Channel estimation method and computer device of load management system

Abstract

The application relates to a channel estimation method of a load management system and a computer device, and belongs to the field of channel signal estimation. The application uses a channel estimation model composed of an SRGAN and a DnGAN in combination, when a signal is recovered by using an FBMC system, selected pilot signals are processed and used as the input of the channel estimation model, the high-resolution signal image is generated through the processing of the SRGAN, and the generated high-resolution signal image is processed through the DnGAN to generate a denoised high-resolution signal image.

Description

Technical Field

[0001] This invention relates to a channel estimation method and computer device for a load management system, belonging to the field of channel signal estimation. Background Technology

[0002] In response to changes in energy production and consumption, and to promote the upgrading of the power grid to an energy internet, the application of wireless mobile communication in power generation, transmission, distribution, substation, and consumption is booming. Examples include distributed wireless sensing and data acquisition devices, transmission line inspection drones, distribution automation intelligent terminals, substation wireless intelligent inspection robots, mobile operation terminals for maintenance and repair, and smart meter data acquisition terminals. Due to its resistance to frequency-selective fading and strong anti-interference capabilities, Orthogonal Frequency Division Multiplexing (OFDM) technology is widely used in mobile communication, wireless local area networks, and power line communication.

[0003] In the context of the 6G era of the Internet of Everything, Filter Bank Multi Carrier (FBMC) technology, as one of the candidate physical layer waveforms, can greatly reduce the sidelobe leakage and low bandwidth utilization of OFDM. It can meet the enhanced mobile broadband requirements of intelligent inspection robots and remote augmented reality inspection systems, the massive terminal communication requirements of distributed photovoltaic power stations and distributed sensors, and the low latency communication requirements of intelligent fusion terminals and distributed power distribution automation terminals.

[0004] FBMC employs a set of parallel filter banks and allows for custom filter design, effectively truncating sidelobes in the frequency domain and significantly reducing out-of-band power leakage. Unlike OFDM with a cyclic prefix, FBMC uses Offset Quadrature Amplitude Modulation (OQAM) to reduce inter-symbol interference (ISI) and inter-carrier interference (ICI), and achieves higher bandwidth utilization due to the absence of a cyclic prefix. However, OQAM modulation disrupts inter-carrier orthogonality, and each symbol is susceptible to ISI and ICI from surrounding elements, commonly referred to as imaginary interference, which affects the recovery of the received signal. Summary of the Invention

[0005] The purpose of this invention is to provide a channel estimation method and computer device for a load management system, in order to solve the problem of inaccurate signal recovery in FBMC systems due to the presence of imaginary interference.

[0006] To achieve the above objectives, the present invention includes:

[0007] A channel estimation method for a load management system is characterized by employing an FBMC system to segment the received signal, extracting the pilot response portion, and preprocessing the pilot response portion to obtain a signal matrix; extracting features from the information in the signal matrix to obtain feature maps of signal strength and phase information, and reshaping and adjusting the obtained feature maps to generate input data for a channel estimation model; processing the input data using the constructed channel estimation model to obtain the channel estimate of the FBMC system, wherein the channel estimation model includes SRGAN and DnGAN, SRGAN extracts channel features from the input data to generate a high-resolution signal image, and DnGAN denoises the extracted channel features to generate a denoised high-resolution signal image; the training data used during the training of the channel estimation model includes selected original pilot signals and the corresponding pilot response portion after passing through the FBMC system.

[0008] Furthermore, both SRGAN and DnGAN are equipped with the MLAM framework, which includes a spatial attention mechanism module and a channel attention mechanism module.

[0009] Furthermore, the channel attention mechanism module includes pooling layers, a multilayer perceptron, activation functions, and channel attention. The channel attention mechanism module first compresses the input feature map globally using max pooling and average pooling layers to obtain a one-dimensional vector. This one-dimensional vector is then processed by the multilayer perceptron, and an activation function is used to obtain channel attention weights. Finally, the channel attention weights are multiplied by the original feature map to obtain the feature map processed by the channel attention mechanism module. The spatial attention mechanism module includes pooling layers, convolutional layers, activation functions, and spatial attention. The channel attention mechanism module first processes the feature map obtained from the channel attention mechanism module using max pooling and average pooling layers to obtain the spatial information of the feature map. Then, it passes through convolutional layers and uses an activation function to obtain spatial attention weights. Finally, the spatial attention weights are multiplied by the feature map to obtain the feature map processed by the spatial attention mechanism module.

