Feature enhancement multi-task neural network training and complex target parameter prediction method and system

Abstract

The application discloses a feature enhancement multi-task neural network training and complex target parameter prediction method and system, and belongs to the technical field of artificial intelligence neural networks. The method acquires limited feedback observation samples and evaluation data, quantizes, normalizes and preselects the length of the observation samples, and constructs an input matrix; the input matrix is input into a feature enhancement multi-task neural network, fusion features are obtained through convolution feature extraction, residual enhancement, channel attention weighting and local feature and global statistical feature fusion; a parameter index and a prediction matrix parameter are output by using a classification branch and a regression branch, an unsupervised loss function is constructed according to the output result and the evaluation data, and network parameters are updated; the trained network is used to infer the limited feedback observation samples, and target parameters are output. The application does not need optimal parameter matrix supervision labels, can reduce data labeling cost and feedback overhead, and improves parameter prediction accuracy under limited feedback conditions.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and machine learning, specifically to a method and system for feature-enhanced multi-task neural network training and complex target parameter prediction. The method is applicable to feature extraction, parameter learning, performance optimization, and online inference under conditions of limited observation data. Background Technology

[0002] With the development of artificial intelligence and machine learning technologies, neural networks have been widely applied to tasks such as complex parameter prediction, nonlinear mapping, and data-driven optimization. For application scenarios with high-dimensional input data and complex relationships between variables, traditional methods based on manual feature design or fixed mathematical models often struggle to fully extract local features and global correlation information from the data, thus limiting prediction accuracy and adaptability.

[0003] Existing neural network parameter prediction methods typically rely on a large number of training samples labeled with optimal parameters, and complete network training through supervised learning. Obtaining the optimal parameters often requires additional iterative optimization, matrix factorization, or exhaustive search, which is computationally complex and has high data generation costs. When the length of the input observation data is limited, the feedback information contains quantization errors, or the sample distribution changes, the prediction stability of existing networks is also prone to decline.

[0004] Furthermore, some existing methods primarily employ a single feature extraction structure or a single output task, making it difficult to simultaneously handle discrete parameter indices and continuous matrix parameters. During training, they typically rely solely on parameter label errors for optimization, failing to fully utilize the performance evaluation information of the predicted parameters in real-world tasks. Therefore, it is necessary to propose a method and system capable of extracting multi-level features from limited feedback data and achieving target parameter prediction through collaborative prediction of classification and regression, unsupervised performance optimization, and online inference. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for feature-enhanced multi-task neural network training and complex target parameter prediction, in order to solve the problems of existing parameter prediction methods that rely heavily on complete observation data and manual supervision labels, have insufficient effective feature extraction from limited feedback data, and have difficulty in meeting performance constraints in the prediction results.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A limited number of feedback observation samples and corresponding performance evaluation data are obtained; the limited feedback observation samples are quantized, normalized, and have a preset length selected to construct an input matrix; the input matrix is ​​input into a feature-enhanced multi-task neural network, and fused features are obtained through convolutional feature extraction, residual enhancement feature extraction, channel attention weighting, and fusion of local and global statistical features; the fused features are input into a classification branch and a regression branch respectively, the classification branch outputs the prediction probability corresponding to each candidate discrete parameter index, and the regression branch outputs the real and imaginary parameters of the complex prediction matrix; candidate target parameters are generated according to each candidate discrete parameter index, and combined with the complex prediction matrix... The system calculates downstream task performance metrics corresponding to each candidate target parameter based on the complex prediction matrix and performance evaluation data. It then constructs a probability-weighted performance loss based on the downstream task performance metrics and the predicted probabilities output by the classification branch, a prediction matrix performance loss based on the complex prediction matrix and performance evaluation data, determines the performance ranking results based on the downstream task performance metrics, and constructs a classification branch training loss. The system combines the probability-weighted performance loss, prediction matrix performance loss, classification branch training loss, and constraint penalty terms according to preset weights to obtain an unsupervised loss function. This function is then updated using backpropagation to update the network parameters of the feature-enhanced multi-task neural network. Finally, the trained feature-enhanced multi-task neural network is used to perform forward inference on new finite feedback observation samples, outputting the corresponding complex target parameters.

