Radio tomography target detection method and system based on deep learning, and electronic equipment
By combining convolutional neural networks and reinforcement learning networks with self-attention mechanisms, the problems of large computational complexity and low precision in wireless tomography technology are solved, and high-precision, low-complexity device-free target positioning is achieved.
Patent Information
- Application Number
- CN202510859195.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-25
- Publication Date
- 2025-09-26
AI Technical Summary
Existing wireless tomography technology has problems such as large computational complexity, limited reconstruction accuracy, and susceptibility to noise interference. In addition, traditional methods cannot achieve target positioning without carrying equipment.
A wireless tomography method based on deep learning is adopted, combined with a convolutional neural network (CNN) and a reinforcement learning network with a self-attention mechanism. RSS data is collected using a multi-sensor network. The model is trained through adaptive data enhancement and reinforcement learning optimization algorithm to improve positioning accuracy and multi-target detection capabilities.
It significantly improves positioning accuracy, reduces computational complexity, enhances multi-target detection capabilities, and realizes target positioning without equipment.
Smart Images
Figure CN120703684A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a passive target detection technology for wireless sensor networks, and in particular to a wireless tomography target detection method, system and electronic equipment based on deep learning. Background Art
[0002] Positioning technologies based on received signal strength (RSS) can be categorized as ranging-based and range-free. Wireless detection technologies based on RSS ranging require the deployment of multiple wireless sensor nodes in space, with the target carrying a wireless node. The target's location is determined by measuring the RSS between the unknown node and a beacon node, inferring the distance between them. However, this technology requires the target to carry additional equipment, making it impossible to achieve device-free positioning.
[0003] Radio Tomography (RTI) is a low-cost target detection technology that exploits the attenuation characteristics of wireless signals. Traditional methods are primarily based on received signal strength (RSS) ranging models or image reconstruction algorithms, such as Tikhonov regularization. However, these methods suffer from high computational complexity, limited reconstruction accuracy, and susceptibility to noise.
[0004] In recent years, deep learning methods, particularly CNN and Transformer networks, have achieved significant breakthroughs in the field of RTI. However, using CNNs alone can easily lose global features, and while Transformers offer global perception capabilities, they come with high computational overhead. Therefore, combining the local feature extraction capabilities of CNNs with the global information fusion capabilities of Transformers, while simultaneously optimizing the network structure through reinforcement learning, is crucial for improving target detection performance in wireless tomography. Summary of the Invention
[0005] In response to the above situation, the embodiments of the present application propose a wireless tomography target detection method, system and electronic device based on deep learning. This application significantly improves positioning accuracy, reduces computational complexity, and enhances multi-target detection capabilities.
[0006] In a first aspect, an embodiment of the present application provides a wireless tomography target detection method based on deep learning, the method comprising: Step S10, deploying a wireless sensor node link network in the monitoring area; Step S20, obtaining an RSS signal data set on the wireless sensor node link network; Step S30, constructing a deep learning neural network model; Step S40, using a reinforcement learning optimization algorithm to train a deep learning neural network model to obtain a reinforcement learning neural network model; Step S50, inputting the new RSS signal data set into the reinforcement learning neural network model to obtain the probability distribution of the target position; Step S60 determines the target position according to the output probability distribution of the target position.
[0007] In one embodiment, the method further includes step S70 of mapping the probability distribution of the target location to a heat map.
[0008] In one embodiment, step S20 further includes: The wireless sensor node link network is discretized into a pixel grid, and each pixel records the change of the RSS signal to form a multidimensional matrix. The RSS signal data set is represented as a multidimensional matrix.
[0009] In one embodiment, the algorithm adopted by the reinforcement learning optimization algorithm is a PPO algorithm, and the PPO algorithm adopts a reward function, which calculates the reward function based on the difference between the probability distribution of the obtained target position and the probability distribution of the actual target position.
[0010] In one embodiment, the reward function is the inverse of the squared reconstruction error, and the loss function is the squared reconstruction error: L_target=||p_{predict}-p_{truth}||2 The probability distribution of the final output target position is p_{predict}, and the probability distribution of the true target position is p_{truth}.
[0011] In one embodiment, step S60 further includes: determining the target position using an argmax method, where the argmax method directly selects a pixel point with a maximum probability value as the most probable position of the target.
[0012] In one embodiment, the method further includes: performing data enhancement processing on the RSS signal data set, wherein the data enhancement processing method includes: Gaussian noise perturbation, data smoothing, data rotation, translation, and scaling.
[0013] In one embodiment, the deep learning neural network model includes an input layer, multi-layer convolutional layers, an attention layer, and a fully connected layer connected in sequence, wherein activation functions and pooling layer structures are used between the convolutional layers, and the attention layer adopts a multi-head attention mechanism.