[0010] Furthermore, SRGAN includes a generator network and a discriminator network. After the first activation function in the SRGAN generator network and the first activation function in the SRGAN discriminator network, an MLAM framework is introduced. Through the spatial attention mechanism module and the channel attention mechanism module in the MLAM framework, SRGAN generates high-resolution signal images. DnGAN also includes a generator network and a discriminator network. After the first activation function in the DnGAN generator network and the first activation function in the DnGAN discriminator network, an MLAM framework is introduced. Through the spatial attention mechanism and the channel attention mechanism in the MLAM framework, DnGAN removes noise from the high-resolution signal images generated by SRGAN.

[0011] Furthermore, the preprocessing in step S1 includes: performing a discrete Fourier transform on the pilot response to obtain a complex matrix; then separating the real and imaginary parts of the complex matrix; combining the real and imaginary parts to form a real matrix; and then normalizing the real matrix to obtain a signal matrix of the same dimension.

[0012] Furthermore, in step S2, an interpolation network is used to extract features from the signal matrix to obtain a feature map after concatenating the signal strength and phase information.

[0013] Furthermore, step S2 also includes reshaping the dimension of the adjusted feature map using a reshaping layer to generate a feature map with the same dimension as the input data of the channel estimation model.

[0014] Furthermore, during training, the channel estimation model employs an optimization algorithm to enable gradient descent and continuously adjusts the network weights and biases through backpropagation until the loss function meets the requirements.

[0015] Furthermore, the loss function is the mean squared error loss function.

[0016] A computer device includes a processor for executing a computer program to implement the steps of a channel estimation method for a load management system.

[0017] The beneficial effects of this invention are as follows: This invention is an improved invention. This invention uses a channel estimation model composed of SRGAN and DnGAN. When using the FBMC system to recover the signal, the selected pilot signal is processed and used as the input of the channel estimation model. After processing by SRGAN, a high-resolution signal image is generated. The generated high-resolution signal image is then processed by DnGAN to generate a denoised high-resolution signal image. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the FBMC communication system;

[0019] Figure 2 This is a schematic diagram of the SRGAN model used in this invention;

[0020] Figure 3 This is a schematic diagram of the DnGAN model used in this invention;

[0021] Figure 4 This is a schematic diagram of the MLAM framework;

[0022] Figure 5 This is a flowchart illustrating the channel estimation method of the FBMC system of the present invention;

[0023] Figure 6 This is a schematic diagram showing the ideal minimum mean square error, channel network, and mean square error performance of the proposed model under different signal-to-noise ratios when the pilot number is 8.

[0024] Figure 7 This is a schematic diagram showing the ideal minimum mean square error, channel network, and mean square error results of the proposed model under different signal-to-noise ratios when the pilot number is 16.

[0025] Figure 8 This is a schematic diagram showing the ideal minimum mean square error, channel network, and mean square error results of the proposed model when the signal-to-noise ratio is 10dB and the number of pilots is different. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0027] The concept of this invention is to employ neural network models based on SRGAN (Super-Resolution Generation Adversarial Networks) and DnGAN (Denoising Generation Adversarial Networks), and to establish an Adaptive Denoising and Super-Resolution GAN with Multi-layered Attention Mechanism (DSMGAN) by introducing the MLAM (Multi-layered attention mechanism) framework. Considering that the dimension of the feature-extracted signal data may conflict with the dimension of the DSMGAN input data, it is necessary to reshape and adjust the feature-extracted signal data, and then use the reshaped and adjusted signal data as the input data of DSMGAN.

[0028] Method Implementation Examples:

[0029] This embodiment provides a channel estimation method for a load management system, including the following steps:

[0030] S1. The received signal is segmented using the FBMC system to extract the pilot response portion, and the pilot response portion is preprocessed to obtain the signal matrix.

