Deep learning based wireless tomographic target localization method and system

Abstract

The application discloses a wireless tomography target positioning method and system based on deep learning. The application generates an RSS data set based on a wireless sensor network for passive positioning, then trains the neural network by using the RSS data set, and finally inputs the measured RSS attenuation value data of the wireless sensor network for passive positioning into the neural network for wireless tomography target positioning. The application realizes the neural network for wireless tomography passive positioning based on CNN and Transformer, saves the positioning time, and improves the positioning accuracy. The application reduces the number of wireless tomography target positioning sensor nodes, reduces the complexity, and saves the cost.

Description

Technical Field

[0001] This invention belongs to the field of passive target localization technology, specifically, it relates to a wireless tomographic target localization method and system based on deep learning. Background Technology

[0002] Positioning technologies based on Received Signal Strength (RSS) can be divided into range-based positioning and range-free positioning.

[0003] RSS-based wireless detection technology requires the deployment of multiple wireless sensor nodes in space, with the target carrying one wireless node. The distance between the unknown node and the beacon node is deduced by measuring the RSS distance between them, thus determining the target's location. The problem with this technology is that the target must carry additional equipment, making device-free positioning impossible.

[0004] RSS-based ranging-free positioning technologies include fingerprint positioning and radio tomographic imaging (RTI). For fingerprint positioning, the fingerprint database needs to be updated whenever the environment changes, resulting in high maintenance costs and making it unsuitable for real-time positioning. Radio tomographic imaging (RTI) is an emerging low-power, low-cost wireless detection technology. Its principle is to construct the relationship between RSS changes and target location using a shadow weight model, and then derive the target location image by solving unwell-posed equations. Specifically, multiple wireless nodes are placed around the monitoring area, and the monitoring area is imaged by measuring the shadow fading of each link, thereby inferring the target's location. This method is widely used in the positioning field due to its intuitive presentation.

[0005] In solving unbalanced equation problems, Tikhonov regularization is widely used. However, the image reconstruction accuracy obtained by this method is insufficient, with problems such as inaccurate reconstruction of target contour edges and target wake aliasing. It is difficult to restore the contour information of objects and targets in the environment, and its ability to distinguish targets is limited in multi-target localization scenarios. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to improve the positioning performance of wireless tomographic target localization, increase the accuracy of target contours and the resolution of multi-target localization, and reduce time complexity, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a passive localization method for wireless tomography based on deep learning, including the steps of generating an RSS dataset using a wireless sensor network based on passive localization, building a neural network based on CNN and TransformerEncoder and training the neural network using the RSS dataset, measuring RSS attenuation data using a wireless sensor network based on passive localization, and localizing the target in wireless tomography.

[0008] In the steps of generating an RSS dataset based on a passive positioning wireless sensor network, n wireless sensor nodes are arranged around the positioning area, forming a wireless sensor network consisting of n(n-1) wireless links. The wireless sensor nodes send and receive signals to locate targets within the positioning area. The RSS data includes pixel matrix data and RSS attenuation value data. The pixel matrix data is matrix data in which the positioning area is divided into grids, with each grid representing one pixel. Grids containing targets and those not containing targets are represented by different values. The RSS attenuation value data is vector data composed of wireless link attenuation values. The RSS dataset is obtained through actual measurements or generated using a wireless signal attenuation model. The RSS dataset includes an RSS training set and an RSS validation set.

[0009] In the step of building a neural network based on CNN and Transformer Encoder and training the neural network using the RSS dataset, the neural network built based on CNN and Transformer Encoder includes an input layer, a first CNN layer, a first Transformer Encoder layer, a first batch normalization and activation layer, a second CNN module, a second Transformer Encoder module, a second batch normalization and activation layer, a flattening and deactivation layer, and a fully connected and activation function layer connected in sequence. The RSS decay value data in the RSS dataset is used as the input of the neural network, and the pixel matrix data in the RSS dataset is used as the output of the neural network for training. The number of neurons in the first layer of the neural network is equal to the dimension of the RSS decay value data, and the number of neurons in the last layer of the neural network is equal to the dimension of the pixel matrix data.

[0010] In the target localization step of wireless tomography, a trained neural network is used. The measured RSS data obtained in the step of measuring the RSS attenuation data of the wireless sensor network based on passive localization is used as the input, and the output is the pixel matrix of the reconstructed image. The center of the pixel with the highest brightness in the pixel matrix is ​​the estimated target position.

[0011] Furthermore, the RSS dataset is generated using a wireless signal attenuation model, and the steps include:

[0012] In step S101 of setting initialization parameters, the positioning area is evenly divided into P pixels, and the number of wireless sensor nodes is set to n. The n wireless sensor nodes are arranged around the positioning area, and each node sends and receives signals, thereby forming L = n(n-1) wireless links.

[0013] Step S102 involves generating pixel matrix data; in this step, a total of N pixels are generated. total The pixel matrix data of the image is generated as follows: For each image, m targets are randomly generated in the positioning area. Each target is a square with a side length of h pixels. The target positions are randomly generated in the positioning area, thereby generating pixel matrix data Δx, where m is a random natural number not greater than 5 and h is a natural number not greater than 5.

[0014] Step S103 involves generating RSS attenuation data Δr using a wireless signal attenuation model. In this step, RSS attenuation data Δr, which consists of RSS attenuation values ​​for all links, is generated based on the pixel matrix data Δx of each image. The RSS attenuation data Δr is generated using the following wireless signal attenuation model:

[0015] Δr=WΔx+χ η

[0016] Among them, Δr=[Δr1, Δr2,..., Δr L ] T , Δr i Let Δx represent the RSS attenuation value of the i-th link, where Δx = [Δx1, Δx2, ..., Δx]. P ] T Δx j Let represent the value of the j-th pixel in the localization region, and W be an L×P dimensional weight matrix, where [W] i，j =w i，j w i，j χ represents the influence factor of the value of pixel j on the attenuation value of link i; η It is measurement noise that follows a Gaussian distribution, χ η ~N(0, η) 2 );

[0017] In step S104, the attenuation value data Δr of the RSS is normalized. The formula for normalizing the attenuation value data Δr of the RSS in this step is:

[0018]

[0019] In the formula, Δr i This represents the RSS attenuation value of the i-th link;

[0020] Step S105: Generating the RSS dataset; in this step, N totalThe pixel matrix data Δx and the RSS decay value data Δr of the image are used as the RSS dataset. The RSS dataset is divided proportionally, with one part used as the training set and the other part used as the validation set for training the neural network.

