RRU multi-band passive intermodulation elimination method based on sequence-to-sequence learning model

Abstract

The invention discloses an RRU multi-band passive intermodulation elimination method based on a sequence-to-sequence learning model, and the method comprises the steps: obtaining a transmitting signal and a receiving signal of a multi-band RRU, carrying out the feature extraction of the signals through adaptive transformation, and obtaining the multi-band intermodulation interference feature data; generating a nonlinear intermodulation feature vector based on the multi-band intermodulation interference feature data; training a PIM signal prediction model by using a sequence-to-sequence learning model, and performing residual compensation on a prediction result based on an error optimization mechanism; performing adaptive filtering processing on the predicted PIM interference signal to generate an optimized interference compensation signal; and dynamically updating the PIM signal prediction model by using an online learning mechanism based on the signal data after interference suppression. According to the method, the multi-band nonlinear intermodulation characteristics can be modeled in a self-adaptive manner, the PIM prediction precision is improved, and the calculation complexity is reduced.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technologies, and specifically to a method for eliminating multi-band passive intermodulation of RRU based on a sequence-to-sequence learning model. Background Art

[0002] In a wireless communication system, a remote radio unit (RRU) is widely used in 5G / 6G base stations to support multi-band coexistence and large-scale MIMO technologies. However, in an environment of high-power transmission and multi-band signal coexistence, the non-linear characteristics of radio frequency devices will cause passive intermodulation (PIM) interference, resulting in a decline in the quality of received signals and affecting the communication performance of the base station.

[0003] Traditional PIM interference cancellation methods mainly rely on physical optimization (such as antenna layout adjustment, radio frequency device optimization) or interference compensation methods based on mathematical models (such as polynomial modeling, regression analysis). However, these methods have the following problems:

[0004] Limitations of static modeling: Methods based on fixed mathematical models are difficult to adapt to the complex non-linear intermodulation characteristics of multi-bands and cannot dynamically adjust the interference suppression strategy.

[0005] Real-time performance and computational complexity: Traditional signal processing methods have high computational complexity and are difficult to achieve real-time PIM suppression in high-rate, low-latency 5G / 6G scenarios.

[0006] Insufficient environmental adaptability: Existing PIM suppression schemes lack an adaptive mechanism and are difficult to cope with changes in intermodulation characteristics caused by environmental changes (such as temperature drift, device aging, etc.), resulting in poor long-term stability. Summary of the Invention

[0007] The purpose of the present invention is to provide a method for eliminating multi-band passive intermodulation of RRU based on a sequence-to-sequence learning model to solve the problems raised in the above background art.

[0008] To achieve the above purpose, the present invention provides the following technical solution: A method for eliminating multi-band passive intermodulation of RRU based on a sequence-to-sequence learning model, including the following steps:

[0009] S100. Obtain the transmitted signal and received signal of the multi-band RRU, and perform feature extraction on the signals by using an adaptive transformation to obtain multi-band intermodulation interference feature data;

[0010] S200. Based on the multi-band intermodulation interference feature data, use an autoencoder to perform deep learning on the PIM signal features to generate non-linear intermodulation feature vectors;

[0011] S300. Based on the non - linear inter - modulation feature vector and the multi - band transmission signal, use a sequence - to - sequence learning model for training to obtain a PIM signal prediction model;

[0012] S400. Input the multi - band transmission signal into the PIM signal prediction model to generate a predicted PIM interference signal, and use an error optimization mechanism to perform residual compensation on the prediction result;

[0013] S500. Perform adaptive filtering on the predicted PIM interference signal to generate an optimized interference compensation signal, and superimpose the compensation signal on the received signal to suppress PIM interference;

[0014] S600. Based on the signal data after interference suppression, use an online learning mechanism to dynamically update the PIM signal prediction model;

[0015] S700. Combine the model update parameters of the online learning and the long - term environmental feedback data, use a reinforcement learning method to optimize the hyperparameters of the PIM signal prediction model, and adjust the feature extraction strategy based on the feedback.

[0016] Preferably, the step S100 includes:

[0017] S110. Obtain the RRU multi - band transmission signal and the received signal, and perform time - domain and frequency - domain transformations on the signals to obtain the time - frequency distribution matrix of the signals;

[0018] S120. Extract the inter - modulation frequency components of the multi - band signals from the time - frequency distribution matrix, and enhance the inter - modulation frequency components based on an adaptive transformation method to obtain multi - band inter - modulation interference feature data;

[0019] S130. Perform feature screening on the multi - band inter - modulation interference feature data, remove background noise and irrelevant interference components, and generate an inter - modulation interference feature sequence to be input into the model.

[0020] Preferably, the step S200 includes:

[0021] S210. Input the inter - modulation interference feature sequence into an auto - encoder, and extract features of the sequence through an encoder network to generate a hidden state vector of the multi - band inter - modulation signal;

[0022] S220. Perform dimensionality reduction on the hidden state vector, and reconstruct the inter - modulation signal feature sequence through a decoder network, calculate the reconstruction error, and optimize the parameters of the auto - encoder;

[0023] S230. Extract the final non - linear inter - modulation features based on the optimized auto - encoder, and normalize the features to generate a non - linear inter - modulation feature vector.

[0024] Preferably, step S300 includes:

[0025] S310. Based on the non - linear inter - modulation feature vector and the multi - band transmission signal, construct a sequence - to - sequence learning model and initialize the model parameters;

[0026] S320. Conduct supervised training on the sequence - to - sequence learning model, so that the model learns the mapping relationship between the non - linear inter - modulation features and the PIM signal, and generates a PIM signal prediction model;

[0027] S330. Based on the error feedback mechanism of the training process, adjust the structure and parameters of the sequence - to - sequence learning model, and optimize the PIM signal prediction model to make it adapt to the multi - band inter - modulation interference environment.

[0028] Preferably, step S400 includes:

[0029] S410. Input the multi - band transmission signal into the PIM signal prediction model, and generate a preliminarily predicted PIM interference signal according to the model output;

[0030] S420. Calculate the residual between the predicted PIM interference signal and the received signal based on the error signal, and adjust the error compensation parameters through an optimization module;

[0031] S430. Based on the residual optimization mechanism, perform compensation adjustment on the predicted PIM interference signal to generate an optimized PIM interference signal.