[0031] S2. Extract features from the signal matrix obtained in S1 to obtain feature maps of the signal's phase and amplitude, and reshape and adjust the obtained feature maps to generate input data of a set size.

[0032] S3. The input data is processed using the constructed channel estimation model to obtain the channel estimate of the FBMC system.

[0033] FBMC communication system such as Figure 1 As shown:

[0034] In an FBMC communication system, the transmitted signal at the transmitter is x(t). Since there are multiple users, each equipped with a single antenna, N parallel input complex signals are transmitted through N sub-channel filters and then superimposed. The i-th complex signal transmitted on the n-th subcarrier can be expressed as:

[0035]

[0036] Where α and β represent the real and imaginary parts of the I-th complex signal on the n-th subcarrier, respectively, and the real and imaginary parts of the symbol are interleaved in the time domain. Where T is the FBMC symbol period. The nth subcarrier signal, after filtering, can be expressed as:

[0037]

[0038] Where h(t) represents the impulse response of the prototype filter. Then, the data is modulated by a subcarrier modulator. Considering that the FBMC signal contains M symbols, it can be expressed as:

[0039]

[0040] Where N represents the number of subcarriers, f c M represents the carrier frequency, and M represents the number of symbols. This represents the subcarrier spacing, with bandwidth B = f·N, x m,n t represents the communication information to be transmitted on the nth subcarrier of the mth symbol. b Indicates the symbol period, a n Let represent the complex weights transmitted on the nth subcarrier, then the transmit power of the nth subcarrier is...

[0041] In an FBMC system, the bit sequence transmitted at the transmitter undergoes radix conversion, pilot insertion, serial-to-parallel conversion, modulation, addition of a cyclic prefix, and parallel-to-serial conversion. In real-world channels, perfect orthogonality does not exist, leading to inter-symbol and inter-carrier interference, as well as the effects of fading channels and noise. Therefore, the received signal y(t) in the time domain of an FBMC system is represented as:

[0042] y(t) = x(t) + w(t)

[0043] in, It is Gaussian white noise.

[0044] In a preferred embodiment, the signal y(t) includes a pilot response portion y h (t).

[0045] Specifically, S1 in this embodiment includes:

[0046] S11 and FBMC extract the pilot response based on the known original pilot y(t);

[0047] S12. Perform Discrete Fourier Transform processing, extract the real and imaginary parts, and normalize the pilot response part from step S11.

[0048] Preferably, the S12 process includes:

[0049] S121. Discrete Fourier Transform Processing: Transform the pilot response from the time domain to the frequency domain to obtain a complex matrix;

[0050] S122. Extracting the real and imaginary parts: Separate the real and imaginary parts of the complex matrix to obtain two real matrices;

[0051] S123. Normalization: Normalize each of the two real matrices. The normalization calculation expression can be written as:

[0052]

[0053] Where μ(X) and σ(X) are the mean and variance of the matrix, and 1e-15 is a constant used for numerical stability;

[0054] Specifically, S2 in this embodiment is:

[0055] The signal matrix obtained by S1 is processed by an interpolation network to extract signal intensity and phase information. The signal intensity and phase information are then concatenated to form a feature map with more obvious features. The feature map is then reshaped and adjusted by a reshaping layer so that it can be used as input to DSMGAN.

[0056] Specifically, S3 in this embodiment includes:

[0057] S31. Construct the channel feature extraction model - SRGAN. The feature extraction module of the SRGAN network includes: convolutional layer, residual block, upsampling layer, and activation function.

[0058] S32. Constructing the channel denoising model - DnGAN. The denoising module of the DnGAN network includes: convolutional layer, residual block, discriminator, and activation function.

[0059] S33. Introduce the MLAM framework into SRGAN and DnGAN networks. The MLAM framework includes a spatial attention mechanism module and a channel attention mechanism module.

[0060] S34. Construct the initial channel estimation model, namely the initial DSMGAN model. The initial DSMGAN model consists of SRGAN and DnGAN introduced into the MLAM framework.