[0021] Furthermore, in step S103, w i,j The calculation is performed using an elliptic model, i.e.:

[0022]

[0023] In the formula, It is the distance between the transmitting and receiving nodes on the i-th link. and λ represents the distance from the center of the j-th pixel to both ends of the i-th link's transceiver node, and λ is a controllable variable that adjusts the length of the minor axis of the ellipse.

[0024] Furthermore, the steps of training the neural network using the RSS dataset include forward propagation step S210 and backward propagation step S220.

[0025] Step S210, in which the neural network performs forward propagation, includes:

[0026] Step S211, extract N from the RSS dataset total The RSS decay value data Δr is fed into the input layer in batches, with each batch containing N RSS decay value data Δr. After being fed into the input layer, the data is reconstructed into an image vector.

[0027] Step S212: The reshaped image vector is fed into the first CNN layer for two-dimensional convolution to extract features and the extracted features are reshaped into the input vector of the first Transformer Encoder layer.

[0028] In step S213, the input vector sequentially passes through the second feature extraction of the first Transformer Encoder layer, the first batch normalization and activation operation of the first batch normalization and activation layer (BN&ReLU), the third feature extraction of two-dimensional convolution in the second CNN layer, the fourth feature extraction of the second Transformer Encoder layer, the second batch normalization and activation operation of the batch normalization and activation layer (BN&ReLU), and finally is sent to the flattening and deactivation layer for flattening and deactivation processing.

[0029] Step S214: The flattened and deactivated vector is fed into a fully connected layer and activation function layer. Through the fully connected layer and the Sigmoid activation function, the final output vector is transformed into a probability representation, thus completing the classification of the reconstructed pixel matrix data.

[0030] In step S220, the neural network is trained by backpropagation and the hyperparameters are adjusted. In this step, a loss function is used to train the neural network by backpropagation. After training, a trained neural network is obtained. Then, the hyperparameters of the constructed neural network are adjusted according to training experience. The hyperparameters are finally determined based on the performance on the validation set to achieve the optimal localization effect of the neural network.

[0031] Furthermore, in step S211, the vector of each batch of images is in the form of (N, C, H, W), where N is the size of the batch, C is the number of channels and C = 1, H is the height of the image and H = n, and W is the width of the image and W = n-1.

[0032] In step S212, the first CNN layer uses a convolutional kernel of size (a, a-1) and has C1 output channels. The input vector of the first TransformerEncoder layer is in the form of (N, S, E), where S = (n-a+1). 2 , E = C1.

[0033] Furthermore, the first Transformer Encoder layer and the second Transformer Encoder layer have the same network structure, both including a multi-head self-attention layer and a feedforward layer connected in sequence, and residual connections are used between the input and output of each attention layer and the feedforward layer.

[0034] Furthermore, the second CNN layer in step S213 is a multi-layer convolutional layer, with a batch normalization (BN) layer used between every two convolutional layers to improve the network convergence speed and mitigate gradient vanishing during training.

[0035] Furthermore, in step S220, the loss function is binary cross-entropy (BCE).

[0036] This invention also provides a passive localization system for wireless tomography based on deep learning, characterized by comprising a passive localization wireless sensor network, an RSS data measurement device, an RSS dataset generation and recording device, and a neural network computing device based on CNN and Transformer Encoder; wherein,

[0037] The passive positioning wireless sensor network consists of wireless sensor nodes. n wireless sensor nodes are arranged around the positioning area to form a network consisting of n(n-1) wireless links. The wireless sensor nodes send and receive signals to locate the target in the positioning area. The wireless sensor nodes are connected to the RSS data measurement device.

[0038] The RSS data measurement device is used to control the wireless sensor node to send and receive signals, measure the RSS attenuation value data of each wireless link arranged around the positioning area, and send the RSS attenuation value data to the RSS dataset generation and recording device.

[0039] The RSS data generation and recording device is used to generate or record the RSS dataset for training neural networks, as well as the RSS attenuation data of the positioning area where the target to be located is located. The RSS dataset includes pixel matrix data and RSS attenuation data. The pixel matrix data is matrix data in which the positioning area is divided into grids, with each grid representing one pixel. Grids containing the target and those not containing the target are represented by different values. The RSS attenuation data is vector data composed of wireless link RSS attenuation data. The RSS attenuation data is generated in two ways: firstly, through actual measurement using an RSS data measurement device and a passive positioning wireless sensor network; and secondly, through calculation using a wireless signal attenuation model. The RSS dataset includes an RSS training set and an RSS validation set.

[0040] The neural network based on CNN and Transformer Encoder includes, in sequence, an input layer, a first CNN layer, a first Transformer Encoder layer, a first batch normalization and activation layer, a second CNN module, a second Transformer Encoder module, a second batch normalization and activation layer, a flattening and deactivation layer, and a fully connected and activation function layer. The neural network is trained using an RSS dataset, with RSS decay values ​​from the RSS dataset as input and pixel matrix data from the RSS dataset as output. The number of neurons in the first layer of the neural network equals the dimension of the RSS decay values, and the number of neurons in the last layer equals the dimension of the pixel matrix data. The trained neural network takes RSS decay values ​​of the target location region as input and outputs a pixel matrix containing the target location.

[0041] Furthermore, the RSS data generation and recording device includes the following modules for generating RSS datasets: an initialization parameter setting module, a pixel matrix generation module, an RSS attenuation value data generation module, an RSS attenuation value data normalization module, and an RSS dataset generation module; among which...

[0042] The initialization parameter setting module is used to evenly divide the positioning area into P pixels, set the number of wireless sensor nodes to n, and arrange the n wireless sensor nodes around the positioning area. Each node sends and receives signals, thereby forming L = n(n-1) wireless links.

[0043] The pixel matrix data generation module is used to generate N totalThe pixel matrix data of the image is generated as follows: For each image, m targets are randomly generated in the positioning area. Each target is a square with a side length of h pixels. The target positions are randomly generated in the positioning area, thereby generating pixel matrix data Δx, where m is a random natural number not greater than 5 and h is a natural number not greater than 5.

