A neural network-based large-scale internet of things sampling and reconstruction method and system

Abstract

The application provides a large-scale Internet of Things sampling and reconstruction method and system based on a neural network, which firstly constructs a sampling point selection network by using a Gumbel-Softmax skill, then constructs a data reconstruction network by using a full connection layer, and simultaneously trains the sampling point selection network and the data reconstruction network by using Internet of Things node data of a training set in a training stage through a reasonable design of a loss function, so that the sampling point selection strategy and the data reconstruction strategy are given at one time. The application performs sampling point selection and signal reconstruction on large-scale Internet of Things data based on a neural network. Compared with previous methods, the application is a data-driven method, does not need a graph topology structure as prior knowledge, can only output the sampling point selection strategy and the strategy of reconstructing complete data according to sampling node data by inputting Internet of Things node data, and has certain advantages in reconstruction performance under the condition of a low sampling rate.

Description

Technical Field

[0001] This invention belongs to the field of graph signal processing technology, specifically relating to a method and system for large-scale Internet of Things (IoT) sampling and reconstruction based on neural networks. Background Technology

[0002] The Internet of Things (IoT) has experienced explosive growth in recent years, attracting widespread attention in both industry and academia. It is projected that by 2030, the connection density of IoT will reach 10 million devices per square kilometer, making data acquisition from such a massive network a challenging task. Unlike the independent signals of devices in cellular networks, data in sensor networks typically possess an inherent structure, which can serve as crucial prior knowledge for signal acquisition and processing. Utilizing network structure information, the overall network data can be acquired through signal sampling and reconstruction with minimal device data. In this context, Graph Signal Processing (GSP), as a powerful mathematical abstraction of networks and readily representing network structures, is emerging as a powerful tool for sensor network data sampling and reconstruction, gaining increasing attention as a research area. It defines traditional signal processing methods such as translation, filtering, and convolution from a graph perspective. When the graph structure is known, in graph signal processing, sampled signals from sensors can be sent to a data center, and the complete signal can then be reconstructed using graph signal processing methods.

[0003] Currently, there are graph signal sampling and reconstruction methods based on graph signal processing theory, such as minSpec, minForb, and FastSSS. minSpec and minForb treat graph signal processing as a sampling and reconstruction problem of band-limited signals in a linear space; minSpec minimizes the mean square error of the reconstructed signal, while minForb minimizes the maximum error. FastSSS, on the other hand, considers maximizing the use of atoms in the signal's representation dictionary. These methods share two main characteristics: 1) they require knowledge of the graph topology of the sensor nodes, which is difficult to obtain in practice; 2) they require eigenvalue decomposition or matrix inversion, which increases complexity when there are many sensor nodes, i.e., a large matrix size. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for large-scale Internet of Things sampling and reconstruction based on neural networks, so as to solve the problem of requiring known graph topology in the prior art and to make some improvements to performance.

[0005] To achieve the above objectives, the present invention employs the following technical solution:

[0006] A large-scale IoT sampling and reconstruction method based on neural networks includes the following steps:

[0007] Step 1: Normalize the randomly initialized neural network training parameters using the softmax function to obtain the class distribution of the neural network training parameters; establish random variables that follow a class distribution; establish a sampling matrix, which is obtained by sampling according to the distribution of the random variables using the Gumbel-Max method; perform differentiable estimation on the sampled matrix using the softmax function to obtain a sampling matrix that can be trained by the neural network, which is the sampling network.

[0008] Step 2: Build a DNN network as the reconstruction network. The DPP network has P layers, and each layer adopts a neural network with a linear structure.

[0009] Step 3: Train the sampling network and the reconstruction network using the training set data. During the training process, obtain the loss function and correct the parameters of the sampling network and the reconstruction network using the loss function to obtain the trained sampling network and the reconstruction network.

[0010] Step 4: Test the trained sampling network and reconstruction network using test set data to obtain the final sampling network and reconstruction network.

[0011] A further improvement of the present invention is that:

[0012] Preferably, in step 1, the normalization process is achieved by the following formula (5):

[0013]

[0014] In the formula, π i =[π i,0 ,π i,1 ,…,π i,N-1 ] represents the class distribution of the training parameters of the neural network, π i The sum of the elements in the middle is 1; (η) i,j ) is η i The j-th element, (η) i,j′ In order to represent η i All elements, where N is the total number of nodes.

[0015] Preferably, in step 1, the random variable is c. i ~Cat(N,π) i ), where π i The class distribution of the training parameters of the neural network; each row vector s in the sampling matrix i for:

[0016] s i =one_hot(c i (6)

[0017] In the formula, one_hot(·) represents the one-hot vector encoding operation.

[0018] Preferably, in step 1, the sampling matrix is ​​sampled according to the random variables using the following formula:

[0019]

[0020] In the formula, g i,j It is an independent and identically distributed random variable that follows a Gumbel(0,1) distribution. i,j These are the sampling network training parameters;

[0021] The differentiable estimation of the sampling matrix after distribution sampling is achieved by the softmax function using the following equation (8).

[0022]

[0023] in, Represents the distribution of the i-th category, π i,j Represents π i The j-th element The vector represents the total number of sampling points, and τ is the temperature coefficient. The closer the value of τ is to 0, the closer the vector is to the unique heat vector. The larger the value of τ, the closer the vector is to a uniform distribution.

[0024] Preferably, in step 2, the structural formula of the p-th layer is:

[0025]

[0026] in, It is a weight matrix. Both are offset vectors and trainable parameters, h (p) φ represents the number of neurons in the p-th layer. (p) Let x represent the activation function of the p-th layer. (p-1) This represents the output of layer p-1.

[0027] Preferably, 6. The method for large-scale IoT sampling and reconstruction based on neural networks according to claim 1, characterized in that, in step 2, the expression of the reconstructed network is:

[0028]

[0029] In the formula, W (i) ,b (i) Let represent the network parameters of the i-th layer, and P represent the total number of layers. This represents the sampled data.

[0030] Preferably, in step 3, the loss function is:

[0031]

[0032] In the formula, π i,j These are the sampling network training parameters.

[0033] Preferably, in step 3, the calculation formula for correcting the parameters of the sampling network and the reconstruction network using the loss function is as follows:

[0034]

[0035] In the formula, η represents the training parameters of the sampling network, W and b represent the training parameters of the reconstruction network, and ρ represents the penalty factor. The loss function is based on the mean squared error. This is the loss function after incorporating entropy, used to adjust the importance of different loss functions.

[0036] Preferably, in steps 3 and 4, both the training set data and the test set data are preprocessed using a factor analysis model.

[0037] A large-scale IoT sampling and reconstruction system based on neural networks, comprising:

[0038] A sampling network unit is used to normalize randomly initialized neural network training parameters using a softmax function to obtain the class distribution of the neural network training parameters; establish random variables that follow a class distribution; establish a sampling matrix, which is obtained by sampling the random variables according to the distribution using the Gumbel-Max method; and perform differentiable estimation on the sampled matrix using a softmax function to obtain a sampling matrix that can be used to train the neural network.

[0039] The reconstructed network unit is used to build the DNN network as a reconstruction network. The DPP network has a total of P layers, and each layer adopts a neural network with a linear structure.

[0040] The training unit is used to train the sampling network and the reconstruction network using training set data. During the training process, a loss function is obtained, and the parameters of the sampling network and the reconstruction network are corrected using the loss function to obtain the trained sampling network and the reconstruction network.

