Large-scale MIMO channel estimation method fusing dual attention mechanism and TCN-BiLSTM network

Abstract

The application provides a large-scale MIMO channel estimation method fusing a double attention mechanism and a TCN-BiLSTM network, a time convolution network (TCN) is combined with a bidirectional long short-term memory network (BiLSTM), and a double attention mechanism is introduced to further enhance the feature extraction and focusing ability of the model. Compared with other advanced deep learning models, the method can greatly reduce the prediction error in various channel environments; the method has excellent performance and stability in the non-line-of-sight scene with severe multipath effect and interference; by providing more accurate CSI, the method enables the large-scale MIMO base station to generate more accurate beams, more effectively transmits signal energy to the target user, minimizes the interference on other users, brings higher single-user data rate, and allows the system to serve more users at the same time in the same frequency spectrum resource, thereby significantly improving the throughput and spectrum utilization efficiency of the whole network.

Description

Technical Field

[0001] This invention pertains to mobile communication systems, specifically relating to a large-scale MIMO channel estimation method that integrates a dual attention mechanism with a TCN-BiLSTM network. Background Technology

[0002] With the continuous development of mobile communication systems towards higher speeds, lower latency, and higher reliability, Massive MIMO (Massive Multiple-Input Multiple-Output) technology has attracted widespread attention as one of the key supporting technologies for fifth-generation (5G) and future sixth-generation (6G) communication networks. To realize the performance potential of Massive MIMO, accurately estimating channel state information (CSI) has become a core task in system design.

[0003] Traditional channel estimation methods, such as least squares (LS) and minimum mean square error (MMSE), while theoretically sound and easy to implement, heavily rely on ideal statistical priors and linear channel assumptions. Under real-world, complex, dynamic, and non-stationary channel conditions, these methods suffer from insufficient accuracy and poor robustness. Especially in non-steady-state environments such as high-speed movement, frequency-selective fading, and sudden noise changes, these methods struggle to adapt to the rapid evolution and uncertainty of the channel.

[0004] To overcome the aforementioned bottlenecks, researchers have introduced deep learning methods for channel modeling. Long Short-Term Memory (LSTM) networks have become one of the mainstream approaches due to their strong sequence modeling capabilities. However, the standard LSTM structure suffers from high parameter coupling and a single memory mechanism, making it difficult to balance modeling depth and training efficiency. Therefore, Extended Long Short-Term Memory (XLSTM) networks have been proposed, introducing multi-channel gating and deep structures to improve the fitting ability to steady-state high-dimensional channels, making it particularly suitable for channel estimation in stable periodic scenarios.

[0005] While XLSTM achieves good performance in steady-state scenarios, it still lacks sufficient sensitivity to rapidly changing and nonlinearly drastic non-steady-state channels, particularly in modeling short-term disturbances and sudden dynamic characteristics. Therefore, improving the response capability to local changes and sudden disturbances has become a key research focus for next-generation non-steady-state channel modeling methods.

[0006] Against this backdrop, some researchers have attempted to introduce convolutional neural networks (CNNs) into communication modeling, but these lack global temporal memory capabilities. In contrast, temporal convolutional networks (TCNs), while preserving convolutional parallelism and training stability, introduce causal convolutional structures, resulting in stronger short-term dependency modeling capabilities. Combining TCNs with bidirectional LSTMs (BiLSTMs) further enhances the ability to model preceding and following contexts, providing a technical foundation for handling non-stationary channel variations. Summary of the Invention

[0007] The problem this invention aims to solve is: how to construct an efficient deep neural network with both local and global modeling capabilities and a key feature focusing mechanism to achieve accurate CSI estimation under rapidly changing and non-stationary MIMO channel conditions. This invention provides a large-scale MIMO channel estimation method based on a dual attention mechanism, TCN-BiLSTM. This method aims to address the problems of low accuracy and poor robustness caused by overly simplified model assumptions in traditional channel estimation algorithms, as well as the limitations of single deep learning models in capturing complex spatiotemporal characteristics of channels. Therefore, it achieves a channel estimation solution with higher estimation accuracy, stronger generalization ability, and better robustness in complex and dynamic wireless environments.

