Adaptive symbol reconstruction method for hplc channel transient interference and system thereof

Abstract

The present application relates to an adaptive symbol reconstruction method and system for HPLC channel transient interference, first, the received signal in the HPLC channel is acquired and the transient interference feature is extracted, then the manifold dimension is determined according to the QAM modulation order, the tangent space base vector is generated by calculating the partial derivative of the received sample in each dimension, the metric tensor parameter matrix is calculated based on the inner product of the base vector, the interference mapping relationship is established, then the manifold perception generator of the multi-layer LSTM network and the manifold preserving discriminator of the convolutional neural network are constructed, the discriminator and the generator are alternately trained to realize the antagonistic balance, finally, the received signal is reconstructed into the original symbol based on the trained generator, the present application can effectively resist the arc noise interference with the amplitude of 3 times the signal strength.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of communication technology, in particular to an adaptive symbol reconstruction method for HPLC channel transient interference and a system thereof, and particularly to a dual-channel autoencoder adversarial training symbol reconstruction technical solution based on information geometric manifold learning. BACKGROUND

[0002] As a communication method that uses existing power lines for data transmission, high-power line communication (HPLC) systems have been widely used in smart grids, industrial control and other fields. However, since power lines are not designed as transmission media for data transmission, HPLC systems face many challenges, of which the most notable is various types of transient interference. These interferences, especially arc noise interference, can cause signal distortion and symbol misjudgment, seriously affecting communication quality and reliability.

[0003] Traditional methods for processing HPLC channel transient interference mainly include filtering techniques based on statistical signal processing, error correction techniques based on channel coding, and channel compensation techniques based on equalizers. However, these methods have the following problems when dealing with complex and variable transient interference:

[0004] 1. Limited ability to handle nonlinear transient interference, especially arc noise and other interference with complex time-frequency characteristics;

[0005] 2. Traditional filtering methods often lose effective signal information while suppressing interference;

[0006] 3. Conventional equalizers lack the ability to adapt to rapid changes in the channel and are difficult to track the characteristics of transient interference;

[0007] 4. Existing signal reconstruction methods are usually based on Euclidean space models, ignoring the intrinsic geometric structure of the signal space.

[0008] With the development of deep learning technology, some neural network-based signal processing methods have begun to be applied to communication systems, but there is still no special solution for the characteristics of HPLC channel transient interference, especially a systematic method that combines information geometry and deep learning technology. SUMMARY

[0009] To overcome the shortcomings of the prior art, the present application proposes an adaptive symbol reconstruction method for HPLC channel transient interference and a system thereof, specifically a dual-channel autoencoder adversarial training symbol reconstruction method based on information geometric manifold learning and a system thereof, aiming to solve the problem of symbol reconstruction under HPLC channel transient interference and improve the anti-interference ability and signal recovery quality of the communication system.

[0010] The adaptive symbol reconstruction method for HPLC channel transient interference provided by the application establishes the geometric characterization of the signal space by parameterizing the QAM symbol set as a Riemannian manifold and calculating the metric tensor parameter matrix, adopts a multi-layer LSTM network to construct a manifold-aware generator and a convolutional neural network to construct a manifold-preserving discriminator, realizes the adversarial balance by alternately training the discriminator and the generator, and finally reconstructs the interference signal into the original symbol. The method comprises the following steps: acquiring the received signal in the HPLC channel and extracting the transient interference features; constructing a signal manifold model based on information geometry; performing two-channel autoencoder adversarial training; and performing adaptive symbol reconstruction.

[0011] When constructing the signal manifold model, the manifold dimension is determined according to the QAM modulation order, the received samples of the known training sequence are collected, the tangent space basis vectors are generated by calculating the partial derivatives of the received samples in each dimension, the metric tensor parameter matrix is obtained by calculating the inner product matrix based on the tangent space basis vectors, and the influence of the transient interference on the metric tensor is analyzed to establish the interference mapping relationship. This method based on the coordinate system and the partial derivative calculation converts the abstract manifold concept into a calculable parameter matrix.

[0012] In the adversarial training phase, the manifold-aware generator adopts a multi-layer LSTM network structure of an encoder-manifold mapping layer-a decoder, and the manifold-preserving discriminator adopts a convolutional neural network structure containing a manifold distance calculation layer. The discriminator training and the generator training are alternately performed: when the discriminator training is fixed, the discriminator is updated by minimizing the discrimination loss function; when the generator training is fixed, the generator is updated by minimizing the generation loss function containing the adversarial loss and the manifold distance loss. The training is terminated when the discriminator output probability is stable or reaches the preset iteration number. This specific network structure design and training step ensures the implementability of the adversarial training.

[0013] The application specifically comprises:

[0014] The adaptive symbol reconstruction method for HPLC channel transient interference comprises the following steps:

[0015] Acquiring the received signal in the HPLC channel and extracting the transient interference features;

[0016] Constructing a signal manifold model based on information geometry, comprising:

[0017] Determining the manifold dimension parameter according to the QAM modulation order, for M-order QAM modulation, the manifold dimension is set to log2(M);

[0018] Collecting the received samples of the known training sequence and constructing a local coordinate system;

[0019] Generating a tangent space basis vector set by calculating the partial derivatives of the received samples in each dimension in the coordinate system;

[0020] Based on the tangent space basis vector, the inner product matrix between coordinate directions is calculated to generate a metric tensor parameter matrix describing the local geometric relationship between symbols;

[0021] The influence of transient interference on the metric tensor parameter matrix is analyzed to establish a mapping relationship between the symbol distribution before and after the interference;

[0022] Performing double-channel autoencoder adversarial training, comprising:

[0023] Constructing a manifold-aware generator using a multi-layer LSTM network structure, including an encoder part for extracting the time sequence features of the interference signal, a manifold mapping layer for mapping the features to the symbol manifold coordinate system, and a decoder part for reconstructing the original symbol point;

