A power line channel modeling method based on deep learning
By modeling the power line channel using autoencoder and hybrid generative adversarial network in deep learning technology, the problem that existing statistical modeling methods rely on theoretical knowledge is solved, and a more accurate and direct channel modeling effect is achieved.
Patent Information
- Application Number
- CN202210381643.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-04-12
- Publication Date
- 2025-02-11
- Estimated Expiration
- 2042-04-12
AI Technical Summary
The existing statistical power line channel modeling methods rely on the understanding of electromagnetic fields, transmission line theory and actual power line networks. When the assumption or approximate deviation is large, the model cannot represent the real channel.
Using a deep learning-based approach, power line channels are modeled using autoencoder (AE) and hybrid generative adversarial network (MGAN), and directly learning hidden distributions from real channel data, reducing dependence on electromagnetic signal propagation and transmission line theory.
This method can more accurately model complex power line channels, reduce the need for theoretical knowledge, avoid mathematical approximations and assumptions, learn channel characteristics directly from the data, and support the evaluation of PLC technology and the development of advanced PLC solutions.
Smart Images
Figure CN114781251B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of power communication and relates to a power line channel modeling method based on deep learning. Background Art
[0002] PLC technology uses the existing distribution network and transmission facilities as communication media to achieve data transmission. Therefore, PLC technology has incomparable advantages: wide coverage and low cost, and it is inevitable to become the basic communication technology of smart grid. However, with the upgrading of demand, the concept of "smart" emerges in an endless stream. Smart cities and smart homes put forward higher requirements for PLC, forcing BPLC to operate at a higher frequency.
[0003] The existing statistical power line channel modeling method makes certain assumptions about the actual environment and then establishes the corresponding power line channel model. However, the traditional statistical modeling method strongly relies on the understanding of electromagnetic fields, transmission line theory, and actual power line networks. When the assumptions or approximations have large deviations, the model cannot represent the real channel. Summary of the invention
[0004] In view of this, an object of the present invention is to provide a power line channel modeling method based on deep learning.
[0005] In order to achieve the above object, the present invention provides the following technical solutions:
[0006] A power line channel modeling method based on deep learning, the method comprising the following steps:
[0007] S1: Establish autoencoder AE and hybrid generative adversarial network MGAN;
[0008] S2: Power line channel modeling by joint AE and MGAN.
[0009] Optionally, S1 is specifically: the autoencoder AE includes an encoder and a decoder; the encoder learns a potential representation y of an input sample x; the encoder reconstructs x at the output end using information inside y;
[0010] The general learning formula for deterministic AE is:
[0011]
[0012] Among them, θ opt is the optimal parameter of the neural network, δ is the distance measure, while E and F represent the encoder and decoder functions, θ1 and θ2 are the parameters of the encoder and decoder, respectively;
[0013] When the feature space is subject to energy constraint EC, a regularization term is added to the cost function to track this restriction. The cost function for training AE is:
[0014]
[0015] Among them, A and b are the weight matrix and bias vector in the linear transformation, respectively, and both are usually constants; ||·||2 represents the l2 norm, that is, the Euclidean distance; the square of the l2 norm is the classic mean square error MSE;
[0016] MGAN uses a mixture of multiple distributions instead of a single distribution in the standard GAN to approximate the real data distribution, while expanding the divergence of multiple distributions to cover different data modes; MGAN is equivalent to K generators G 1:K , the game between the discriminator D and the classifier C; each generator G k Map z to x k =G k (z), thus inducing a single distribution p Gk ; K generators together get a mixture of K distributions, that is, p in the data space model Sample G u (z) is selected as the output, and the index u is sampled from the multinomial distribution Mult(π), where π = [π1,π2,...,π K ] is a predefined mixed coefficient; the discriminator D is used to distinguish fake samples from real samples; C is a multi-classifier used to determine the index tag of the generator corresponding to the fake sample; in order to minimize the number of parameters, except for the input layer, the K generators share parameters; except for the output layer, the multi-classifier C and the discriminator D share parameters;
[0017] The training of MGAN is expressed as the following minimax objective function:
[0018]
[0019] Among them, C k (x) indicates x comes from G k The probability of β>0 is the diversity hyperparameter, p Gk It's G k The probability distribution of generated data; the first two terms represent the interaction between the generator and the discriminator in the standard GAN; the last term represents the interaction between the generator and the classifier, which is the standard softmax loss of the multi-classifier, and its purpose is to maximize the entropy of the multi-classifier, that is, to maximize the difference between the data distribution generated by each generator and the data distribution generated by other generators; the strength of the interaction is controlled by β.
[0020] Optionally, S2 is specifically as follows: let x be the input data, E:x→y be the function that maps x to y, then there is a deterministic relationship between the distributions of x and y; suppose that prior knowledge of the complexity of x is provided, that is, the distribution p(x), and the purpose becomes the distribution p(y) for low-dimensional features; if the function E is differentiable, p(x) is obtained from p(y) according to the variable change rule:
[0021]
[0022] Among them, det(·) is an operator for finding the determinant of a matrix;
[0023] The power line channel data H is used as the input of AE, and the encoder block of the trained AE is used as the transformation function E(·) to distribute the feature y p feature (y) Modeling: The new feature y generated by MGAN g As the input of the decoder block F(·), the generated statistical distribution is p data (H) New channel H g ;
[0024] Select the loss function of formula (2) to train AE, where δ is the MSE; add a regularization term in formula (2), that is:
[0025]
[0026] Among them, the second term is the normalized mean square error NMSE, which enhances the effect of data with smaller values;
[0027] Make changes to the network training:
[0028] (1) According to the average channel gain ACG, the power line channel data H is divided into K categories, and y is divided into K categories;
[0029] (2) The training of AE remains unchanged. When training MGAN, we proportionally extract features from each class y according to the predefined mixing coefficient. k A small batch is extracted from the batch to form a large batch;
[0030] (3) Determine y at C g When the corresponding generator index is marked, let C also recognize the real feature y k categories, prompting each G k Can generate matches for a certain class y k The new features of MGAN become:
[0031]
[0032] in, is the distribution of the original features of the kth class; where is the distribution of the original features of the kth class.
[0033] Optionally, S2 specifically includes: the training method of AE and MGAN in the power line channel is:
[0034] S21: train AE using power line channel data H;
[0035] S22: Extract low-dimensional features y of the power line channel data H using the encoder of AE: y=E(H);
[0036] S23: Save the decoder module;
[0037] S24: Use the original feature y to train MGAN and save the generator module;
[0038] S25: From the distribution p used when training MGAN z The sampling noise z;
[0039] S26: Generate new features y using the MGAN generator g :y g =G(z);
[0040] S27: Using the decoder of AE to obtain new power line channel data H g :H g =F(y g ).
[0041] The beneficial effects of the present invention are: using auto-encoder (AE) and hybrid generative adversarial network (MGAN) to model complex power line channels, using real channel data to train the network and obtain network parameters. This solution minimizes the need for knowledge such as electromagnetic signal propagation and transmission line theory, and does not require complex theoretical analysis or data processing, avoiding mathematical approximations and assumptions, and directly learning hidden distributions from real channel data. This method can provide support for the evaluation of PLC technology and the development of advanced PLC solutions.
[0042] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:
[0044] Figure 1 It is the autoencoder structure;
[0045] Figure 2 This is the basic architecture of MGAN. DETAILED DESCRIPTION
[0046] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.
[0047] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.
[0048] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.
[0049] 1. The principles of autoencoders and hybrid generative adversarial networks
[0050] AE is a special neural network consisting of an encoder and a decoder. Figure 1 As shown, the encoder tries to learn a latent representation y of the input sample x, usually a feature in a low-dimensional space. The encoder uses the information inside y to reconstruct x at the output.
[0051] The general learning formula for deterministic AE is:
[0052]
[0053] Among them, θ opt is the optimal parameter of the neural network, δ is the measure of distance, while E and F represent the encoder and decoder functions, respectively, and θ1 and θ2 are the parameters of the encoder and decoder, respectively.
[0054] When the feature space is subject to energy constraints (EC), a regularization term can be added to the cost function to track this limitation. The cost function for training AE is:
[0055]
[0056] Among them, A and b are the weight matrix and bias vector in the linear transformation, respectively, and both are usually constants. ||·||2 represents the l2 norm, that is, the Euclidean distance. The square of the l2 norm is the classic mean squared error (MSE).
[0057] MGAN uses a mixture of multiple distributions instead of a single distribution in standard GAN to approximate the real data distribution, while expanding the divergence of multiple distributions to cover different data modes. Figure 2 As shown, MGAN is equivalent to K generators G 1:K , the game between the discriminator D and the classifier C. Each generator G k Map z to x k =G k (z), thus inducing a single distribution p Gk ; K generators together get a mixture of K distributions, that is, p in the data space model Sample G u (z) is selected as the output, and the index u is sampled from the multinomial distribution Mult(π), where π = [π1,π2,...,π K ] is the coefficient of the predefined mixture. The discriminator D aims to distinguish fake samples from real samples. C is a multi-classifier that aims to judge the index label of the generator corresponding to the fake sample. On the other hand, in order to minimize the number of parameters, the K generators share parameters except the input layer; the multi-classifier C and the discriminator D also share parameters except the output layer. Compared with the standard GAN, MGAN only adds minimal complexity, and its training process is still efficient. In addition, this parameter sharing scheme enables the network to utilize its common information, such as low-level features close to the data layer, which helps to train the model.
[0058] The training of MGAN can be expressed as the following minimax objective function:
[0059]
[0060] Among them, Ck (x) indicates x comes from G k The probability of β>0 is the diversity hyperparameter, p Gk It's G k The probability distribution of generated data. The first two terms represent the interaction between the generator and the discriminator in a standard GAN. The last term represents the interaction between the generator and the classifier, which can be a standard softmax loss for multiple classifiers, and its goal is to maximize the entropy of the multiple classifiers, that is, to maximize the difference between the data distribution generated by each generator and the data distribution generated by other generators. The strength of this interaction is controlled by β.
[0061] 2. Power line channel modeling method based on combined AE and MGAN
[0062] Let x be the input data, E:x→y be the function that maps x to y, then there is a deterministic relationship between the distributions of x and y. Assuming that we have prior knowledge of the complexity of x, i.e., the distribution p(x), the goal becomes the distribution p(y) for low-dimensional features. If the function E is differentiable, we can get p(x) from p(y) according to the variable change rule:
[0063]
[0064] Here, det(·) is an operator for finding the determinant of a matrix.
[0065] Based on the above analysis, the power line channel data H can be used as the input of AE, and the encoder block of the trained AE is used as the transformation function E(·). feature (y) model. Then, the new feature y generated by MGAN g As the input of the decoder block F(·), the generated statistical distribution is p data (H) New channel H g Table 1 summarizes the training methods of AE and MGAN for generating power line channels.
[0066] Table 1 Power line channel generation algorithm combining AE and MGAN
[0067]
[0068] It is hoped that the AE can keep the energy of feature y as low as possible when reconstructing the PLC channel H. Therefore, the present invention selects the loss function described in formula (2) to train the AE, where δ is the MSE. However, because the value of some PLC channel data is very small, this loss function is not enough to ensure that the AE network learns the features of this part of the data, so a regularization term is added to formula (2), that is:
[0069]
[0070] Among them, the second item is the normalized mean squared error (NMSE), which enhances the effect of data with smaller values.