[0007] Compared with existing technologies, this invention improves the feature representation capability of limited feedback data through convolutional feature extraction, residual enhancement, channel attention weighting, and fusion of local and global statistical features; it enhances the joint prediction capability of multi-type target parameters by collaboratively outputting parameter indices and prediction matrix parameters through classification and regression branches; it reduces the dependence on the optimal parameter matrix supervision label by using an unsupervised loss function based on performance indicators, thereby reducing the cost of training data construction; and it reduces the online computational complexity by enabling the trained network to perform fast inference on new limited feedback observation samples. Attached Figure Description

[0008] Figure 1 This is a diagram showing the relationship between the system modules of the present invention.

[0009] Figure 2 This is a flowchart of the method of the present invention.

[0010] Figure 3 This is a diagram of the feature-enhanced multi-task neural network structure of the present invention.

[0011] Figure 4 This is a comparison chart of the performance indicators of the present invention under different signal-to-noise ratio conditions in a downstream application of wireless communication. Detailed Implementation

[0012] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Equivalent substitutions or modifications can be made by those skilled in the art without departing from the concept of the present invention, and all such substitutions or modifications should fall within the scope of protection of the present invention.

[0013] Example 1: System Structure and Training Framework. The feature-enhanced multi-task neural network training and target parameter prediction system of this embodiment includes an input data construction module, a feature enhancement module, a multi-task prediction module, a target parameter generation module, an unsupervised training module, and an online inference module. The input data construction module quantizes, normalizes, and selects a preset length for limited feedback observation samples to construct an input matrix. The feature enhancement module obtains fused features through convolutional feature extraction, residual enhancement, channel attention weighting, and the fusion of local and global statistical features. The multi-task prediction module outputs parameter indices and the real and imaginary parameters of the prediction matrix through classification and regression branches, respectively. The target parameter generation module generates target parameters based on the above outputs and calculates performance indicators using evaluation data. The unsupervised training module constructs an unsupervised loss function, updates network parameters through gradient calculation and backpropagation, and obtains the trained neural network. The online inference module performs forward inference on preprocessed new observation samples and outputs the corresponding target parameters. The system module relationships described in this embodiment can be combined with... Figure 1 understand.

[0014] Example 2: Neural Network Training and Target Parameter Prediction Method. This example first acquires limited feedback observation samples and corresponding evaluation data. The observation samples are quantized, normalized, and have a preset length. An input matrix is ​​constructed according to a preset arrangement. Then, the input matrix is ​​fed into a feature-enhanced multi-task neural network. Fusion features are obtained through convolutional feature extraction, residual enhancement, channel attention weighting, and the fusion of local and global statistical features. These fusion features are then input into the classification and regression branches respectively. The classification branch outputs parameter indices, and the regression branch outputs the real and imaginary parameters of the prediction matrix. Target parameters are generated based on the parameter indices and the prediction matrix. Performance indicators are calculated using the evaluation data. An unsupervised loss function is constructed using the weighted sum of the negative values ​​of the performance indicators and constraint penalty terms. The gradient of the unsupervised loss function relative to the network parameters is calculated. The network parameters are updated through backpropagation until a preset training termination condition is met. After training, the same preprocessing as in the training phase is performed on new limited feedback observation samples. The obtained input matrix is ​​then fed into the trained neural network for forward inference, outputting the corresponding target parameters. The method flow described in this example can be combined with… Figure 2 understand.

[0015] Example 3: Feature Enhancement Multi-Task Neural Network. The input matrix is ​​input into the A²FP-Net feature enhancement multi-task neural network. First, initial features are obtained through convolutional mapping, batch normalization, and nonlinear activation in the first convolutional layer. Then, the initial features are processed by the residual enhancement feature extraction module, and the input features are added to the residual features using cross-layer identity connections to obtain residual enhanced features. These residual enhanced features are further processed through the second and third convolutional layers to extract deeper local features. Simultaneously, channel weights are generated through global average pooling, fully connected mapping, and activation functions. These channel weights are then used to adaptively weight different feature channels.