[0024] Constructing a manifold-preserving discriminator using a convolutional neural network structure, including a feature extraction layer, a manifold distance calculation layer for calculating the distance between the reconstructed symbol and the ideal symbol on the manifold, and a discriminant output layer for outputting the discrimination probability;

[0025] Alternately performing the discriminator training step and the generator training step:

[0026] In the discriminator training step, the generator parameters are fixed, and the discriminator parameters are updated by minimizing the discrimination loss function, so that the discrimination output of the real symbol sample is close to 1, and the discrimination output of the generator reconstructed symbol is close to 0;

[0027] In the generator training step, the discriminator parameters are fixed, and the generator parameters are updated by minimizing the generation loss function, so that the discrimination output of the reconstructed symbol is close to 1, and the geodesic distance between the reconstructed symbol and the real symbol on the manifold is minimized;

[0028] When the discriminator output probability stabilizes within a predetermined range, or reaches a predetermined number of iterations, the training is terminated;

[0029] Performing adaptive symbol reconstruction, reconstructing the received signal into the original transmitted symbol based on the trained generator model.

[0030] As a preferred, the receiving signal in the HPLC channel is obtained, and the transient interference features are extracted, specifically comprising:

[0031] Receiving the OFDM modulated signal and performing short-time Fourier transform;

[0032] Extracting the arc noise time-frequency matrix in the 0.5-12MHz frequency band;

[0033] Dividing the time-frequency matrix into multiple subbands and calculating the energy distribution of each subband;

[0034] Distinguishing between narrowband arc and wideband arc features, and constructing an interference feature vector;

[0035] The interference degree index of the data symbol is calculated, and when the interference degree sum exceeds a preset threshold, the data packet is marked as an interfered data packet.

[0036] Preferably, the inner product matrix between the coordinate directions is calculated, specifically:

[0037] The inner product of the i-th coordinate direction basis vector and the j-th coordinate direction basis vector of the point on the manifold is calculated to generate the i-th row and j-th column component of the metric tensor;

[0038] All coordinate direction combinations are traversed to generate a complete metric tensor parameter matrix.

[0039] Preferably, the adaptive symbol reconstruction method for HPLC channel transient interference further comprises:

[0040] The transient interference in the HPLC channel is represented as a Lie group transformation set;

[0041] The invariant feature subspace of the Lie group transformation is identified;

[0042] Based on the invariant feature subspace, an isometry constraint condition is constructed;

[0043] According to the isometry constraint condition, the manifold parameters and the network training strategy are optimized.

[0044] Preferably, the discriminant loss function includes a real symbol discriminant loss term and a generated symbol discriminant loss term, the generated loss function includes an adversarial loss term and a manifold distance loss term, and the manifold distance in the manifold distance loss term is calculated by solving a geodesic equation.

[0045] Preferably, the generator adopts a multi-layer LSTM network structure, including:

[0046] The encoder part extracts the time sequence features of the interference signal;

[0047] The manifold mapping layer maps the features to the symbol manifold coordinate system;

[0048] The decoder part reconstructs the original symbol point;

[0049] Wherein, the input of the generator is the time-frequency feature of the interference signal, and the output is the reconstructed symbol.

[0050] Preferably, the discriminator adopts a convolutional neural network structure, including:

[0051] The feature extraction layer extracts the signal features;

[0052] The manifold distance calculation layer calculates the distance between the reconstructed symbol and the ideal symbol on the manifold;

[0053] discrimination output layer, evaluating the reconstruction quality, outputting the discrimination probability;

[0054] Wherein, the manifold distance adopts geodesic distance on Riemann manifold, instead of Euclidean distance.

[0055] As preferred, the adaptive symbol reconstruction is performed, specifically including:

[0056] Receiving OFDM modulated QAM signal;

[0057] Extracting signal time-frequency features, inputting the trained generator;

[0058] The generator outputs the reconstructed symbol point;

[0059] Calculating the confidence score of the reconstructed symbol;

[0060] According to the confidence score, selecting the optimal mapping constellation point;

[0061] When symbol misjudgment occurs, performing feedback correction on the channel noise estimation matrix;

[0062] Based on the current interference characteristics, dynamically adjusting the QAM constellation point mapping rule.

[0063] As preferred, the adaptive symbol reconstruction method for HPLC channel transient interference further includes:

[0064] Monitoring symbol reconstruction performance indicators;

[0065] When the performance indicators are lower than the preset threshold, triggering model parameter update;

[0066] Adjusting the training sample distribution, increasing the sample proportion of the current interference type;

[0067] Retraining the generator and discriminator, updating the model parameters;

[0068] Deploying the updated model to the symbol reconstruction engine.

[0069] The adaptive symbol reconstruction system for HPLC channel transient interference is used to realize the method, including:

[0070] Channel transient interference sensing module, for acquiring the received signal in the HPLC channel, and extracting the transient interference features;

[0071] Signal manifold modeling module, for constructing a signal manifold model based on information geometry, parameterizing the QAM symbol set as a Riemann manifold, and calculating the information metric tensor on the manifold;

[0072] A dual-channel adversarial training module is used to train the manifold-aware generator and the manifold-preserving discriminator, and the adversarial balance between the generator and the discriminator is realized through alternating optimization.

[0073] An adaptive symbol reconstruction engine is used to reconstruct the received signal into the original transmitted symbol based on the trained generator model.

[0074] A parameter feedback adjustment module is used to monitor system performance, dynamically optimize model parameters and reconstruction strategies.

[0075] The present application also introduces Lie group theory to analyze the invariance of interference transformation, designs an isometry constraint optimization mechanism, solves the instability problem of adversarial training, and can effectively resist arc noise interference with an amplitude of 3 times the signal strength.