[0041] The testing unit is used to test the trained sampling network and reconstructed network using test set data to obtain the final sampling network and reconstructed network.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] This invention proposes a method and system for large-scale IoT sampling and reconstruction based on neural networks. The method first constructs a sampling point selection network using the Gumbel-Softmax technique, and then constructs a data reconstruction network using fully connected layers. By designing a reasonable loss function, both the sampling point selection network and the data reconstruction network are trained simultaneously using IoT node data from the training set during the training phase, providing both sampling point selection and data reconstruction strategies in one go. This invention performs sampling point selection and signal reconstruction on large-scale IoT data based on neural networks. Compared to previous methods, this invention is a data-driven method that does not require graph topology as prior knowledge. It can output both a sampling point selection strategy and a strategy for reconstructing complete data based on the sampling node data, simply by inputting IoT node data. In terms of reconstruction performance, this invention has certain advantages under low sampling rate conditions. Simulation experiments demonstrate that this invention has certain reconstruction performance advantages compared to existing graph signal processing-based methods when the IoT node topology is unknown and the sampling rate is low. This method solves the problem of sensor node sampling point selection and signal data reconstruction when the graph topology is unknown, extracting useful information from graph signal data and providing sampling point selection and graph signal reconstruction strategies.

[0044] Furthermore, this method is based on Large Scale IoT, treating IoT nodes as nodes in a graph topology and IoT node data as graph data, abstracting the entire system into a Graph Signal Processing (GSP) model. Under this model, this invention uses multiple sets of complete frequency-domain band-limited graph data as prior knowledge, designs appropriate neural network structures and loss functions, and performs supervised learning through neural networks. Finally, it provides sampling node selection and signal reconstruction strategies for Large Scale IoT, aiming to select appropriate IoT nodes and reconstruct all node data even without knowing the graph topology. Attached Figure Description

[0045] Figure 1 It is an Internet of Things (IoT) node model.

[0046] Figure 2 This is a schematic diagram of the network structure.

[0047] Figure 3 This is a comparison chart of sensor graph reconstruction error performance.

[0048] Figure 4 This is a performance comparison chart of Community graph reconstruction error.

[0049] Figure 5 This is a comparison chart of ER plot reconstruction error performance.

[0050] Figure 6 This is a comparison chart of BA diagram reconstruction error performance.

[0051] Figure 7 This is a comparison chart of the reconstruction error performance of Random Regular graphs.

[0052] Figure 8 This is a 15-point sampled map of minSpec on the Senor plot.

[0053] Figure 9 This is a 15-point sampled plot of minForb on the Senor plot.

[0054] Figure 10 This is a 15-point sampled plot of FastSSS on the Senor plot.

[0055] Figure 11 This is a 15-point sampling diagram on the Sensor image of the present invention.

[0056] Figure 12 This is a 15-point sampling plot of minSpec on the Community graph.

[0057] Figure 13 This is a 15-point sampling plot of minForb on the Community graph.

[0058] Figure 14 This is a 15-point sampled graph of FastSSS on the Community map.

[0059] Figure 15 This is a 15-point sampling diagram on the Community map of the present invention. Detailed Implementation

[0060] The present invention will now be described in further detail with reference to the accompanying drawings:

[0061] This invention discloses a method and system for large-scale Internet of Things (IoT) sampling and reconstruction based on neural networks. Figure 2 This is a schematic diagram of the network structure. The actual problem that this invention aims to solve is:

[0062]

[0063] in Represents the sampling matrix, Represents the sampled data. This represents the reconstructed sensor node data. The desired image signal reconstructed using deep learning methods is... To get as close as possible to the original graph signal f, i.e., their mean square error It should be as small as possible. For example... Figure 2As shown, the present invention constructs a network consisting of two parts, specifically a sampling network and a reconstruction network.

[0064] Figure 2 In the diagram, solid arrows represent the forward propagation process, and long dashed arrows represent the backward propagation process. The output of the sampling network is a sampling matrix. The output of the reconstructed network is the reconstructed graph signal.

[0065]

[0066] Where η is the training parameter of the neural network. It is the sampling matrix The approximation achieved by reconstructing the network implements a mapping Rec: Make:

[0067]

[0068] Specifically, the method includes the following steps:

[0069] Step 1: Data preprocessing and synthesis.

[0070] The system model of this invention is based on Large Scale Internet of Things (IoT), treating IoT nodes as nodes in a graph topology and IoT node data as graph data, thus abstracting the entire system into a graph signal processing model. For data preprocessing, an extended form of a traditional factor analysis model is used to represent the smoothed graph signal. Specifically, assuming there are N nodes in the IoT, the graph signal can be represented as follows: Its corresponding Graph Fourier Transform (GFT) is in It is the Laplace matrix The eigenvector matrix. In classic factor analysis models, a Gaussian distribution is used as the coefficients. The prior distribution of, in particular, assume It follows a descending, zero-mean, multivariate Gaussian distribution, with the following specific distribution form:

[0071]

[0072] in Representing the eigenvalue matrix The pseudo-inverse of , where Λ is the eigenvalue matrix of the Laplace matrix L. Combining the definition of the matrix U and its interpretation of the graphical Fourier transform, the above for The assumption actually indicates that the energy of the generated synthetic data signal is mainly concentrated in the low-frequency part, improving the smoothness of the graph signal. This invention generates multiple sets of test and training graph data using the above method.

[0073] See Figure 1 , Figure 1 It is an IoT node model. GSP theory usually treats each node as a graph node, node data as graph data, and assigns weights with certain physical meanings to the graph as edges. Graph data sampling and reconstruction actually aims to obtain the best possible reconstruction performance by selecting as few nodes as possible.

[0074] Step 2, constructing the sampling network, specifically the process of constructing trainable sampling network coefficients:

[0075] First, randomly initialize the neural network training parameters η. i =[η i,0 ,η i,1 ,…,η i,N-1 Let it follow a Gaussian distribution, η i It can be viewed as an unnormalized probability distribution. In order to make it conform to a probability distribution, the softmax function is used to normalize it. The normalization process is shown in the following equation (5):

[0076]

[0077] The obtained probability distribution π i =[π i,0 ,π i,1 ,…,π i,N-1 The sum of the elements is 1, that is It can be viewed as a categorical distribution and used as the training parameters of the network.

[0078] Then, let the random variable c i Follows the category distribution π i , denoted as c i ~Cat(N,π) i In order for the sampling network to follow the random variable c i The distribution generates a desired sampling matrix. Each row of the sampling matrix S is defined as

[0079] s i =one_hot(c i (6)

[0080] Here, one_hot(·) represents the one-hot vector encoding operation. However, neural networks cannot directly sample from the distribution. In order for the neural network to be able to follow the probability distribution πi =[π i,0 ,π i,1 ,…,π i,N-1 Sampling is performed using the Gumbel-Max technique. For c i ~Cat(N,π) i Sampling according to a probability distribution can be performed using the following formula:

[0081]

[0082] Among them, g i,j It is an independent and identically distributed random variable that follows a Gumbel(0,1) distribution. i,j This involves sampling network training parameters. However, the argmax operation is non-differentiable, so it cannot be used for backpropagation training of neural networks.

[0083] Finally, the argmax operation is estimated to be differentiable using softmax, making it trainable for the neural network, as shown in the following equation.

[0084]

[0085] Where τ is the temperature coefficient, the closer τ is to 0, the closer the vector is to the one-hot vector, and the larger the τ value, the closer the vector is to a uniform distribution. In the following training process, τ is gradually made to approach 0, so that the final vector will approximate the form of equation (6), as shown in the following equation.

[0086]

[0087] Obtain the sampling network

[0088] Step 3, Reconstructing the Network Establishment

[0089] like Figure 2 As shown, the DNN network needs to implement a mapping from the sampled signal to the reconstructed signal. Rec: Assume the DNN has P layers, each using a linear neural network structure, with the structure of the p-th layer as shown in the equation. (9)

[0090]

[0091] in, It is a weight matrix. Both are offset vectors and trainable parameters, h (p) φ represents the number of neurons in the p-th layer. (p) Let represent the activation function of the p-th layer.