[0016] Global pooling is performed on the channel-weighted features to obtain global statistical features. These global statistical features are then concatenated with deep local features to obtain fused features. These fused features serve as input to both the classification and regression branches. By using residual enhancement, channel attention weighting, and the fusion of local and global statistical features, the network's ability to extract effective features and correlation information from limited feedback data is improved. The network structure described in this embodiment can be combined with... Figure 3 understand.

[0017] Example 4: Multi-task prediction and unsupervised training. In this example, the fused features obtained in Example 3 are input into the classification branch and the regression branch, respectively. The classification branch outputs the predicted probabilities corresponding to each candidate discrete parameter index through a fully connected mapping and a classification activation function. The regression branch outputs the real and imaginary parameters of the complex prediction matrix through a fully connected mapping, and the two are combined to form a complex prediction matrix. Candidate target parameters are generated for each candidate discrete parameter index, and the downstream task performance index corresponding to each candidate target parameter is calculated by combining the complex prediction matrix and the performance evaluation data corresponding to the limited feedback observation samples.

[0018] Based on the downstream task performance metrics corresponding to each candidate target parameter and the predicted probabilities output by the classification branch, a probability-weighted performance loss is constructed; a prediction matrix performance loss is constructed based on the complex prediction matrix and performance evaluation data; the performance ranking result is determined based on the downstream task performance metrics corresponding to each candidate target parameter, and a classification branch training loss is constructed based on the performance ranking result; the probability-weighted performance loss, prediction matrix performance loss, classification branch training loss, and constraint penalty term used to measure the degree of satisfaction of target parameter constraints are combined according to preset weights to obtain an unsupervised loss function. During training, the gradient of the unsupervised loss function with respect to the network parameters in the convolutional layer, residual enhancement structure, channel attention structure, feature fusion structure, classification branch, and regression branch is calculated, and the network parameters are updated through backpropagation until the loss function converges, the preset number of training rounds is reached, or the performance metrics meet the preset conditions. The training process does not require obtaining the optimal complex parameter matrix as a supervision label in advance; the training objective of the classification branch is determined by the performance evaluation results of the candidate target parameters.

[0019] Example 5: Exemplary Application of Target Parameter Prediction. This example is an exemplary downstream application of the feature-enhanced multi-task neural network prediction framework described in this invention in the task of generating precoding parameters for wireless communication, and does not constitute a limitation on the application scenario or target parameter type of this invention. Received signal strength indicators corresponding to different training signals are used as finite feedback observation samples, and the corresponding channel samples are used as evaluation data. The measured values ​​are quantized, normalized, and pre-selected for a preset length, and an input matrix is ​​constructed according to the user's order. The input matrix is ​​input into the trained feature-enhanced multi-task neural network, where the classification branch outputs parameter indices, and the regression branch outputs the real and imaginary parameters of the prediction matrix. The channel samples are used to calculate performance indicators during the training phase and are not used as supervision labels for the optimal precoding matrix.

[0020] The analog precoding matrix A is obtained by querying the predefined codebook using the parameter index output by the classification branch. The real and imaginary parameters output by the regression branch are combined to obtain the all-digital precoding prediction matrix U. The all-digital precoding prediction matrix U is then mapped to a digital baseband precoding matrix using the Moore-Penrose pseudoinverse of the analog precoding matrix A. :

[0021]

[0022] The analog precoding matrix A and the digital baseband precoding matrix Multiplying them together, we obtain the unnormalized combined precoding matrix W0:

[0023]

[0024] The combined precoding matrix W0 is power normalized according to the preset transmit power P to obtain the final combined precoding matrix W:

[0025]

[0026] in, This represents the Moore-Penrose pseudoinverse of the analog precoding matrix A, where P represents the preset transmit power. The Frobenius norm is represented. A performance index is calculated using the combined precoding matrix W and evaluation data, and this performance index is used for unsupervised network training. During the online inference phase, only new, finitely observed samples require the same preprocessing. The combined precoding matrix W is generated using the trained neural network's output parameter index and prediction matrix, following the matrix mapping and power normalization process described above. This embodiment is only used to illustrate one specific application of the target parameter prediction results.