[0061] S35. Train the initial DSMGAN model to obtain the best-performing neural network model, i.e., the constructed channel estimation model. Then deploy it to the FBMC system. Then input the reshaped and adjusted signal data from step S2 into the constructed channel estimation model to finally obtain the estimated channel signal.

[0062] Specifically, S34 in this embodiment is:

[0063] First, the S31 model is constructed. Then, the MLAM framework is introduced after the first activation function of the generator and discriminator networks in the model, as follows: Figure 2As shown, in the SRGAN generator network incorporating the MLAM framework, the input is first passed through a 9×9×64 convolutional layer to the subsequent PReLU activation function to generate a feature map. Then, the input feature map is compressed globally by max pooling and average pooling layers to obtain a one-dimensional vector. This one-dimensional vector is then processed by a multilayer perceptron, and an activation function is used to obtain channel attention weights. These channel attention weights are then multiplied with the original feature map to obtain the feature map processed by the channel attention mechanism module. The feature map obtained by the channel attention mechanism module is then processed to obtain the spatial information of the feature map. It is then passed through a convolutional layer and an activation function to obtain spatial attention weights. These spatial attention weights are then multiplied with the feature map to obtain the feature map processed by the spatial attention mechanism module. In the main body, each convolutional layer is followed by a batch renormalization (BRN) layer, followed by a PReLU activation function, and then another convolutional layer and a BRN. Its main structure is: convolutional layer + BRN + PReLU activation function + convolutional layer + BRN, where the convolutional layer is 3×3×64 and uses skip connections to form residual blocks. Skip connections allow the input and output to be added together. In the SRGAN discriminator network, "convolutional layer + BRN + LeakyReLU activation function" is adopted, where the size of the convolutional layer is still 3×3×64. The final output layer uses the Sigmoid activation function for calculation, which is used to output a higher resolution image.

[0064] Similarly, in the S32 model, the MLAM framework is introduced after the first activation function of the generator and discriminator networks, such as... Figure 3 As shown, in the DnCNN generator network incorporating the MLAM framework, the input is first passed through a 3×3×64 convolutional layer to the subsequent PReLU activation function to generate a feature map. Then, the MLAM framework processes the data to obtain a more prominent feature map. In the main body, each convolutional layer is followed by a Batch Renormalization (BRN) layer, and then another PReLU activation function. Its main structure is: convolutional layer + BRN + PReLU activation function, where the convolutional layer is 3×3×64 and uses skip connections to form residual blocks. Skip connections allow the input and output to be added together. In the SRGAN discriminator network, a "convolutional layer + BRN + LeakyReLU activation function" approach is adopted, where the convolutional layer size remains 3×3×64. The final output layer uses the Sigmoid activation function for computation, improving noise reduction capabilities.

[0065] Figure 4This is a schematic diagram of the MLAM framework, which includes a channel attention mechanism module and a spatial attention mechanism module. The channel attention mechanism module includes pooling layers, a multilayer perceptron, activation functions, and channel attention weights. The spatial attention mechanism module includes pooling layers, convolutional layers, activation functions, and spatial attention weights. Then, the SRGAN and DnGAN network models introduced into the MLAM framework are combined, as shown below. Figure 5 The overall network model architecture of the combined DSMGAN network is shown below. The interpolation network is used for feature extraction, and then the extracted features are concatenated to form a larger feature map. The role of the reshaping layer is to adjust the shape and dimension of the feature map after concatenation by the interpolation network so that it can be used as input to the SRGAN network.

[0066] Specifically, an MLAM framework is added to the generator and discriminator networks of SRGAN to extract key content based on the features of the input image and to guide the generator network to better reconstruct high-resolution images through attention weights.

[0067] Specifically, the MLAM framework is introduced into the generator and discriminator networks of DnGAN. This allows MLAM to be applied to the input feature maps of each deconvolutional layer, helping the model better focus on the details and edges of the reconstructed image. By introducing the MLAM framework during the reconstruction process, noise can be removed more effectively, resulting in a clearer image.