[0092] The overall reconstructed network structure can be represented as:

[0093]

[0094] In the actual implementation, let P = 4.

[0095] Step 3: Training Set Training Phase. The training set data generated in Step 1 is input into the neural network of this invention for training, and the resulting network model parameters are saved. The neural network can be divided into two parts: a sampling network and a reconstruction network. The sampling network uses the Gumbel-Softmax technique, and the reconstruction network uses a fully connected DNN network.

[0096] Table 1. Node Data Sampling Point Selection and Signal Reconstruction Learning Algorithm

[0097]

[0098]

[0099] Table 1 shows the steps of the node data sampling point selection and signal reconstruction learning algorithm. From the above description, it can be seen that the learning parameters of the neural network are... W = {W (p) ,p=0,...,P-1}, and b={b (p) ,p=0,...,P-1}. Where η is the training parameter of the sampling network, and W and b are the training parameters of the reconstruction network. During training, the reconstructed graph signal output by the network is reduced. The network parameters are trained using the mean square error between the signal f and the original image signal. The loss function based on the mean square error can be written as:

[0100]

[0101] Where, N samples This represents the number of samples. To improve s i To improve the convergence performance towards a one-hot vector, we can add an entropy term to the loss function, aiming for a higher entropy π. i =[π i,0 ,π i,1 ,…,π i,N-1 More are penalized, and the loss function takes the form:

[0102]

[0103] During training, the parameters of both the sampling network and the reconstructed network are updated simultaneously. The final optimization problem is:

[0104]

[0105] Where ρ is the penalty factor, used to adjust the importance of different loss functions. In practice, this invention uses the Adam optimizer to obtain the final training parameters η, W, and b. Furthermore, during training, the sampling network and the reconstruction network have different learning rates. α is used respectively... η and α W,b This means that during the training phase, two optimizers will be set up. In one iteration, when η is updated using the backpropagation algorithm, W and b will remain unchanged, and vice versa. α is set... η >α W,b As mentioned earlier, the temperature coefficient τ in softmax will gradually approach 0 during training. τ is initialized to 5 and gradually decreased to 0.5 during training. The η parameter is randomly initialized using a Gaussian distribution.

[0106] Table 2 shows the experimental parameter settings. The parameters for the training process are listed.

[0107] Table 2 Experimental parameter settings

[0108]

[0109] Step 3: Test Set Validation Phase. First, the sampling network obtained in Step 2... It's just an approximation of the sampling matrix; we use a threshold function to... The sampled data is transformed into a strict sampling matrix S, which allows sampling of the test set. The sampled test set data is then input into the reconstructed network model saved in step 2 to verify the effectiveness of the method of the present invention.

[0110] This invention also discloses a large-scale IoT sampling and reconstruction system based on neural networks for the above-described method, comprising:

[0111] A sampling network unit is used to normalize randomly initialized neural network training parameters using a softmax function to obtain the class distribution of the neural network training parameters; establish random variables that follow a class distribution; establish a sampling matrix, which is obtained by sampling the random variables according to the distribution using the Gumbel-Max method; and perform differentiable estimation on the sampled matrix using a softmax function to obtain a sampling matrix that can be used to train the neural network.

[0112] The reconstructed network unit is used to build the DNN network as a reconstruction network. The DPP network has a total of P layers, and each layer adopts a neural network with a linear structure.

[0113] The training unit is used to train the sampling network and the reconstruction network using training set data. During the training process, a loss function is obtained, and the parameters of the sampling network and the reconstruction network are corrected using the loss function to obtain the trained sampling network and the reconstruction network.

[0114] The testing unit is used to test the trained sampling network and reconstructed network using test set data to obtain the final sampling network and reconstructed network.