[0032] Preferably, step S500 includes:

[0033] S510. Based on the optimized PIM interference signal, construct an adaptive filter and set the initial weights of the filter;

[0034] S520. Use the received signal and the optimized PIM interference signal for filtering calculation, and adjust the weight parameters of the adaptive filter to obtain an optimized interference compensation signal;

[0035] S530. Superimpose the optimized interference compensation signal and the received signal to suppress PIM interference and extract the purified effective signal.

[0036] Preferably, step S600 includes:

[0037] S610. Obtain the signal data after interference suppression and calculate the current model prediction error;

[0038] S620. Adjust the parameters of the sequence-to-sequence learning model based on the error signal, and perform incremental training to improve the adaptability of the PIM signal prediction model to environmental changes;

[0039] S630. Update the PIM signal prediction model based on the incremental training result, and store the optimized model parameters.

[0040] Preferably, the step S700 includes:

[0041] S710. Based on the model update parameters of the online learning, construct a reinforcement learning environment, and initialize the reinforcement learning policy;

[0042] S720. Extract features from the long-term environmental feedback data, and train the reinforcement learning policy to optimize the hyperparameters of the sequence-to-sequence learning model;

[0043] S730. Based on the hyperparameters optimized by the reinforcement learning policy, adjust the adaptive transformation weights of the feature extraction method, and update the feature extraction policy.

[0044] Preferably, the step of dynamically updating the PIM signal prediction model by using the online learning mechanism includes:

[0045] Obtain the signal data after interference suppression, and calculate the current model prediction error;

[0046] Adjust the parameters of the sequence-to-sequence learning model based on the error signal, and perform incremental training to improve the adaptability of the model to environmental changes;

[0047] Update the PIM signal prediction model based on the incremental training result, and store the optimized model parameters.

[0048] Preferably, the step of optimizing the hyperparameters of the PIM signal prediction model by using the reinforcement learning method includes:

[0049] Based on the model update parameters of the online learning, construct a reinforcement learning environment, and initialize the reinforcement learning policy;

[0050] Extract features from the long-term environmental feedback data, and train the reinforcement learning policy to optimize the hyperparameters of the sequence-to-sequence learning model;

[0051] Based on the hyperparameters optimized by the reinforcement learning policy, adjust the adaptive transformation weights of the feature extraction method, and update the feature extraction policy to enhance the long-term stability of the system.

[0052] Compared with the prior art, the present invention realizes real-time prediction and suppression of multi-band complex intermodulation interference through deep learning technology combined with signal processing optimization means, and has the following beneficial effects:

[0053] 1. Adaptive PIM modeling based on deep learning improves the prediction accuracy. Traditional PIM interference modeling relies on fixed mathematical models (such as polynomial regression), which are difficult to accurately capture the multi-band non-linear intermodulation characteristics. The present invention automatically extracts the non-linear characteristics of PIM signals through an autoencoder and trains a PIM signal prediction model using a sequence-to-sequence (Seq2Seq) learning model, enabling the system to accurately model PIM interference characteristics and improve the prediction accuracy.

[0054] 2. Adaptive error optimization mechanism enhances the robustness of PIM signal prediction. Due to the complex non-linear effects in the wireless channel, traditional PIM prediction models are vulnerable to device drift and environmental changes. The present invention adopts an error optimization mechanism to calculate the residual between the predicted PIM signal and the actual received signal, and adaptively adjusts the error compensation parameters, enabling the prediction model to maintain high accuracy during long-term operation and avoiding the problem of model accuracy degradation over time.

[0055] 3. Multi-band joint signal processing improves the interference suppression ability. The present invention supports the joint input of multi-band RRU signals, extracts PIM intermodulation characteristics through joint time-frequency domain analysis, and overcomes the computational redundancy problem caused by independent modeling for each frequency band in traditional methods. Through multi-band feature learning, it is able to capture the intermodulation interference between different frequency bands and improve the interference suppression ability of the system.

[0056] 4. Adaptive filtering processing improves the real-time performance of PIM interference suppression. Traditional PIM interference suppression relies on static filters, which are difficult to cope with dynamic signal environments. The present invention adopts an adaptive filter (LMS filter) combined with an optimized filter weight adjustment strategy to dynamically update the filter parameters based on the error signal, making the suppression of PIM signals more accurate and improving the elimination efficiency of PIM signals.

[0057] 5. Online learning mechanism improves the model's environmental adaptability. In a complex wireless environment, PIM intermodulation characteristics change over time. Traditional methods are difficult to cope with long-term effects such as device aging and temperature drift. The present invention dynamically updates the PIM prediction model based on the signal data after interference suppression through an online learning mechanism (OnlineLearning), enabling the system to adapt to different environments and improve long-term stability.

[0058] 6. Optimize the hyperparameters of the model using reinforcement learning to improve long-term stability. The hyperparameters of traditional deep learning models usually rely on manual adjustment and cannot adapt dynamically to different communication environments. This invention combines the reinforcement learning (RL) method to optimize key hyperparameters such as the learning rate and feature extraction weights of the PIM prediction model based on long-term environmental feedback data, ensuring the stability and adaptability of the model during long-term operation. Description of the Drawings

[0059] Figure 1 It is a schematic flowchart of a method for eliminating multi-band passive intermodulation of RRU based on a sequence-to-sequence learning model provided by an embodiment of this application;

[0060] Figure 2 It is a structural block diagram of a system for eliminating multi-band passive intermodulation of RRU based on a sequence-to-sequence learning model provided by an embodiment of this application. Detailed Embodiments

[0061] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0062] Embodiment 1: Refer to Figure 1 , which is a schematic flowchart of a method for eliminating multi-band passive intermodulation of RRU based on a sequence-to-sequence learning model provided by an embodiment of the present invention. This process can at least include steps S100-S700:

[0063] S100. Obtain the multi-band RRU transmission signal and reception signal, and perform feature extraction on the signal using adaptive transformation to obtain multi-band intermodulation interference feature data;

[0064] S200. Based on the multi-band intermodulation interference feature data, use an autoencoder to perform deep learning on the PIM signal features to generate a non-linear intermodulation feature vector;

[0065] S300. Based on the non-linear intermodulation feature vector and the multi-band transmission signal, use a sequence-to-sequence learning model for training to obtain a PIM signal prediction model;

[0066] S400. Input the multi-band transmission signal into the PIM signal prediction model to generate a predicted PIM interference signal, and use an error optimization mechanism to perform residual compensation on the prediction result;

[0067] S500. Perform adaptive filtering on the predicted PIM interference signal to generate an optimized interference compensation signal, and superimpose the compensation signal on the received signal to suppress PIM interference;

[0068] S600. Based on the signal data after interference suppression, use an online learning mechanism to dynamically update the PIM signal prediction model to enhance the model's environmental adaptability;

[0069] S700. Combine the model update parameters of the online learning with the long-term environmental feedback data, use the reinforcement learning method to optimize the model hyperparameters, and adjust the feature extraction strategy based on the feedback.

