signal. The online update control module 6 then updates the signal based on the cancelled signal. The interference residue of the signal and the fluctuation of antenna isolation are used to dynamically adjust the weight update frequency of the LSTM model. The entire device forms a closed-loop control from acquisition to output: the quality of the preprocessing module 2 affects the fitting accuracy of the LSTM, the output of the LSTM affects the calibration and cancellation effect, and the residue after cancellation, in turn, determines the speed of model update. This close cooperation allows the device to be directly embedded into the digital baseband processing unit of existing repeaters without significant changes to the hardware architecture. It also supports both FPGA and ASIC implementations, allowing for flexible selection based on cost and power consumption in engineering. In practical use, operators do not need to manually adjust parameters; the device can automatically track channel changes and isolation fluctuations, and will not experience self-oscillation during long-term operation, significantly reducing maintenance workload.

[0031] Please refer to Figure 1 One embodiment of the present invention is as follows: A self-excitation elimination method based on LSTM includes the following steps: S1. Acquire the mixed signal from the repeater receiver and the self-excited interference reference signal from the repeater coupling end. The mixed signal includes the mobile communication useful signal and the self-excited interference signal. The two signals are acquired synchronously with synchronized sampling clocks to ensure timing consistency. The acquisition duration is 10 seconds to provide sufficient data support for subsequent preprocessing and model calculation.

[0032] S2. Preprocess the mixed signal and the self-excited interference reference signal respectively to obtain the preprocessed mixed signal and the self-excited interference reference signal; The preprocessing in step S2 includes baseband digitization, adaptive Kalman filter noise reduction, and amplitude normalization. The two acquired signals undergo baseband digitization processing. The analog signal is demodulated to baseband using an RF demodulation module, and then converted to a digital signal using a 12-bit analog-to-digital converter (ADC). The signal sampling rate is set to 10MHz. The core formula for the ADC is: ; in, The digitized signal value. The original analog signal value, This is the maximum input voltage for analog-to-digital conversion (1.2V in this embodiment). The minimum input voltage for analog-to-digital conversion (taken as -1.2V in this embodiment), 2 12 -1 represents the maximum quantization value for a 12-bit analog-to-digital converter.

[0033] The core formula for adaptive Kalman filter noise reduction is: State prediction equation: ; Covariance prediction equation: ; Kalman gain equation: ; State update equation: ; Covariance update equation: ; in, for The predicted state value at time 10:00. Here is the state transition matrix. for State estimate at time 10:00 For the control matrix, for Control input value at any time, Predict the covariance at time k. For process noise covariance, for Kalman gain at time step For the observation matrix, To observe the noise covariance, for The observed value at time, It is the identity matrix; An adaptive Kalman filter algorithm is used to reduce noise in the digital signal, with the filter window set to 512 points; the Kalman filter parameters are: State transition matrix Control matrix Observation matrix Process noise covariance Observation noise covariance Initial covariance After removing abnormal data such as radio frequency noise and pulse interference, the signal-to-noise ratio (SNR) is improved to over 35 dB. The SNR calculation formula is as follows: ; in, Signal-to-noise ratio (unit: dB) Useful signal This is a noise signal. The signal length is used; the amplitude of the denoised signal is normalized. The amplitude normalization process uses the min-max normalization algorithm, the core formula of which is: ; in, The normalized signal amplitude mapped to the interval [-1, 1] The original signal amplitude, The maximum amplitude of the original signal. , This represents the minimum amplitude of the original signal. Mapping the signal amplitude to the [-1,1] interval can avoid excessive differences in signal amplitude that could affect the training and computation accuracy of the LSTM model. This completes signal preprocessing, and the preprocessed signal has no obvious noise or abnormal fluctuations, so it can be directly input into the LSTM model for computation.

[0034] S3. Input the preprocessed mixed signal and the self-excited interference reference signal into a preset LSTM model to generate a cancellation signal; The LSTM model uses a single-layer or multi-layer long short-term memory network structure. This embodiment takes a two-layer model as an example. A two-layer LSTM model was constructed, with 2 neurons in the input layer, corresponding to the preprocessed mixed signal and the self-excited interference reference signal. The first LSTM hidden layer had 128 neurons, and the second LSTM hidden layer had 64 neurons. The sigmoid activation function was used for the input gate, forget gate, and output gate, while the tanh activation function was used for cell state updates. The core gating formulas and parameter values ​​of the LSTM model are as follows: The gating structure formula of the LSTM model includes: Forget gate operation formula: ; in, , ; Input gate operation formula: ; in, , ; Cell state candidate value formula: ; in, , ; Cell state update formula:

[0035] Output gate operation formula: ; in, , ; Hidden state update formula: ; The parameters of the second LSTM layer are the same as those of the first layer, except that the dimensions of the weight matrix are adjusted. The bias dimension is adjusted to ; in, The output value of the forget gate. The output value of the input gate. The output value of the output gate. It is the sigmoid activation function. The hyperbolic tangent activation function is used. Here is the weight matrix for the forget gate. Here is the weight matrix of the input gate. This is the weight matrix for candidate cell states. This is the weight matrix of the output gate. For the bias term of the forget gate, For the bias term of the input gate, This is a bias term for candidate cell state values. This is the bias term for the output gate. for The hidden state at all times for The hidden state at all times for Input signal at time, for Candidate values ​​of cell state at time t. for Cellular state at any given moment for Cellular state at any given moment This is an element-wise multiplication operation.

[0036] The LSTM model also includes a dropout regularization layer, and the calculation formula for the dropout regularization layer is as follows: ; in, This is the hidden state after dropout processing. For dropout probability, For a mask matrix that follows a Bernoulli distribution, To suppress model overfitting.

[0037] If the dropout probability is set to 0.2, then the formula for the dropout regularization layer is: ; The preprocessed time-series signal is input frame by frame into the LSTM model with a frame length of 256 points and a step size of 128 points. The frame length is calculated using the following formula: ; in, For frame length, The signal sampling rate, The duration of a single frame signal is given. The long and short-term timing features of the self-excited interference signal are extracted using the gating structure of the LSTM model. The dynamic transmission characteristics of the self-excited interference channel are fitted. The LSTM model operation delay is ≤0.8ms. The output is a cancellation signal that perfectly matches the amplitude, phase, and timing of the actual self-excited interference signal. The output formula for the cancellation signal in step S3 is: ; in, for The cancellation signal output at any time, This is the output layer weight matrix. , For output layer bias terms, , To suppress model overfitting.

[0038] S4. After the cancellation signal is calibrated for amplitude and phase, a time-domain subtraction operation is performed with the preprocessed mixed signal to obtain a pure and useful communication signal. In step S4, the amplitude and phase calibration uses the minimum mean square error calibration algorithm, with the number of iterations set to 100. The core formula of the minimum mean square error calibration algorithm includes: Calibration error objective function: ; Amplitude calibration formula: ,in, The value range is [0.95, 1.05]; Phase calibration formula: ,in The value range is [-3°, 3°]; in, For calibration error, To calibrate the data length, For ideal cancellation signal, The original cancellation signal output by the LSTM model. The amplitude of the calibrated cancellation signal. The amplitude of the cancellation signal before calibration. For amplitude calibration coefficient, The phase of the calibrated cancellation signal. The phase of the cancellation signal before calibration. This is for phase calibration deviation.

[0039] After calibration, the amplitude deviation of the corrected cancellation signal is ≤0.3dB and the phase deviation is ≤3°. The formula for calculating the amplitude deviation is: ; in The amplitude deviation is expressed in dB. To obtain the ideal cancellation signal amplitude (equal to the amplitude of the self-excited interference signal), the calibrated cancellation signal is subtracted from the preprocessed mixed signal in the time domain. The core formula for the time-domain subtraction operation is: ; in, For pure and useful communication signals, This is the preprocessed mixed signal. To cancel the signal; after the self-excited interference is eliminated, the residual interference of the canceled signal is ≤-45dB.

[0040] S5. Based on the residual interference of the pure useful communication signal, perform online weight updates on the LSTM model, and adjust the weight update frequency of the LSTM model when the residual interference exceeds a preset threshold or the antenna isolation fluctuation exceeds a preset range.

[0041] The formula for calculating the residual amount of interference in step S5 is: ; in, The residual interference level (unit: dB). To compensate for the residual interference signal after cancellation, The original mixed signal, For signal data length, , Substituting the data from this embodiment (the amplitude of the original mixed signal is 1.2V, the amplitude of the calibrated canceled signal is 1.197V, and the amplitude of the canceled residual interference signal is 0.003V), the calculation yields... The residual interference level is completely consistent with that in the effect verification, ensuring that the calculation logic is consistent and the data is matched.

[0042] An online real-time training and weight update mechanism for the LSTM model is established, with the residual interference of the canceled signal as the optimization objective. The Adam optimization algorithm is used to update the weights of the LSTM model online. The core formula of the Adam optimization algorithm is: First-order moment estimation: ; Second-order moment estimation: ; First-order moment deviation correction: ; Second-order moment deviation correction: ; Weight update formula: ; in, This is a first-moment estimate of the gradient. For the second moment estimation of the gradient, For momentum parameters, =0.9, For momentum parameters, =0.999, for The gradient of the loss function at time step 1. The corrected first moment, The corrected second moment, for Model weights at time step for Model weights at time step For learning rate, The initial value is 0.001~0.01. To prevent tiny constants with a denominator of 0, =10 -8 .

[0043] The initial learning rate is set to 0.005. When the channel interference intensity increases (interference amplitude increases by 10 dB, the calculation formula for interference amplitude increase is...), the learning rate is adjusted accordingly. Substituting the data yields When the interference amplitude increases from 0.5V to 1.58V or the antenna isolation fluctuation exceeds the -60dB threshold, the learning rate automatically increases to 0.01; when the residual interference is ≤-50dB, the learning rate automatically decreases to 0.001 to reduce computational power consumption. No offline pre-training is required throughout the process; the LSTM model iteratively optimizes in real time, updating the weights after each frame of signal processing. This quickly adapts to dynamic channel changes and antenna isolation fluctuations, achieving closed-loop adaptive self-excitation elimination. The antenna isolation calculation formula is: ; in, Antenna isolation (unit: dB) For the output power of the transmitting antenna, This refers to the transmit power coupled to the receiving antenna.

[0044] An FPGA hardware deployment scheme can be adopted, using Xilinx Zynq-7000 series FPGA chips. The logic of LSTM algorithm, signal preprocessing, signal calibration, subtraction cancellation, and weight update is ported through Verilog language. The LSTM model operation module is optimized in parallel and a pipelined architecture is adopted to improve the operation efficiency and ensure that the signal processing latency is ≤0.8ms, which meets the real-time processing requirements of 5G mobile communication. The FPGA chip is integrated into the digital baseband processing unit of the wireless repeater, which is compatible with the existing baseband system and can be directly embedded into the existing repeater equipment. The engineering transformation difficulty is low and the hardware power consumption is controlled within 5W.

[0045] Alternatively, a customized ASIC chip design can be adopted, integrating LSTM models, signal preprocessing, signal calibration, and other algorithm logic into the ASIC chip. The chip uses a 28nm process, further reducing hardware power consumption (power consumption ≤2W) and improving computing efficiency (signal processing latency ≤0.5ms). It is suitable for the miniaturized deployment requirements of indoor distributed repeaters and can be directly embedded into small repeater baseband modules without occupying additional hardware resources. Other technical parameters are the same as in Example 1, and the self-excitation cancellation effect is the same (interference residue ≤-45dB, convergence speed is improved by more than 70% compared with traditional algorithms).

[0046] Compared with traditional LMS and NLMS adaptive filtering ICS elimination algorithms, this scheme relies on the gating structure advantages and long-short-term time series feature capture capabilities of LSTM neural networks, and has the following significant beneficial effects, fully demonstrating the novelty, inventiveness and practicality of the invention: First, the convergence speed is greatly improved. The LSTM model effectively solves the gradient vanishing problem of traditional RNN through the gating structure. It can quickly capture the long and short-term time-series characteristics of self-excited interference signals. The speed of fitting the dynamic changes of the channel is more than 70% faster than traditional algorithms. It can quickly respond to channel changes and antenna isolation fluctuations and avoid self-excited oscillations. Second, the self-excitation cancellation accuracy is significantly improved. It can accurately fit the transmission characteristics of nonlinear and multipath time-varying channels and generate a cancellation signal that perfectly matches the actual self-excitation interference. The interference residue (negative value) is reduced by 16dB~18dB compared with the traditional algorithm. The interference suppression effect is significantly better than the traditional algorithm. The communication signal distortion is ≤0.1%, which meets the requirements of 5G high-quality communication. Third, the system robustness is significantly enhanced. Through the dropout regularization layer and online weight update mechanism, it can stably adapt to complex outdoor / indoor communication environments and antenna isolation fluctuations, without the need for manual parameter adjustment, and the continuous stable operation time of the repeater is increased by more than 80%. Fourth, it has wider engineering adaptability, is compatible with various wireless repeater equipment and all mobile communication frequency bands, and has flexible hardware deployment. It can be implemented through FPGA / ASIC solutions and can be directly embedded into existing repeater baseband systems. It is easy to modify and the cost is controllable. Fifth, the working gain of the repeater is effectively improved. Under the same antenna deployment conditions, the equipment gain can be increased by 10dB~18dB, reducing the stringent requirements for antenna isolation during engineering installation (isolation requirement is reduced from -80dB to -60dB), simplifying the construction process and reducing engineering construction costs. Sixth, the model has strong generalization ability and can adapt to self-excited interference scenarios of different intensities and types. There is no need to retrain the model for specific scenarios, which significantly improves its practicality. It is an extension of the advantages of LSTM models in high-precision time series prediction scenarios.

[0047] Please refer to Figure 2 One embodiment of the present invention is as follows: A self-excitation cancellation device for performing the above-described LSTM-based self-excitation cancellation method includes: Signal acquisition module 1 is used to acquire the mixed signal at the receiver end of the repeater and the self-excited interference reference signal at the coupling end of the repeater. The mixed signal includes mobile communication useful signal and self-excited interference signal. Preprocessing module 2, connected to signal acquisition module 1, is used to preprocess the mixed signal and self-excited interference reference signal, and output the preprocessed mixed signal and self-excited interference reference signal; The LSTM interference prediction module 3 is connected to the preprocessing module 2 and is used to receive the preprocessed mixed signal and the self-excited interference reference signal to generate a cancellation signal. The amplitude and phase calibration module 4 is connected to the LSTM interference prediction module 3 and is used to perform amplitude and phase calibration on the cancellation signal and output the calibrated cancellation signal. The time-domain subtraction cancellation module 5 is connected to the preprocessing module 2 and the amplitude and phase calibration module 4 respectively. It is used to perform time-domain subtraction operation on the calibrated cancellation signal and the preprocessed mixed signal to output a clean and useful communication signal. The online update control module 6 is connected to the time-domain subtraction cancellation module 5 and the LSTM interference prediction module 3, respectively. It is used to update the weights of the LSTM model online according to the interference residue of the clean useful communication signal, and adjust the weight update frequency of the LSTM model when the interference residue exceeds a preset threshold or the antenna isolation fluctuation exceeds a preset range.