[0115] Figure 3 — Figure 7 This is a demonstration of the data reconstruction performance of the present invention. The horizontal axis represents the number of sampling points, and the vertical axis represents the mean square error between the reconstructed data and the original data. Figure 3 This is a comparison chart of sensor graph reconstruction error performance. Figure 4 This is a performance comparison chart of Community graph reconstruction error. Figure 5 This is a comparison chart of ER plot reconstruction error performance. Figure 6 This is a comparison chart of BA diagram reconstruction error performance. Figure 7 This is a performance comparison chart of Random Regular graph reconstruction errors. The graph topology generation used the PyGSP toolkit, a Python version of the MATLAB graph signal processing toolkit GSPBox. Experiments showed numerical results for the algorithm's performance on Sensor, Random Regular, Community, Barbasi Albert (BA), and Erdos Renyi (ER) graphs. minSpec, minForb, and FastSSS were selected as comparison methods; these three methods use the graph topology as prior knowledge, while this invention uses complete IoT node data as prior knowledge. Specifically, the parameters k = 10 for the Random Regular graph, where k represents the degree of each node; m0 = 6 and m = 3 for the BA graph, where m0 represents the initial number of nodes and m represents the number of nodes generated each time; p = 0.01 for the ER graph, representing the node connection probability; and Nc = 10 for the Community graph, where Nc represents the number of communities. All other parameters for all graphs are set to default values. Except for the Sensor and Community graphs, which are weighted graphs, the other graphs are unweighted graphs.

[0116] from Figures 3-8The results show that the proposed method achieves good reconstruction performance on Sensor graphs. On Community graphs, the slope of the proposed method's curve is relatively low, giving it an advantage when there are few sampling points. However, its performance advantage decreases as the number of sampling points increases, but it still outperforms minSpec. FastSSS shows relatively good performance on Community graphs. On ER, BA, and Random Regular graphs, the proposed method's performance is close to minSpec and minForb, but slightly worse. Overall, deep neural network methods show better performance on weighted graphs, especially when there are few sampling points (i.e., a low sampling rate), where deep neural network methods are optimal. However, the slope decreases and performance deteriorates as the number of sampling points increases. On unweighted graphs, the proposed method is slightly worse than minSpec and minForb, with a generally consistent trend.

[0117] Figures 9-15 The paper demonstrates different sampling point selection strategies for the Sensor and Community. The results show that minSpec and minForb have very similar point selection strategies, resulting in consistent performance. However, at lower sampling rates, this invention can select sampling points more evenly across the entire image, thus outperforming other methods in data reconstruction performance at lower sampling rates.

[0118] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A large-scale IoT sampling and reconstruction method based on neural networks, characterized in that, Includes the following steps: Step 1: Normalize the randomly initialized neural network training parameters using the softmax function to obtain the class distribution of the neural network training parameters; establish random variables that follow a class distribution; establish a sampling matrix, which is obtained by sampling according to the distribution of the random variables using the Gumbel-Max method; perform differentiable estimation on the sampled matrix using the softmax function to obtain a sampling matrix that can be trained by the neural network, which is the sampling network. Step 2: Build a DNN network as the reconstruction network. The DPP network has P layers, and each layer adopts a neural network with a linear structure. In step 2, the first The structural formula of the layer is: (5) in, It is a weight matrix. Both are offset vectors and are trainable parameters. Indicates the first The number of neurons in a layer Indicates the first The activation function of the layer, represent p -1 layer output; In step 2, the expression for reconstructing the network is: (6) In the formula, Representing the i The network parameters of the layer, P Represents the total number of floors. This represents the sampled data; Step 3: Train the sampling network and the reconstruction network using the training set data. During the training process, obtain the loss function and correct the parameters of the sampling network and the reconstruction network using the loss function to obtain the trained sampling network and the reconstruction network. Step 4: Test the trained sampling network and reconstruction network using test set data to obtain the final sampling network and reconstruction network; The training and test sets consist of IoT nodes, which are considered as nodes in a graph topology. The data from these IoT nodes is treated as graph data, and the entire system is abstracted as a graph signal processing model. The data preprocessing process is as follows: Assuming there are N nodes in the IoT, the graph signal can be represented as... Its corresponding graphical Fourier transform is ,in It is the Laplace matrix The eigenvector matrix; in the classic factor analysis model, a Gaussian distribution is used as the coefficients. The prior distribution, assuming It follows a descending, zero-mean, multivariate Gaussian distribution, with the following specific distribution form: (4) in, Representing the eigenvalue matrix The pseudo-reversal, among which It is the Laplace matrix eigenvalue matrix; The neural network-based large-scale IoT sampling and reconstruction method is used for selecting sensor node sampling points and reconstructing signal data when the graph topology is unknown.