[0068] Specifically, the modules introduced into the MLAM framework are different and have different uses. In SRGAN, the MLAM framework is introduced to enhance the resolution effect, while in DnGAN, the MLAM framework is introduced to improve the noise reduction capability. Therefore, the weight allocation of spatial attention and channel attention will be different in the two frameworks.

[0069] Specifically, the spatial attention module helps the model focus on different regions in an image. Weight allocation can be calculated based on the correlation matrix of the input feature map, which reflects the degree of correlation between different spatial locations; higher correlation indicates a significant spatial relationship between these locations. The correlation matrix can be used to normalize the weights to ensure that the sum of the weights is 1. The specific weight allocation ratio can be determined experimentally; for example, linear or non-linear mappings can be attempted based on the maximum and minimum values ​​of the correlation matrix.

[0070] Specifically, the channel attention module helps the model focus on different channels in the input feature map. Weight assignment can be calculated based on global statistics. A common approach is to use global average pooling to extract global statistics from the feature map, and then use a fully connected layer to generate the channel attention weights. Weight assignment can be adjusted by introducing appropriate non-linear activation functions into the fully connected layer. The specific weight allocation ratio can be determined experimentally; for example, linear or non-linear mappings based on the maximum and minimum values ​​of global statistics can be tried.

[0071] Specifically, S35 in this embodiment is:

[0072] The key network parameters of the initial DSMGAN model are set, including batch size, learning rate, and number of training epochs. The network is trained using the training set. An adaptive moment estimation optimization algorithm is used to achieve gradient descent and the network weights and biases are continuously adjusted through backpropagation. The network model is then adjusted by calculating the loss function (mean squared error loss function). If the loss function value does not meet the requirements (the loss function value is not small enough), the network model parameters are adjusted and training continues until the loss function is small enough to meet the requirements. The network model is then tested and optimized using the test set. The network model parameters are further adjusted to obtain the optimal neural network model. The training set and the test set are pre-collected datasets that have undergone feature extraction and reshaping and adjustment. The training set and the test set are divided into datasets in an 8:2 ratio.

[0073] Specifically, when the FBMC system receives a signal from the user, it segments and preprocesses the signal to obtain a signal matrix. Figure 5 U1~U k The obtained signal matrix is ​​then input into an interpolation network, which extracts the signal strength and phase information from the signal matrix. Figure 5 In Then the features are concatenated to form a feature map with more distinct characteristics. Figure 5 The feature map (h) is reshaped and adjusted by the reshaping layer and can be used as input data for the network model. Figure 5 In SRGAN and DnGAN, which incorporate the MLAM framework, are used to generate estimated signals.

[0074] In this embodiment, to demonstrate the beneficial effects of the present invention, a comparative experiment was conducted between the proposed method and the ideal minimum mean square error value and the channel network model. (See attached diagram). Figure 6 , Figure 7 , Figure 8As shown in the diagram. Here, Pilots represents the number of pilots, snr represents the signal-to-noise ratio, MSE represents the mean squared error, blue represents the ideal minimum mean squared error, green represents the ChannelNet based on Super-Resolution Convolutional Neural Network (SRCNN) and Denoising Convolutional Neural Network (DnCNN), and pink represents the network model provided in this invention. Figure 6 and Figure 7 The mean square error (MSE) results for the three models are shown for pilot frequencies of 8 and 16, under different signal-to-noise ratios (SNRs). Simulation results show that the MSE values ​​of all three models decrease with increasing SNR. ChannelNet tends to plateau after an SNR greater than 15 dB, while the model proposed in this embodiment tends to plateau after an SNR greater than 25 dB. It can also be seen that the MSE value of the model proposed in this embodiment is lower than that of the ChannelNet model and is relatively close to the ideal minimum MSE value. The experimental results when the SNR is fixed at 10 dB are shown below. Figure 8 As shown, the mean square error decreases with the increase of the number of pilots. The results for different pilot numbers show that the results of this embodiment are better than those of ChannelNet.