[0014] In a second aspect, an embodiment of the present application further provides a wireless tomography target detection system based on deep learning, comprising: The data acquisition unit deploys a wireless sensor node link network in the monitoring area and obtains the RSS signal data set on the wireless sensor node link network; The model construction training unit constructs a deep learning neural network model, uses a reinforcement learning optimization algorithm to train the deep learning neural network model, and obtains the reinforcement learning neural network model; a model calculation unit, inputting the new RSS signal data set into the reinforcement learning neural network model to obtain a probability distribution of the target position; The result calculation unit determines the target position according to the output probability distribution of the target position.
[0015] In a third aspect, an embodiment of the present application further provides an electronic device, comprising: a processor; and a memory arranged to store computer-executable instructions, wherein the executable instructions, when executed, enable the processor to perform any of the above-mentioned wireless tomography target detection methods.
[0016] At least one of the above technical solutions adopted in the embodiments of the present application can achieve the following beneficial effects: This application first uses a multi-sensor network to collect received signal strength (RSS) data from an environment and constructs a dataset using adaptive data augmentation techniques. It then builds a network combining a convolutional neural network (CNN) and an adaptive reinforcement learning network based on an attention mechanism to optimize signal reconstruction and target localization performance. Compared to traditional methods, this application significantly improves positioning accuracy, reduces computational complexity, and enhances multi-target detection capabilities. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings: Figure 1 A schematic flow chart of a wireless tomography target detection method based on deep learning according to an embodiment of the present application is shown.
[0018] Figure 2 Schematic diagram of the structure of a wireless tomography target detection system based on deep learning according to an embodiment of the present application. DETAILED DESCRIPTION
[0019] In recent years, artificial intelligence technology has been widely used due to its advantages in efficiency and accuracy, which provides a new way to accurately and quickly identify persistent precipitation.
[0020] In order to make the purpose, technical solutions and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with the specific embodiments of this application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. The technical solutions provided by each embodiment of this application are described in detail below in conjunction with the drawings.
[0021] This application proposes a wireless tomography target detection method and system based on reinforcement learning and deep neural networks. This application first utilizes a multi-sensor network to collect received signal strength (RSS) data from the environment and constructs a dataset using adaptive data augmentation techniques. Next, a convolutional neural network (CNN) is constructed that combines an adaptive reinforcement learning network based on an attention mechanism to optimize signal reconstruction and target localization performance. Compared to traditional methods, this application significantly improves positioning accuracy, reduces computational complexity, and enhances multi-target detection capabilities.
[0022] The present invention provides a wireless tomography target detection method based on deep learning, which mainly includes: 1. Construct a high-quality RSS dataset: Acquire RSS data based on wireless sensor networks and use data enhancement techniques to improve data diversity.
[0023] 2. Deep reinforcement learning model training: Build a reinforcement learning network that combines CNN and self-attention mechanism to optimize object detection performance.
[0024] 3. Target positioning and reconstruction: The trained model is used to process the actual measured RSS data and output the target location distribution.
[0025] The wireless tomography target detection method based on deep learning is specifically as follows: Step S10: deploying a wireless sensor node link network in the monitoring area.
[0026] Evenly distribute n wireless sensor nodes in the target detection area or according to actual needs, and establish a link between any two nodes to form a total of n 2 A wireless sensor node network consists of 100 wireless sensor node links (or a different number of links depending on the specific network topology). The wireless sensor nodes are used to send and receive signals to locate targets within the monitoring area. Each wireless sensor node link collects raw signal strength (RSS) data in real time to determine signal attenuation along different paths.
[0027] The following strategies can be considered for the deployment of wireless sensor nodes: Uniform grid layout: ensures that every sub-area within the detection area is covered by sensor nodes, which is suitable for scenarios where balanced monitoring of the entire area is required.
[0028] Adaptive deployment: Based on prior information or pre-detection results, more nodes are deployed in key areas to improve resolution.
[0029] Step S20, obtaining an RSS signal data set on the wireless sensor node link network; Step S20 further includes: The wireless sensor node link network is discretized into a pixel grid. Each wireless sensor node link corresponds to a pixel point, representing a sampling point. Each pixel point represents the RSS signal collected from each wireless sensor node link.
[0030] For each wireless link, the changes in the RSS signal are recorded at each pixel to form a multidimensional matrix X. Each row or layer of the matrix represents the RSS signal distribution of different links in the entire area. The RSS signal dataset of this application is represented as a multidimensional matrix.
[0031] For example, assuming that the wireless sensor node link network is discretized into an M×N pixel grid, the matrix can be constructed:
[0032] