Deep learning DD channel estimation system and method combined with adaptive CIR truncation mechanism

Abstract

The invention provides a deep learning DD channel estimation system and method combined with an adaptive CIR truncation mechanism, and the system comprises a pilot frequency estimation module which carries out the channel estimation according to a received pilot frequency sequence, and outputs a pilot frequency assisted channel estimation value; the deep learning prediction module is used for predicting the channel state of the next moment according to the historical channel estimation sequence; the data auxiliary channel estimation module is used for balancing and decoding the data symbols based on the channel state, and obtaining data auxiliary channel estimation after reconstruction; the self-adaptive CIR truncation module takes data-assisted channel estimation as input and outputs a channel estimation value after truncation; and the periodic pilot frequency correction module is used for triggering pilot frequency updating and correcting the pilot frequency assisted channel estimation value. According to the method, the influence of a data judgment error is actively considered, the tap weight is dynamically output by using the lightweight neural network, the historical sequence is intelligently cleaned and cut off, error propagation is effectively inhibited, and the stability and robustness of prediction are improved.

Description

Technical Field

[0001] This invention relates to the field of wireless communication, and more specifically, to a deep learning DD channel estimation system and method that incorporates an adaptive CIR truncation mechanism. Background Technology

[0002] In modern wireless communication systems, channel estimation is one of the key technologies for achieving reliable transmission. The receiver needs to accurately estimate the Channel State Information (CSI) to perform equalization, detection, and decoding operations, ensuring that data can be correctly recovered even under conditions of multipath fading, noise interference, and time-varying channels. The accuracy of channel estimation directly determines the system's bit error rate, throughput, and spectral efficiency. Traditional methods mainly rely on periodic pilot symbols for channel estimation and interpolation. While this provides relatively robust estimates, excessive pilots consume limited spectrum resources, reducing the effective bandwidth available for data transmission and system throughput. To reduce pilot overhead, existing technologies have proposed Decision-Directed (DD) channel estimation methods, which utilize data signals as "virtual pilots" to participate in channel estimation based on initial pilot estimation, thereby improving spectral efficiency. However, existing DD schemes tend to treat erroneous symbols as virtual pilots when data detection errors occur, introducing estimation bias that accumulates in subsequent time recursion or prediction processes, leading to error propagation. This phenomenon is particularly prominent under low signal-to-noise ratio and non-stationary channel conditions, severely impacting system reliability and decoding performance. Although deep learning-based channel prediction methods have been applied in DD channel estimation to adapt to errors caused by channel time-varying characteristics, most methods still rely directly on data detection-assisted channel estimation results and lack effective suppression of data detection error accumulation, leading to instability in prediction results.

[0003] Current research and patents have proposed a scheme that combines decision-guided (DD) channel estimation with deep learning prediction methods to improve channel prediction performance.

[0004] Patent document CN112968847B discloses a channel estimation method based on deep learning and data pilot assistance. In this implementation, the system first completes the initial channel estimation using pilot symbols; the obtained channel estimation sequence is input into a deep learning model (such as a recurrent neural network or a convolutional neural network) for learning and extrapolation to predict the channel state information at the next moment, thereby completing the time-varying channel tracking; then, the prediction results are used for data detection and decoding and symbol reconstruction, and the reconstructed symbols are used for channel estimation to obtain the data-assisted channel estimation result as the prediction input for the next time.

[0005] DFT channel estimation is similar in concept to adaptive CIR truncation. DFT channel estimation mainly relies on converting the frequency domain channel response into the time domain channel impulse response through inverse discrete Fourier transform (IDFT), and then truncating the time domain taps based on a fixed threshold or simple rules (such as an energy threshold). Its core function is denoising, that is, removing low-energy taps caused by noise.

[0006] The disadvantages of existing technologies include: 1. Existing technologies combine direct-dial (DD) based channel estimation with deep learning prediction: for example, using recurrent neural networks to predict historical channel sequences and updating them based on data-assisted estimation. However, these methods often only focus on learning and extrapolating historical sequences, lacking consideration and handling for potential errors in data-driven decisions, and are easily affected by error propagation.

[0007] 2. Error propagation suppression methods for DFT channel estimation lack targeted handling for dynamically introduced pseudopaths (such as false multipaths caused by data detection errors), which easily accumulate errors under the DD framework, especially performing poorly in non-stationary channel environments.

[0008] 3. No solutions have yet been found that specifically incorporate channel characteristics, thus lacking targeted optimization for different channel environment types.

[0009] Overall, existing methods lack an adaptive error propagation suppression mechanism for data detection errors. Therefore, current methods cannot effectively solve the error propagation problem of DD structures, nor can they balance pilot efficiency with the robustness of deep learning predictions. Summary of the Invention

[0010] To address the shortcomings of existing technologies, the purpose of this invention is to provide a deep learning DD channel estimation method and system that combines an adaptive CIR truncation mechanism.

[0011] A deep learning DD channel estimation system combining an adaptive CIR truncation mechanism, provided by the present invention, includes: The pilot estimation module is used to perform channel estimation based on the received pilot sequence and output pilot-aided channel estimation values. The deep learning prediction module is used to predict the channel state at the next moment based on the historical channel estimation sequence. The data-assisted channel estimation module equalizes and decodes data symbols based on the channel state predicted by the deep learning prediction module, and then reconstructs the data-assisted channel estimate. The adaptive CIR truncation module takes data-aided channel estimation as input and outputs truncated channel estimation values. The periodic pilot correction module is used to trigger pilot updates and correct the pilot-aided channel estimates.

[0012] Preferably, each module is connected through a feedback loop; the prediction output is fed into the data estimate, and the data estimate is truncated and fed back to the prediction, forming a closed loop.

[0013] Preferably, the input to the pilot estimation module includes the pilot set P and the received sequence of the t-th transmission block. The output includes pilot-aided channel estimation. , where t=0 represents the initial estimation stage.

[0014] Preferably, the input to the deep learning prediction module includes the historical channel estimate of the t-th transport block. The output includes a prediction of the channel state at the next time step. .

[0015] Preferably, the input to the data-assisted channel estimation module includes the channel prediction for the (t+1)th transport block. The output includes data-aided frequency domain channel estimation. Decoding bit log-likelihood ratio (LLR); utilizing The data symbols of the (t+1)th transport block are equalized and decoded to reconstruct the soft decision result. ,use As virtual pilots, more accurate data-assisted channel estimation can be calculated. .

[0016] Preferably, the input to the adaptive CIR truncation module includes data-aided channel frequency domain response estimation. The decoded bit log-likelihood ratio (LLR), signal-to-noise ratio (SNR), and current pilot spacing length (M) are calculated. The output includes a channel estimate after adaptive CIR truncation. The tap weights of each path are output by a lightweight neural network, and the CIR is adaptively truncated to filter out pseudo-paths and reduce noise.

[0017] Preferably, the input of the periodic pilot correction module includes periodically inserted pilot symbols; every [period] Each transport block triggers a pilot update to correct the accumulated error under iterative channel prediction and data-assisted channel estimation.

[0018] A deep learning-based DD channel estimation method combining an adaptive CIR truncation mechanism, provided by the present invention, includes: Step S1: Perform channel estimation based on the received pilot sequence and output the pilot-aided channel estimation value; Step S2: Predict the channel state at the next moment based on the historical channel estimation sequence, and output the prediction result; Step S3: Equalize and decode the data symbols based on the prediction results, and obtain the data-assisted channel estimate after reconstruction; Step S4: Take the data-assisted channel estimation as input and output the truncated channel estimation value; Step S5: Periodically trigger pilot updates to correct the channel estimation values ​​assisted by the pilots.

[0019] Preferably, the multipath components of the temporal impulse response are dynamically truncated by using the output tap weights / masks of the neural network ACT-Net.

[0020] Preferably, the adaptively truncated channel estimate is input into the deep learning prediction model, and the channel prediction result for the next time step is output for subsequent data detection and decoding.

[0021] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention introduces an adaptive CIR truncation mechanism (ACT-Net) to proactively consider the impact of data decision errors in the DD-based channel estimation and deep learning prediction framework. It utilizes a lightweight neural network to dynamically output tap weights to intelligently clean and truncate historical sequences, effectively suppressing error propagation and improving the stability and robustness of prediction.

[0022] 2. Compared to the suppression mechanism in DFT channel estimation, the adaptive truncation mechanism utilizes a lightweight neural network to dynamically output delay tap weights, incorporating real-time conditions and reliability information. This not only effectively reduces noise but also intelligently filters spurious paths and long-tailed energy. Through adaptive weighting or truncation, it suppresses error propagation caused by incorrect decisions, maintaining the sparse structure of the channel. The advantage of this mechanism lies in its intelligent adaptability and robustness: it can automatically adjust the strategy according to current channel conditions, significantly reducing pilot overhead, improving prediction accuracy, and achieving higher system throughput and reliability in sparse multipath environments such as low VHF. In contrast, traditional DFT methods rely on manual parameter settings and have weaker adaptability to complex scenarios. Attached Figure Description

[0023] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a block diagram of the receiving end functional modules of the present invention.

[0024] Figure 2 This is a time series diagram for channel estimation and prediction.

[0025] Figure 3 This is a diagram of the lightweight neural network architecture (ATC-Net) used for adaptive CIR truncation.

[0026] Figure 4 This is a diagram of the LSTM network architecture used for channel prediction.

[0027] Figure 5 This is a comparison chart of the simulation performance of the present invention. Detailed Implementation

[0028] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0029] Reference Figure 1 As shown, a deep learning DD channel estimation system combining an adaptive CIR truncation mechanism includes: The pilot estimation module is used to perform channel estimation based on the received pilot sequence and output pilot-aided channel estimation values. The deep learning prediction module is used to predict the channel state at the next moment based on the historical channel estimation sequence. The data-assisted channel estimation module equalizes and decodes data symbols based on the channel state predicted by the deep learning prediction module, and then reconstructs the data-assisted channel estimate. The adaptive CIR truncation module takes data-aided channel estimation as input and outputs truncated channel estimation values. The periodic pilot correction module is used to trigger pilot updates and correct the pilot-aided channel estimates.

[0030] In one embodiment, the above module specifically includes: 1. Pilot estimation module Input: Pilot set P, received sequence of the t-th transport block

[0031] Output: Pilot-aided channel estimation (t=0 indicates the initial estimation stage) Function: The system uses periodic pilot sequences to perform least squares (LS) or linear minimum mean square error (LMMSE) estimation of the frequency domain channel in the initial and periodic triggering phases. This provides the system with basic channel information and a priori information for subsequent processing.

[0032] 2. Deep Learning Prediction Module Input: Historical channel estimate of the t-th transport block (Channel estimation final output based on data from each preceding transport block or pilot-aided channel estimation) Output: Predict the channel state at the next time step. .

[0033] Function: Input historical channel estimation sequences into an LSTM network to predict the channel state at the next moment and track the time-varying nature of the channel.

[0034] 3. Data-assisted channel estimation module Input: Channel prediction for the (t+1)th transport block

[0035] Output: Data-assisted frequency domain channel estimation Decoded bit log-likelihood ratio LLR Function: Utilize The data symbols of the (t+1)th transport block are equalized and decoded to reconstruct the soft decision result. ,use As virtual pilots, more accurate data-assisted channel estimation is calculated.

[0036] 4. Adaptive CIR truncation module Input: Data-assisted channel frequency response estimation The decoded bit log-likelihood ratio (LLR), SNR, and current pilot spacing length M Output: Channel estimation after adaptive CIR truncation

[0037] Function: Adaptively truncating and denoising the CIR by outputting tap weights for each path using a lightweight neural network. The adaptively truncated CIR... The historical sequence values ​​are used for channel prediction at the next transport block time.

[0038] 5. Periodic pilot correction module Input: Periodically inserted pilot symbols Function: every Each transmission block triggers a pilot update for correction. This helps avoid the long-term accumulation of errors.

[0039] These modules are connected through a feedback loop: the prediction output is fed into the data estimate, and the data estimate is truncated and fed back into the prediction, forming a closed loop.

[0040] A deep learning-based DD channel estimation method incorporating an adaptive CIR truncation mechanism includes: Step S1: Perform channel estimation based on the received pilot sequence and output the pilot-aided channel estimation value; Step S2: Predict the channel state at the next moment based on the historical channel estimation sequence, and output the prediction result; Step S3: Equalize and decode the data symbols based on the prediction results, and obtain the data-assisted channel estimate after reconstruction; Step S4: Take the data-assisted channel estimation as input and output the truncated channel estimation value; Step S5: Periodically trigger pilot updates to correct the channel estimates to pilot-assisted values.

[0041] Example 1 This embodiment describes a single-carrier communication system as an example, where the receiver performs signal processing in the baseband complex domain. Ultra-shortwave channels have narrowband characteristics and a finite number of multipath components. The receiver needs to accurately estimate and predict the time-varying channel under the condition of finite pilot signals. Considering single-carrier frequency domain equalization (SC-FDE) technology, assuming each transmission block has a length of N, a cyclic prefix (CP) with a length of N is added to the beginning of the block. ,satisfy ,in Let be the number of valid taps in the CIR. Let the transmitted symbol sequence of the t-th transport block be... This can be a sequence of pilot symbols or data symbols. The frequency domain model of the received signal after removing the pilot symbol (CP) and performing an N-point FFT is as follows:

[0042] in Let be the frequency domain receive vector of the t-th transmission block. Let be the frequency domain transmission vector of the t-th transmission block. Let be the channel frequency domain response, represented as a diagonal matrix, for the t-th transport block. It is additive white Gaussian noise, with each element having a mean of 0 and a variance of . , independent and identically distributed complex Gaussian random variables.

[0043] The specific implementation steps are as follows.

[0044] Step 1: When Send symbol sequence Carrying pilot sequence, for the received pilot signal Channel estimation, taking the LS method as an example:

[0045] During the initial channel estimation phase, multiple transport blocks can be used to send pilot sequences and accumulate sufficient historical prior information.

[0046] Step 2: Estimate the frequency response using historical channel data. Perform channel prediction for the next transport block, where In the first case, the pilot-assisted channel estimation result from step 1 is used; otherwise, the data-assisted channel estimation result after adaptive CIR truncation correction from step 4 is used. First, extract... The diagonal elements are then separated into real and imaginary parts to form a matrix:

[0047] in, express The k-th diagonal element is considered as a temporal sequence of length N, where each "time step" (row) is a 2D feature vector (real part + imaginary part). LSTM networks are sequence models, adept at handling temporal dependencies. The input is fed into an LSTM network, and the model outputs a matrix. Then remapped to Channel prediction for the next transport block.

[0048] Step 3: Use the prediction channel Channel equalization, taking zero-forcing equalization as an example:

[0049] Equalized symbol sequence After an N-point IFFT, the data is returned to the time domain and fed into a soft-decision and decoding unit (e.g., a Turbo or LDPC decoder) to obtain the corresponding LLR sequence, which is then reconstructed to obtain the data symbol vector. The LLR sequence is output from soft decoding, with a length of M ( , where b is the number of bits for each symbol), the LLR of the m-th bit is denoted as LLRm. Using As a virtual pilot, update the data to aid estimation. ,in .

[0050] Step 4: Estimate the data-assisted frequency domain response Transform to time-domain impulse response using IDFT Each element corresponds to a complex value of a time-delay tap. Then, a lightweight neural network is used to generate the tap weight vector. The network input is a concatenated matrix of the real and imaginary parts of the time-domain impulse response. The real and imaginary parts are separated from the time-domain vector to form a matrix:

[0051] in, express The i-th element.

[0052] Considering the focus on lightweight design and reliability, this example calculates the average of the absolute values ​​of the LLR. The SNR and the interval length M from the previous pilot block are used as global feature inputs to represent the overall data decision reliability (low). SNR indicates a high risk of error, guiding the network to make more conservative cuts, where:

[0053]

[0054] Features can also be represented as signed LLR vectors and aggregated into tap-level LLRs for finer local adjustments, thereby further suppressing error propagation. , M is used as an online feature input to the Feature-wise Linear Modulation (FiLM) parameter generation part to obtain the scaling and bias parameters:

[0055] It performs linear modulation on the intermediate output of the main network ATC-Net, incorporating external condition information. For the ATC-Net network, the final output... The process can be represented as:

[0056] Application weight Perform soft truncation:

[0057] Finally Transformed to the channel frequency domain response via DFT This mechanism adaptively and effectively suppresses spurious paths (false multipaths introduced by incorrect decisions) and long-tail energy (noise propagation), reducing the error propagation rate.

[0058] Step 5: Feed the ATC-Net output back to the data assistance module, repeat the equalization and reconstruction steps 3 and 4, and input it into the ATC-Net for truncation until convergence. The convergence condition can adopt some classic LLR-based stopping criteria in iterative decoding (see Shao, Rose Y., Shu Lin, and Marc PC Fossorier. "Two simple stopping criteria for turbo decoding." IEEE Transactions on Communications 47.8 (1999): 1117-1120.).

[0059] Step 6: If If not, proceed to step 1; otherwise, proceed to steps 2-5. As input for the next prediction.

[0060] Timing diagram as follows Figure 2 The diagram illustrates the processing flow for a complete transmission cycle (e.g., 10 transmission blocks): initial pilot stage (t=0-2), continuous DD stage (t=3-8), and pilot correction stage (t=9).

[0061] ATC-Net, a lightweight neural network for data-aided estimation with adaptive truncation, such as... Figure 3 As shown, at the input end, the temporal impulse response vector obtained by IFFT from the data-assisted channel estimation result is received. This vector is a complex sequence of length N. To facilitate network processing, it is split into amplitude and phase components and concatenated to form an input feature tensor of size (N,2). Simultaneously, global reliability features are extracted, including the absolute value of the average log-likelihood ratio of the current block. The signal-to-noise ratio (SNR) estimate and the interval length M from the previous pilot block are concatenated at input and mapped to a high-dimensional space through a fully connected layer. The entire network employs a CNN network structure incorporating Feature-wise Linear Modulation (FiLM) mechanism. The FiLM branch dynamically adjusts the main network through linear modulation. 1. The network first captures the local energy clusters and phase continuity features of CIR through the main processing path. Specifically, this path starts with the input tensor, passes through a one-dimensional convolutional layer (with a kernel length of 5, 2 input channels, and 16 output channels), and the obtained features are activated by GeLU. Then, it passes through another one-dimensional convolutional layer with a kernel length of 7 and the number of channels from 16 to 32 to further extract the temporal correlation across taps. After GeLU activation and layer normalization, a preliminary feature representation is generated.

[0062] 2. Simultaneously, the conditional processing path inputs three scalar reliability features (|LLR|, SNR, T) as 3-dimensional vectors into two fully connected layers, mapping the dimension from 3 to 64, and then from 64 to 32. These vectors undergo ReLU activation and layer normalization to obtain the conditional vector. This conditional vector is further mapped to a scaling factor γ and a bias parameter β for subsequent FiLM adjustment.

[0063] 3. After generating convolutional features in the main processing path, the network introduces a FiLM layer. This layer directly applies the γ and β output from the conditional processing path to the convolutional feature channels of the main path, performing channel-by-channel linear modulation (feature affine transformation: feature × γ + β). This dynamically adjusts the amplitude of the convolutional features, automatically enhancing suppression in low-reliability scenarios (e.g., lower |LLR| or SNR) (e.g., more conservative weight allocation to suppress spurious paths) and preserving more impulse response information in high-reliability scenarios. This FiLM embedding mechanism enhances the network's robustness to noise and error propagation, making the overall processing more adaptive.

[0064] 4. The FiLM-adjusted feature representation is generated by mapping the 32 channels to 1 using a one-dimensional convolutional layer, followed by Softmax activation to obtain a weight vector of length N. Physically, this weight vector assigns soft weights to each time delay tap, allowing for truncation and retention. Subsequently, the weight vector is applied to the time-domain impulse response for weighted processing, and then transformed back to the frequency-domain channel response using FFT for subsequent channel prediction and data detection.

[0065] The training process of this network includes the following aspects: 1. Data Preparation: Channel data preparation combines publicly available reference measurement data and its derived simulation forms, and generates the received signal according to the set modulation scheme and additive noise level. This allows us to obtain training samples with known real channels. The input-label pair is the data-assisted channel estimate obtained from link simulation within the DD channel estimation framework. The corresponding time-domain impulse response Reliability information obtained during the decoding process SNR and M, with labels representing the true values. .