[0075] Device Example:

[0076] This embodiment provides a computer device including a processor and a computer program running on the processor. When estimating a received signal, the processor program executes a channel estimation method for a load management system as described in the method embodiment. Since the method is already sufficiently clear, it will not be described again here.

[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A channel estimation method for a load management system, characterized in that, The process includes the following steps: S1. The received signal is segmented using an FBMC system to extract the pilot response portion, and the pilot response portion is preprocessed to obtain a signal matrix; S2. Feature extraction is performed on the information in the signal matrix to obtain feature maps of signal strength and phase information, and the obtained feature maps are reshaped and adjusted to generate input data for the channel estimation model; S3. The input data is processed using the constructed channel estimation model to obtain the channel estimate of the FBMC system. The channel estimation model includes SRGAN and DnGAN. The SRGAN extracts channel features from the input data to generate a high-resolution signal image, and the DnGAN denoises the extracted channel features to generate a denoised high-resolution signal image. The training data used during the training of the channel estimation model includes selected original pilot signals and the corresponding pilot response portion after passing through the FBMC system.

2. The channel estimation method for a load management system according to claim 1, characterized in that, Both SRGAN and DnGAN are equipped with MLAM frameworks, which include spatial attention mechanism modules and channel attention mechanism modules.

3. The channel estimation method for a load management system according to claim 2, characterized in that, The channel attention mechanism module includes a pooling layer, a multilayer perceptron, an activation function, and channel attention. The channel attention mechanism module first compresses the input feature map globally through a max pooling layer and an average pooling layer to obtain a one-dimensional vector. Then, the one-dimensional vector is processed by the multilayer perceptron, and then the activation function is used to obtain the channel attention weights. Finally, the channel attention weights are multiplied by the original feature map to obtain the feature map after processing by the channel attention mechanism module. The spatial attention mechanism module includes a pooling layer, a convolutional layer, an activation function, and spatial attention. The channel attention mechanism module first processes the feature map obtained by the channel attention mechanism module through a max pooling layer and an average pooling layer to obtain the spatial information of the feature map. Then, it passes through a convolutional layer and uses an activation function to obtain the spatial attention weights. Finally, the spatial attention weights are multiplied by the feature map to obtain the feature map after processing by the spatial attention mechanism module.

4. The channel estimation method for a load management system according to claim 2, characterized in that, The SRGAN includes a generator network and a discriminator network. An MLAM framework is introduced after the first activation function of the SRGAN generator network and the first activation function of the SRGAN discriminator network. The spatial attention and channel attention mechanisms within the MLAM framework enable the SRGAN to generate high-resolution signal images. The DnGAN also includes a generator network and a discriminator network. An MLAM framework is introduced after the first activation function of the DnGAN generator network and the first activation function of the DnGAN discriminator network. The spatial attention and channel attention mechanisms within the MLAM framework enable the DnGAN to remove noise from the high-resolution signal images generated by the SRGAN.

5. The channel estimation method for a load management system according to claim 1, characterized in that, The preprocessing described in step S1 includes: performing a discrete Fourier transform on the pilot response to obtain a complex matrix; then separating the real and imaginary parts of the complex matrix; combining the real and imaginary parts to form a real matrix; and then normalizing the real matrix to obtain a signal matrix of the same dimension.

6. The channel estimation method for a load management system according to claim 1, characterized in that, In step S2, an interpolation network is used to extract features from the signal matrix to obtain a feature map after concatenating the signal strength and phase information.

7. The channel estimation method for a load management system according to claim 1, characterized in that, Step S2 further includes reshaping the dimension of the adjusted feature map using a reshaping layer to generate a feature map with the same dimension as the input data of the channel estimation model.

8. The channel estimation method for a load management system according to claim 1, characterized in that, The channel estimation model employs an optimization algorithm during training to achieve gradient descent and continuously adjusts the network weights and biases through backpropagation until the loss function meets the requirements.

9. The channel estimation method for a load management system according to claim 8, characterized in that, The loss function is the mean squared error loss function.

10. A computer device, comprising a processor, characterized in that, The processor is used to execute a computer program to implement the steps of the channel estimation method of the load management system according to any one of claims 1 to 9.

Images