[0048] In summary, the self-oscillation cancellation method and device based on LSTM provided by this invention acquires the mixed signal from the repeater receiver and the self-oscillation interference reference signal from the repeater coupling end, and preprocesses the two signals to obtain the preprocessed mixed signal and the self-oscillation interference reference signal. These two signals are then input into a preset LSTM model, enabling the model to simultaneously learn the correlation between the interference characteristics in the mixed signal and the pure reference signal, thereby generating a cancellation signal whose amplitude, phase, and timing match the actual self-oscillation interference. After amplitude and phase calibration, the cancellation signal is subtracted from the preprocessed mixed signal in the time domain. The residual interference in the resulting clean and useful communication signal is monitored in real time, and the LSTM model is updated online accordingly. When the residual interference exceeds a preset threshold or the antenna isolation fluctuation exceeds a preset range, the weight update frequency is further adjusted. In this method, the generation effect of the LSTM model directly affects the accuracy of the cancellation calibration, while the residual interference after cancellation controls the update intensity of the LSTM model in a closed loop, enabling the entire method to adaptively track channel changes and significantly improve convergence speed and interference cancellation accuracy.

[0049] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A self-excitation elimination method based on LSTM, characterized in that, Includes the following steps: S1. Collect the mixed signal from the receiver of the repeater and the self-excited interference reference signal from the coupling end of the repeater. The mixed signal includes the mobile communication useful signal and the self-excited interference signal. S2. Preprocess the mixed signal and the self-excited interference reference signal respectively to obtain the preprocessed mixed signal and the self-excited interference reference signal; S3. Input the preprocessed mixed signal and the self-excited interference reference signal into a preset LSTM model to generate a cancellation signal; S4. After the cancellation signal is calibrated for amplitude and phase, a time-domain subtraction operation is performed with the preprocessed mixed signal to obtain a pure and useful communication signal. S5. Based on the residual interference of the pure useful communication signal, perform online weight updates on the LSTM model, and adjust the weight update frequency of the LSTM model when the residual interference exceeds a preset threshold or the antenna isolation fluctuation exceeds a preset range.

2. The self-excitation elimination method based on LSTM according to claim 1, characterized in that, The preprocessing in step S2 includes baseband digitization, adaptive Kalman filter noise reduction, and amplitude normalization. The core formula for adaptive Kalman filter noise reduction is: State prediction equation: ; Covariance prediction equation: ; Kalman gain equation: ; State update equation: ; Covariance update equation: ; in, for The predicted state value at time 10:

00. Here is the state transition matrix. for State estimate at time 10:00 For the control matrix, for Control input value at any time, Predict the covariance at time k. For process noise covariance, for Kalman gain at time step For the observation matrix, To observe the noise covariance, for The observed value at time, It is the identity matrix; The amplitude normalization process uses the min-max normalization algorithm, the core formula of which is: ； in, The normalized signal amplitude mapped to the interval [-1, 1] The original signal amplitude, The maximum amplitude of the original signal. This represents the minimum amplitude of the original signal.

3. The self-excitation elimination method based on LSTM according to claim 1, characterized in that, The LSTM model employs a single-layer or multi-layer long short-term memory network structure, and the gating structure formula of the LSTM model includes: Forget gate operation formula: ; Input gate operation formula: ; Cell state candidate value formula: ; Cell state update formula: Output gate operation formula: ; Hidden state update formula: ; in, The output value of the forget gate. The output value of the input gate. The output value of the output gate. It is the sigmoid activation function. The hyperbolic tangent activation function is used. Here is the weight matrix for the forget gate. Here is the weight matrix of the input gate. This is the weight matrix for candidate cell states. This is the weight matrix of the output gate. For the bias term of the forget gate, For the bias term of the input gate, This is a bias term for candidate cell state values. This is the bias term for the output gate. for The hidden state at all times for The hidden state at all times for Input signal at time, for Candidate values ​​of cell state at time t. for Cellular state at any given moment for Cellular state at any given moment This is an element-wise multiplication operation.