[0071] The traditional power line channel model contains multiple reference channels, which are usually divided by fading. Therefore, this application also divides the channels into several categories according to attenuation. Inspired by the above facts, this application makes changes to the network training: (1) According to the size of the average channel gain (ACG), the power line channel data H is divided into K categories (y is also divided into K categories). (2) The training of AE remains unchanged. When training MGAN, according to the predefined mixing coefficient, the features y from each category are proportionally added. k Instead of extracting a large batch directly from all the features y, we extract a small batch from y to form a large batch. (3) In C, we determine y g When the corresponding generator index is marked, let C also recognize the real feature y k categories, prompting each G k Can generate matches for a certain class y k The objective function of MGAN becomes:
[0072]
[0073] in, is the distribution of the original features of the kth class. is the distribution of the original features of the kth class.
[0074] Specific Example 1: Training of AE and MGAN
[0075] This application divides the channels into four categories according to the size of ACG at intervals of 20 dB, namely, the first category: (-20, 0]dB, the second category: (-40, -20]dB, the third category: (-60, -40]dB, and the fourth category: (-∞, -60]dB. In other words, a large amount of real signal-to-noise data is obtained through special measurement activities, usually the channel frequency response (Channel Frequency Response, CFR). The CFR can be divided into four categories according to the above criteria, and the mixing coefficient π is set to the proportion of each type of channel. If the frequency range and resolution are set to 1.8~100MHz and 50KHz, each CFR sample is a 1965-dimensional complex vector. However, neural networks cannot process complex numbers directly, so this chapter converts the 1965-dimensional complex vector into a 3930-dimensional real vector, with the first 1965 entries corresponding to the real part of the CFR and the last 1965 entries corresponding to the imaginary part of the CFR.
[0076] Table 2 lists the architecture of AE, where the input layer has the same dimension as the training sample. In order to reduce the complexity of training MGAN, the dimension of the feature space is selected to be one eighth of the input space dimension. Table 3 gives the hyperparameters for training AE.
[0077] Table 2 Network architecture of AE
[0078]
[0079] Table 3 Hyperparameters for training AE
[0080]
[0081] Table 4 is the network architecture of MGAN. BN and SN are the abbreviations of Batch Normalization and Spectral Normalization, respectively. SN can make the discriminator satisfy the 1-Lipschitz constraint and enhance the stability of training. '? ' represents whether to choose, '√' represents yes, and '×' represents no. The 4 / 1 feature map in the last layer of C and D means that two separate fully connected layers are applied to the penultimate layer, C outputs 4 logical values, and D outputs 1 logical value. This also corresponds to Softmax / Sigmod. Table 5 is the hyperparameters for training MGAN. Since the indoor power line channels are divided into four categories, the number of generators is also set to 4.
[0082] Table 4 Network architecture of MGAN
[0083]
[0084] Table 5 Hyperparameters for training MGAN
[0085]
[0086]
[0087] Specific embodiment 2: Generation of power line channel instance
[0088] According to the specific embodiment 1, after the network training is completed, the decoder of AE and the generator of MGAN can be used to synthesize the power line channel instance.
[0089] The decoder of AE contains 1 input layer, 1 output layer and 3 hidden layers; the generator of MGAN contains 3 shared hidden layers, 1 shared output layer and 4 independent input layers. The 4 independent input layers of MGAN are equivalent to MGAN having 4 generators. The first generator synthesizes features associated with the power line channel with ACG range of (-20,0]dB, and the second, third and fourth generators correspond to the power line channels with ACG range of (-40,-20], (-60,-40], (-∞,-60]dB respectively.
[0090] If you need an indoor power line channel with an ACG range of (-40,-20]dB, you can obtain an instance by following the steps below:
[0091] (1) Draw a 100-dimensional noise vector z from the standard normal distribution N(0,1).
[0092] (2) Use the second generator of MGAN to map the noise vector z to the new feature y g .
[0093] (3) Use the AE decoder to transform the feature y g Rebuilt as new indoor power line channel H g .
[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.
