A method for enhancing star flash communication signals based on deep neural network

Abstract

The present invention belongs to the field of industrial data processing, and in particular relates to a method for enhancing star flash communication signals based on deep neural networks. The method first uses distributed star flash nodes to collect signal samples and collect relevant indicators; then constructs an improved residual recursive neural network model, and enhances the signal by preprocessing, introducing gated residual units, multi-layer stacked networks and multi-scale feature fusion; then quantitatively evaluates the enhancement effect based on peak signal-to-noise ratio, mean square error and structural similarity index; constructs a reinforcement learning environment optimization for substandard signals, takes signal quality indicators as states, adjusts network parameters as actions, and trains and updates Q-value functions according to specific strategies. The present invention solves the shortcomings of traditional methods such as feature redundancy, information loss, gradient problems and poor adaptability, and can effectively retain signal features, suppress noise, improve reconstruction accuracy, achieve adaptive enhancement, and improve the robustness and generalization ability of the enhancement effect.

Description

Technical Field

[0001] The present invention belongs to the field of industrial data processing, and in particular relates to a method for enhancing star flash communication signals based on a deep neural network. Background Art

[0002] While Starflash technology excels in signal acquisition and transmission, it still has shortcomings in the area of signal enhancement. In real-world communication environments, signal quality is often affected by a variety of factors, such as environmental noise, interfering signals, and multipath effects. Especially in complex interference environments, relying solely on Starflash technology's hardware synchronization and distributed acquisition is insufficient to fundamentally address the problem of degraded signal quality. Random noise in the environment can mask important characteristics of the target signal, and multipath propagation of the signal can cause severe attenuation and distortion. These issues not only reduce signal transmission stability but also potentially impact the overall performance of the communication system. Existing signal enhancement methods primarily rely on simple neural networks. These methods fail to account for the relatively simple architecture of simple neural networks and lack the ability to deeply analyze the complex characteristics of signals. This makes them particularly vulnerable to dynamic signal changes under multiple interference factors. Furthermore, while removing noise, some useful signals may be mistakenly eliminated as noise. Alternatively, while preserving signal characteristics, noise interference cannot be effectively suppressed, resulting in a high noise level in the processed signal and poor signal enhancement. Summary of the Invention

[0003] In response to the technical problems existing in the above-mentioned background technology, the present invention proposes a star flash communication signal enhancement method based on deep neural network, which is simple in method and strong in theoretical basis.

[0004] In order to achieve the above object, the technical solution adopted by the present invention includes the following steps:

[0005] S1. First, distributed star flash nodes are used to collect signal samples and collect the signal-to-noise ratio, delay, and bit error rate of the signal;

[0006] S2. Construct a signal enhancement model based on an improved residual recurrent neural network;

[0007] S3. Compare and analyze the enhanced signal with the expected signal standard to quantitatively evaluate the effect of signal enhancement;

[0008] S4. Build a reinforcement learning environment to enhance and optimize the signals that do not meet the quantitative evaluation requirements. The specific implementation is as follows:

[0009] S41. First, the signal enhancement process is abstracted into a reinforcement learning framework. State S t is the current peak signal-to-noise ratio, mean square error, and structural similarity index of the signal, action A is the adjustment of the enhanced network parameters; the reward function is Rt =0.4ΔPSNR-0.4ΔMSE+0.2ΔSSIM;

[0010] S42, the input layer is in state S t , using a two-layer fully connected network with 128 neurons in each layer, the activation function is ReLU, each action corresponds to an action value, and the optimal action is selected;

[0011] S43, start training, initialize the reinforcement learning model parameters, use the ∈-greedy strategy to select actions, and enhance the signal according to the selected action, and calculate the next state S t+1 , using the reward function R t Calculate reward value;

[0012] S44. Update the Q value function and optimize the strategy as follows: in For the next state S t+1 The maximum Q value of all possible actions A′;

[0013] S45, if E total If it is greater than 40, the enhancement stops, otherwise it continues;

[0014] S5. Finally, the enhanced signal data is obtained.

[0015] Preferably, the construction of the signal enhancement model of the improved residual recurrent neural network in step S2 is:

[0016] S21. First, the input signal x(t) is preprocessed to remove noise and the signal data is standardized to meet the input requirements of the model;

[0017] S22, introduce the gated residual unit to control the information flow of the signal, and the state of the recursive step is updated as follows: where h t is the hidden state at time step t, is the residual mapping function, which captures the dependency between the current signal and the historical signal. g(t) is the gate function. is the denoised input signal, ⊙ represents the element-by-element product;

[0018] S23, using a multi-layer stacked residual recurrent neural network where each layer uses the output of the previous layer as input and processes it through a residual unit;

[0019] S24. Finally, the signal reconstruction performance is optimized through multi-scale feature fusion.

