Search

CN-122026955-A - HPLC channel self-adaptive coding modulation and anti-noise system based on deep learning

CN122026955ACN 122026955 ACN122026955 ACN 122026955ACN-122026955-A

Abstract

The invention discloses an HPLC channel self-adaptive coding modulation and anti-noise system based on deep learning, which relates to the technical field of coding modulation and anti-noise, and comprises the steps of deploying a multi-sensor to collect power lines and environment data, extracting 48-dimensional feature vectors through a multi-scale convolution attention network, constructing a twin network by BiLSTM and GCN, calculating dynamic weights by combining GRU, improving channel prediction accuracy, adopting a layered DDPG architecture, balancing throughput, bit error rate and energy consumption, generating anti-noise signals by a dual-branch GAN structure, combining regularization optimization training, quickly adapting to new scenes through meta-learning, setting hysteresis thresholds to reduce switching times, and combining time-frequency analysis and fuzzy logic to evaluate channel quality. The invention obviously improves the communication performance of the HPLC through the multi-module collaborative optimization, reduces the channel prediction error and the error rate, realizes the real-time processing of the edge end and comprehensively enhances the reliability of the system.

Inventors

  • ZHENG WENPING
  • LI WENLONG

Assignees

  • 北京鼎亚科技有限公司

Dates

Publication Date
20260512
Application Date
20251209

Claims (10)