[0076] The present application has the following beneficial effects:

[0077] 1. By modeling the signal space as a Riemannian manifold, the intrinsic geometric structure of the symbol space is captured, which is more accurate than the traditional Euclidean space model in representing signal characteristics and improving symbol reconstruction accuracy.

[0078] 2. The dual-channel autoencoder adversarial training strategy is adopted, combined with the principle of information geometry, to realize adaptive processing of complex nonlinear transient interference, and the symbol error rate is reduced by more than 60% compared with traditional methods in a harsh environment with SNR of -5dB.

[0079] 3. The Lie group theory is introduced to analyze the invariance of interference transformation, and an isometry constraint optimization mechanism is designed to solve the instability problem of adversarial training, which improves the training convergence speed by 3 times and reduces the sample demand by 70%.

[0080] 4. The system realizes adaptive processing capability for multiple interference types (narrowband arc, wideband arc), supports multiple modulation modes from 64QAM to 4096QAM, and greatly improves the communication reliability of the HPLC system in harsh environments.

[0081] 5. Through modular design and optimization implementation, the processing delay is controlled within 100us, supporting real-time data transmission of more than 100Mbps, meeting the needs of industrial applications. BRIEF DESCRIPTION OF DRAWINGS

[0082] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings, in which:

[0083] Figure 1 Figure 1 is a schematic diagram of the overall architecture of the HPLC channel transient interference adaptive symbol reconstruction system of the present application. DETAILED DESCRIPTION

[0084] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0085] Example 1

[0086] like Figure 1 As shown, the adaptive symbol reconstruction system for transient interference in HPLC channels proposed in this invention mainly includes a channel transient interference sensing module 1, a signal manifold modeling module 2, a dual-channel adversarial training module 3, an adaptive symbol reconstruction engine 4, and a parameter feedback adjustment module 5. These modules form a closed-loop optimization system, realizing the entire process of interference sensing, model training, and symbol reconstruction.

[0087] In one embodiment of the present invention, the channel transient interference sensing module 1 is responsible for acquiring the received signal in the HPLC channel and extracting transient interference features through techniques such as short-time Fourier transform; the signal manifold modeling module 2 parameterizes the QAM symbol set into a Riemannian manifold and constructs a signal space model based on information geometry; the dual-channel adversarial training module 3 trains the manifold sensing generator and the manifold preserving discriminator to improve the symbol reconstruction accuracy through an adversarial mechanism; the adaptive symbol reconstruction engine 4 reconstructs the received signal into the original transmitted symbol based on the trained generator model; and the parameter feedback adjustment module 5 monitors the system performance and dynamically optimizes the model parameters and reconstruction strategy.

[0088] The adaptive symbol reconstruction method for transient interference in the HPLC channel of the present invention is described in detail below.

[0089] This invention first acquires the received signal in the HPLC channel through the channel transient interference sensing module 1 and extracts transient interference characteristics. Specific steps include:

[0090] First, an OFDM modulated signal is received, and the sampling rate is set to 2.5 times the signal bandwidth. In a preferred embodiment of the invention, for an HPLC signal with a bandwidth of 10 MHz, the sampling rate is set to 25 MSPS.

[0091] Secondly, a short-time Fourier transform is performed on the received signal. The window length can be dynamically adjusted according to the signal characteristics, typically set between 256 and 1024 points. The short-time Fourier transform can be expressed as:

[0092] ,

[0093] in, Let be the time-frequency representation, representing the th The first time frame Complex values ​​at each frequency point; This is a time-domain signal, representing the sampled received signal; For the window function, usually use the Hanning window or Hamming window to reduce the spectrum leakage; For the FFT length, usually an integer power of 2; For the frame shift, usually set to Or ; For the time index, indicating the th time frame; For the frequency index, indicating the th frequency point; For the imaginary unit, ; For the basis function of Fourier transform.

[0094] Then, extract the arc noise time-frequency matrix in the 0.5-12MHz frequency band. Divide the frequency band into 240 sub-bands, each with a width of 5kHz, and calculate the energy distribution and burst characteristic parameters for each sub-band.

[0095] In addition, narrowband arc and wideband arc characteristics are distinguished. Narrowband arc usually has a concentrated energy distribution in the range of 0.5MHz-1.6MHz, while wideband arc has energy distribution in the range of 1.6MHz-13MHz or even wider. According to the energy distribution characteristics, an interference feature vector is constructed, including interference type, intensity, frequency domain distribution and time domain variation characteristics.

[0096] Finally, calculate the interference degree index of data symbols , defined as:

[0097] ,

[0098] Where, is the interference degree index, indicating the proportion of time-frequency points affected by interference; is the measurement time, in milliseconds; is the energy at time-frequency point , in watts; is the interference detection threshold, dynamically set according to the background noise level, usually 3-5 times the background noise energy; is an indicator function, taking the value of 1 when the condition in the parentheses is met, otherwise 0; represents the sum over all time points. When the total interference degree exceeds the preset threshold (usually 10%), the data packet is marked as an interfered data packet.

[0099] The present application constructs a signal manifold model based on information geometry through a signal manifold modeling module 2. This module converts the abstract manifold concept into a calculable metric tensor parameter matrix through specific coordinate system construction, partial derivative calculation and inner product operation, realizing the numerical characterization of the geometric structure of the signal space. The specific steps include:

[0100] First, determine the manifold dimension according to the QAM modulation order. For QAM modulation, the manifold dimension is set as For example, for 64QAM modulation, the manifold dimension is set as 6; for 256QAM modulation, the manifold dimension is set as 8.

[0101] Second, collect the received samples of the known training sequence. In the preferred embodiment of the present application, 1000-5000 known symbol points are selected as training samples, covering the signal characteristics under different interference conditions.