[0020] Preferably, after introducing a gated residual unit to control the information flow of the signal, the complexity index of the signal is calculated based on the key features of the signal: Where SNR max To preset the maximum signal-to-noise ratio threshold, SNR current is the current signal-to-noise ratio, τ threshold represents the preset delay tolerance threshold, τ current is the current signal delay, BER current is the current bit error rate, BER threschold is the preset bit error rate threshold; by calculating the complexity index of the signal, if C signal >1 triggers model adjustment, that is, increasing the number of gated residual units in the residual recurrent neural network and expanding the network depth.

[0021] Preferably, the gating function g(t) in step S22 is learned by an independent network module, which dynamically determines the signal transmission mode according to the characteristics of the current signal and is calculated by the following formula: Where W g , U g , b g is a trainable parameter and σ is the Sigmoid activation function.

[0022] Preferably, it is characterized in that the steps of optimizing the signal reconstruction performance by multi-scale feature fusion in step S24 are:

[0023] S241, first output the hidden state of each layer of residual recurrent neural network Composing a multi-scale feature set;

[0024] S242. Empower different features through the attention mechanism and integrate them into a unified representation: where β (l) The importance score of the feature at a specific scale is automatically learned by network training;

[0025] S243, the fused feature f t The reconstructed signal is generated by processing through linear transformation and nonlinear activation function: Where W f ,b f are the weight matrix and bias respectively.

[0026] Preferably, the integrated evaluation score after fusion of the quantitative evaluation signal enhancement effect in step S3 is: E total =0.4PSNR-0.4MSE+0.2SSIM, where PSNR, MSE, and SSIM are the peak signal-to-noise ratio, mean square error, and structural similarity index of the signal, respectively. If E total A value greater than 40 indicates that the enhancement is effective.

[0027] Compared with the existing technology, the advantages and positive effects of the present invention are that, in terms of model construction, a gated residual unit is introduced to dynamically control the propagation path of signal features, effectively retain useful features, suppress interference from irrelevant information, and solve the problem of feature redundancy and information loss in traditional models; through multi-scale feature fusion technology, the attention mechanism is used to weight and fuse features of different scales, comprehensively capture the multi-level features of the signal, and improve the accuracy and robustness of signal reconstruction. During training, the residual structure avoids gradient vanishing or explosion, ensuring stable convergence of training. In terms of optimization strategy, a reinforcement learning environment is constructed, and signal enhancement is abstracted as a reinforcement learning problem. With signal quality indicators as states, model parameter adjustments as actions, and enhancement effects as rewards, adaptive enhancement is achieved, improving the robustness and generalization ability of the enhancement effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following is a brief introduction to the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0029] Figure 1 The figure is a schematic diagram of the structural flow of the present invention. DETAILED DESCRIPTION

[0030] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0031] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways than those described herein. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0032] In modern communication systems, due to the presence of noise interference, multipath effects and signal attenuation in complex communication environments, signal quality often fails to meet the requirements of high-performance communication. Especially in high-speed, long-distance or multi-interference source scenarios, signal distortion and bit errors increase significantly, greatly limiting the practical application capabilities of the system. In order to overcome the above problems, the present invention proposes a method for enhancing star flash communication signals based on deep neural networks. The implementation process is as follows: Figure 1 shown.

[0033] To improve signal acquisition accuracy and spatiotemporal consistency, a distributed Starflash collaborative acquisition solution is employed. Multiple Starflash acquisition nodes are deployed within the communication area to collect signals and ensure synchronization accuracy. Clocks between nodes are aligned using Starflash's ultra-low-latency synchronization link. Master-slave synchronization is employed, with one high-precision node acting as the master and the remaining nodes acting as slaves. Slave nodes adjust their acquisition times based on the master's clock information, allowing the entire acquisition network to function as a tightly coordinated whole, acquiring signal samples with extremely high spatiotemporal resolution and measuring the signal-to-noise ratio, latency, and bit error rate.

[0034] Considering traditional recurrent neural networks, due to the lack of dynamic control mechanisms, traditional models are prone to introducing irrelevant features or losing key signal information, resulting in unsatisfactory signal enhancement effects. Their ability to capture multi-scale features is weak, and they cannot effectively restore complex signals. In addition, traditional models are easily limited by gradient vanishing or exploding during deep training, which affects the model's training effect and performance. The present invention introduces a gated residual unit, which can dynamically control the propagation path of signal features, effectively retain useful features, and suppress the interference of irrelevant information, thus solving the problems of signal feature redundancy and information loss in traditional models. In addition, through multi-scale feature fusion technology, the model uses an attention mechanism to weight and unify features of different scales, enabling it to more comprehensively capture the multi-level characteristics of the signal, significantly improving the accuracy and robustness of signal reconstruction. In deep network training, the improved model adopts a residual structure, avoiding the gradient vanishing or exploding problems that occur in traditional recurrent neural networks when multiple layers are stacked, ensuring training stability and convergence. The signal enhancement model of the improved residual recurrent neural network is constructed as follows: first, the input signal x(t) is preprocessed to denoise and the signal data is normalized to meet the model's input requirements. The gated residual unit is introduced to control the information flow of the signal, and the state update of the recursive step is: where h t is the hidden state at time step t, is the residual mapping function, which captures the dependency between the current signal and the historical signal. g(t) is the gating function, which dynamically determines the signal transmission mode according to the characteristics of the current signal. It is calculated by the following formula: Where W g , U g , b g is a trainable parameter, σ is the Sigmoid activation function, is the denoised input signal, and ⊙ represents the element-by-element product. After introducing the gated residual unit to control the information flow of the signal, the complexity index of the signal is calculated based on the key features of the signal: Where SNR max To preset the maximum signal-to-noise ratio threshold, SNR currentis the current signal-to-noise ratio, τ threshold represents the preset delay tolerance threshold, τ current is the current signal delay, BER current is the current bit error rate, BER threschold is the preset bit error rate threshold; by calculating the complexity index of the signal, if C signal >1 triggers model adjustment, that is, increasing the number of gated residual units in the residual recurrent neural network and expanding the network depth. Then, a multi-layer stacked residual recurrent neural network is used, and each layer uses the output of the previous layer as input and processes it through residual units. Finally, the signal reconstruction performance is optimized by multi-scale feature fusion. The implementation steps are as follows: First, the hidden state of each layer of the residual recurrent neural network is output Form a multi-scale feature set; empower different features through the attention mechanism and fuse them into a unified representation: where β (l) Indicates the importance score of a specific scale feature, which is automatically learned by network training; the fused feature f t The reconstructed signal is generated by processing through linear transformation and nonlinear activation function: Where W f ,b f are the weight matrix and bias respectively.

[0035] Next, the peak signal-to-noise ratio, mean square error, and structural similarity index of the signal are used to compare and analyze the enhanced signal with the expected signal standard to quantitatively evaluate the effect of signal enhancement. total =0.4PSNR-0.4MSE+0.2SSIM, where PSNR, MSE, and SSIM are the peak signal-to-noise ratio, mean square error, and structural similarity index of the signal, respectively. If E total If it is greater than 40, it means that the enhancement is effective. This evaluation method intuitively reflects the actual effect of signal enhancement and can accurately guide the subsequent optimization process. total For those less than 40, an optimization model based on reinforcement learning is used.

[0036] The present invention constructs a reinforcement learning environment, abstracts the signal enhancement process into a reinforcement learning problem, uses signal quality indicators (such as peak signal-to-noise ratio, mean square error and structural similarity index) as states, regards model parameter adjustment as actions, and uses the enhancement effect as a reward. The model can dynamically optimize parameter settings and achieve adaptive enhancement. First, the signal enhancement process is abstracted into a reinforcement learning framework. State S t is the current peak signal-to-noise ratio, mean square error, and structural similarity index of the signal, action A is the adjustment of the enhanced network parameters; the reward function is R t =0.4ΔPSNR-0.4ΔMSE+0.2ΔSSIM; the input layer is state St , use a two-layer fully connected network with 128 neurons in each layer, the activation function is ReLU, each action corresponds to an action value, and the optimal action is selected; start training, initialize the reinforcement learning model parameters, use the ∈-greedy strategy to select actions, and enhance the signal according to the selected action, and calculate the next state S t+1 , using the reward function R t Calculate the reward value; update the Q value function, and the optimization strategy is: in For the next state S t+1 The maximum Q value of all possible actions A′; if E total If the value is greater than 40, enhancement is stopped; otherwise, enhancement continues. This design enables the model to flexibly adjust its strategy in diverse and complex environments, improving the robustness and generalization of the enhancement effect. Furthermore, the reinforcement learning model uses an ∈-greedy strategy to explore the optimal action combination, enabling the signal enhancement process to not only achieve local optimization capabilities but also maximize long-term effects. Furthermore, this optimization framework incorporates a deep neural network as a policy model, enabling efficient processing of multidimensional complex signals and ensuring the stability and accuracy of the optimization process.

[0037] The above description is merely a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any person skilled in the art may utilize the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes for application in other fields. However, any simple modification, equivalent change, and modification of the above embodiments made in accordance with the technical essence of the present invention without departing from the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A method for enhancing star flash communication signals based on deep neural networks, characterized in that: The following steps are involved: S1. First, distributed star flash nodes are used to collect signal samples and collect the signal-to-noise ratio, delay, and bit error rate of the signal; S2. Construct a signal enhancement model based on an improved residual recurrent neural network; S3. Compare and analyze the enhanced signal with the expected signal standard to quantitatively evaluate the effect of signal enhancement; S4. Build a reinforcement learning environment to enhance and optimize the signals that do not meet the quantitative evaluation requirements. The specific implementation is as follows: S41. First, the signal enhancement process is abstracted into a reinforcement learning framework. State S t is the current peak signal-to-noise ratio, mean square error, and structural similarity index of the signal, action A is the adjustment of the enhanced network parameters; the reward function is R t =0.4ΔPSNR-0.4ΔMSE+0.2ΔSSIM; S42, the input layer is in state S t , using a two-layer fully connected network with 128 neurons in each layer, the activation function is ReLU, each action corresponds to an action value, and the optimal action is selected; S43, start training, initialize the reinforcement learning model parameters, use the ∈-greedy strategy to select actions, and enhance the signal according to the selected action, and calculate the next state S t+1 , using the reward function R t Calculate reward value; S44. Update the Q value function and optimize the strategy as follows: in For the next state S t+1 The maximum Q value of all possible actions A′; S45, if E total If it is greater than 40, the enhancement stops, otherwise it continues; S5. Finally, the enhanced signal data is obtained.

2. The method for enhancing star flash communication signals based on deep neural network according to claim 1, characterized in that: The construction of the signal enhancement model of the improved residual recurrent neural network in step S2 is: S21. First, the input signal x(t) is preprocessed to remove noise and the signal data is standardized to meet the input requirements of the model; S22, introduce the gated residual unit to control the information flow of the signal, and the state of the recursive step is updated as follows: where h t is the hidden state at time step t, is the residual mapping function, which captures the dependency between the current signal and the historical signal. g(t) is the gate function. is the denoised input signal, ⊙ represents the element-by-element product; S23, using a multi-layer stacked residual recurrent neural network where each layer uses the output of the previous layer as input and processes it through a residual unit; S24. Finally, the signal reconstruction performance is optimized through multi-scale feature fusion.

3. The method for enhancing star flash communication signals based on deep neural network according to claim 2, characterized in that: After introducing the gated residual unit to control the information flow of the signal, the complexity index of the signal is calculated based on the key characteristics of the signal: Where SNR max To preset the maximum signal-to-noise ratio threshold, SNR current is the current signal-to-noise ratio, τ threshold represents the preset delay tolerance threshold, τ current is the current signal delay, BER current is the current bit error rate, BER threschold is the preset bit error rate threshold; by calculating the complexity index of the signal, if C signal >1 triggers model adjustment, that is, increasing the number of gated residual units in the residual recurrent neural network and expanding the network depth.

4. The method for enhancing star flash communication signals based on deep neural network according to claim 2, characterized in that: The gating function g(t) in step S22 is learned by an independent network module. It dynamically determines the signal transmission mode according to the characteristics of the current signal and is calculated by the following formula: Where W g , U g , b g is a trainable parameter and σ is the Sigmoid activation function.

5. The method for enhancing star flash communication signals based on deep neural network according to claim 2, characterized in that: The steps for optimizing the signal reconstruction performance by multi-scale feature fusion in step S24 are as follows: S241, first output the hidden state of each layer of residual recurrent neural network h t (l) Composing a multi-scale feature set; S242. Empower different features through the attention mechanism and integrate them into a unified representation: where β (l) The importance score of the feature at a specific scale is automatically learned by network training; S243, the fused feature f t The reconstructed signal is generated by processing through linear transformation and nonlinear activation function: Where W f ,b f are the weight matrix and bias respectively.

6. The method for enhancing star flash communication signals based on deep neural network according to claim 1, characterized in that: The comprehensive evaluation score after fusion of the quantitative evaluation signal enhancement effect in step S3 is: total =0.4PSNR-0.4MSE+0.2SSIM, where PSNR, MSE, and SSIM are the peak signal-to-noise ratio, mean square error, and structural similarity index of the signal, respectively. If E total A value greater than 40 indicates that the enhancement is effective.

Images