  1. 1. An HPLC channel adaptive coding modulation and anti-noise system based on deep learning, comprising: The heterogeneous multidimensional channel feature depth perception module is used for synchronously acquiring power line voltage, current waveform, environment temperature and humidity and electromagnetic noise intensity data by deploying an electromagnetic interference sensor, a temperature sensor, a humidity sensor and a sampler, carrying out space-time feature fusion on original data by adopting a feature extraction network based on a convolution attention mechanism MCA-Net, and outputting a vector group containing 48-dimensional features; The dynamic time-varying channel twin modeling module comprises a twin network architecture formed by a two-way long-short-term memory network BiLSTM and a graph rolling network GCN, a main network for carrying out time sequence modeling on historical channel characteristics, an auxiliary network for mining channel relevance among nodes in a power line topological structure through the GCN, a dynamic weight updating mechanism, a gate control circulation unit GRU and a dynamic weight updating module, wherein the dynamic weight updating mechanism is introduced for calculating influence factors of the channel characteristics at all moments based on the gate control circulation unit GRU, and the calculation mode is as follows Wherein The weight of the moment t is used for measuring the influence degree of the moment channel characteristics on the model; H t is the hidden layer state of the GRU at the time t and contains the information of the historical channel characteristics; Is a bias vector; For activating the function, mapping the weighted result to between 0 and 1; The multi-objective reinforcement learning joint decision module is used for dividing a coding modulation decision into two subtask levels of modulation mode selection and coding rate adjustment based on a decision maker of hierarchical depth deterministic strategy gradient H-DDPG, constructing a three-dimensional reward function comprising throughput, bit error rate and energy consumption Wherein R is a comprehensive rewarding value for evaluating the advantages and disadvantages of a decision strategy, T is throughput for reflecting the efficiency of data transmission, E is error rate for reflecting the accuracy of data transmission, C is energy consumption for measuring the energy consumption of system operation, alpha, beta, And storing historical decision data through an experience playback pool, and combining a dual-delay network structure to prevent strategy overfitting.
  2. 2. The deep learning based HPLC channel adaptive coding modulation and anti-noise system of claim 1, further comprising: the anti-generation anti-noise enhancement module comprises a dual-branch generation anti-network DB-GAN constructed at a transmitting end, a main branch generator which adopts a coding and decoding structure based on a transducer, learns noise characteristics from noise distribution and generates anti-noise signals, an auxiliary branch which learns a bidirectional mapping relation between an original signal and the anti-noise signals through a loop generation anti-noise network CycleGAN, and an anti-training regularization term introduced Wherein L reg is a regularization loss term used to constrain the training of the discriminator to prevent it from overfitting; controlling the weight of regularization loss as a regularization coefficient; representing the gradient of the discriminator D to the input signal x for measuring the sensitivity of the discriminator to the input signal; For L2 norm, the module length for calculating the gradient; the code rate dynamic adaptation module based on the enhanced migration establishes an enhanced migration learning model, accelerates the code rate adaptation in a new scene by utilizing the history decision experience in a similar channel scene, adopts a quick weight initialization strategy based on meta-learning and passes through a formula Adjusting model parameters, wherein theta new is the model parameters in the new scene, and theta base is the model parameters in the basic scene; for learning rate, controlling the step length of parameter updating; setting a dynamic hysteresis threshold mechanism, and self-adaptively adjusting code rate switching sensitivity according to the channel change speed; The multi-scale performance joint evaluation and feedback module comprises a performance evaluation module, a model light collaborative optimization module, a teacher network, a temperature scaling soft tag knowledge migration and structural pruning technology, wherein the performance evaluation module is constructed at a receiving end, a time-frequency analysis method combining short-time Fourier transform and wavelet transform is adopted, instantaneous signal distortion is detected at a time scale, error rate trends are analyzed at a second time scale through a long-short-term memory network LSTM; The multi-source data space-time alignment and fusion module adopts a data time alignment algorithm based on Dynamic Time Warping (DTW), calculates similarity distances of different sensor data sequences, adopts a self-adaptive weighted fusion algorithm, dynamically distributes weights according to reliability of sensor historical data, and adopts a fusion formula of Wherein F is the result of the fused data, n represents the number of sensors participating in the fusion, wi is the weight of the ith sensor data, the numerical value is dynamically determined according to the reliability evaluation of the historical sensor data, and Si is the data acquired by the ith sensor.
  3. 3. The deep learning based HPLC channel adaptive coding modulation and anti-noise system of claim 1, further comprising: The channel state prediction uncertainty quantization module is used for carrying out uncertainty estimation on a model prediction result by introducing a Monte Carlo dropouout method in dynamic time-varying channel twin modeling, obtaining prediction result distribution through random forward propagation, calculating a prediction interval [ l, u ], and obtaining a quantization formula as follows Wherein l is the lower limit of the prediction interval, and u is the upper limit of the prediction interval; Reflecting the predicted central trend for the average value of the predicted results of each random forward propagation; And k is a confidence coefficient, and is set according to the required confidence level for adjusting the width of the prediction interval.
  4. 4. The deep learning based HPLC channel adaptive coding modulation and anti-noise system of claim 1, wherein the strategy update formula for the layered depth deterministic strategy gradient H-DDPG in the multi-objective reinforcement learning joint decision module is Wherein Determining a decision strategy of an intelligent agent for parameters of a strategy network; controlling the step length of parameter updating for the learning rate of the strategy network; As an objective function Network parameters relating to policies The method comprises the steps of (1) guiding the direction of parameter updating, maximizing an objective function, and reducing strategy exploration space by hierarchically optimizing different subtask strategies.
  5. 5. The deep learning-based HPLC channel adaptive coding modulation and anti-noise system of claim 1, wherein a total loss function of a dual-branch generated anti-network DB-GAN in the anti-noise enhancement module is Ltotal = Ladv + Lrec + Lcyc, wherein Ltotal is a total loss of DB-GAN and is used for measuring a total error of model training, ladv is an anti-loss and reflects an anti-game degree between a generator and a discriminator, the anti-noise signal generated by the generator is trained, lrec is a reconstruction loss and is used for constraining the anti-noise signal generated by the generator to realize reduction of information of an original signal, lcyc is a cyclic consistency loss and ensures that a bidirectional mapping relation between the original signal and the anti-noise signal improves quality of the generated signal, and the enhancement signal is realized to suppress noise while retaining the original information by balancing each loss term optimization.
  6. 6. The deep learning-based HPLC channel adaptive coding modulation and anti-noise system of claim 1, wherein the meta gradient calculation formula for meta learning weight update in the enhanced migration-based code rate dynamic adaptation module is Wherein A meta gradient of the loss function L with respect to a meta parameter theta meta is used for guiding the updating of the meta parameter; Gradient of the loss function L with respect to the model parameter θ; for the gradient of the model parameter theta relative to the meta parameter theta meta , the weight initialization and parameter adjustment based on meta learning are realized through meta gradient calculation.
  7. 7. The deep learning based HPLC channel adaptive coding modulation and anti-noise system of claim 1, wherein membership functions of fuzzy logic synthesis evaluations in the performance joint evaluation and feedback module employ gaussian functions Wherein For inputting performance index x belonging to membership degree of a fuzzy set, the value range is 0-1, which is used for describing the similarity degree of the performance index and the fuzzy set, x is the input performance index including error rate and throughput, c is the central value of Gaussian function, and the central position of membership function is determined; the width of the Gaussian function is used for controlling the shape of the membership function, and the fuzzy degree of the fuzzy set is reflected; and obtaining the channel quality score through fuzzy rule reasoning.
  8. 8. The deep learning based HPLC channel adaptive coding modulation and anti-noise system of claim 1, wherein the sparsification constraint function of structured pruning in the model lightweight co-optimization module is Wherein L sparse is a sparsification loss function used for restricting sparsity of model parameters; For the sparsification coefficient, controlling the weight of the sparsification loss, and adjusting the sparsification degree of the model; W is the weight set of the model, and W represents the absolute value of the weight W, so that unimportant weights in the model are approaching to 0 by minimizing a loss function, structured pruning is realized, redundant connection is reduced, and the calculated amount of the model is reduced.
  9. 9. The deep learning based HPLC channel adaptive coding modulation and anti-noise system of claim 1, further comprising: constructing a federal transfer learning-based cross-regional communication network, wherein each node adopts a transfer learning method based on feature alignment when locally training a model, measures the feature distribution difference between a source domain and a target domain through the maximum average difference MMD, and the formula is as follows Wherein In the feature space Under, source domain With the target domain The maximum average difference value of the two domains is used for measuring the similarity degree of the characteristic distribution of the two domains; Representing mathematical expectations, x1 and x2 are the slave source fields respectively And a target domain And k (x 1, x 2) is a kernel function, and the data samples are mapped to a high-dimensional feature space to calculate distribution difference.
  10. 10. The deep learning based HPLC channel adaptive coding modulation and anti-noise system of claim 1, further comprising: The hardware-algorithm collaborative acceleration module is used for designing a neural network acceleration chip architecture, mapping matrix multiplication operation to a hardware parallel unit by adopting a pulse array and data rearrangement technology, and combining a model quantization technology to quantize floating point parameters into 8-bit fixed point numbers through a formula The method comprises the steps of calculating precision loss, wherein q is a quantized fixed-point numerical value, w is an original floating-point model parameter, s is a scaling factor and is used for mapping the floating-point parameter into a representation range of a fixed-point number; To round the function, the result of the calculation is rounded to the nearest integer.