[0044] The RSS attenuation data generation module is used to generate RSS attenuation data Δr using a wireless signal attenuation model. This module generates RSS attenuation data Δr for all links based on the pixel matrix data Δx of each image. The RSS attenuation data Δr is generated using the following wireless signal attenuation model:

[0045] Δr=WΔx+χ η

[0046] Among them, Δr=[Δr1, Δr2,..., Δr L ] T , Δr i Let Δx represent the RSS attenuation value of the i-th link, where Δx = [Δx1, Δx2, ..., Δx]. P ] T Δx i Let represent the value of the j-th pixel in the localization region, and W be an L×P dimensional weight matrix, where [W] i，j =w i，j w i，j χ represents the influence factor of the value of pixel j on the attenuation value of link i; η It is measurement noise that follows a Gaussian distribution, χ η ~N(0, η) 2 );

[0047] The RSS attenuation value data normalization module, in step S104, normalizes the RSS attenuation value data Δr. The formula for normalizing the RSS attenuation value data Δr in this module is:

[0048]

[0049] In the formula, Δr i This represents the RSS attenuation value of the i-th link;

[0050] The RSS dataset generation module is used to generate RSS dataset 5; in this module, N total The pixel matrix data Δx and the RSS decay value data Δr of the image are used as the RSS dataset. The RSS dataset is divided proportionally, with one part used as the training set and the other part used as the validation set for training the neural network.

[0051] Beneficial effects

[0052] This invention uses a small number of nodes to classify each pixel in the positioning area, resulting in clearer image reconstruction capabilities. This invention reduces the number of sensor nodes, lowers the complexity of the positioning system, and saves positioning costs.

[0053] This invention can create datasets using a shadow fading model and can also learn features directly from data without system model information, exhibiting stronger generalization capabilities. Furthermore, the created dataset focuses only on the location of the target, without imaging the entire environment. This signal attenuation model can filter out signal propagation loss, thus offering greater applicability compared to other attenuation models.

[0054] This invention implements a neural network for passive localization in wireless tomography based on CNN and Transformer, achieving a significant improvement in overall average localization error compared to the optimal average localization error. Attached Figure Description

[0055] Figure 1 This is a flowchart of the passive localization method for wireless tomography based on deep learning, as described in this invention.

[0056] Figure 2 This is a schematic diagram of the passive positioning wireless sensor network in this invention.

[0057] Figure 3 This is a diagram of the neural network framework structure based on Transformer Encoder and CNN in this invention.

[0058] Figure 4 This is a structural diagram of the neural network in a specific embodiment of the present invention.

[0059] Figure 5 This is the specific internal structure of the Transformer Encoder layer in this invention.

[0060] Figure 6 This invention relates to the specific internal structure of the second CNN layer.

[0061] Figure 7 This is a real-world diagram of the target being measured in an embodiment of the present invention.

[0062] Figure 8 This is a schematic diagram of the positioning results in an embodiment of the present invention.

[0063] Figure 9 This is a block diagram of the wireless tomography passive localization system based on deep learning, as described in this invention. Detailed Implementation

[0064] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0065] Example 1

[0066] This embodiment is used to describe in detail the passive localization method for wireless tomography based on deep learning of the present invention.

[0067] like Figure 1 As shown, the passive localization method for wireless tomography based on deep learning of the present invention includes the following steps: step S100, generating an RSS dataset based on a passive localization wireless sensor network; step S200, building a neural network based on CNN and TransformerEncoder and training the neural network using the RSS dataset; step S300, measuring RSS attenuation data based on a passive localization wireless sensor network; and step S400, localizing the target in wireless tomography.

[0068] In step S100 of generating an RSS dataset based on a passive positioning wireless sensor network, n wireless sensor nodes are arranged around the positioning area, forming a wireless sensor network consisting of n(n-1) wireless links. The wireless sensor nodes send and receive signals to locate targets within the positioning area. The RSS data includes pixel matrix data and RSS attenuation value data. The pixel matrix data is matrix data in which the positioning area is divided into grids, with each grid representing one pixel. Grids containing targets and those not containing targets are represented by different values. The RSS attenuation value data is vector data composed of wireless link attenuation values. The RSS dataset is obtained through actual measurement or generated using a wireless signal attenuation model. The RSS dataset includes an RSS training set and an RSS validation set.

[0069] In some specific embodiments of the present invention, the steps of generating the RSS dataset using a wireless signal attenuation model include: step S101 of setting initialization parameters, step S102 of generating pixel matrix data, step S103 of generating RSS attenuation value data Δr through the wireless signal attenuation model, step S104 of normalizing the RSS attenuation value data Δr, and step S105 of generating the RSS dataset.

[0070] In step S101 of setting initialization parameters, the positioning area is uniformly divided into P pixels, for example... Figure 2 The square region shown can be divided into P = p 2 There are n pixels, where p is the side length of the square. There are n wireless sensor nodes, which are arranged around the positioning area. Each node sends and receives signals, thus forming L = n(n-1) wireless links.

[0071] In step S102, which generates pixel matrix data, a total of N are generated. totalThe pixel matrix data of the image is generated as follows: For each image, m targets are randomly generated in the positioning area. Each target is a square with a side length of h pixels. The target positions are randomly generated in the positioning area. This generates pixel matrix data Δx, where m is a random natural number not greater than 5, i.e., m is a random natural number between 1 and 5, and h is a natural number not greater than 5, i.e., 1≤h≤5.

[0072] In step S103, which generates RSS attenuation data Δr using a wireless signal attenuation model, RSS attenuation data Δr, which is an element representing the RSS attenuation value of all links, is generated based on the pixel matrix data Δx of each image. The RSS attenuation data Δr is generated using the following wireless signal attenuation model:

[0073] Δr=WΔx+χ η

[0074] Among them, Δr=[Δr1, Δr2,..., Δr L ] T , Δr i Let Δx represent the RSS attenuation value of the i-th link, where Δx = [Δx1, Δx2, ..., Δx]. P ] T Δx j Let represent the value of the j-th pixel in the localization region, and W be an L×P dimensional weight matrix, where [W] i，j =w i，j w i,j χ represents the influence factor of the value of pixel j on the attenuation value of link i; η It is measurement noise that follows a Gaussian distribution, χ η ~N(0, η) 2 This invention focuses only on the location of the target to be located, without imaging the environment of the entire area. The signal attenuation model can filter out the propagation loss of the signal, thus having stronger applicability compared with other attenuation models.