4. The self-excitation elimination method based on LSTM according to claim 1, characterized in that, The output formula for the cancellation signal in step S3 is: ； in, for The cancellation signal output at any time, This is the output layer weight matrix. For output layer bias terms, To suppress model overfitting.

5. The self-excitation elimination method based on LSTM according to claim 1, characterized in that, The LSTM model also includes a dropout regularization layer, and the calculation formula for the dropout regularization layer is as follows: ； in, This is the hidden state after dropout processing. For dropout probability, For a mask matrix that follows a Bernoulli distribution, To suppress model overfitting.

6. The self-excitation elimination method based on LSTM according to claim 1, characterized in that, In step S4, the amplitude and phase calibration uses the minimum mean square error calibration algorithm. The core formula of the minimum mean square error calibration algorithm includes: Calibration error objective function: ; Amplitude calibration formula: ; Phase calibration formula: ; in, For calibration error, To calibrate the data length, For ideal cancellation signal, The original cancellation signal output by the LSTM model. The amplitude of the calibrated cancellation signal. The amplitude of the cancellation signal before calibration. For amplitude calibration coefficient, The phase of the calibrated cancellation signal. The phase of the cancellation signal before calibration. This is for phase calibration deviation.

7. The self-excitation elimination method based on LSTM according to claim 1, characterized in that, The core formula for the time-domain subtraction operation in step S4 is: ； in, For pure and useful communication signals, This is the preprocessed mixed signal. This is to cancel out the signal.

8. The self-excitation elimination method based on LSTM according to claim 1, characterized in that, The formula for calculating the residual amount of interference in step S5 is: ； in, The residual interference level (unit: dB). To compensate for the residual interference signal after cancellation, The original mixed signal, This represents the length of the signal data.

9. The self-excitation elimination method based on LSTM according to claim 1, characterized in that, In step S5, the LSTM model is updated online using the Adam optimization algorithm. The core formula of the Adam optimization algorithm is: First-order moment estimation: ; Second-order moment estimation: ; First-order moment deviation correction: ; Second-order moment deviation correction: ; Weight update formula: ; in, This is a first-moment estimate of the gradient. For the second moment estimation of the gradient, For momentum parameters, =0.9, For momentum parameters, =0.999, for The gradient of the loss function at time step 1. The corrected first moment, The corrected second moment, for Model weights at time step for Model weights at time step For learning rate, The initial value is 0.001~0.

01. To prevent tiny constants with a denominator of 0, =10 -8 .

10. A self-excitation elimination device for performing the LSTM-based self-excitation elimination method according to any one of claims 1 to 9, characterized in that, include: The signal acquisition module is used to acquire the mixed signal from the receiver of the repeater and the self-excited interference reference signal from the coupling end of the repeater. The mixed signal includes mobile communication useful signals and self-excited interference signals. A preprocessing module, connected to the signal acquisition module, is used to preprocess the mixed signal and the self-excited interference reference signal, and output the preprocessed mixed signal and the self-excited interference reference signal. An LSTM interference prediction module, connected to the preprocessing module, is used to receive the preprocessed mixed signal and the self-excited interference reference signal, and generate a cancellation signal. An amplitude and phase calibration module, connected to the LSTM interference prediction module, is used to perform amplitude and phase calibration on the cancellation signal and output the calibrated cancellation signal. The time-domain subtraction cancellation module is connected to the preprocessing module and the amplitude and phase calibration module respectively. It is used to perform time-domain subtraction operation on the calibrated cancellation signal and the preprocessed mixed signal to output a clean and useful communication signal. The online update control module is connected to the time-domain subtraction cancellation module and the LSTM interference prediction module, respectively. It is used to update the weights of the LSTM model online according to the interference residue of the clean useful communication signal, and adjust the weight update frequency of the LSTM model when the interference residue exceeds a preset threshold or the antenna isolation fluctuation exceeds a preset range.

Images