[0008] To achieve the above objectives, this invention proposes a hybrid network architecture based on deep learning. The core of this architecture is a novel combined model that organically integrates a Temporal Convolutional Network (TCN) with a Bidirectional Long Short-Term Memory (BiLSTM) network. Furthermore, it enhances the model's feature extraction and focusing capabilities by introducing a dual attention mechanism (including temporal and spatial attention).

[0009] A large-scale MIMO channel estimation method integrating a dual attention mechanism and a TCN-BiLSTM network includes the following steps:

[0010] Step 1, data preprocessing and feature engineering, including complex-to-real number conversion, data normalization, sliding window construction, and feature extraction, to obtain feature vectors as input to the model:

[0011] Step 2, Model and Data Flow: Establish a TCN-BiLSTM hybrid model that combines a temporal convolutional network (TCN) and a bidirectional long short-term memory network (BiLSTM). In addition, a dual attention mechanism, including temporal attention and spatial attention, is introduced into the model for data processing.

[0012] Step 3: Perform model training and parameter configuration; optimize network parameters by minimizing the error between predicted and true values.

[0013] Step 4: Put the TCN-BiLSTM hybrid model, which has been trained and adjusted in Step 3, into use. Input the feature vector processed in Step 1, and output the predicted value of the future channel state to complete the channel estimation.

[0014] The beneficial effects achieved by this invention are as follows:

[0015] (1) Significantly improves channel estimation accuracy: Experimental results show that the method of this invention can significantly reduce prediction error in various channel environments compared with other advanced deep learning models. For example, in the most challenging outdoor non-line-of-sight scenarios, the mean square error (MSE) is reduced by more than 78%. This improvement in accuracy directly translates into more effective precoding and beamforming, which is the basis for improving system performance.

[0016] (2) Enhanced robustness in complex environments: The method of this invention exhibits superior performance and stability in non-line-of-sight (NLOS) scenarios with severe multipath effects and interference. Its architecture design enables it to effectively cope with rapid and drastic changes in the channel, ensuring the reliability of communication links in complex environments such as urban areas and indoor spaces, and reducing communication interruptions or quality degradation caused by channel estimation errors.

[0017] (3) Improved system throughput and spectrum efficiency: By providing more accurate CSI, this invention enables large-scale MIMO base stations to generate more precise beams, delivering signal energy more effectively to target users while minimizing interference to other users. This directly results in higher single-user data rates and allows the system to serve more users simultaneously within the same spectrum resources, thereby significantly improving the overall network throughput and spectrum utilization efficiency.

[0018] (4) Improve the reliability and stability of communication systems: Higher estimation accuracy and robustness mean more stable communication connections. For end users, this means fewer dropped calls, smoother HD video experiences, and lower latency interactive applications (such as online games and remote control). This is crucial for meeting the high reliability requirements of 5G and future communication systems. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the TCN feature extraction process based on the attention mechanism in a specific embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of the overall model structure of the TCN-BiLSTM channel estimation method based on the dual attention mechanism in a specific embodiment of the present invention.

[0021] Figure 3 This is a comparison chart showing the fitting relationship between the predicted values ​​and the actual channel values ​​of the method of the present invention and the comparison method in an indoor non-line-of-sight (NLOS) scenario according to a specific embodiment of the present invention.

[0022] Figure 4This is a comparison chart showing the fitting relationship between the predicted values ​​and the actual channel values ​​of the method of the present invention and the comparison method in an outdoor non-line-of-sight (NLOS) scenario according to a specific embodiment of the present invention.

[0023] Figure 5 This is a comparison chart showing the fitting relationship between the predicted values ​​and the actual channel values ​​of the method of the present invention and the comparison method under urban non-line-of-sight (NLOS) conditions in a specific embodiment of the present invention.

[0024] Figure 6 This is a comparison chart showing the fitting relationship between the predicted values ​​and the actual channel values ​​of the method of the present invention and the comparison method in a high-mobility-NLOS scenario according to a specific embodiment of the present invention. Detailed Implementation

[0025] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0026] This invention provides a specific implementation flow for a large-scale MIMO channel estimation method based on a dual attention mechanism TCN-BiLSTM. This method aims to learn and predict future channel states from historical channel data using an end-to-end deep learning model.