[0102] Then, based on the received samples, construct a local coordinate system and define the tangent space on the manifold. Specifically, for a point on the manifold , generate a set of tangent space basis vectors by calculating the partial derivatives of the point in each coordinate direction . The tangent space at point on the manifold can be represented as:

[0103] ,

[0104] where is the tangent space at point , representing the local linear approximation space of the manifold at that point; is the basis vector of the coordinate direction, representing the partial derivative along the th coordinate direction at point , which is calculated by numerical differentiation method; represents the linear space spanned by the vectors; is the manifold dimension, equal to the logarithm (base 2) of the QAM modulation order; is the coordinate index, from 1 to . In actual calculation, for the received sample point , take a small increment along the th coordinate direction, calculate the ratio of and to obtain the approximate value of the partial derivative in that direction.

[0105] Next, calculate the metric tensor on the manifold to generate a parameter matrix describing the local geometric characteristics. Specifically, for any two basis vectors of the coordinate direction in the tangent space and , calculate their inner product to obtain the th component of the metric tensor. The Riemannian metric tensor can be represented as:

[0106] ​​​ ,

[0107] in, To measure the components of a tensor, it is a A matrix, representing points Coordinate direction and The inner product; The inner product operation is represented by the inner product, which is defined as follows for complex vectors: ,in for The conjugate of complex numbers; and The first and the The basis vectors in each coordinate direction are obtained using the aforementioned partial derivative calculation method. The metric tensor describes the local scaling characteristics in different directions on the manifold and is the foundation for subsequent geodesic distance calculations. Traversing all... combination( ), generate complete Metric tensor parameter matrix.

[0108] Finally, the deformation effect of transient disturbances on the manifold is analyzed, and the symbol distribution mapping relationship before and after the disturbance is established. The disturbance can be regarded as a mapping on the manifold. By comparing the changes in the metric tensor before and after interference, the impact of interference on the geometric structure of the symbol space is analyzed. Its deformation effect can be measured using the pullback metric. analyze:

[0109] ,

[0110] in, To pull back the components of the metric, represent the mapping. Measuring The result when pulled back to the original coordinate system; For mapping The elements of the Jacobian matrix represent the mapping. The The component is related to the first... Partial derivatives of each independent variable; For the metric tensor components on the target manifold; For manifold dimension; Indicates all and From 1 to The values ​​are summed. This formula quantifies the degree of influence of disturbances on the symbol distribution by calculating the product of the Jacobian matrix of the mapping and the metric tensor.

[0111] In an embodiment of the present application, the transient interference in HPLC channel is also represented as a set of Lie group transformations. For interference such as arc noise, its representation on Lie algebra can be identified, and a set of generators of transformation group is constructed. By analyzing the invariant characteristic subspace of Lie group transformation, an equivariance constraint condition is constructed:

[0112] ,

[0113] where, is the generator function, representing the mapping from interference signal to original symbol; is the interference transformation, representing the effect of channel interference on signal; is the input signal, representing the original signal without interference; represents applying interference first and then reconstruction; represents reconstructing first and then applying interference; represents approximately equal, allowing small errors. This equivariance constraint ensures that the generator has stable reconstruction ability for a specific type of interference.

[0114] The present application trains the dual-channel autoencoder through the dual-channel adversarial training module 3. Through the explicit network structure design (multi-layer LSTM generator and convolutional neural network discriminator) and specific alternating training steps, the implementable adversarial training system is realized. The specific steps include:

[0115] First, construct the training data set, which contains interference signals and corresponding ideal symbols. In the preferred embodiment of the present application, the training data set size is 10000-50000 pairs of samples, covering different intensity and type of interference scenarios.

[0116] Second, block the interference signal time-frequency matrix to generate the training sample matrix. Divide the 5kHz data into 64 blocks, each containing 256 sampling points, and randomly sample to generate a sample matrix of .

[0117] Then, construct the specific network structure of the manifold-aware generator and the manifold-preserving discriminator:

[0118] (1) The manifold-aware generator adopts a multi-layer LSTM network structure:

[0119] The encoder part contains 3-5 layers of LSTM network, each layer containing 64-256 neurons, for extracting the time sequence features of the interference signal. The input is the time-frequency feature vector of the interference signal, and the output is the encoded feature representation. The hidden state dimension of the LSTM layer decreases layer by layer, for example, the first layer has 256 neurons, the second layer has 128 neurons, and the third layer has 64 neurons.

[0120] The manifold mapping layer is a fully connected layer that maps the feature vector output by the encoder to the manifold coordinate system. The output dimension of this layer is equal to the manifold dimension For 64QAM modulation, the output dimension is 6. The mapping layer uses a tanh activation function to ensure that the output values are within a reasonable range.

[0121] The decoder part contains a 3-5 layer fully connected network with the number of neurons decreasing layer by layer, and finally outputs reconstructed symbols with the same dimension as the original symbols. For example, the first layer has 128 neurons, the second layer has 64 neurons, and the output layer has 2 outputs for the real and imaginary parts of the QAM symbols. The decoder uses a ReLU activation function, and the output layer does not use an activation function.

[0122] (2) The manifold preserving discriminator uses a convolutional neural network structure:

[0123] The feature extraction layer contains 3-5 convolutional layers, each using a 3x3 convolution kernel with the number of channels gradually increasing from 32 to 256. The first layer has 32 channels, the second layer has 64 channels, the third layer has 128 channels, and the fourth layer has 256 channels. Each convolutional layer is followed by a BatchNorm layer and a LeakyReLU activation function.

[0124] The manifold distance calculation layer is a specially designed network layer for calculating the geodesic distance between the reconstructed symbols and the ideal symbols on the manifold. This layer first maps the input symbols to the manifold coordinate system, and then calls the geodesic solver to calculate the distance. The calculation of geodesic distance involves solving the geodesic equation, which is the innovative algorithm of the present invention and will be described in detail later.