[0027] Example 6: Performance Verification. To verify the effectiveness of the limited feedback data processing, feature enhancement structure, and unsupervised training mechanism, the A²FP-Net of this invention is compared with AFP-Net, HBF-Net, and traditional baseline methods under the same dataset, training sample partitioning, and main training parameters. AFP-Net and HBF-Net are neural network prediction methods; Maximum Specific Transmission (MRT), Zero-Fork (ZF), and Genetic Algorithm (GA) are used as comparison methods; and the all-digital precoding results obtained using complete channel state information are used as the upper bound for performance reference.

[0028] In one simulation implementation, an 8×8 uniform planar array is used, the number of observation objects is set to 4, the finite feedback observation length is set to 32, the quantization bit depth is set to 8, and the discrete phase precision is set to 2 bits. Except for the observation noise condition under investigation, all methods use the same simulation environment and evaluation conditions. The above parameters are only used to illustrate one embodiment of the present invention and can be adjusted according to the actual data scale and application requirements.

[0029] like Figure 4 As shown, the system reachability and speed of each method increase overall with the increase of the signal-to-noise ratio parameter. Within the tested range, the system reachability and speed of the A²FP-Net of this invention are higher than those of AFP-Net, HBF-Net, and comparative methods such as MRT, ZF, and GA, and are closer to the performance reference upper bound obtained using complete channel state information. These results indicate that residual enhancement, channel attention, and the fusion of local and global statistical features can improve the feature representation capability of limited feedback observation samples, and that multi-task prediction and unsupervised training mechanisms can improve the accuracy and stability of target parameter prediction. The performance index comparison results of this embodiment can be combined with... Figure 4 understand.

Claims

1. A method for training a feature-enhanced multi-task neural network and predicting complex target parameters, characterized in that, Includes the following steps: (1) Obtain a limited feedback observation sample and the performance evaluation data corresponding to the limited feedback observation sample, wherein the performance evaluation data is used to calculate the downstream task performance index corresponding to the complex target parameter; The finite feedback observation samples are quantized, normalized, and have a preset length selected to construct an input matrix; (2) Input the input matrix into the feature-enhanced multi-task neural network, and obtain the fused features through convolutional feature extraction, residual enhancement feature extraction, channel attention weighting, and fusion of local features and global statistical features; (3) Input the fusion features into the classification branch and the regression branch, output the predicted probability of each candidate discrete parameter index through the classification branch, and output the real part parameter and imaginary part parameter of the complex prediction matrix through the regression branch; (4) Generate candidate target parameters according to each candidate discrete parameter index, and input the candidate target parameters and the complex prediction matrix into the performance evaluation process. Calculate the downstream task performance index corresponding to each candidate target parameter based on the performance evaluation data. (5) Construct a probability-weighted performance loss based on the downstream task performance indicators corresponding to each candidate target parameter and the predicted probability output by the classification branch; construct a prediction matrix performance loss based on the complex prediction matrix and the performance evaluation data; determine the performance ranking result based on the downstream task performance indicators corresponding to each candidate target parameter, and construct a classification branch training loss based on the performance ranking result. The probability-weighted performance loss, prediction matrix performance loss, classification branch training loss, and constraint penalty term are combined according to preset weights to obtain an unsupervised loss function; the gradient of the unsupervised loss function with respect to the network parameters is calculated, and the network parameters of the feature-enhanced multi-task neural network are updated through backpropagation. When training the model, this neural network does not need to prepare optimal complex form parameter matrix samples as supervision labels in advance. (6) Use the trained feature-enhanced multi-task neural network to perform forward inference on the limited feedback observation samples and output the corresponding complex target parameters.

2. The method according to claim 1, characterized in that, When the limited feedback observation sample selects the received signal strength indication measurement value of the wireless communication scenario, step (1) includes: converting the received signal strength indication measurement values ​​corresponding to multiple observation signals into quantized values ​​according to a preset quantization bit number; selecting data from the quantized values ​​according to a preset measurement length; normalizing the selected data and constructing an input matrix according to the arrangement order of multiple measurement objects.