[0075] In some specific embodiments of the present invention, in step S103, w i，j The calculation is performed using an elliptic model, i.e.:

[0076]

[0077] In the formula, It is the distance between the transmitting and receiving nodes on the i-th link. and λ represents the distance from the center of the j-th pixel to both ends of the i-th link transceiver node, and λ is a controllable variable that adjusts the length of the minor axis of the ellipse. The value is adjusted according to the size and shape of the positioning area, and in indoor passive positioning scenarios, it is generally set between 0.5 and 1m.

[0078] In step S104, the formula for normalizing the RSS attenuation data Δr is as follows:

[0079]

[0080] In the formula, Δr i This represents the RSS attenuation value of the i-th link;

[0081] In step S104 of generating the RSS dataset, N total Magnitude (e.g., N) total The pixel matrix data Δx and RSS decay value data Δr of the image (=40000) are used as the RSS dataset. The RSS dataset is divided proportionally, with one part used as the training set and the other part as the validation set for training the neural network. For example, 90% is used as the training set and 10% as the validation set, with the proportion limited to achieve the training objective.

[0082] In step S200, which involves building a neural network based on CNN and Transformer Encoder and training the neural network using the RSS dataset, as follows: Figure 3 As shown, the neural network built based on CNN and Transformer Encoder includes an input layer, a first CNN layer, a first Transformer Encoder layer, a first batch normalization and activation layer, a second CNN module, a second Transformer Encoder module, a second batch normalization and activation layer, a flattening and deactivation layer, and a fully connected and activation function layer connected in sequence. The RSS decay value data in the RSS dataset is used as the input of the neural network, and the pixel matrix data in the RSS dataset is used as the output of the neural network for training. The number of neurons in the first layer of the neural network is equal to the dimension of the RSS decay value data, and the number of neurons in the last layer of the neural network is equal to the dimension of the pixel matrix data.

[0083] This invention employs a network structure combining CNN and Transformer Encoder to improve the model's generalization ability. Both CNN and Transformer Encoder's attention mechanisms play a role in feature extraction. The difference lies in their approaches: CNN is essentially a simplified version of Self-Attention, as it only considers information from the local receptive field; while Self-Attention is more like a complex version of CNN, automatically learning which vectors are relevant to the current vector, taking into account global information. When training data is limited, Self-Attention is prone to overfitting, while CNN performs better; when training data is abundant, Self-Attention can utilize more information, while CNN performs relatively worse. Therefore, this invention effectively combines CNN and Transformer Encoder to achieve good model performance regardless of the amount of training data. Multi-Head Attention, due to the presence of multiple queries (Q), performs Self-Attention operations in parallel, and then concatenates the output vectors, effectively executing multiple attention mechanisms. Each time, the parameters differ, allowing the extraction of different information; in simpler terms, different focuses result in different perceptions. Therefore, it can capture richer feature information and performs better than Self-Attention. In addition, the feedforward neural network in the Transformer Encoder maps the attention results at each location to a larger-dimensional feature space, then uses ReLU to introduce non-linearity for filtering, and finally restores the original dimension. Therefore, the Transformer Encoder provides a non-linear transformation unit compared to simple Self-Attention.

[0084] The neural networks used in some specific embodiments of the present invention are as follows: Figure 4 As shown. See also Figure 3 and Figure 4 The steps for training the neural network using the RSS dataset include forward propagation step S210 and backward propagation step S220.

[0085] Step S210, in which the neural network performs forward propagation, includes:

[0086] Step S211, extract N from the RSS dataset totalThe RSS attenuation data Δr is fed into the input layer in batches, with N RSS attenuation data Δr in each batch. After being fed into the input layer, it is reconstructed into an image vector. In some specific embodiments of the present invention, the input layer reconstructs the RSS attenuation data Δr into a vector of the form (N, C, H, W), where N is the batch size, C is the number of channels, H is the height of the image, and W is the width of the image; the dimension of the RSS attenuation data Δr is L×1, which can be reconstructed into a vector of size (N, 1, n, n-1), for example... Figure 4 The input shown is (1*12*11).

[0087] Step S212: The reshaped image vector is fed into the first CNN layer for two-dimensional convolution to extract features, and the extracted features are then reshaped into the input vector of the first Transformer Encoder layer. In some specific embodiments of the present invention, the first CNN layer uses a convolution kernel of size (a, a-1) and has C1 output channels, for example, Figure 4 In the second module Conv,3*2,n64s1, 3*2 represents the size of the convolution kernel; n64 represents the number of output channels as 64; s1 represents the stride of the convolution as 1, and the other parameters are default values. First, the vector from step S211 is reshaped into an image vector with equal height and width, meaning the output image vector of the first CNN layer has a size of (N, C1, n-a+1, n-a+1). Then, this image vector with equal height and width is reshaped into the form (N, S, E) to be input into the first Transformer Encoder layer. The first Transformer Encoder module can provide better input for subsequent CNN modules. Here, S is the number of input vectors, E is the dimension of each vector, and S = (n-a+1). 2 , E = C1.

[0088] In step S213, the input vector sequentially passes through the second feature extraction of the first Transformer Encoder layer, the first batch normalization and activation operation of the first batch normalization and activation layer (BN&ReLU), the third feature extraction of two-dimensional convolution in the second CNN layer, the fourth feature extraction of the second Transformer Encoder layer, the second batch normalization and activation operation of the batch normalization and activation layer (BN&ReLU), and finally is sent to the flattening and deactivation layer for flattening and deactivation processing.

[0089] Step S214: The flattened and deactivated vector is fed into a fully connected layer and activation function layer. Through the fully connected layer and the Sigmoid activation function, the final output vector is transformed into a probability representation, thus completing the classification of the reconstructed pixel matrix data.

[0090] In step S220, the neural network is trained by backpropagation and the hyperparameters are adjusted. In this step, a loss function is used to train the neural network by backpropagation. After training, a trained neural network is obtained. Then, the hyperparameters of the constructed neural network are adjusted according to training experience. The hyperparameters are finally determined based on the performance on the validation set to achieve the optimal localization effect of the neural network.