[0070] Step S100 includes at least steps S110 - S130:

[0071] S110. Obtain the RRU multi-band transmission signal and the received signal, and perform time-domain and frequency-domain transformations on the signals to obtain the time-frequency distribution matrix of the signals.

[0072] Specifically, collect the RRU downlink transmission signal S TX (t) and the uplink received signal S RX (t). Among them, S TX (t) is the transmission signal of the base station on multiple frequency bands, and S RX (t) includes the true signal and the PIM interference signal at the receiving end.

[0073] Perform time-domain transformation on the signals. In the time domain, the signals can be expressed as:

[0074]

[0075] Among them, A i represents the amplitude of the i-th carrier, f i is the corresponding frequency, φ i is the phase, and N represents the total number of transmission frequency bands.

[0076] Perform frequency-domain transformation to obtain the time-frequency distribution matrix. Perform time-frequency analysis on the signals through the short-time Fourier transform (STFT) or wavelet transform (WT) to obtain the time-frequency distribution matrix M STFT :

[0077]

[0078] Among them, W(t) is the analysis window function, f represents the frequency, τ is the time offset, j is the imaginary unit, and this transformation can be used to identify the non-linear components and intermodulation components of the signals.

[0079] S120. Extract the intermodulation frequency components of the multi-band signals from the time-frequency distribution matrix, and enhance the intermodulation frequency components based on the adaptive transformation method to obtain the multi-band intermodulation interference feature data.

[0080] Specifically, extract the intermodulation frequency component f STFT from the time-frequency distribution matrix M IM , where:

[0081]

[0082] where m and n represent the intermodulation orders (such as third-order intermodulation IM3: 2f i -f j ), f i and f j are the base station transmission frequencies respectively, and f IM is the generated intermodulation frequency.

[0083] Enhance the intermodulation frequency components based on the adaptive transformation method. Further, use wavelet packet decomposition (WPD) to enhance the signal features to make the intermodulation frequency features more prominent:

[0084]

[0085] where ψ k (t) is the wavelet basis function, α k represents the corresponding transformation coefficient, and k is the number of wavelet decomposition layers. This method can be used to amplify the PIM features while reducing the background noise interference.

[0086] S130. Screen the features of the multi-band intermodulation interference feature data, remove the background noise and irrelevant interference components, and generate the intermodulation interference feature sequence to be input into the model.

[0087] Specifically, evaluate the correlation between the signals of different frequency bands and the PIM features based on the mutual information (MI) method, and screen out the most representative intermodulation interference features:

[0088]

[0089] where p(x, y) is the joint probability distribution, and p(x) and p(y) are the marginal probability distributions of signals X and Y respectively. By calculating the mutual information value, select the feature subset with the greatest impact on the intermodulation interference.

[0090] Remove the background noise and irrelevant interference components. Further, use the principal component analysis (PCA) method to remove the background noise and irrelevant interference components and reduce the data dimension:

[0091] X' = XW

[0092] Wherein, X is the original feature matrix, W is the feature vector matrix, and X' is the feature data after dimensionality reduction. This step can effectively remove irrelevant information and improve the accuracy of PIM signal prediction.

[0093] Generate the intermodulation interference feature sequence to be input into the model. Finally, reconstruct the screened and dimensionally reduced feature data into a sequence format and input it into the subsequent deep learning model:

[0094] F seq ={f IM，1 , f IM，2 ,..., f IM,T}

[0095] Wherein, T is the number of time steps, and f IM，t represents the intermodulation interference feature at the t-th moment. This feature sequence will be used as input to enter the autoencoder for deep learning.

[0096] Explanation of the connection between the previous and subsequent steps

[0097] S110 provides the time-frequency distribution matrix of the signal for S120 to extract the intermodulation frequency components in the subsequent step.

[0098] After being enhanced, the intermodulation interference feature data generated by S120 is transmitted to S130 for dimensionality reduction and feature screening.

[0099] The intermodulation interference feature sequence F seq generated by S130 will be used as input to enter S200 (deep learning with an autoencoder) to optimize the learning and modeling ability of PIM features.

[0100] Step S200 includes at least steps S210 - S230:

[0101] S210: Input the intermodulation interference feature sequence into the autoencoder, and extract features from the sequence through the encoder network to generate the hidden state vector of the multi-band intermodulation signal.

[0102] Input the intermodulation interference feature sequence calculated by S130 into the autoencoder. Specifically, use the feature sequence F seq calculated by S130 as the input of the autoencoder, and define the input feature matrix Where:

[0103] T represents the number of time steps, corresponding to the time series dimension of the multi-band signal;

[0104] M represents the feature dimension, corresponding to the number of intermodulation frequencies.

[0105] At the encoder end, feature extraction of the input data is performed through an encoder network. A multi-layer neural network (MLP) or a long short-term memory network (LSTM) is used to perform dimensionality reduction on the input data. Specifically, it is expressed as follows:

[0106] h t = f θ (X t )

[0107] where:

[0108] represents the hidden state vector at the t-th time step;

[0109] f θ (π) is the encoder network, trained by the neural network parameters θ;

[0110] D is the dimension of the hidden state vector, satisfying D < M, for dimensionality reduction.

[0111] Furthermore, the hidden state vectors of all time steps are combined to obtain a complete hidden state matrix:

[0112]

[0113] This matrix is used to represent the non-linear features of the multi-band intermodulation signal and provides input for subsequent reconstruction.

[0114] Step connection description:

[0115] The input of this step is the intermodulation interference feature sequence F calculated in S130 seq ;

[0116] The output H of this step is used as the input data for the dimensionality reduction processing and reconstruction in S220.

[0117] S220. Perform dimensionality reduction on the hidden state vector, and reconstruct the intermodulation signal feature sequence through a decoder network, calculate the reconstruction error, and optimize the parameters of the autoencoder.

[0118] Specifically, use dimensionality reduction methods (such as PCA, t-SNE, or Transformer attention mechanism) to further extract effective features to obtain the dimensionality-reduced feature representation:

[0119] Z = g φ (H)

[0120] where:

[0121] is the dimensionality-reduced feature matrix;

[0122] g φ (π) is the dimensionality reduction mapping function, and φ is the dimensionality reduction parameter;

[0123] D' < D, further reducing the feature dimension and improving the computational efficiency.