Claims
1. A power line channel modeling method based on deep learning, characterized in that: The method comprises the following steps: S1: Establish autoencoder AE and hybrid generative adversarial network MGAN; S2: Power line channel modeling with joint AE and MGAN; Specifically, S1 includes: the autoencoder AE includes an encoder and a decoder; the encoder learns the potential representation y of the input sample x; the encoder reconstructs x at the output end using the information inside y; The general learning formula for deterministic AE is: Among them, θ opt is the optimal parameter of the neural network, δ is the distance measure, while E and F represent the encoder and decoder functions, θ1 and θ2 are the parameters of the encoder and decoder, respectively; When the feature space is subject to energy constraint EC, a regularization term is added to the cost function to track this restriction. The cost function for training AE is: Among them, A and b are the weight matrix and bias vector in the linear transformation, respectively, and both are usually constants; ||·||2 represents the l2 norm, that is, the Euclidean distance; the square of the l2 norm is the classic mean square error MSE; MGAN uses a mixture of multiple distributions instead of a single distribution in the standard GAN to approximate the real data distribution, while expanding the divergence of multiple distributions to cover different data modes; MGAN is equivalent to K generators G1: K , the game between the discriminator D and the classifier C; each generator G k Map z to x k =G k (z), thus inducing a single distribution K generators together get a mixture of K distributions, i.e., p in the data space model ; Sample G u (z) is selected as the output, and the index u is sampled from the multinomial distribution Mult(π), where π = [π1, π2, ..., π K ] is a predefined mixed coefficient; the discriminator D is used to distinguish fake samples from real samples; C is a multi-classifier used to determine the index tag of the generator corresponding to the fake sample; in order to minimize the number of parameters, except for the input layer, the K generators share parameters; except for the output layer, the multi-classifier C and the discriminator D share parameters; The training of MGAN is expressed as the following minimax objective function: Among them, C k (x) indicates x comes from G k The probability of , β>0 is the diversity hyperparameter, It's G k The probability distribution of generated data; the first two terms represent the interaction between the generator and the discriminator in the standard GAN; the last term represents the interaction between the generator and the classifier, which is the standard softmax loss of the multi-classifier, and its purpose is to maximize the entropy of the multi-classifier, that is, to maximize the difference between the data distribution generated by each generator and the data distribution generated by other generators; the strength of the interaction is controlled by β; Specifically, S2 is: let x be the input data, E:x→y be the function that maps x to y, then there is a deterministic relationship between the distributions of x and y; let the prior knowledge of the complexity of x be provided, that is, the distribution p(x), and the purpose becomes the distribution p(y) for low-dimensional features; if the function E is differentiable, according to the variable change rule, p(x) is obtained from p(y): Among them, det(·) is an operator for finding the determinant of a matrix; The power line channel data H is used as the input of AE, and the encoder block of the trained AE is used as the transformation function E(·) to distribute the feature y p feature (y) Modeling: The new feature y generated by MGAN g As the input of the decoder block F(·), the generated statistical distribution is p data (H) New channel H g ; Select the loss function of formula (2) to train AE, where δ is the MSE; add a regularization term in formula (2), that is: Among them, the second term is the normalized mean square error NMSE, which enhances the effect of data with smaller values; Make changes to the network training: (1) According to the average channel gain ACG, the power line channel data H is divided into K categories, and y is divided into K categories; (2) The training of AE remains unchanged. When training MGAN, we proportionally extract features from each class y according to the predefined mixing coefficient. k A small batch is extracted from the batch to form a large batch; (3) Determine y at C g When the corresponding generator index is marked, let C also recognize the real feature y k categories, prompting each G k Can generate matches for a certain class y k The new features of MGAN become: in, is the distribution of the original features of the kth class; where is the distribution of the original features of the kth class; The S2 is specifically: the training method of AE and MGAN in the power line channel is: S21: train AE using power line channel data H; S22: Extract low-dimensional features y of the power line channel data H using the encoder of AE: y=E(H); S23: Save the decoder module; S24: Use the original feature y to train MGAN and save the generator module; S25: From the distribution p used when training MGAN z The sampling noise z; S26: Generate new features y using the MGAN generator g :y g =G(z); S27: Using the decoder of AE to obtain new power line channel data H g :H g =F(y g ).