3. The method according to claim 1, characterized in that, Step (2) includes: performing convolutional mapping, batch normalization, and nonlinear activation on the input features to obtain residual features; adding the input features and the residual features through cross-layer identity connections to obtain residual enhanced features; performing global average pooling on the residual enhanced features, generating channel weights through fully connected mapping and activation functions, and using the channel weights to adaptively weight different feature channels; performing global pooling on the channel-weighted features to obtain global statistical features, and concatenating the global statistical features with local features to obtain fused features.

4. The method according to claim 1, characterized in that, Steps (3) to (5) include: outputting the predicted probabilities corresponding to each candidate discrete parameter index through the classification output layer, and outputting the real and imaginary parameters of the complex prediction matrix through the regression branch; generating candidate target parameters according to each candidate discrete parameter index, and calculating the downstream task performance index corresponding to each candidate target parameter in combination with the complex prediction matrix and performance evaluation data; constructing a probability-weighted performance loss based on the downstream task performance index and the predicted probabilities output by the classification branch, constructing a prediction matrix performance loss based on the complex prediction matrix and performance evaluation data, and determining the performance ranking result based on the downstream task performance index to construct the classification branch training loss; combining the probability-weighted performance loss, prediction matrix performance loss, classification branch training loss, and constraint penalty term according to preset weights to obtain an unsupervised loss function, calculating the gradient of the unsupervised loss function relative to the network parameters, and updating the network parameters of the feature-enhanced multi-task neural network through backpropagation.

5. The method according to claim 1, characterized in that, When the complex target parameters are used to generate precoding parameters for a wireless communication system: an analog parameter matrix A is obtained based on the discrete parameter index, and a fully digital parameter matrix U is obtained based on the complex prediction matrix; the Morse-Penrose pseudoinverse of the analog parameter matrix A is multiplied by the fully digital parameter matrix U to obtain a digital baseband parameter matrix; the analog parameter matrix A is multiplied by the digital baseband parameter matrix and normalized according to a preset power constraint to obtain a combined precoding matrix W.

6. A feature-enhanced multi-task neural network training and complex target parameter prediction system, characterized in that, It includes an input data construction module, a feature enhancement module, a multi-task prediction module, a target parameter generation module, an unsupervised training module, and an online inference module; the input data construction module is used to acquire finite feedback observation samples and performance evaluation data corresponding to the finite feedback observation samples, and to quantize, normalize, and select a preset length for the finite feedback observation samples to construct an input matrix; The feature enhancement module is used to input the input matrix into the feature enhancement multi-task neural network, and sequentially perform convolutional feature extraction, residual enhancement feature extraction, channel attention weighting, and fusion of local features and global statistical features to obtain fused features; the multi-task prediction module is used to input the fused features into the classification branch and the regression branch respectively, and output the prediction probability corresponding to each candidate discrete parameter index and the real part and imaginary part parameters of the complex prediction matrix; the target parameter generation module is used to generate candidate target parameters according to each candidate discrete parameter index, combine the candidate target parameters with the complex prediction matrix, and calculate the downstream task performance index corresponding to each candidate target parameter according to the performance evaluation data; The unsupervised training module is used to construct a probability-weighted performance loss based on the downstream task performance indicators and the predicted probabilities output by the classification branch corresponding to each candidate target parameter; construct a prediction matrix performance loss based on the complex prediction matrix and performance evaluation data; determine the performance ranking results and construct a classification branch training loss based on the downstream task performance indicators corresponding to each candidate target parameter; combine the probability-weighted performance loss, prediction matrix performance loss, classification branch training loss, and constraint penalty terms according to preset weights to obtain an unsupervised loss function; calculate the gradient of the unsupervised loss function with respect to the network parameters; and update the network parameters of the feature-enhanced multi-task neural network through backpropagation. The online inference module is used to input new finite feedback observation samples into the trained feature-enhanced multi-task neural network and output the corresponding complex target parameters.

7. An electronic device comprising a processor and a memory, characterized in that, The memory stores a computer program, which, when executed by the processor, implements the method described in any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 5.

Images