[0124] Furthermore, the decoder network (Decoder) is used to restore the dimension-reduced feature Z to the original signal feature:

[0125]

[0126] Where:

[0127] is the reconstructed feature matrix;

[0128] h ψ (·) is the decoder network, and the parameter ψ is optimized through training.

[0129] Calculate the error between the original input X and the reconstruction result and optimize it using the mean square error (MSE) loss:

[0130]

[0131] Where:

[0132] L MSE is the reconstruction error;

[0133] X i，j and respectively represent the values of the original data and the reconstructed data at i, j.

[0134] Explanation of step connection:

[0135] The input of this step is the hidden state matrix H generated by S210;

[0136] The output of this step is the optimized hidden state vector Z, which serves as the input data for S230.

[0137] S230, based on the optimized autoencoder, extracts the final non-linear intermodulation features and normalizes the features to generate a non-linear intermodulation feature vector.

[0138] Specifically, the dimension-reduced feature Z calculated by S220 is used as the final intermodulation feature representation:

[0139]

[0140] Where:

[0141] is the finally extracted non-linear intermodulation feature vector;

[0142] is the optimized feature transformation matrix.

[0143] Normalize the said features

[0144] To ensure data consistency, perform normalization on the final feature vector:

[0145]

[0146] Where:

[0147] F norm Is the non - linear intermodulation feature after normalization;

[0148] μ and σ are respectively The mean and standard deviation in each dimension.

[0149] Finally, the normalized feature vector F norm Is used as the input data for subsequent S300 (sequence - to - sequence learning model training).

[0150] Explanation of step connection:

[0151] The input of this step is the feature matrix Z after dimensionality reduction in S220;

[0152] The output F of this step norm Is used as the input data for S300 (PIM signal prediction model training).

[0153] Step S300 includes at least steps S310 - S330:

[0154] S310. Based on the non - linear intermodulation feature vector and the multi - band transmission signal, construct a sequence - to - sequence learning model and initialize the model parameters.

[0155] Specifically, establish a Seq2Seq learning framework based on an encoder - decoder (Encoder - Decoder), where:

[0156] The encoder (Encoder) is used to extract the temporal features of the input signal;

[0157] The decoder (Decoder) is used to predict the PIM interference signal.

[0158] Let the input sequence be:

[0159] X seq ={F norm , S TX (t)}

[0160] Where:

[0161] Is the non - linear intermodulation feature vector calculated in S230;

[0162] is a multi - band transmission signal matrix;

[0163] T is the number of time steps, D′ is the feature dimension, and M is the transmission signal dimension.

[0164] Furthermore, at the encoder end, a bidirectional LSTM (BiLSTM) is used to extract temporal features:

[0165] H t = BiLSTM(X seq , W enc )

[0166] where:

[0167] is the hidden state generated by the encoder;

[0168] W enc is the encoder parameter set;

[0169] D k is the hidden layer dimension.

[0170] At the decoder end, an attention mechanism is adopted to optimize the PIM signal prediction:

[0171]

[0172] where:

[0173] C t is the context vector;

[0174] α t，i is the attention weight, defined as follows:

[0175] e t，i is the matching score, calculated from trainable parameters.

[0176] Step connection description:

[0177] The input of this step is the non - linear intermodulation feature vector F norm ;

[0178] The output of this step is the initialized Seq2Seq model parameters W enc , W dec , used for model training in S320.

[0179] S320. Supervise and train the sequence - to - sequence learning model so that the model learns the mapping relationship between non - linear intermodulation features and PIM signals, and generates a PIM signal prediction model.

[0180] Specifically, the Seq2Seq model is supervised and trained using the transmitted signal and received signal data collected by S100:

[0181] Let the true PIM signal S IM (t) be: S IM (t) = S RX (t) - S TX (t)

[0182] Where:

[0183] S RX (t) is the uplink received signal;

[0184] S TX (t) is the downlink transmitted signal;

[0185] S IM (t) is the true label data of the PIM interference signal.

[0186] Supervise and train the Seq2Seq model

[0187] During the training process, optimize the output of the decoder to minimize the error:

[0188]

[0189] Where:

[0190] L is the mean squared error (MSE) loss;

[0191] is the PIM signal predicted by the decoder.

[0192] Use the Adam optimizer to adjust the parameters of the encoder and decoder:

[0193]

[0194] Where:

[0195] η is the learning rate;

[0196] respectively represent the parameters of the encoder and decoder in the nth round of training.

[0197] Explanation of step connection:

[0198] The input of this step is the Seq2Seq model parameters W enc , W dec ;

[0199] The output of this step is the trained PIM signal prediction model fθ (X) is used for the S330 to optimize the model.

[0200] The S330, based on the error feedback mechanism of the training process, adjusts the model structure and parameters to optimize the PIM signal prediction model to adapt it to the multi-band intermodulation interference environment.

[0201] Further, after each training epoch ends, calculate the error gradient:

[0202] ΔL = L (n) -L (n-1)

[0203] If ΔL > 0, it means the model has not converged and the hyperparameters need to be adjusted.

[0204] If the error converges slowly or cannot be reduced, then adjust the Seq2Seq model structure:

[0205] Optimize the LSTM using the attention mechanism: where β t，i is the optimized attention weight;

[0206] Improve the stability using Residual Connection: This method can improve the accuracy of PIM signal prediction.

[0207] Adopt forward propagation and the error feedback mechanism to update the optimized model:

[0208]

[0209] where W are the parameters of the updated Seq2Seq model.

[0210] Description of step connection:

[0211] The input of this step is the PIM signal prediction model obtained by S320 training;

[0212] The output of this step is the optimized PIM signal prediction model, which is used by S400 to generate and compensate for PIM interference signals.

[0213] Finally, the PIM signal prediction model can be used to accurately predict PIM interference signals and ensure the adaptability of the model in different frequency bands and interference environments through the optimization mechanism.

[0214] Step S400 includes at least steps S410 - S430:

[0215] S410: Input the multi-band transmission signal into the PIM signal prediction model and generate a preliminarily predicted PIM interference signal according to the model output.

[0216] Specifically, the multi-band transmission signal S collected by S100 TX (t) is input into the PIM signal prediction model f trained by S330 θ to generate a predicted PIM interference signal:

[0217]

[0218] Where:

[0219] is the multi-band transmission signal of the base station;

[0220] f θ is a sequence-to-sequence (Seq2Seq) learning model optimized through training;

[0221] is the preliminary PIM interference signal obtained by model prediction;

[0222] T represents the number of time steps, and M represents the number of multi-band signals.