[0091] Specifically, in steps S212 and S213, the two-dimensional convolution of the first CNN layer and the second CNN layer is as follows: for an input size of (N, C) in H in W in The input is (N, C), and the output size is (N, C). out H out W out The operational relationships are as follows:

[0092]

[0093] In the formula, * is the two-dimensional cross-correlation operator.

[0094] In addition, the input dimensions (N, C) in H in W in ) and output dimensions (N, C) out H out W out The calculation relationship is as follows:

[0095]

[0096]

[0097] In the formula, kernel_size is the size of the convolution kernel; stride is the stride of the convolution, which defaults to 1; padding is the size of the padding added to all four sides of the input, which defaults to 0; and dilation is the spacing between kernel elements, which defaults to 1.

[0098] Specifically, see Figure 5 The first and second Transformer Encoder layers in step S213 include a cascaded multi-head attention layer and a feedforward layer, constituting a Transformer Encoder layer for capturing features. The multi-head attention layer uses a multi-head attention mechanism; the feedforward layer is a unidirectional, multi-layered neural network where neurons are arranged hierarchically, with each neuron connected only to neurons in the previous layer, and there is no feedback between layers. Furthermore, a residual connection is used between the input and output of each multi-head attention layer and the feedforward layer. The residual connection helps solve the problems of vanishing gradients and network degradation. For example, the final output O of the multi-head attention layer is composed of:

[0099] O=γAttn(I)+(1-γ)I

[0100] Where I and Attn(I) are the input and output of the multi-head attention layer, respectively, and γ is a trainable scalar with an initial value of 0.5. The attention mechanism preceding Convolutional Neural Networks (CNNs) can provide a better representation of the input to CNNs.

[0101] The multi-head attention layer first linearly projects each input into three sets of trainable query Q, key K, and value V. Then, it performs a dot product operation on all Q values ​​with respect to K, divides by a scaling factor, and passes this result through a Softmax activation function to obtain an attention weight map. Finally, it performs matrix multiplication between the obtained attention weight map and V to obtain a new vector. The scaling factor is used to prevent QK from being over-multiplied. T If the inner product is too large, the resulting large value will fall into the saturation region when input into Softmax, causing the gradient to vanish and making training impossible.

[0102] The calculation formula for the attention mechanism is as follows:

[0103]

[0104] In the formula: Q represents the query matrix, K represents the key matrix, and V represents the value matrix; It is a scaling factor, and d can be the dimension of the hidden layer, such as the dimension of the key matrix K.

[0105] The calculation steps for the multi-head attention layer are as follows;

[0106] 1) Convert multiple input vectors separately using the formula Two results, Attn1 and Attn2, are obtained from the calculation.

[0107] 2) Multiply the results Attn1 and Attn2 by V respectively to obtain Attn. V1 and Attn V2 ;

[0108] 3) Place the Attn V1 and Attn V2 By splicing, we get Attn V12 ;

[0109] 4) Place the Attn V12 Dimensional transformation and passing through a fully connected layer yield the attention mechanism result Attn.

[0110] The expression for the Softmax function is:

[0111]

[0112] In the formula: x i The input to the Softmax layer is represented by N, which represents the dimension of the output vector. The function solver then calculates the probability of the output vector.

[0113] Furthermore, in some specific embodiments of the present invention, the normalization in the Transformer Encoder structure adopts Layer-Normalization, that is, normalizing all neurons in the same layer of the same sample, which can solve the gradient vanishing problem and the weight matrix degradation problem.

[0114] Specifically, the second CNN layer in step S213 is a multi-layer CNN layer. For example, a 4-layer CNN layer, such as... Figure 6 As shown, a batch normalization (BN) layer is applied between every two convolutional layers to improve network convergence speed and mitigate gradient vanishing during training. In addition, the ReLU activation function can be used for non-linear activation.

[0115] The ReLU activation function is used to increase the non-linearity of the neural network, i.e., f(z) = max(0, z), where z is the value of the input to the hidden layer after linear computation, and f(z) represents the value of z after activation. Other activation functions can also be used, such as LeakyReLU.

[0116] In step S213, dropout is used to address the overfitting problem. Specifically, some neurons are randomly deactivated. The probability of random deactivation is a manually defined parameter, emb_dropout. This probability directly affects the number of neurons lost; too many neurons will degrade model performance, while too few will make training the entire model difficult. In this invention, a random deactivation probability of 0.2 is used.

[0117] The expression for the Sigmoid activation function in step S214 is:

[0118]

[0119] In step 220, the loss function of the neural network is binary cross-entropy (BCE). In some specific embodiments of the present invention, the Adam optimizer with an initial learning rate of 0.001 is used to train the model. The batch size is set to 32, and the network is trained for 60 epochs. Alternatively, other optimizers can be selected, and the batch size and training epochs can be changed according to actual needs.

[0120] More specifically, for a batch of data containing N samples, each sample may have M labels, the loss function is calculated as follows:

[0121]

[0122] in,

[0123]

[0124] Among them, l n This is the loss corresponding to the nth sample, where the loss for the i-th pixel is...

[0125]

[0126] Among them, w i These are hyperparameters used to address the issue of imbalanced samples between labels. For a batch of RSS training data, if a certain label appears less frequently, it should be given a higher weight when calculating the loss; x n These are the actual labels, corresponding to Δx and y in the pixel matrix data. n This is the model output.

[0127] In step S300 of the test RSS attenuation data based on passive positioning wireless sensor network, RSS data in the real environment is collected through wireless sensor nodes, and data preprocessing is performed to construct a relevant test dataset. Specific steps include...

[0128] Step S301: Place n wireless nodes at equal intervals around the positioning area, all nodes are on a plane. The coordinates of each node are known. Each node can send and receive signals, and these wireless nodes form L = n(n-1) wireless links.

[0129] Step S302: When there is no target, measure the RSS value r of each link. l l∈[1,2,...,L];

[0130] Step S303: At time t, the target enters the monitoring area, and the RSS value r of each link is measured at this time. l，t Since l∈[1,2,...,L], the change value Δr of RSS can be obtained from this. l,t =r l，t -r l ;Δr l,t The shadow decay in LOS is caused by the occlusion of the target.

[0131] Step S303: Normalize and process the acquired RSS data to serve as input for the subsequent neural network.