[0066] 2. Loss Function: The impulse response is weighted by the output tap weight vector. By performing tap-by-tap weighting, we obtain The actual impulse response provided by the simulated channel model For comparison, mean squared error is used as the regression loss:

[0067] Ideal weights can also be generated based on the actual impulse response, and the generated weights and ideal weights can be measured using the binary cross-entropy (BCE) metric. Differences between them:

[0068] 3. Standard configuration of optimizer and hyperparameters: The optimizer uses the Adaptive Moment Estimation (Adam) algorithm with an initial learning rate of 10⁻³ and a decay of 0.9. The batch size is chosen to be 16-32, which ensures stable gradient estimation while adapting to conventional computing resources. The epochs are chosen to be 50-100, and an early stopping threshold is set based on the validation MSE.

[0069] In embodiments of the present invention, the overall structure of the adaptive CIR truncated neural network is designed to be as lightweight as possible. Taking a typical configuration as an example, the morphological branch uses two layers of one-dimensional convolutions, increasing the number of channels from 2 to 32; the conditional branch uses two layers of fully connected layers to map 3D global features to a 32-dimensional conditional vector; and the output contains only one mask generation branch. With this configuration, the total number of network parameters is approximately 8 × 10³, allowing for real-time operation on conventional embedded processors or software radio platforms.

[0070] LSTM network architecture for channel prediction, such as Figure 4 As shown, this network takes the channel estimate of a single transport block as input, uses the temporal sequence characteristics of symbols for prediction, and outputs the predicted channel response for the next transport block. The input layer is a fully connected layer that maps the 2D vector at each time step to the hidden dimension D, generating an input sequence of shape (N, D). This layer does not introduce memory but provides feature representation for subsequent LSTM. The hidden layer uses N layers of stacked LSTM units, each with a hidden dimension D, forming a memorized recurrent structure. The stacked design allows shallow layers to capture local sequence patterns (such as frequency correlations between adjacent symbols), while deeper layers retain long-term memory (such as time-varying channel trends) through the layer-by-layer propagation of hidden states. The sequence output of the last LSTM unit, with shape (N, D), encodes the memorized representation of the entire input sequence. The output layer is a fully connected layer that maps the hidden state of each time step of the last LSTM layer back to the 2D output.

[0071] The training process of this network includes the following aspects: 1. Data Preparation: The dataset is derived from simulated data calculated based on the VHF time-varying channel model. The input-label pair is a real matrix representing the current transport block channel mapping. The label is the true channel matrix mapping for the next transport block time. .

[0072] 2. Loss function: The mean squared error (MSE) is used, expressed as...

[0073] 3. Optimizer and Hyperparameter Configuration: The Adam algorithm is selected as the optimizer, with an initial learning rate of 10⁻³ and a decay rate of 0.9. The batch size is set to 32-64, the epochs to 50-100, and an early stopping threshold is set based on the validation MSE.

[0074] To evaluate the effectiveness of the proposed Adaptive CIR Truncation Mechanism (ACT-Net) in suppressing error propagation, the figure below shows the performance results of simulations conducted on a low VHF band single-carrier communication system. The channel model simulates typical sparse multipath propagation in a VHF environment, with a maximum delay spread of L=10 and slow time-varying characteristics (Doppler spread corresponds to vehicle movement at 50 km / h). Key simulation parameters include: block length N=256, sufficient cyclic prefix length to cover delay spread, QPSK modulation for data symbols, LS estimation for channel estimation, and ZF equalization for channel equalization. SNR varies in 2 dB steps within the range of 0 to 30 dB. Performance is quantified by the normalized mean square error (NMSE) of the frequency domain channel estimation, calculated using the following formula:

[0075] On average, calculations were performed across 100 independent implementations, each containing 1000 transport blocks to capture steady-state behavior under DD operations.

[0076] The baseline pure DD estimation (labeled DD) does not employ any truncation or suppression mechanism, directly using the frequency domain channel estimation result assisted by channel prediction data combined with LSTM. It is susceptible to pseudopaths induced by detection errors and error propagation. DD estimation under the DFT channel estimation method is labeled DD-DFT. The mechanism of this invention is labeled DD-ACT, employing a lightweight CNN-FiLM architecture (as described in the embodiments) to dynamically generate the weights for each tap. Figure 5 As shown, DD-ACT achieves stable estimation accuracy across the entire SNR range, particularly excelling in the low-to-medium SNR region. Notably, in the low SNR (0–10 dB) region, where DD errors dominate and propagate actively, the pure DD curve drops sharply from approximately 0 dB to around -8 dB; the DD-DFT gradually decreases from near -12 dB but remains above -14 dB (i.e., does not reach a lower NMSE); DD-ACT decreases slightly from -10 dB and flattens out, stabilizing between -10 and -11 dB, effectively suppressing spurious multipath and noise artifacts through conditionally aware adaptation. At high SNR (>20 dB), the three curves tend to converge, but the flatness of DD-ACT highlights its robustness in maintaining the sparse structure of the channel without excessive truncation.