[0223] Furthermore, calculate the main parameters of the predicted PIM signal in the time domain and frequency domain:

[0224] Calculate the power of the predicted signal:

[0225] Perform short-time Fourier transform (STFT) to obtain frequency domain characteristics: where W(π) is the window function, represents the time-frequency distribution.

[0226] Specifically, take as the input data for subsequent S420 to calculate the residual, and record its spectral information for subsequent optimization.

[0227] Step connection description:

[0228] The input of this step is the PIM signal prediction model f optimized by S330 θ and the multi-band transmission signal S collected by S100 TX (t);

[0229] The output of this step is as the input data for S420 to calculate the error signal.

[0230] S420 calculates the residual between the predicted PIM interference signal and the actual received signal based on the error signal, and adjusts the error compensation parameters through the optimization module.

[0231] Specifically, based on the uplink received signal S collected by S100RX (t), calculate the error signal:

[0232]

[0233] Where:

[0234] e(t) is the predicted error signal;

[0235] S RX (t) is the actual received signal;

[0236] is the predicted PIM signal.

[0237] Furthermore, calculate the mean square error (MSE) of the error signal:

[0238]

[0239] If L MSE exceeds the set error threshold, the error compensation parameter needs to be adjusted.

[0240] Adopt an adaptive weight adjustment mechanism to optimize the error compensation parameter λ:

[0241]

[0242] Where:

[0243] λ is the error compensation parameter;

[0244] η is the learning rate;

[0245] n represents the number of iterations.

[0246] Step connection description:

[0247] The input of this step is the predicted PIM signal generated by S410 and the uplink received signal S collected by S100 RX (t);

[0248] The output of this step is the optimized error compensation parameter λ, which is used for PIM signal compensation adjustment in S430.

[0249] S430, based on the residual optimization mechanism, compensates and adjusts the predicted PIM interference signal to generate an optimized PIM interference signal.

[0250] Specifically, perform error compensation on the predicted PIM signal generated by S410 :

[0251]

[0252] Where:

[0253] is the finally optimized PIM signal;

[0254] λ is the error compensation parameter;

[0255] e(t) is the error signal.

[0256] Furthermore, calculate the mean square error of the optimized PIM signal:

[0257]

[0258] If is lower than the preset threshold, it is considered that the PIM signal optimization is successful, otherwise, it is necessary to return to S420 to readjust the compensation parameter.

[0259] Specifically, store and transmit it to S500 for PIM interference suppression.

[0260] Step connection description:

[0261] The input of this step is the error compensation parameter λ and the error signal e(t) calculated by S420;

[0262] The output of this step is the optimized PIM signal for S500 to perform PIM interference cancellation.

[0263] Finally, this step ensures the optimized compensation of the PIM signal, improves the accuracy of PIM interference signal prediction, and enhances the subsequent interference suppression effect.

[0264] Step S500 includes at least steps S510 - S530:

[0265] S510. Based on the optimized PIM interference signal, construct an adaptive filter and set the initial weight of the filter.

[0266] Specifically, based on the optimized PIM signal calculated by S430 construct a least mean square (LMS) adaptive filter:

[0267]

[0268] Where:

[0269] y(t) is the filtered output signal;

[0270] w i is the filter weight;

[0271] x(t) is the input signal, that is, the optimized PIM signal

[0272] N is the order of the filter.

[0273] Furthermore, set the initial weights of the filter usually initialized to zero:

[0274]

[0275] This initialization method ensures that the filter does not introduce additional errors in the initial stage and optimizes the weights through subsequent iterations.

[0276] Explanation of step connection:

[0277] The input of this step is the optimized PIM interference signal calculated by S430

[0278] The output of this step is the initialized adaptive filter for S520 to perform weight optimization and filtering calculation.

[0279] S520. Use the received signal and the optimized PIM interference signal to perform filtering calculation, and adjust the weight parameters of the adaptive filter to obtain the optimized interference compensation signal.

[0280] Specifically, based on the adaptive filter constructed in S510, calculate the filtered output signal:

[0281]

[0282] Where:

[0283] y(t) is the filtered interference compensation signal;

[0284] is the filter weight after the nth iteration;

[0285] is the optimized PIM interference signal.

[0286] Furthermore, based on the uplink received signal S RX (t) collected by S100, calculate the filtering error:

[0287] e(t) = S RX (t) - y(t)

[0288] Where:

[0289] e(t) is the filtering error, representing the difference between the actual received signal and the filtered compensation signal.

[0290] Use the LMS (Least Mean Squares) algorithm to update the filter weights:

[0291]

[0292] Wherein:

[0293] η is the learning rate, which controls the weight update step size.

[0294] Specifically, store the updated filter weights and the final filtered compensation signal y(t) for signal superposition processing by S530.

[0295] Step connection description:

[0296] The input of this step is the filter weights initialized by S510 and the received signal S collected by S100 RX (t);

[0297] The output of this step is the optimized filtered compensation signal y(t), which is used for PIM interference suppression by S530.

[0298] S530 superimposes the optimized interference compensation signal and the received signal to suppress PIM interference and extract the purified effective signal.

[0299] Specifically, superimpose the optimized interference compensation signal y(t) calculated by S520 and the received signal S RX (t) to generate a purified signal:

[0300] S clean (t) = S RX (t) - y(t)

[0301] Wherein:

[0302] S clean (t) is the purified signal after suppressing PIM interference.

[0303] Furthermore, calculate the signal power ratio (IMR, Intermodulation Rejection Ratio) before and after PIM suppression:

[0304]

[0305] Wherein:

[0306] P RX is the original received signal power;

[0307] P clean is the power of the purified signal.

[0308] If the IMR exceeds the set threshold, it is considered that PIM suppression is successful; otherwise, it is necessary to return to S520 to adjust the filter weights.

[0309] Specifically, output S clean (t) to the signal processing unit of the wireless communication base station for further demodulation and data processing by the base station.

[0310] Description of step connection:

[0311] The input of this step is the optimized interference compensation signal y(t) calculated by S520 and the received signal S RX (t) collected by S100;

[0312] The output of this step is the purified signal S clean (t) for further data processing by the base station.