Description

HPLC channel self-adaptive coding modulation and anti-noise system based on deep learning Technical Field The invention relates to the technical field of coded modulation and noise immunity, in particular to an HPLC channel self-adaptive coded modulation and noise immunity system based on deep learning. Background In the field of Power Line Communication (PLC), the high-speed power line carrier (HPLC) technology has the advantages of no need of additional wiring, wide coverage, etc., and becomes an important means for data transmission of smart power grids. However, HPLC channels have complex characteristics such as strong noise, time-variability, and multipath effects. Transformer noise, switching arc interference, and electromagnetic noise generated by industrial equipment present in the power system can severely degrade signal quality. For example, in an area with a complex electromagnetic environment such as an industrial park, the HPLC communication error rate is often 20% or more, which results in frequent data retransmission and a significant decrease in communication efficiency. In addition, the power line load change, weather factors and the like can cause the channel characteristics to be obviously changed in a short time, the traditional fixed parameter coding modulation mode is difficult to adapt to the dynamic change, and the stability and the reliability of communication cannot be ensured. The existing HPLC channel processing technology mostly adopts the traditional digital signal processing method and the fixed code modulation strategy. For example, a Cyclic Prefix (CP) is adopted to perform channel error correction against multipath effects and convolutional codes, but these methods rely on a priori knowledge of the channel, and have limited performance in a complex and variable practical environment. With the application of deep learning in the communication field, partial research attempts to utilize a neural network to perform channel estimation and modulation recognition, but the schemes often independently process links such as channel modeling, code modulation decision and the like, and lack collaborative optimization among the links. For example, when channel prediction is performed using deep learning alone, the prediction results fail to effectively guide dynamic adjustment of the code modulation strategy, resulting in limited overall communication performance improvement. Meanwhile, the existing deep learning model has large parameter quantity and high calculation complexity, and is difficult to run on the power line communication terminal equipment with limited resources in real time. In the anti-noise aspect, the traditional filtering algorithm and equalization technology can inhibit noise to a certain extent, but has poor processing effect on non-stationary and burst noise. In recent years, noise-resistant methods based on generation of a countermeasure network (GAN) lack of targeted optimization of HPLC channel characteristics, and thus have problems of signal distortion, unstable noise resistance and the like in practical applications. In addition, the existing system lacks quantitative evaluation of the uncertainty of the channel state, risk cannot be effectively weighed when coding modulation decision is made, and communication resource waste or communication interruption is easily caused. Therefore, there is a need for an HPLC channel adaptive code modulation and anti-noise technique that can adapt to complex channel environments and achieve multi-loop co-optimization. Disclosure of Invention The invention provides an HPLC channel self-adaptive code modulation and anti-noise system based on deep learning, which solves the problems in the prior art. In order to achieve the purpose, the invention adopts the following technical scheme that the HPLC channel adaptive code modulation and anti-noise system based on deep learning comprises: The heterogeneous multidimensional channel feature depth perception module is used for synchronously acquiring power line voltage, current waveform, environment temperature and humidity and electromagnetic noise intensity data by deploying an electromagnetic interference sensor, a temperature sensor, a humidity sensor and a sampler, carrying out space-time feature fusion on original data by adopting a feature extraction network based on a convolution attention mechanism MCA-Net, and outputting a vector group containing 48-dimensional features; The dynamic time-varying channel twin modeling module comprises a twin network architecture formed by a two-way long-short-term memory network BiLSTM and a graph rolling network GCN, a main network for carrying out time sequence modeling on historical channel characteristics, an auxiliary network for mining channel relevance among nodes in a power line topological structure through the GCN, a dynamic weight updating mechanism, a gate control circulation unit GRU and a dynamic weight updating module, wherei