[0125] The discriminator output layer is a fully connected network containing 2 layers, with 128 neurons in the first layer and 1 neuron in the output layer. The output layer uses a Sigmoid activation function to output a discrimination probability between 0 and 1, representing the probability that the input symbol is a real symbol.

[0126] Next, the specific steps of the alternating training are performed:

[0127] (1) Discriminator training steps:

[0128] Fix all parameters (weights and biases) of the generator so that they are not updated during the training process.

[0129] Sample a batch of real symbol samples and the corresponding interference signals from the training data set.

[0130] Generate reconstructed symbols using the generator, at which time the generator serves as a forward calculation tool and does not update its parameters.

[0131] The real symbol and the reconstructed symbol are input into the discriminator respectively, and the discrimination output and are obtained.

[0132] The discrimination loss function is calculated:

[0133] ,

[0134] wherein, is the discriminator loss function, representing the training target of the discriminator; is the discriminator function, the input is the symbol, and the output is a real number between 0 and 1, representing the probability that the input symbol is the real symbol; is the generator function, the input is the interference signal, and the output is the reconstructed symbol; is the real symbol, from the training data set; is the interference signal, from the training data set; is the probability distribution of the real symbol; is the probability distribution of the interference signal; represents the expected value of the real symbol distribution, which is approximated by the average value of the batch samples in actual calculation; represents the expected value of the interference signal distribution; is the natural logarithm function. The first term of the loss function makes the discrimination output of the real symbol close to 1, and the second term makes the discrimination output of the generated symbol close to 0.

[0135] The gradient of the discrimination loss function with respect to the discriminator parameters is calculated using the backpropagation algorithm , wherein represents all trainable parameters of the discriminator (including convolutional layer weights, fully connected layer weights, and biases).

[0136] The discriminator parameters are updated using an optimizer (such as the Adam optimizer): , wherein is the learning rate of the discriminator, preferably 0.0002.

[0137] (2) Generator training steps:

[0138] Fix all parameters of the discriminator so that they are not updated during the training process.

[0139] Sample a batch of interference signals and corresponding real symbols from the training data set.

[0140] Generate the reconstructed symbol .

[0141] reconstruct the symbol input the discriminator to get the discriminant output .

[0142] calculate the manifold distance , i.e. the geodesic distance of the reconstructed symbol and the real symbol on the manifold, the calculation method is as follows.

[0143] calculate the generation loss function :

[0144] ,

[0145] wherein, is the generator loss function, representing the training target of the generator; is the discriminator function; is the generator function; is the interference signal; is the corresponding real symbol; is the probability distribution of the interference signal; represents the expected value of the interference signal distribution; is the natural logarithm function; is the trade-off factor, used to balance the adversarial loss and the manifold distance loss, usually set between 0.1-0.5, preferably 0.3; is the geodesic distance on the manifold, representing the distance of the reconstructed symbol and the real symbol on the manifold. The first term (adversarial loss term) of this loss function makes the discriminant output of the reconstructed symbol close to 1, and the second term (manifold distance loss term) minimizes the manifold distance between the reconstructed symbol and the real symbol, ensuring that the reconstruction result conforms to the geometric structure of the manifold.

[0146] calculate the gradient of the generation loss function to the generator parameters using the back propagation algorithm , wherein represents all trainable parameters of the generator (including LSTM layer weights, manifold mapping layer weights, fully connected layer weights and biases).

[0147] update the generator parameters using an optimizer (such as Adam optimizer): , wherein is the learning rate of the generator, preferably 0.0002.

[0148] (3) Training termination condition:

[0149] monitor the discriminant output probability When the probability stabilizes within a preset range (e.g., between 0.45 and 0.55, indicating that the discriminator cannot effectively distinguish between real and generated symbols, reaching an adversarial balance), the training is considered to have converged, and training is terminated.

[0150] Alternatively, training can be terminated when the preset maximum number of iterations is reached, such as 5000 iterations.

[0151] During training, the discriminator and generator are trained alternately, typically one training cycle for the discriminator followed by one training cycle for the generator, to maintain training balance.

[0152] Geodesic distance The calculation involves solving the geodesic equations on the manifold, which is one of the core innovative algorithms of this invention:

[0153] A geodesic is the shortest path between two points on a manifold, and its equation is:

[0154] ,

[0155] in, For the first on the geodesic line Each coordinate component ; These are geodesic parameters, with a value range of [value range missing]. ,in Corresponding starting point (real symbol) Corresponding endpoint (reconstructed symbol); for right The second derivative of represents the acceleration of the geodesic; and They are respectively and right The first derivative of represents the velocity of the geodesic; Christoffel symbols describe the local curvature properties of a manifold by measuring the tensor. Calculated; For manifold dimension; Indicates all and From 1 to The summation of . This equation is a system of second-order ordinary differential equations that describes how geodesics proceed along the flattest direction of the manifold.

[0156] Christopher Symbol The calculation formula is:

[0157] ,

[0158] in, Christoffel symbols, denoted as, represent the change in coordinate basis when moving from a point to a nearby point on the manifold, and are key quantities in describing the degree of curvature of the manifold; are the inverse matrix elements of the metric tensor, satisfying where is the Kronecker delta function, which is 1 when and 0 otherwise; , , are the metric tensor components, calculated using the aforementioned methods; denote the metric tensor components with respect to coordinates , which are calculated using numerical differentiation methods; is the dimension of the manifold; denotes the summation over all from 1 to . This formula calculates the Christoffel symbols using the metric tensor and its partial derivatives, reflecting the geometric structure of the manifold.

[0159] In practical calculations, the geodesic equation is solved using numerical methods:

[0160] Initialization: Set the starting point (true symbol), the ending point (reconstructed symbol), and the initial velocity .

[0161] Discretization: Uniformly divide the parameter from 0 to 1 into steps, typically , with a step size .