[0313] Step S600 includes at least steps S610 - S630:

[0314] S610. Obtain the signal data after interference suppression and calculate the current model prediction error

[0315] Specifically, take the purified signal S clean (t) calculated by S530 as the reference signal after PIM signal suppression:

[0316] S clean (t) = S RX (t) - y(t)

[0317] Where:

[0318] S clean (t) is the purified signal after PIM interference suppression;

[0319] S RX (t) is the uplink received signal collected by S100;

[0320] y(t) is the optimized interference compensation signal calculated by S520.

[0321] Furthermore, calculate the error generated by the current PIM signal prediction model f θ when the same multi - band transmission signal S TX (t) is input:

[0322] e model (t) = S clean (t) - f θ (S TX (t))

[0323] Where:

[0324] e model (t) is the model prediction error signal;

[0325] f θ(S TX (t)) is the PIM signal predicted by the current model.

[0326] Specifically, calculate the mean square error of the error signal:

[0327]

[0328] Where:

[0329] L MSE represents the magnitude of the model prediction error;

[0330] T is the number of time steps.

[0331] If L MSE exceeds the set error threshold, model parameter adjustment and incremental training are required.

[0332] Step connection description:

[0333] The input of this step is the purified signal S clean (t) calculated by S530, and the multi-band transmission signal S TX (t) collected by S100;

[0334] The output of this step is the calculated error e model (t), which is used for S620 to adjust the model parameters and perform incremental training.

[0335] S620. Based on the error signal, adjust the parameters of the sequence-to-sequence learning model and perform incremental training to improve the model's adaptability to environmental changes.

[0336] Specifically, use the gradient descent method (Gradient Descent) to adjust the parameters θ of the PIM signal prediction model:

[0337]

[0338] Where

[0339] θ(n) represents the model parameters of the nth round of training;

[0340] η is the learning rate, which controls the parameter update step size.

[0341] Furthermore, use the error signal e model (t) calculated by S610 as incremental training data to update the Seq2Seq model:

[0342]

[0343] Where:

[0344] The model after incremental training;

[0345] λ is the error compensation factor to ensure a moderate adjustment amplitude of the model.

[0346] Furthermore, calculate the error of the updated model:

[0347]

[0348] If is lower than the preset threshold, it is considered that the model has adapted to the new interference environment, otherwise, incremental training needs to continue.

[0349] Description of step connection:

[0350] The input of this step is the error signal e model (t) calculated by S610, and the PIM signal prediction model f trained by S330 θ ;

[0351] The output of this step is the model after incremental training for S630 to store and update the model parameters.

[0352] S630. Based on the incremental training result, update the PIM signal prediction model and store the optimized model parameters

[0353] Specifically, take the incremental training model calculated by S620 as the new PIM signal prediction model:

[0354]

[0355] Where:

[0356] is the finally updated PIM prediction model.

[0357] Furthermore, store the model parameter θ new into the system database:

[0358] Save(θ new )

[0359] Among them, Save represents the storage operation to ensure that the subsequent communication system can still use the latest model parameters when the environment changes.

[0360] Specifically, use the new PIM signal prediction model to test the newly input multi-band transmission signal. If the test error is still within the acceptable range, the update is completed, otherwise, return to S620 to continue incremental training.

[0361] Description of step connection:

[0362] The input of this step is the incremental training model calculated by S620.

[0363] The output of this step is the stored optimized model parameter θ new for subsequent PIM signal prediction.

[0364] Step S700 includes at least steps S710 - S730:

[0365] S710. Update the parameters of the model based on the online learning, construct a reinforcement learning environment, and initialize the reinforcement learning policy.

[0366] Specifically, based on the PIM signal prediction model calculated by S630 construct a reinforcement learning state space S:

[0367] s = {θ new , PIM level , IMR prev}

[0368] where:

[0369] θ new is the model parameter optimized by the current online learning;

[0370] PIM level is the current PIM interference level, determined by the interference signal power calculated by S530:

[0371] IMR prev is the intermodulation interference rejection ratio (IMR) of the previous PIM suppression.

[0372] Furthermore, set the action set A of the reinforcement learning agent, that is, the adjustable hyperparameters:

[0373] A = {η, λ, d}

[0374] where:

[0375] η is the model learning rate;

[0376] λ is the error compensation factor;

[0377] d is the adaptive transformation weight for feature extraction.

[0378] Adopt the Q-learning method to initialize the Q function:

[0379]

[0380] where:

[0381] Q(S, A) represents the expected return of performing action A in state S;

[0382] R is the immediate return, defined as follows:

[0383] IMR new is the new PIM interference suppression ratio;

[0384] is the error of the new model;

[0385] α, β are weight factors.

[0386] Step connection description:

[0387] The input of this step is the model parameter θ calculated by S630 new and the PIM signal power calculated by S530;

[0388] The output of this step is the reinforcement learning environment S, the action space A, and the Q-learning policy Q(S, A), which are used for reinforcement learning training in S720.

[0389] S720 extracts features from the long-term environmental feedback data and trains the reinforcement learning policy to optimize the hyperparameters of the sequence-to-sequence learning model.

[0390] Specifically, the following features are extracted from the long-term PIM interference suppression records:

[0391] F env ={PIM trend , IMR trend , L trend}

[0392] Where:

[0393] PIM trend is the PIM interference trend in the past T time windows;

[0394] IMR trend is the change in IMR in the past T time windows;

[0395] L trend is the change in the model error in the past T time windows.

[0396] Furthermore, based on the historical feedback feature F env , the DQN (Deep Q-Network) is used to optimize the reinforcement learning policy:

[0397]

[0398] Where:

[0399] Q θ The Q function parameterized by the neural network;

[0400] α is the learning rate;

[0401] γ is the discount factor;

[0402] A′ is the best action for the next step.

[0403] Adopt a strategy to select the optimal hyperparameters:

[0404]

[0405] where A * is the optimized set of hyperparameters.

[0406] Explanation of step connection:

[0407] The input of this step is the reinforcement learning environment S set by S710 and the feedback data of long-term PIM suppression;

[0408] The output of this step is the optimized set of hyperparameters A * , which is used for S730 to adjust the feature extraction strategy.

[0409] S730, based on the hyperparameters optimized by the reinforcement learning strategy, adjusts the adaptive transformation weights of the feature extraction method and updates the feature extraction strategy to enhance the long-term stability of the system.

[0410] Specifically, based on the optimized hyperparameters A obtained from the reinforcement learning training * , adjust the adaptive transformation weight d of the feature extraction method * :

[0411] W new = W old + d * ·ΔW

[0412] where:

[0413] W new is the updated feature transformation weight;

[0414] W old is the original feature transformation weight;

[0415] ΔW is the adjustment amount calculated by the reinforcement learning.