[0027] The technical solution proposed in this invention aims to achieve collaborative and in-depth modeling of the spatiotemporal characteristics of the channel through its prediction principle and architecture design, thereby obtaining excellent prediction accuracy.

[0028] In time series modeling, there is an inherent challenge: how to efficiently capture both local, high-frequency mutation features and global, long-term dependencies simultaneously. The hybrid architecture of this invention cleverly solves this problem by combining the complementary advantages of TCN and BiLSTM.

[0029] Temporal Convolutional Networks (TCNs) are used for local and short-term feature extraction: TCN modules utilize their convolutional structure to process input time series data in parallel and efficiently. The TCN used in this invention incorporates two key technologies:

[0030] Causal convolution ensures that the output at prediction time t depends only on the input at the current and past times (t, t-1), strictly adhering to the causal relationship of the time series and avoiding the leakage of future information. Its mathematical expression is:

[0031]

[0032] Here, the subscript i represents the index of the current convolutional kernel on the time-series input, and ti represents the index of the kernel on the past i-th time-series input. x is the input sequence, f is the convolutional kernel, and k is the kernel size. This structure ensures the effectiveness of the prediction.

[0033] Dilated Convolution: By introducing a fixed gap (dilation factor d) between the elements of the convolution kernel, TCN can exponentially expand its receptive field with relatively low computational cost, thereby capturing dependencies over longer time spans. Its mathematical expression is:

[0034]

[0035] By progressively increasing the expansion factor d (e.g., 1, 2, 4, 8, ...), TCN can efficiently learn multi-scale temporal characteristics, and is particularly adept at capturing local, short-term, drastic changes in the channel, such as rapid signal fading.

[0036] Residual Connections: To enable the construction of deeper networks that learn more complex features, TCN introduces residual connections. These connections directly add the input x to the output F(x) of the current layer block via skip connections, i.e., y = Activation(x + F(x)). This structure greatly alleviates the vanishing gradient problem in deep networks, ensuring efficient information transfer and stable model training.

[0037] The Attention-TCN feature extraction module first expands the receptive field of convolution operations by combining causal convolution and dilated convolution, effectively extracting local and global features from the sequence data. After the convolution operation, the output of each convolutional layer is processed by a specially designed attention layer for feature weighting. The attention mechanism obtains attention weights by calculating the correlation between sequence features, making the model more focused on the time segments with significant influence in the channel response sequence. The combination of causal convolution and dilated convolution, with an attention layer added before feeding into the BiLSTM, weights the TCN output sequence, highlighting key time segments and weakening irrelevant features.

[0038] exist Figure 1 In this process, the attention weight vector v is obtained from the TCN output and along the time dimension using Softmax. The specific calculation process is as follows: First, the feature sequence matrix F extracted by the TCN... t =[f1,f2,…,f n The feature attention correlation coefficient e is obtained after nonlinear transformation. t =[e1,e2,…,e n The formula is: e t =tanh(We f t +b e ), among which, W e Let b represent the attention weight matrix. e This represents the bias vector.

[0039] Next, the attention coefficients are normalized using the Softmax function to obtain the attention weights V. t =[v1,v2,…,v n The formula is:

[0040] Secondly, the high-dimensional feature vector F t With feature attention weight V t Multiplying them yields feature vectors S of different dimensions. t =[s1,s2,…,s n The formula is: S t =[v1f1,v2f2,…,s n ].

[0041] Finally, the TCN feature extraction model with the added attention mechanism can automatically learn the correlation between different features in the data. The feature extraction module based on Attention-TCN, by combining the local feature capture capability of convolutional neural networks with the feature selection capability of attention mechanisms, can extract features more efficiently and accurately from large-scale MIMO channel data.

[0042] Bidirectional Long Short-Term Memory Network (BiLSTM) for global and long-term context understanding: After TCN extracts efficient local features, this invention uses BiLSTM to perform deeper global context modeling on these features.

[0043] LSTM units can selectively remember and forget historical information through their internal forget gate, input gate, and output gate, thereby effectively capturing long-term dependencies.

[0044] BiLSTM consists of a forward LSTM and a backward LSTM. The forward LSTM processes information from the beginning of the sequence to the end t = (1,,T), capturing historical context; the backward LSTM processes information from the end of the sequence to the beginning (t = T,,1), capturing future context. At time t, the output of BiLSTM... It is a forward-hidden state. and backward hidden state splicing: This bidirectional processing mechanism enables the model to fully utilize complete information from all preceding and following moments when judging the channel state at any given time, thereby gaining a deeper understanding of the overall evolution trend of the channel.

[0045] The combination of TCN and BiLSTM creates a powerful synergistic effect: TCN, as an efficient local feature extractor, provides BiLSTM with high-quality, information-intensive input; while BiLSTM, based on the features extracted by TCN, performs global contextual association and logical inference. This architecture enables the model to respond quickly to instantaneous changes in the channel while accurately grasping its long-term variation patterns.

[0046] To enable the model to intelligently focus on the most critical information in the channel data, this invention introduces a dual attention mechanism, which acts like a learnable, adaptive spatiotemporal filter.

[0047] Temporal Attention: This mechanism is applied to the output sequence of a BiLSTM to dynamically evaluate the contribution of each time step in the sequence to the final prediction. It calculates the attention weight α for each time step. t The model can give higher attention to moments containing key changes (such as the onset of fading or the emergence of disturbances) while ignoring relatively stable moments with less information. The weighted time features are represented as follows: This allows the model to more accurately capture key nodes of dynamic channel changes.

[0048] Spatial Attention: In massive MIMO systems, channel information is multi-dimensional, corresponding to paths between different antenna pairs. Spatial attention operates on the feature dimension, aiming to identify which antenna paths (or feature dimensions) are primary and reliable at the current moment, and which are secondary or severely interfered with. This is achieved by calculating an attention weight β for each feature dimension. j The model can adaptively amplify the signal contribution of important paths while suppressing the effects of noisy or interfering paths. This is crucial for improving the purity and accuracy of channel estimation in complex multipath environments.

[0049] The introduction of the dual attention mechanism enables the model to no longer treat all spatiotemporal information equally, but instead learns to "target" and concentrate limited computing resources on the most critical information, thereby significantly improving the robustness and prediction accuracy of the model.

[0050] The TCN-BiLSTM model proposed in this invention is designed to be most suitable for channel environments with strong time-varying characteristics, complex multipath effects and non-stationary features, especially non-line-of-sight (NLOS) propagation scenarios in urban environments.

[0051] The reasons are as follows: NLOS scenarios are among the most challenging environments in wireless communication. Signals cannot reach the receiver directly, but instead converge through multiple paths such as reflection, diffraction, and scattering from buildings, resulting in severe and rapid fading (fast fading) of the signal, and the channel state exhibits high nonlinearity and nonstationarity.

[0052] The TCN module of this invention, with its powerful local feature capture capability, can accurately identify and model these rapid and intense signal fluctuations.

[0053] The BiLSTM module can learn the underlying long-term change patterns and contextual relationships from these seemingly random fluctuations.

[0054] The dual attention mechanism is particularly crucial in this environment: spatial attention helps the model distinguish and focus on a few main, high-energy reflection paths while suppressing interference from a large number of weak and unstable scattering paths; temporal attention helps the model capture critical channel state transitions caused by user movement or environmental changes.

[0055] Experimental data fully demonstrate this point. In NLOS scenarios, the performance advantage of the model of this invention is most significant compared to other models, with substantial reductions in metrics such as mean square error (MSE) and mean absolute error (MAE), proving the high degree of compatibility between its architecture and the complex characteristics of NLOS channels.

[0056] In summary, this invention constructs a powerful model capable of deeply understanding and accurately predicting complex channel spatiotemporal dynamics through the collaborative architecture of TCN-BiLSTM and adaptive filtering with dual attention mechanism, thereby fundamentally solving the limitations of traditional methods and achieving a breakthrough improvement in channel estimation performance.

[0057] This invention introduces the TCN-BiLSTM structure into the non-steady MIMO channel estimation scenario for the first time; it innovatively introduces a dual attention mechanism (time and channel) to improve the ability to capture key features; the fusion mechanism is reasonably designed to adapt to short-term and long-term CSI dynamic changes; and the complete training framework and data flow construction scheme are suitable for deployment in actual communication systems.

[0058] The core of this invention lies in proposing an innovative TCN-BiLSTM hybrid network architecture. This architecture is not a simple model concatenation, but a deep collaborative design. It utilizes a Temporal Convolutional Network (TCN) to efficiently capture local and short-term abrupt changes in the channel in parallel, while simultaneously employing a Bidirectional Long Short-Term Memory Network (BiLSTM) to perform global context modeling on the features extracted by the TCN, thereby capturing long-term dependencies. This architecture cleverly addresses the inherent limitations of a single model when simultaneously processing high-frequency details and low-frequency trends.

[0059] This invention introduces a novel dual spatiotemporal attention mechanism, essentially a learnable, data-driven adaptive filter. This mechanism allows the model to dynamically focus on the most critical information in both temporal and spatial dimensions: the temporal attention mechanism identifies the historical moments that contribute the most to the prediction, while the spatial attention mechanism identifies the antenna paths with the strongest and most reliable signals. This enables the model to intelligently amplify effective signals and suppress noise and interference.

[0060] This invention provides a complete end-to-end deep learning framework from raw data to final prediction. This framework automatically learns the inherent complex spatiotemporal correlations directly from high-dimensional channel data, completely eliminating the reliance of traditional methods (such as MMSE) on unrealistic prior channel statistics, making it more suitable for real and ever-changing wireless communication environments.

[0061] The invention was implemented on a standard deep learning experimental platform. The hardware environment included an Intel Core i7-14700KF processor, 64GB of RAM, and an NVIDIA GeForce RTX 4080 SUPER graphics card to ensure the efficient parallel computing power required for model training. The software environment was based on the Python language and used TensorFlow 2.11 as the deep learning framework.

[0062] To verify the effectiveness and robustness of the method of this invention, this embodiment uses the internationally recognized COST2100 wireless channel simulation dataset. This dataset can simulate a variety of real wireless propagation environments. This embodiment selects four typical communication scenarios for model training and evaluation: Indoor Non-Line-of-Sight (NLOS), Outdoor Non-Line-of-Sight (NLOS), Urban Non-Line-of-Sight (NLOS), and High-Mobility Non-Line-of-Sight (NLOS). These scenarios cover different channel conditions from stable to drastically changing, and can comprehensively test the performance of the method of this invention.

[0063] Step 1: Data Preprocessing and Feature Engineering

[0064] To make the raw channel data suitable for deep learning models, a series of preprocessing and feature engineering steps are required:

[0065] 1. Conversion from complex to real numbers: The original CSI data is a complex matrix containing real and imaginary parts. First, each complex value is split into its real and imaginary parts, and then these parts are concatenated along the feature dimension to form a pure real matrix.

[0066] 2. Data Normalization: To accelerate model convergence and improve training stability, the Z-score normalization method is used to normalize the data. For each feature dimension, the formula x′=(x-μ) / σ is applied for calculation, where μ and σ are the mean and standard deviation of the feature on the training set, respectively.

[0067] 3. Sliding Window Construction: To transform time series data into supervised learning samples, a sliding window technique is employed. A fixed window length is set (e.g., 10 time steps), and this window is slid along the time axis. The sequence within the window is used as the model input, and the data from the next time step after the window is used as the prediction target.

[0068] 4. Feature Extraction: To provide the model with richer information, additional statistical and frequency domain features are extracted from each sliding window. These new features are then concatenated with the original sequence features to form the final model input vector. The specific implementation process is as follows:

[0069] First, within each sliding window L, the following statistical features need to be extracted: mean: variance: Next, the time-domain signal needs to be converted to the frequency domain using a Fast Fourier Transform: Finally, the features need to be concatenated: According to the above definition, let the original sequence window length be L, the input feature dimension be d, and based on the m (mean, variance, etc.) and n (dominant frequency, spectral energy, etc.) indicators obtained from the statistical features and frequency domain features respectively, the final input vector for each sliding window is obtained: z = [x 1:L ,s 1:m ,f 1:n ], where x 1:L ,s 1:m ,f 1:n These are the original sequence features, statistical feature vectors, and frequency domain feature vectors, respectively.