[0077] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0078] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A deep learning DD channel estimation system combining an adaptive CIR truncation mechanism, characterized in that, include: The pilot estimation module is used to perform channel estimation based on the received pilot sequence and output pilot-aided channel estimation values. The deep learning prediction module is used to predict the channel state at the next moment based on the historical channel estimation sequence. The data-assisted channel estimation module equalizes and decodes data symbols based on the channel state predicted by the deep learning prediction module, and then reconstructs the data-assisted channel estimate. The adaptive CIR truncation module takes data-aided channel estimation as input and outputs truncated channel estimation values. The periodic pilot correction module is used to trigger pilot updates and correct the pilot-aided channel estimates.

2. The deep learning DD channel estimation system combining adaptive CIR truncation mechanism according to claim 1, characterized in that, Each module is connected through a feedback loop; the prediction output is fed into the data estimate, and the data estimate is truncated and fed back to the prediction, forming a closed loop.

3. The deep learning DD channel estimation system combining adaptive CIR truncation mechanism according to claim 1, characterized in that, The input to the pilot estimation module includes the pilot set P and the received sequence of the t-th transmission block. The output includes pilot-aided channel estimation. , where t=0 represents the initial estimation stage.

4. The deep learning DD channel estimation system combining adaptive CIR truncation mechanism according to claim 1, characterized in that, The input to the deep learning prediction module includes the historical channel estimate of the t-th transport block. The output includes a prediction of the channel state at the next time step. .

5. The deep learning DD channel estimation system combining adaptive CIR truncation mechanism according to claim 1, characterized in that, The input to the data-assisted channel estimation module includes the channel prediction for the (t+1)th transmission block. The output includes data-aided frequency domain channel estimation. Decoding bit log-likelihood ratio (LLR); utilizing The data symbols of the (t+1)th transport block are equalized and decoded to reconstruct the soft decision result. ,use As virtual pilots, more accurate data-assisted channel estimation can be calculated. .

6. The deep learning DD channel estimation system combining adaptive CIR truncation mechanism according to claim 1, characterized in that, The input to the adaptive CIR truncation module includes data-assisted channel frequency response estimation. The decoded bit log-likelihood ratio (LLR), signal-to-noise ratio (SNR), and current pilot spacing length (M) are calculated. The output includes a channel estimate after adaptive CIR truncation. The tap weights of each path are output by a lightweight neural network, and the CIR is adaptively truncated to filter out pseudo-paths and reduce noise.

7. The deep learning DD channel estimation system combining adaptive CIR truncation mechanism according to claim 1, characterized in that, The input to the periodic pilot correction module includes periodically inserted pilot symbols; every [period] Each transport block triggers a pilot update to correct the accumulated error under iterative channel prediction and data-assisted channel estimation.

8. A deep learning-based DD channel estimation method combining an adaptive CIR truncation mechanism, characterized in that, include: Step S1: Perform channel estimation based on the received pilot sequence and output the pilot-aided channel estimation value; Step S2: Predict the channel state at the next moment based on the historical channel estimation sequence, and output the prediction result; Step S3: Equalize and decode the data symbols based on the prediction results, and obtain the data-assisted channel estimate after reconstruction; Step S4: Take the data-assisted channel estimation as input and output the truncated channel estimation value; Step S5: Periodically trigger pilot updates to correct the channel estimation values ​​assisted by the pilots.

9. The deep learning DD channel estimation method combining adaptive CIR truncation mechanism according to claim 8, characterized in that, By utilizing the output tap weights / masks of the neural network ACT-Net, the multipath components of the time-domain impulse response are dynamically truncated.

10. The deep learning DD channel estimation method combining adaptive CIR truncation mechanism according to claim 8, characterized in that, The adaptively truncated channel estimate is input into the deep learning prediction model, which outputs the channel prediction result for the next time step, and is used for subsequent data detection and decoding.

Images