[0416] Furthermore, based on the strategy trained by S720, optimize the signal feature extraction method:

[0417] If the PIM signal trend fluctuates greatly, increase the length of the feature extraction window;

[0418] If the change in IMR is small, reduce the dynamic adjustment weight of the filter;

[0419] If the MSE error decreases slowly, increase the learning rate η of the adaptive transform * 。

[0420] Store the optimized feature extraction strategy and apply it to subsequent signal processing. Specifically, store the optimized parameter W new and the feature extraction strategy, and apply them during subsequent signal processing to enhance the long-term stability of the system.

[0421] Explanation of step connection:

[0422] The input of this step is the optimized hyperparameter A obtained by training S720 * ;

[0423] The output of this step is the optimized feature extraction strategy W new for subsequent signal processing.

[0424] Embodiment 2: Figure 2 Show a structural block diagram of an RRU multi-band passive intermodulation cancellation system based on a sequence-to-sequence learning model according to an embodiment of the present invention. As Figure 2 shown, this structure may include:

[0425] The data acquisition and preprocessing module 10 is used to acquire the downlink transmission signal and the uplink reception signal in the RRU (Remote Radio Unit) multi-band wireless communication system. This module can monitor the multi-band RRU signal in real time and capture key data in the wireless communication environment through high-precision sampling equipment to ensure signal integrity and data synchronization. Specifically, this module includes:

[0426] The multi-band signal acquisition unit: used to acquire the downlink transmission signal S TX (t) and the uplink reception signal S RX (t) of the RRU in multiple bands, ensuring that the data acquisition range covers all bands that may generate passive intermodulation (PIM) interference.

[0427] The time-frequency transformation unit: performs time-domain and frequency-domain analysis on the acquired signal, and uses the short-time Fourier transform (STFT) or wavelet transform (WT) to extract the intermodulation signal characteristics and generate the time-frequency distribution matrix M STFT (t, f).

[0428] The signal feature extraction unit: extracts the intermodulation frequency components from the time-frequency distribution matrix and uses an adaptive transformation method to enhance the intermodulation signal to improve the recognizability of the PIM interference signal.

[0429] The PIM signal modeling and feature learning module 20 is mainly responsible for performing deep learning on the PIM signal features and constructing a feature representation suitable for multi-band PIM prediction. Specifically, this module includes:

[0430] Autoencoder feature learning unit: Input the intermodulation interference features extracted by the data acquisition module into an autoencoder, and extract the non-linear intermodulation feature vector F IM .

[0431] Feature dimensionality reduction and denoising unit: Use principal component analysis (PCA) and denoising autoencoder to reduce the feature dimensions, remove irrelevant interference components, and ensure that the data input into the model has high predictive value.

[0432] The PIM signal prediction model training module 30 is responsible for training a deep neural network model based on sequence-to-sequence (Seq2Seq) learning based on the feature data extracted by the feature learning module to achieve PIM signal prediction. This module includes:

[0433] Seq2Seq encoder unit: Use a bidirectional LSTM or Transformer network to encode the input intermodulation feature vector F IM and extract the temporal pattern of PIM interference.

[0434] Seq2Seq decoder unit: Based on the output of the encoder, generate the predicted PIM signal

[0435] Supervised training unit: Calculate the prediction error based on the received signal S RX (t), and use gradient descent to optimize the loss function: where S IM (t) = S RX (t) - S TX (t) is the true PIM signal.

[0436] The PIM error optimization and compensation module 40 is used to calculate the error between the predicted PIM signal and the actual received signal, and adjust the output of the PIM prediction model based on the error optimization strategy. This module includes:

[0437] Error calculation unit: Calculate the prediction error signal and evaluate the error level of the current PIM prediction model.

[0438] Error compensation unit: Use the residual learning method to perform error compensation on the predicted PIM signal and generate the optimized PIM signal where λ is the error compensation parameter.

[0439] The PIM signal adaptive filtering and suppression module 50 is responsible for filtering the optimized PIM interference signal, and superimposing the filtered interference compensation signal on the received signal to eliminate PIM interference. This module includes:

[0440] Adaptive filter construction unit: Based on the LMS (Least Mean Squares) adaptive algorithm, construct a filter: where w i is the filter weight.

[0441] Filter weight update unit: Adjust the filter weight based on the error signal e(t) = S RX (t) - y(t): where η is the learning rate.

[0442] The online learning and model optimization module 60 is responsible for dynamically adjusting the model parameters based on the long-term change trend of the PIM signal to improve the environmental adaptability of the system. This module includes:

[0443] Online error analysis unit: Calculate the new prediction error If the error exceeds the threshold, trigger the online learning mechanism.

[0444] Incremental training unit: Online update the model based on incremental data to improve the generalization ability of the model in different environments.

[0445] The reinforcement learning optimization and long-term adaptation module 70 is responsible for using reinforcement learning methods to optimize the hyperparameters of the PIM signal prediction model and adjust the feature extraction strategy to enable the system to long-term adapt to complex wireless environments. This module includes:

[0446] Reinforcement learning environment construction unit: Define the state space S, action set A, and reward function R, and use the Q-learning method to train the reinforcement learning strategy.

[0447] Feature extraction strategy adjustment unit: Based on the optimized hyperparameters, adjust the adaptive transformation weight, update the feature extraction method, and improve the extraction accuracy of the PIM signal

[0448] The method of the present invention includes: obtaining the transmitted signal and received signal of a multi-band RRU, and performing feature extraction on the signals by using adaptive transformation to obtain multi-band intermodulation interference feature data; generating a non-linear intermodulation feature vector based on the multi-band intermodulation interference feature data; training a PIM signal prediction model by using a sequence-to-sequence learning model, and performing residual compensation on the prediction result based on an error optimization mechanism; performing adaptive filtering processing on the predicted PIM interference signal to generate an optimized interference compensation signal; dynamically updating the PIM signal prediction model by using an online learning mechanism based on the signal data after interference suppression; combining the model update parameters of the online learning with long-term environmental feedback data, and optimizing the hyperparameters by using a reinforcement learning method and adjusting the feature extraction strategy. The present invention can adaptively model the multi-band non-linear intermodulation features, improve the PIM prediction accuracy, and reduce the computational complexity.