[0162] Numerical integration: Using the Runge-Kutta method or Euler method, start from and step by step calculate the position and velocity at time . At each time step, calculate the acceleration according to the geodesic equation, and update the velocity and position.

[0163] Distance calculation: The geodesic distance is defined as the arc length along the geodesic, and the calculation formula is:

[0164] ,

[0165] where is the geodesic distance, representing the shortest path length on the manifold between the true symbol and the reconstructed symbol ; is the geodesic parameter, ranging from 0 to 1. is the geodesic distance between the points corresponding to the points is the metric tensor component at the point is the metric tensor component at the point and is the velocity component of the geodesic; summed over all coordinate directions; denotes the integral along the geodesic ; denotes the expression inside the integral is the manifold length (norm) of the velocity vector, defined by the metric tensor. In practice, numerical integration methods (such as the trapezoidal rule or Simpson's rule) are used to approximate the integral.

[0166] By the above method, the geodesic distance is incorporated into the loss function of the generator, ensuring that the generator not only pursues the adversarial goal that the discriminator cannot distinguish, but also ensures that the reconstructed symbols are close to the real symbols in the manifold geometry, thereby improving the accuracy and robustness of reconstruction.

[0167] Finally, according to the discriminator output probability during the training process, the training direction and parameters are dynamically adjusted:

[0168] When , it means that the generator reconstruction effect is good, and the training difficulty can be increased or the learning rate can be reduced;

[0169] When , it means that the reconstruction effect is not good, and similar interference samples can be increased or the learning rate can be increased.

[0170] Preferably, the initial value of the learning rate is set to 0.001, and is reduced by 10% every 1000 iterations, with a minimum of not less than 0.00001. The number of training iterations is usually set between 3000-5000 rounds, and the best model is determined according to the performance of the validation set.

[0171] Example 2

[0172] In an embodiment of the present application, the present application performs adaptive symbol reconstruction through an adaptive symbol reconstruction engine 4. The specific steps include:

[0173] First, receive the QAM signal modulated by OFDM. In the preferred embodiment of the present application, multiple modulation modes from 64QAM to 4096QAM are supported.

[0174] Second, extract the time-frequency features of the signal and input them into the trained generator. The feature extraction method is consistent with the training stage to ensure the consistency of data distribution.

[0175] Then, the generator outputs the reconstructed symbol point. For QAM modulation, the output is a complex vector in dimension

[0176] Next, the confidence score of the reconstructed symbol is calculated. The confidence score can be expressed as:

[0177] ,

[0178] wherein, is the confidence score, ranging from 0 to 1, representing the matching degree of the reconstructed symbol with the ideal constellation point; is the geodesic distance on the manifold, representing the distance of the reconstructed symbol with the ideal constellation point on the manifold; is the reconstructed symbol, which is the output of the generator; is the ideal constellation point, from the standard QAM constellation diagram; is the scaling factor, used to adjust the influence degree of distance on the confidence score, usually set to 1.0-5.0; is the natural exponential function.

[0179] In addition, according to the confidence score, the optimal mapping constellation point is selected. In the preferred embodiment of the present application, the top 3 candidate points with the highest confidence score are selected, combined with the channel coding information for the final decision.

[0180] When a symbol misjudgment occurs, feedback correction is performed on the channel noise estimation matrix. The correction method is:

[0181] ,

[0182] wherein, is the corrected noise estimation matrix, with dimension , is the signal space dimension; is the uncorrected noise estimation matrix; is the reconstructed symbol, which is the output of the generator; is the real symbol, obtained through feedback; is the learning rate, controlling the step size of correction, usually set to 0.01-0.1; represents the conjugate transpose operation, used to convert the column vector to the row vector and take the conjugate; represents the vector outer product, the result is a matrix.

[0183] Finally, based on the current interference characteristics, the QAM constellation point mapping rule is dynamically adjusted. For different interference types, different constellation point mapping strategies are adopted:

[0184] For narrowband arc interference, preferentially protect the symbols on the low-frequency subcarriers;

[0185] For wideband arc interference, adopt uniformly distributed constellation point mapping;

[0186] For mixed interference, dynamically adjust the protection level.

[0187] Embodiment 3

[0188] In another embodiment of the present application, a symbol reconstruction performance indicator is also monitored, and when the performance indicator is lower than a preset threshold, the model parameter update is triggered. The performance indicator can be symbol error rate (SER), bit error rate (BER), or throughput, etc. The preset threshold is determined according to the application scenario, for example, for ordinary data transmission, the SER threshold can be set to ; for critical control information, the SER threshold can be set to .

[0189] When the model needs to be updated, the distribution of the training samples is adjusted, and the proportion of samples of the current interference type is increased; the generator and the discriminator are retrained, and the model parameters are updated; the updated model is deployed to the symbol reconstruction engine. The model updating process can be executed in the background, without affecting real-time communication.

[0190] As shown in Figure 1 , the adaptive symbol reconstruction system for HPLC channel transient interference of the present application includes a channel transient interference perception module 1, a signal manifold modeling module 2, a dual-channel adversarial training module 3, an adaptive symbol reconstruction engine 4, and a parameter feedback adjustment module 5.

[0191] The channel transient interference perception module 1 is used to obtain the received signal in the HPLC channel and extract the transient interference features. This module includes a signal receiving unit, a short-time Fourier transform unit, and an interference feature extraction unit. The signal receiving unit collects the OFDM modulated signal in the HPLC channel; the short-time Fourier transform unit performs time-frequency analysis to generate a time-frequency matrix; and the interference feature extraction unit extracts the arc noise and other transient interference features from the time-frequency matrix to generate an interference feature vector.

[0192] The signal manifold modeling module 2 is used to construct a signal manifold model based on information geometry, parameterize the QAM symbol set as a Riemann manifold, and calculate the information metric tensor on the manifold. This module includes a manifold parameterization unit, a metric tensor calculation unit, and a Lie group transformation analysis unit. The manifold parameterization unit maps the symbol space to the manifold structure; the metric tensor calculation unit determines the local geometric properties on the manifold; and the Lie group transformation analysis unit identifies the invariant properties of the interference transformation.