[0070] Step Two: Model Building and Data Flow

[0071] The core of this invention is as follows Figure 1 The data processing flow of the TCN-BiLSTM hybrid model shown is as follows:

[0072] 1. The preprocessed input sequence is first fed into the TCN module. This module consists of multiple stacked dilated causal convolutional layers. Causal convolution ensures that the convolution operation only depends on the current and past time steps and does not leak future information. Dilated convolution introduces a gap on top of causal convolution to expand the receptive field, thereby capturing dependencies over a longer time range. The TCN module is formed by stacking causal and dilated convolutions, with each layer controlled by "causality + dilation factor".

[0073] 2. The TCN outputs local temporal features. An attention layer is added to weight the TCN output sequence, highlighting key time segments and weakening irrelevant features. The feature sequence output by the TCN module is then fed into the BiLSTM module. This module processes the sequence in both forward and backward directions to capture long-term, global contextual dependencies.

[0074] 3. The hidden state output of the BiLSTM module is fed into a dual attention module. This module computes attention weights in both temporal and spatial dimensions in parallel, and then performs a weighted summation of the features based on these weights to generate a more information-dense context vector. For each time step, the output h... t Calculate attention weights: e t =v T tanh(W h h t +b).

[0075] Among them, e t Assign an attention score to time step t, v T W is the attention scoring weight vector. h h is the linear transformation matrix used to calculate the temporal attention score. t Let be the hidden state input at time step t, b be the bias term, and T be the total number of time steps. The final weighted representation is:

[0076]

[0077] H time This represents the overall sequence list after time weighting, a t h represents the attention weights for each time step. t The hidden state input is the time step, and T is the total number of time steps.

[0078] Spatial attention focuses on different components along the feature dimension. Let H∈ T×d Then, for each feature dimension j, calculate its attention weight:

[0079]

[0080] β j For the attention weights of feature dimension j, g j Let W be the attention weight score for feature dimension j, and w be the learned spatial attention score weight vector. j H is a trainable linear transformation matrix. i,j Let b be a vector of dimension T in the j-th column of the input matrix H. j is the bias term, and m is the number of feature dimensions.

[0081] The final weighted average is:

[0082] H space =H·β

[0083] H space H represents the overall sequence after spatial weighting, H is the concatenated BiLSTM output matrix, and β is the attention weight vector for all feature dimensions.

[0084] The output after fusing temporal and spatial attention is represented as follows:

[0085] H final =H time +H space

[0086] The three symbols above represent, in order: the final feature after fusing temporal and spatial attention, the temporally weighted overall sequence, and the spatially weighted overall sequence. The final output feature vector H final The data is fed into the fully connected layer for regression prediction of the channel value.

[0087] 4. Finally, the context vector is fed into a fully connected (dense layer) to perform the final regression calculation and output a prediction of the future channel state.

[0088] Step 3: Model Training and Parameter Configuration

[0089] The training process of the model aims to optimize the network parameters by minimizing the error between the predicted and the true values.

[0090] Optimizer: The Adam optimizer is used, which can adaptively adjust the learning rate of each parameter and has the advantages of fast convergence speed and stable performance.

[0091] Loss function: Mean Squared Error (MSE) is used as the loss function, and its calculation formula is as follows: in It is the predicted value, y i It is the actual value.

[0092] Dataset splitting: The dataset is split into 90% training set and 10% test set in chronological order to ensure the model's generalization ability on unseen data.

[0093] Hyperparameter configuration: The key hyperparameters of the model are set (3 causal convolutional layers, 64 filters per layer, 5 dilated convolutional kernels, dilation factors of 1, 2, and 4, and ReLU activation function). These parameters are the optimal configuration determined after multiple experiments and optimizations. In addition, the number of hidden units in the BiLSTM module is set to 128, the initial learning rate is 0.001, and the training batch size is 64.

[0094] To verify the superiority of the method of the present invention, its performance was compared with that of a baseline model (such as XLSTM) in four typical scenarios.

[0095] Qualitative analysis: such as Figures 3 to 6 The figure shows the fitting of the prediction curves of the proposed method (TCN-BiLSTM) and the comparison method (XLSTM) to the actual channel values. It can be intuitively seen that in all scenarios, the prediction curve of the proposed method more closely matches the actual value curve. In the non-line-of-sight (NLOS) scenario shown in the figure, the channel fluctuates dramatically, and the comparison method's prediction shows significant lag and bias, while the proposed method can still accurately track the rapid changes in the channel, demonstrating excellent robustness.

[0096] Quantitative analysis: To conduct an objective evaluation, mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²) were used. 2 Three metrics were used to quantify the performance. The results are shown in the table below.

[0097] Table: Quantitative Comparison of Model Prediction Performance in Four Types of Scenarios

[0098]

[0099] Comparing the Indoor-NLOS and Outdoor-NLOS scenarios using the table data, although both belong to non-line-of-sight propagation environments, the Outdoor-NLOS scenario is more significantly affected by environmental interference during signal propagation compared to the Indoor-NLOS scenario. XLSTM achieves a determination coefficient of 0.876 in the Indoor-NLOS environment, indicating that it has some predictive ability for channel changes; however, the high prediction error means that the model still cannot fully capture the time-varying characteristics of the environment. In the more challenging Outdoor-NLOS scenario, the performance of the XLSTM model further declines, with the determination coefficient decreasing to 0.853, indicating that its ability to capture temporal features gradually weakens as environmental complexity and signal interference increase.

[0100] The above description is only a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. Any equivalent modifications or changes made by those skilled in the art based on the content disclosed in the present invention should be included within the scope of protection set forth in the claims.

Claims

1. A large-scale MIMO channel estimation method integrating dual attention mechanism and TCN-BiLSTM network, characterized by: The method includes the following steps: Step 1, data preprocessing and feature engineering, including complex-to-real number conversion, data normalization, sliding window construction and feature extraction, to obtain feature vectors as input to the model; Step 2, Model and Data Flow: A hybrid TCN-BiLSTM model combining a Temporal Convolutional Network (TCN) and a Bidirectional Long Short-Term Memory (BiLSTM) network is established. The model incorporates a dual attention mechanism, including both temporal and spatial attention, for data processing. The data processing flow is as follows: The preprocessed input sequence is first fed into the TCN module; this module consists of multiple stacked dilated causal convolutional layers, each controlled by causality and dilation factor; the TCN outputs local temporal features, and an attention layer is added to weight the TCN output sequence, highlighting key time segments and weakening irrelevant features; The feature sequence output by the TCN module is then fed into the BiLSTM module, which processes the sequence in both forward and backward directions to capture long-term, global contextual dependencies. The hidden state output of the BiLSTM module is fed into a dual attention module, which computes attention weights in both time and space dimensions in parallel, and then performs a weighted summation of the features based on these weights to generate a more information-dense context vector. For each time step output... Calculate attention weights: ； in, Assign a score to the attention level at time step t. Assign a weight vector to the attention score. Here is the linear transformation matrix used to calculate the temporal attention score. Let b be the hidden state input at time step t, b be the bias term, and T be the total number of time steps; the final weighted representation is: This represents the overall sequence list after time weighting. For attention weights at time steps, The hidden state input is the time step, and T is the total number of time steps; Spatial attention focuses on different components of the feature dimension; let... Then, for each feature dimension j, calculate its attention weight: The attention weights for feature dimension j are... Let be the attention weight score for feature dimension j, and w be the learned spatial attention weight vector. Let be a trainable linear transformation matrix. Let T be the vector in the j-th column of the input matrix H. Here, m represents the bias term, and m is the number of feature dimensions. The final weighted average is: The overall sequence is weighted according to spatial weights, and H is the concatenated BiLSTM output matrix. This is the attention weight vector for all feature dimensions; The output after fusing temporal and spatial attention is represented as follows: The three parameters above represent, in order: the final feature after fusing temporal and spatial attention, the overall sequence after temporal weighting, and the overall sequence after spatial weighting; the final output feature vector. The signal is fed into the fully connected layer for regression prediction of the channel value; Finally, the context vector is fed into a fully connected layer, which performs the final regression calculation and outputs a prediction of the future channel state. Step 3: Perform model training and parameter configuration; optimize network parameters by minimizing the error between predicted and true values. Step 4: Put the TCN-BiLSTM hybrid model, which has been trained and adjusted in Step 3, into use. Input the feature vector processed in Step 1, and output the predicted value of the future channel state to complete the channel estimation.

2. The large-scale MIMO channel estimation method integrating dual attention mechanism and TCN-BiLSTM network according to claim 1, characterized in that: In step 1, the conversion from complex to real numbers is as follows: the original CSI data is a complex matrix containing a real part and an imaginary part; firstly, each complex value is split into two components, a real part and an imaginary part, and then they are concatenated along the feature dimension to form a pure real matrix.

3. The large-scale MIMO channel estimation method integrating dual attention mechanism and TCN-BiLSTM network according to claim 1, characterized in that: In step 1, data normalization specifically involves using the Z-score normalization method to normalize the data in order to accelerate model convergence and improve training stability; for each feature dimension, the formula is applied... Perform calculations, where and These are the original features Mean and standard deviation on the training set.

4. The large-scale MIMO channel estimation method integrating dual attention mechanism and TCN-BiLSTM network according to claim 1, characterized in that: In step 1, the sliding window construction is specifically as follows: in order to convert time series data into supervised learning samples, the sliding window technique is used; a fixed window length is set, the window is slid along the time axis, the sequence within the window is used as the model input, and the data of the next time step after the window is used as the prediction target.

5. The large-scale MIMO channel estimation method integrating dual attention mechanism and TCN-BiLSTM network according to claim 1, characterized in that: In step 1, feature extraction specifically involves extracting statistical features, including mean and variance, and frequency domain features obtained through Fast Fourier Transform (FFT) from each sliding window, and concatenating these new features with the original sequence features to form the final model input vector.

6. The large-scale MIMO channel estimation method integrating dual attention mechanism and TCN-BiLSTM network according to claim 1, characterized in that: In step 2, BiLSTM consists of a forward Long Short-Term Memory (LSTM) network and a backward Long Short-Term Memory (LSTM) network; the forward LSTM runs from the beginning to the end of the sequence. Processing information and capturing historical context; backward LSTM works from the end of the sequence to the beginning. Process information and capture future context; At any moment BiLSTM output It is a forward-hidden state. and backward hidden state splicing: .

7. The large-scale MIMO channel estimation method integrating dual attention mechanism and TCN-BiLSTM network according to claim 1, characterized in that: In step 3, the Adam optimizer is used to adaptively adjust the learning rate of each parameter; the mean squared error (MSE) is used as the loss function; the dataset is divided into a 90% training set and a 10% test set in chronological order; the key hyperparameters of the model are set as follows: 3 convolutional layers, 64 filters per layer, 5 kernel size, dilation factors of 1, 2, and 4, ReLU activation function, 128 hidden units in the BiLSTM module, initial learning rate of 0.001, and batch size of 64.

8. The large-scale MIMO channel estimation method integrating dual attention mechanism and TCN-BiLSTM network according to claim 1, characterized in that: The specific steps for outputting the predicted value of the future channel state in step 4 are as follows: Step 4-1: The preprocessed model input vector is first fed into the TCN module; this module consists of multiple stacked dilated causal convolutional layers, used to extract local and multi-scale temporal features from the channel data. Step 2-2: The feature sequence output by the TCN module is then fed into the BiLSTM module. This module processes sequences in both forward and backward directions to capture long-term, global contextual dependencies; In steps 2-3, the hidden state output of the BiLSTM module is fed into the dual attention module; This module computes attention weights in both time and space dimensions in parallel, and performs a weighted summation of features based on these weights to generate a more information-dense context vector. Steps 2-4: Finally, the context vector is fed into a fully connected Dense Layer, which performs the final regression calculation and outputs a prediction of the future channel state.

Images