2. The method for large-scale IoT sampling and reconstruction based on neural networks according to claim 1, characterized in that, In step 1, the normalization process is achieved by the following formula (5): (5) In the formula, The class distribution of the training parameters of the neural network. The sum of the elements in the middle is 1; for No. j One element, In order to express All elements, N This represents the total number of nodes.

3. The method for large-scale IoT sampling and reconstruction based on neural networks according to claim 1, characterized in that, In step 1, the random variable is ,in The class distribution of the training parameters of the neural network; each row vector in the sampling matrix for: (6) In the formula, This represents one-hot vector encoding operations.

4. The method for large-scale IoT sampling and reconstruction based on neural networks according to claim 1, characterized in that, In step 1, the sampling matrix is ​​sampled according to the random variables using the following formula: (7) In the formula, They are independent, identically distributed random variables that follow a Gumbel(0,1) distribution. These are the sampling network training parameters; The differentiable estimation of the sampling matrix after distribution sampling is achieved by the softmax function using the following equation (8). (8) in, Represents the distribution of the i-th category. represent No. j element Represents the total number of sampling points. It is the temperature coefficient. The closer the value is to 0, the closer the vector is to a one-hot vector. The larger the value, the closer the vector is to a uniform distribution.

5. The method for large-scale IoT sampling and reconstruction based on neural networks according to claim 1, characterized in that, In step 3, the loss function is: (7) In the formula, These are the sampling network training parameters.

6. The method for large-scale IoT sampling and reconstruction based on neural networks according to claim 1, characterized in that, In step 3, the calculation formula for correcting the parameters of the sampling network and the reconstruction network using the loss function is as follows: (8) In the formula, To sample network training parameters, and To reconstruct the network training parameters, As a penalty factor, The loss function is based on the mean squared error. This is the loss function after incorporating entropy, used to adjust the importance of different loss functions.

7. A large-scale Internet of Things (IoT) sampling and reconstruction system based on neural networks for implementing the method of claim 1, characterized in that, include: The sampling network unit is used to normalize the randomly initialized neural network training parameters using the softmax function to obtain the class distribution of the neural network training parameters; A random variable is established, which follows a categorical distribution; a sampling matrix is ​​established, which is obtained by sampling the random variable according to the distribution using the Gumbel-Max method; the sampling matrix after distribution sampling is obtained by performing differentiable estimation on the sampling matrix after distribution sampling using the softmax function, and a sampling matrix that can be trained by the neural network is obtained. The reconstructed network unit is used to build the DNN network as a reconstruction network. The DPP network has a total of P layers, and each layer adopts a neural network with a linear structure. The training unit is used to train the sampling network and the reconstruction network using training set data. During the training process, a loss function is obtained, and the parameters of the sampling network and the reconstruction network are corrected using the loss function to obtain the trained sampling network and the reconstruction network. The testing unit is used to test the trained sampling network and reconstructed network using test set data to obtain the final sampling network and reconstructed network.

Images