[0193] The dual-channel adversarial training module 3 is used for training the manifold-aware generator and the manifold-preserving discriminator, and the adversarial balance between the generator and the discriminator is realized by alternating optimization. The module includes a generator training unit, a discriminator training unit and a training strategy adjustment unit. The generator training unit optimizes the generator parameters to improve the symbol reconstruction accuracy; the discriminator training unit optimizes the discriminator parameters to improve the discrimination ability of true and false symbols; and the training strategy adjustment unit dynamically adjusts the learning rate and sample distribution according to the training progress.

[0194] The adaptive symbol reconstruction engine 4 is used for reconstructing the received signal into the original transmitted symbol based on the trained generator model. The module includes a feature extraction unit, a symbol reconstruction unit and a constellation mapping unit. The feature extraction unit extracts time-frequency features from the received signal; the symbol reconstruction unit uses the trained generator model to reconstruct the original symbol; and the constellation mapping unit selects the optimal mapping constellation point according to the confidence score.

[0195] The parameter feedback adjustment module 5 is used for monitoring the system performance, dynamically optimizing the model parameters and the reconstruction strategy. The module includes a performance monitoring unit, a parameter optimization unit and a model updating unit. The performance monitoring unit tracks the performance indicators such as symbol error rate; the parameter optimization unit adjusts the system parameters according to the performance feedback; and the model updating unit triggers the model retraining when necessary.

[0196] The system of the application adopts a modular design, and the modules are connected through standardized interfaces to form a complete processing flow. In hardware implementation, a 32-bit ARM processor can be used as a main controller, cooperating with a DSP coprocessor and an FPGA acceleration unit to meet real-time processing requirements. The system storage configuration includes 4-8MB of system ROM, 16-32MB of system RAM and 256MB-1GB of non-volatile storage.

[0197] Preferably, the system adopts a parallel processing strategy, and interference awareness and feature extraction are performed in parallel, manifold calculation and symbol reconstruction are processed in a pipeline, and model training and parameter optimization are performed in the background. Through memory pool technology and data reuse mechanism, the dynamic allocation overhead is reduced, and the memory usage efficiency is optimized.

[0198] The adaptive symbol reconstruction method and system for HPLC channel transient interference of the application have been experimentally verified and exhibit significant performance advantages. The symbol error rate performance comparison under different signal-to-noise ratios shows that:

[0199] Under the harsh environment of SNR of-5dB, the symbol error rate of the method of the application is 0.023, which is reduced by about 63% compared with the traditional method (0.062);

[0200] For arc noise interference with an amplitude of 3 times the signal strength, the method of the application can still maintain a symbol error rate of 0.045, while the symbol error rate of the traditional method reaches 0.187.

[0201] Under the condition of medium signal-to-noise ratio (5-10 dB), the symbol error rate of the method of the application is reduced to below 0.001, meeting the needs of most application scenarios.

[0202] In addition, the processing delay of the system of the application is controlled within 78 mu s, supporting a data transmission rate of up to 120 Mbps, the training convergence speed is about 3.2 times faster than that of the traditional GAN method, and the sample requirement is reduced by about 72%. The system has not shown significant performance degradation in a stable operation test of 10,000 hours in succession, showing excellent long-term stability.

[0203] The adaptive symbol reconstruction method and system for HPLC channel transient interference provided by the application specifically realize the generator and the discriminator by adopting a specific multi-layer LSTM network and a convolutional neural network structure, realize the adversarial balance through explicit alternating training steps and termination conditions, and innovatively solve the problem of transient interference in HPLC communication by fusing the principle of information geometry and the deep learning technology. Compared with the traditional method, the application has higher symbol reconstruction accuracy, stronger anti-interference ability and better system stability, and provides an effective solution for reliable communication of the HPLC system in a harsh environment.

[0204] Embodiment 4

[0205] Based on the same inventive concept, the embodiments of the application further provide a computer device, including a storage medium, a processor and a computer program stored on the storage medium and executable on the processor. The processor executes the computer program to realize the steps of the adaptive symbol reconstruction method for HPLC channel transient interference described in embodiments 1 or 2 or 3.

[0206] Embodiment 5

[0207] Based on the same inventive concept, the embodiments of the application further provide a computer storage medium, which has a computer program stored thereon, and the computer program is executed by a processor to realize the steps of the adaptive symbol reconstruction method for HPLC channel transient interference described in embodiments 1 or 2 or 3.

[0208] Those skilled in the art will understand that the embodiments of the application can be provided as a method, a system, or a computer program product. Therefore, the application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0209] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0210] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0211] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0212] Finally, it should be noted that the above-mentioned embodiments are merely intended for describing the technical solutions of the present application, but not for limiting it. Although the present application is described in detail with reference to the above embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, and any modification or replacement without departing from the spirit and scope of the present application should be covered in the protection scope of the claims of the present application.