[0132] In step S400 of wireless tomographic target localization, a trained neural network is used. The measured RSS data obtained in the step of measuring RSS attenuation data based on passive localization wireless sensor network is used as input, and the output is the pixel matrix of the reconstructed image. The center of the pixel with the highest brightness in the pixel matrix is ​​the estimated target position.

[0133] like Figure 7 As shown, in one embodiment of the present invention, RSS data before and after the target enters the area are collected and input into a trained neural network. Specifically, this embodiment deploys a network containing 12 wireless nodes in the area to be located, with the node devices deployed around a 1.8m × 1.8m square area. ； Each node device is fixed 1m above the ground, with a horizontal spacing of 0.6m between adjacent nodes; the area to be located is divided into 324 pixels. See the specific real-world diagram below. Figure 8 As shown.

[0134] This invention uses ZigBee nodes with a CC2530 chip core as signal acquisition devices. Signals are sent by one node and received by the other nodes, polling them sequentially. RSS measurements are collected on a computer. First, RSS is collected when no one is in the area, serving as the airspace RSS measurement. A second measurement is collected after a target enters the area.

[0135] In the specific implementation process, in order to evaluate the positioning performance of the positioning system, Table 1 shows the comparison of imaging error, positioning error and algorithm running time between the method of the present invention and the regularization algorithm.

[0136] Table 1 Comparison of Deep Learning and Tikhonov Regularization Algorithm

[0137]

[0138] This invention is based on a network model structure combining CNN and Transformer Encoder. Leveraging the powerful nonlinear mapping capabilities of neural networks, it solves unfit equations to achieve high-quality region image reconstruction. Compared to the Tikhonov regularized wireless tomography target localization method, this invention significantly improves localization performance, achieving advantages in contour imaging and multi-target localization, while also possessing lower time complexity.

[0139] This invention is applied to wireless tomography systems to achieve more accurate device-free localization. It introduces a deep learning method that can learn features directly from data without system model information. Compared to traditional regularization methods, this invention significantly improves localization accuracy and greatly reduces time costs.

[0140] Example 2

[0141] This embodiment is used to describe in detail the deep learning-based wireless tomography target localization system of the present invention.

[0142] like Figure 9 As shown, the deep learning-based wireless tomographic target localization system of the present invention includes a passive localization wireless sensor network, an RSS data measurement device, an RSS dataset generation and recording device, and a neural network computing device based on CNN and Transformer Encoder; wherein,

[0143] The passive positioning wireless sensor network consists of wireless sensor nodes. n wireless sensor nodes are arranged around the positioning area to form a network consisting of n(n-1) wireless links. The wireless sensor nodes send and receive signals to locate the target in the positioning area. The wireless sensor nodes are connected to the RSS data measurement device.

[0144] The RSS data measurement device is used to control the wireless sensor node to send and receive signals, measure the RSS attenuation value data of each wireless link arranged around the positioning area, and send the RSS attenuation value data to the RSS dataset generation and recording device.

[0145] The RSS data generation and recording device is used to generate or record the RSS dataset for training neural networks, as well as the RSS attenuation data of the positioning area where the target to be located is located. The RSS dataset includes pixel matrix data and RSS attenuation data. The pixel matrix data is matrix data in which the positioning area is divided into grids, with each grid representing one pixel. Grids containing the target and those not containing the target are represented by different values. The RSS attenuation data is vector data composed of wireless link RSS attenuation data. The RSS attenuation data is generated in two ways: firstly, through actual measurement using an RSS data measurement device and a passive positioning wireless sensor network; and secondly, through calculation using a wireless signal attenuation model. The RSS dataset includes an RSS training set and an RSS validation set.

[0146] The neural network based on CNN and Transformer Encoder includes, in sequence, an input layer, a first CNN layer, a first Transformer Encoder layer, a first batch normalization and activation layer, a second CNN module, a second Transformer Encoder module, a second batch normalization and activation layer, a flattening and deactivation layer, and a fully connected and activation function layer. The neural network is trained using an RSS dataset, with RSS decay values ​​from the RSS dataset as input and pixel matrix data from the RSS dataset as output. The number of neurons in the first layer of the neural network equals the dimension of the RSS decay values, and the number of neurons in the last layer equals the dimension of the pixel matrix data. The trained neural network takes RSS decay values ​​of the target location region as input and outputs a pixel matrix containing the target location.

[0147] The RSS data generation and recording device includes the following modules for generating RSS datasets: an initialization parameter setting module, a pixel matrix generation module, an RSS attenuation value data generation module, an RSS attenuation value data normalization module, and an RSS dataset generation module.

[0148] The initialization parameter setting module is used to evenly divide the positioning area into P pixels, set the number of wireless sensor nodes to n, and arrange the n wireless sensor nodes around the positioning area. Each node sends and receives signals, thereby forming L = n(n-1) wireless links.

[0149] The pixel matrix data generation module is used to generate N total The pixel matrix data of the image is generated as follows: For each image, m targets are randomly generated in the positioning area. Each target is a square with a side length of h pixels. The target positions are randomly generated in the positioning area, thereby generating pixel matrix data Δx, where m is a random natural number not greater than 5 and h is a natural number not greater than 5.

[0150] The RSS attenuation data generation module is used to generate RSS attenuation data Δr using a wireless signal attenuation model. This module generates RSS attenuation data Δr for all links based on the pixel matrix data Δx of each image. The RSS attenuation data Δr is generated using the following wireless signal attenuation model:

[0151] Δr=WΔx+χ η

[0152] Among them, Δr=[Δr1, Δr2,..., Δr L ] T , Δr iLet Δx represent the RSS attenuation value of the i-th link, where Δx = [Δx1, Δx2, ..., Δx]. P ] T Δx j Let represent the value of the j-th pixel in the localization region, and W be an L×P dimensional weight matrix, where [W] i，j =w i，j w i,j χ represents the influence factor of the value of pixel j on the attenuation value of link i; η It is measurement noise that follows a Gaussian distribution, χ η ~N(0, η) 2 );

[0153] The RSS attenuation value data normalization module, in step S104, normalizes the RSS attenuation value data Δr. The formula for normalizing the RSS attenuation value data Δr in this module is:

[0154]

[0155] In the formula, Δr i This represents the RSS attenuation value of the i-th link;

[0156] The RSS dataset generation module is used to generate RSS dataset 5; in this module, N total The pixel matrix data Δx and the RSS decay value data Δr of the image are used as the RSS dataset. The RSS dataset is divided proportionally, with one part used as the training set and the other part used as the validation set for training the neural network.