[0449] The content not detailedly described in this specification belongs to the prior art well-known to those skilled in the art. Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for eliminating RRU multi-band passive intermodulation based on a sequence-to-sequence learning model, characterized in that: It includes the following steps: S100. Obtain the transmitted signal and received signal of the multi-band RRU, and perform feature extraction on the signals by using adaptive transformation to obtain multi-band intermodulation interference feature data; S200. Based on the multi-band intermodulation interference feature data, use an autoencoder to perform deep learning on the PIM signal features to generate a non-linear intermodulation feature vector; S300. Based on the non-linear intermodulation feature vector and the multi-band transmitted signal, use a sequence-to-sequence learning model for training to obtain a PIM signal prediction model; S400. Input the multi-band transmitted signal into the PIM signal prediction model to generate a predicted PIM interference signal, and use an error optimization mechanism to perform residual compensation on the prediction result; S500. Perform adaptive filtering processing on the predicted PIM interference signal to generate an optimized interference compensation signal, and superimpose the compensation signal and the received signal to suppress PIM interference; S600. Based on the signal data after interference suppression, use an online learning mechanism to dynamically update the PIM signal prediction model; S700. Combine the model update parameters of the online learning and the long-term environmental feedback data, use a reinforcement learning method to optimize the hyperparameters of the PIM signal prediction model, and adjust the feature extraction strategy based on the feedback.

2. A method for eliminating RRU multi-band passive intermodulation based on a sequence-to-sequence learning model according to claim 1, characterized in that: The step S100 includes: S110. Obtain the multi-band transmitted signal and received signal of the RRU, and perform time-domain and frequency-domain transformation on the signals to obtain the time-frequency distribution matrix of the signals; S120. Extract the intermodulation frequency components of the multi-band signals from the time-frequency distribution matrix, and enhance the intermodulation frequency components based on the adaptive transformation method to obtain multi-band intermodulation interference feature data; S130. Perform feature screening on the multi-band intermodulation interference feature data, remove background noise and irrelevant interference components, and generate an intermodulation interference feature sequence to be input into the model.

3. A method for eliminating RRU multi-band passive intermodulation based on a sequence-to-sequence learning model according to claim 1, characterized in that: The step S200 includes: S210. Input the intermodulation interference feature sequence into the autoencoder, and perform feature extraction on the sequence through the encoder network to generate a hidden state vector of the multi-band intermodulation signal; S220. Perform dimensionality reduction processing on the hidden state vector, and reconstruct the intermodulation signal feature sequence through the decoder network, calculate the reconstruction error, and optimize the parameters of the autoencoder; S230. Extract the final non-linear intermodulation features based on the optimized autoencoder, and normalize the features to generate a non-linear intermodulation feature vector.

4. A method for eliminating RRU multi-band passive intermodulation based on a sequence-to-sequence learning model according to claim 1, characterized in that: The step S300 includes: S310. Based on the non-linear intermodulation feature vector and the multi-band transmitted signal, construct a sequence-to-sequence learning model and initialize the model parameters; S320. Perform supervised training on the sequence-to-sequence learning model to enable the model to learn the mapping relationship between the non-linear intermodulation features and the PIM signal, and generate a PIM signal prediction model; S330. Based on the error feedback mechanism of the training process, adjust the structure and parameters of the sequence-to-sequence learning model, optimize the PIM signal prediction model, and make it adapt to the multi-band intermodulation interference environment.

5. A method for eliminating RRU multi-band passive intermodulation based on a sequence-to-sequence learning model according to claim 1, characterized in that: The step S400 includes: S410. Input the multi - band transmission signal into the PIM signal prediction model, and generate a preliminarily predicted PIM interference signal according to the model output; S420. Calculate the residual between the predicted PIM interference signal and the received signal based on the error signal, and adjust the error compensation parameter through an optimization module; S430. Based on the residual optimization mechanism, compensate and adjust the predicted PIM interference signal to generate an optimized PIM interference signal.

6. A method for eliminating RRU multi-band passive intermodulation based on a sequence-to-sequence learning model according to claim 1, characterized in that: The step S500 includes: S510. Based on the optimized PIM interference signal, construct an adaptive filter and set the initial weight of the filter; S520. Perform filtering calculation using the received signal and the optimized PIM interference signal, and adjust the weight parameter of the adaptive filter to obtain an optimized interference compensation signal; S530. Superimpose the optimized interference compensation signal and the received signal to suppress PIM interference and extract the purified effective signal.

7. A method for eliminating RRU multi-band passive intermodulation based on a sequence-to-sequence learning model according to claim 1, characterized in that: The step S600 includes: S610. Obtain the signal data after interference suppression and calculate the current model prediction error; S620. Based on the error signal, adjust the parameters of the sequence - to - sequence learning model and perform incremental training to improve the adaptability of the PIM signal prediction model to environmental changes; S630. Based on the incremental training result, update the PIM signal prediction model and store the optimized model parameters.

8. A method for eliminating RRU multi-band passive intermodulation based on a sequence-to-sequence learning model according to claim 1, characterized in that: The step S700 includes: S710. Based on the model update parameters of online learning, construct a reinforcement learning environment and initialize the reinforcement learning strategy; S720. Extract features from the long - term environmental feedback data and train the reinforcement learning strategy to optimize the hyperparameters of the sequence - to - sequence learning model; S730. Based on the hyperparameters optimized by the reinforcement learning strategy, adjust the adaptive transformation weight of the feature extraction method and update the feature extraction strategy.

9. A method for eliminating RRU multi-band passive intermodulation based on a sequence-to-sequence learning model according to claim 1, characterized in that: The steps of dynamically updating the PIM signal prediction model using the online learning mechanism include: Obtain the signal data after interference suppression and calculate the current model prediction error; Based on the error signal, adjust the parameters of the sequence - to - sequence learning model and perform incremental training to improve the adaptability of the model to environmental changes; Based on the incremental training result, update the PIM signal prediction model and store the optimized model parameters.

10. A method for eliminating RRU multi-band passive intermodulation based on a sequence-to-sequence learning model according to claim 1, characterized in that: The steps of optimizing the hyperparameters of the PIM signal prediction model using the reinforcement learning method include: Based on the model update parameters of online learning, construct a reinforcement learning environment and initialize the reinforcement learning strategy; Extract features from the long - term environmental feedback data and train the reinforcement learning strategy to optimize the hyperparameters of the sequence - to - sequence learning model; Based on the hyperparameters optimized by the reinforcement learning strategy, adjust the adaptive transformation weight of the feature extraction method and update the feature extraction strategy to enhance the long - term stability of the system.

Images