Claims

1. An adaptive symbol reconstruction method for transient interference in HPLC channels, characterized in that, include: Acquire the received signal in the HPLC channel and extract transient interference characteristics; Constructing a signal manifold model based on information geometry includes: The manifold dimension parameter is determined based on the QAM modulation order. For M-order QAM modulation, the manifold dimension is set to log2(M). Collect received samples from known training sequences and construct a local coordinate system; By calculating the partial derivatives of the received samples in each dimension of the coordinate system, a set of basis vectors in the tangent space is generated. Based on the tangent space basis vectors, the inner product matrix between coordinate directions is calculated to generate a metric tensor parameter matrix describing the local geometric relationship between symbols; The impact of transient disturbances on the parameter matrix of the metric tensor is analyzed, and the mapping relationship of the symbol distribution before and after the disturbance is established. Perform adversarial training of the dual-channel autoencoder, including: A manifold-aware generator is constructed using a multi-layer LSTM network structure, including an encoder part for extracting temporal features of interference signals, a manifold mapping layer for mapping features to the symbol manifold coordinate system, and a decoder part for reconstructing the original symbol points. A manifold-preserving discriminator is constructed using a convolutional neural network structure, including a feature extraction layer, a manifold distance calculation layer for calculating the distance between the reconstructed symbol and the ideal symbol on the manifold, and a discriminative output layer for outputting the discriminative probability. Alternately execute the discriminator training step and the generator training step: In the discriminator training step, the generator parameters are fixed, and the discriminator parameters are updated by minimizing the discriminant loss function, so that the discriminant output of the real symbol sample is close to 1 and the discriminant output of the generator reconstructed symbol is close to 0. In the generator training step, the discriminator parameters are fixed, and the generator parameters are updated by minimizing the generation loss function to make the discriminative output of the reconstructed symbol close to 1, while minimizing the geodesic distance between the reconstructed symbol and the real symbol on the manifold. Training terminates when the discriminator output probability stabilizes within a preset range or when a preset number of iterations is reached. Adaptive symbol reconstruction is performed, which reconstructs the received signal into the original transmitted symbol based on the trained generator model.

2. The adaptive symbol reconstruction method for transient interference in HPLC channels according to claim 1, characterized in that, The acquisition of the received signal in the HPLC channel and the extraction of transient interference features specifically include: Receive OFDM modulated signals and perform short-time Fourier transform; Extract the time-frequency matrix of arc noise in the 0.5-12MHz frequency band; The time-frequency matrix is ​​divided into multiple sub-bands, and the energy distribution of each sub-band is calculated. Distinguish between the characteristics of narrowband and broadband electric arcs, and construct an interference feature vector; The interference index of data symbols is calculated. When the total interference exceeds a preset threshold, the data packet is marked as an interfered data packet.

3. The adaptive symbol reconstruction method for transient interference in HPLC channels according to claim 1, characterized in that, The calculation of the inner product matrix between coordinate directions is specifically as follows: Perform an inner product operation on the i-th coordinate direction basis vector and the j-th coordinate direction basis vector of a point on the manifold to generate the i-th row and j-th column component of the metric tensor; Iterate through all coordinate direction combinations to generate a complete metric tensor parameter matrix.

4. The adaptive symbol reconstruction method for transient interference in HPLC channels according to claim 3, characterized in that, Also includes: Transient interference in the HPLC channel is represented as a set of Lie group transforms; Identify the invariant feature subspace of the Lie group transform; Based on the invariant feature subspace, equivariant constraints are constructed; Based on the isovariability constraints, optimize the manifold parameters and network training strategy.

5. The adaptive symbol reconstruction method for transient interference in HPLC channels according to claim 1, characterized in that, The discrimination loss function includes a true symbol discrimination loss term and a generated symbol discrimination loss term. The generation loss function includes an adversarial loss term and a manifold distance loss term. The manifold distance in the manifold distance loss term is calculated by solving the geodesic equation.

6. The adaptive symbol reconstruction method for transient interference in HPLC channels according to claim 1, characterized in that, The generator employs a multi-layer LSTM network structure, including: In the encoder section, the timing features of the interference signal are extracted; The manifold mapping layer maps features to a symbolic manifold coordinate system; The decoder part reconstructs the original symbol points; The generator takes the time-frequency characteristics of the interference signal as input and outputs the reconstructed symbols as output.

7. The adaptive symbol reconstruction method for transient interference in HPLC channels according to claim 1, characterized in that, The discriminator employs a convolutional neural network structure, including: Feature extraction layer, extracts signal features; The manifold distance calculation layer calculates the distance between the reconstructed symbol and the ideal symbol on the manifold. Discriminate the output layer, evaluate the reconstruction quality, and output the discrimination probability; The manifold distance is the geodesic distance on the Riemannian manifold, rather than the Euclidean distance.

8. The adaptive symbol reconstruction method for transient interference in HPLC channels according to claim 1, characterized in that, The execution of adaptive symbol reconstruction specifically includes: Receive OFDM modulated QAM signals; Extract the time-frequency features of the signal and input them into the trained generator; The generator outputs the reconstructed symbol points; Calculate the confidence score of the reconstructed symbol; Based on the confidence score, select the optimal mapping constellation point; When a symbol misjudgment occurs, feedback correction is performed on the channel noise estimation matrix; Based on the current interference characteristics, the QAM constellation point mapping rules are dynamically adjusted.

9. The adaptive symbol reconstruction method for transient interference in HPLC channels according to claim 8, characterized in that, Also includes: Monitor symbol reconstruction performance metrics; When the performance metric falls below a preset threshold, the model parameters are updated. Adjust the distribution of training samples to increase the proportion of samples with the current interference type; Retrain the generator and discriminator, and update the model parameters; Deploy the updated model to the symbolic reconstruction engine.

10. An adaptive symbol reconstruction system for transient interference in HPLC channels, used to implement the method according to any one of claims 1-9, characterized in that, include: The channel transient interference sensing module is used to acquire the received signal in the HPLC channel and extract transient interference characteristics; The signal manifold modeling module is used to construct a signal manifold model based on information geometry, parameterize the QAM symbol set into a Riemannian manifold, and compute the information metric tensor on the manifold; A dual-channel adversarial training module is used to train the manifold-aware generator and the manifold-preserving discriminator, and the adversarial balance between the generator and the discriminator is achieved through alternating optimization; An adaptive symbol reconstruction engine is used to reconstruct received signals into original transmitted symbols based on a trained generator model; The parameter feedback adjustment module is used to monitor system performance and dynamically optimize model parameters and reconstruction strategies.

Images