[0157] This invention employs neural networks for deep learning to achieve a method and system for passive localization in wireless tomography. By using a small number of nodes to classify each pixel in the localization area, this invention achieves clearer image reconstruction compared to the Tikhonov regularization method. This invention reduces the number of sensor nodes, lowers the complexity of the localization system, and saves on localization costs.

[0158] This invention can create datasets using a shadow fading model and can also learn features directly from data without system model information, exhibiting stronger generalization capabilities. Furthermore, the created dataset focuses only on the location of the target, without imaging the entire environment. This signal attenuation model can filter out signal propagation loss, thus offering greater applicability compared to other attenuation models.

[0159] This invention implements a neural network for passive localization in wireless tomography based on CNN and Transformer. It achieves a significant improvement over the traditional Tikhonov regularization method in terms of overall average localization error and optimal average localization error, and also saves approximately 44% of the time compared to the Tikhonov regularization method.

[0160] The above are merely preferred embodiments of the invention and are not intended to limit the invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the protection scope of the invention.

Claims

1. A deep learning-based wireless tomographic target positioning method, characterized in that, The steps include generating an RSS dataset of a wireless sensor network based on passive positioning, building a neural network based on CNN and Transformer Encoder and training the neural network using the RSS dataset, measuring RSS attenuation value data of the wireless sensor network based on passive positioning, and positioning a target of wireless tomography; In the step of generating an RSS dataset using a passive positioning-based wireless sensor network, n wireless sensor nodes are arranged around the positioning area, forming a network consisting of... A wireless sensor network consisting of wireless links allows wireless sensor nodes to send and receive signals to locate targets within a positioning area. RSS data includes pixel matrix data and RSS attenuation data. Pixel matrix data is a matrix data where the positioning area is divided into grids, with each grid representing one pixel; grids containing the target and those not containing the target are represented by different values. RSS attenuation data is vector data composed of wireless link attenuation values. The RSS dataset is obtained through actual measurements or generated using a wireless signal attenuation model. The RSS dataset includes an RSS training set and an RSS validation set. In the step of building a neural network based on CNN and Transformer Encoder and training the neural network using the RSS dataset, the neural network built based on CNN and Transformer Encoder includes an input layer, a first CNN layer, a first Transformer Encoder layer, a first batch normalization and activation layer, a second CNN module, a second Transformer Encoder module, a second batch normalization and activation layer, a flattening and deactivation layer, and a fully connected and activation function layer connected in sequence; the RSS attenuation value data in the RSS dataset is taken as the input of the neural network, the pixel matrix data in the RSS dataset is taken as the output of the neural network, the neural network is trained, the number of first-layer neurons of the neural network is equal to the dimension of the RSS attenuation value data, and the number of last-layer neurons of the neural network is equal to the dimension of the pixel matrix data; In the step of positioning a target of wireless tomography, the trained neural network is used, the measured RSS data obtained in the step of measuring RSS attenuation data of the wireless sensor network based on passive positioning is taken as the input, and a pixel matrix of a reconstructed image is output, and the center of the pixel with the maximum brightness in the pixel matrix is the estimation of the target position.

2. The deep learning based wireless tomographic target positioning method of claim 1, wherein, The RSS dataset is generated using a wireless signal attenuation model, and the steps include: A step S101 of setting initialization parameters is provided, in which the positioning area is evenly divided into P pixels, and n wireless sensor nodes are set, n wireless sensor nodes are arranged around the positioning area, each node transmits and receives signals, thereby forming L wireless links; The step S102 of generating pixel matrix data; in this step, m target squares with a side length of h pixels are randomly generated in the positioning area for each image, and the pixel matrix data is generated The pixel matrix data of the image, for each image, m targets are randomly generated in the positioning area, the targets are squares with a side length of h pixels, and the target positions are randomly generated in the positioning area, thereby generating the pixel matrix data m is a random natural number not greater than 5, and h is a natural number not greater than 5; Generating RSS attenuation value data by a wireless signal attenuation model of step S103; in this step, RSS attenuation value data of all links are generated as elements from the pixel matrix data of each image of step S103; in this step, RSS attenuation value data of all links are generated as elements from the pixel matrix data of each image of step S103; in this step, RSS attenuation value data of all links are generated as elements from the pixel matrix data of each image generated by the following wireless signal attenuation model: ； wherein, , denotes the attenuation value of the RSS of the i-th link, , denotes the value of the j-th pixel in the positioning area, is a weight matrix of dimension N x N, wherein, , denotes the influence factor of the value of the j-th pixel on the attenuation value of the i-th link; is a measurement noise following a Gaussian distribution, ; RSS attenuation value data The step S104 of normalizing the RSS attenuation value data The formula for normalizing is: ； In the formula, RSSi represents the RSS attenuation value of the ith link. The step S105 of generating the RSS dataset; in this step, the pixel matrix data of the amplitude image and the attenuation value data of the RSS as the RSS dataset, and the RSS dataset is divided in proportion, part of which is used as the training set and part of which is used as the validation set for the training of the neural network.

3. The deep learning based wireless tomographic target positioning method of claim 2, wherein, In step S103, The elliptical model is used for calculation, i.e. ； wherein is the distance between the i-th link transceiver node, and are the distances from the j-th pixel center to the two ends of the i-th link transceiver node, respectively, is a controllable variable that adjusts the length of the short axis of the ellipse.

4. The deep learning-based wireless tomographic target positioning method according to claim 2 or 3, characterized by, The step of training the neural network using the RSS dataset includes a forward propagation step S210 and a backward propagation step S220, The step S210 of forward propagation of the neural network includes: Step S211, the RSS data set in step S210 is reshaped into a picture vector after being sent into the input layer in batches. RSS attenuation value data Step S211, the RSS data set in step S210 is reshaped into a picture vector after being sent into the input layer in batches. RSS attenuation value data Step S211, the RSS data set in step S210 is reshaped into a picture vector after being sent into the input layer in batches. Step S212, the reshaped picture vector is sent into the first CNN layer for two-dimensional convolution, features are extracted, and the extracted features are reshaped into an input vector of the first Transformer Encoder layer; Step S213, the input vector sequentially passes through second feature extraction of the first Transformer Encoder layer, first batch normalization and activation operation of the first batch normalization and activation layer, third feature extraction of two-dimensional convolution in the second CNN layer, fourth feature extraction of the second Transformer Encoder layer, second batch normalization and activation operation of the second batch normalization and activation layer, and is finally sent into the flattening and deactivation layer for flattening and deactivation processing; Step S214: the vector after flattening and deactivation processing is sent into the fully connected and activation function layer, the final output vector is converted into a probability representation through the fully connected layer and the Sigmoid activation function, and the classification of the reconstructed pixel matrix data is completed; The neural network is trained and the hyperparameters are adjusted in step S220, in which the neural network is trained by back propagation using a loss function, and a trained neural network is obtained after training; the hyperparameters of the constructed neural network are adjusted according to training experience, and the hyperparameters are finally determined according to the performance of the validation set, so as to realize the optimal positioning effect of the neural network.

5. The deep learning-based wireless tomographic target positioning method of claim 4, wherein, In step S211, each batch of picture vectors is in the form where is the size of the batch, is the number of channels and , is the height of the picture and , is the width of the picture and ; Step S212, the first CNN layer adopts a convolution kernel with a size of , the number of convolution output channels is , and the input vector of the first Transformer Encoder layer is in the form of , wherein , .

6. The deep learning based wireless tomographic target positioning method of claim 4, wherein, The first Transformer Encoder layer and the second Transformer Encoder layer have the same network structure and each include a multi-head self-attention layer and a feedforward layer connected in sequence, and a residual connection is used between the input and the output of each attention layer and the feedforward layer.

7. The deep learning based wireless tomographic target positioning method of claim 4, wherein, The second CNN layer in step S213 is a multi-layer convolutional layer, and a batch normalization layer is used between every two convolutional layers to improve the network convergence speed and alleviate the gradient vanishing in the training process.

8. The deep learning based wireless tomographic target positioning method of claim 4, wherein, In step S220, the loss function is binary cross-entropy.

9. A deep learning based wireless tomographic target positioning system, characterized by The wireless sensor network for passive positioning, the RSS data measurement device, the RSS data set generation recording device, and the neural network calculation device based on CNN and Transformer Encoder; wherein, The wireless sensor network for passive positioning is composed of wireless sensor nodes, n wireless sensor nodes are arranged around the positioning area to form a network composed of n(n-1) wireless links, the wireless sensor nodes send and receive signals to position the target in the positioning area, and the wireless sensor nodes are connected with the RSS data measurement device; The RSS data measurement device is used to control the wireless sensor nodes to send and receive signals, measure the RSS attenuation value data of each wireless link arranged around the positioning area, and send the RSS attenuation value data to the RSS data set generation recording device; The RSS data generation recording device is used to generate or record the RSS data set for training the neural network, and record the RSS attenuation value data of the positioning area where the target to be positioned is located; the RSS data set includes pixel matrix data and RSS attenuation value data; the pixel matrix data is a matrix data obtained by dividing the positioning area into a grid, each grid being a pixel, and the target grid and the grid without the target being represented by different values; the RSS attenuation value data is a vector data composed of wireless link RSS attenuation value data; the RSS attenuation value data is obtained by actual measurement through the RSS data measurement device and the wireless sensor network for passive positioning, or is calculated by using a wireless signal attenuation model; the RSS data set includes an RSS training set and an RSS validation set; The neural network based on the CNN and the Transformer Encoder comprises an input layer, a first CNN layer, a first Transformer Encoder layer, a first batch normalization and activation layer, a second CNN module, a second Transformer Encoder module, a second batch normalization and activation layer, a flattening and deactivation layer, and a full connection and activation function layer connected in sequence; the neural network is trained by using an RSS data set, RSS attenuation value data in the RSS data set is taken as input of the neural network, pixel matrix data in the RSS data set is taken as output of the neural network, training of the neural network is performed, the number of first layer neurons of the neural network is equal to the dimension of the RSS attenuation value data, and the number of last layer neurons of the neural network is equal to the dimension of the pixel matrix data; the trained neural network inputs RSS attenuation value data of a positioning area where a target to be positioned is located, and outputs pixel matrix data containing a position of the target to be positioned.

10. The deep learning based wireless tomographic target positioning system of claim 9, wherein, The module for generating the RSS data set in the RSS data generation recording device comprises an initialization parameter setting module, a pixel matrix generation module, an RSS attenuation value data generation module, an RSS attenuation value data normalization module and an RSS data set generation module; wherein, The initialization parameter setting module is used for dividing the positioning area into P pixels uniformly, setting the wireless sensor nodes as n, arranging the n wireless sensor nodes around the positioning area, sending and receiving signals by each node, and thus forming L wireless links; The pixel matrix data generation module is configured to generate pixel matrix data of the images, for each image, m targets are randomly generated in the positioning area, the target is a square with a side length of h pixels, and the target position is randomly generated in the positioning area, so as to generate the pixel matrix data , m is a random natural number not greater than 5, and h is a natural number not greater than 5. The RSS attenuation value data generation module is configured to generate RSS attenuation value data by using a wireless signal attenuation model The module generates RSS attenuation value data of all links as elements according to the pixel matrix data of each image The RSS attenuation value data of all links as elements The RSS attenuation value data The following wireless signal attenuation model is used to generate: ； wherein, , denotes the attenuation value of the RSS of the i-th link, , denotes the value of the j-th pixel in the positioning area, is a weight matrix of dimension N x N, wherein, , denotes the influence factor of the value of the j-th pixel on the attenuation value of the i-th link; is a measurement noise following a Gaussian distribution, ; The RSS attenuation value data normalization module is configured to normalize the RSS attenuation value data The step S104 of performing normalization processing, in which the RSS attenuation value data The formula for performing normalization processing is: ； In the formula, RSSi represents the RSS attenuation value of the ith link. An RSS dataset generation module for generating an RSS dataset 5; in this module, pixel matrix data of the amplitude image and the attenuation value data of the RSS as the RSS dataset, and the RSS dataset is divided in proportion, part of which is used as a training set and part of which is used as a validation set for the training of the neural network.

Images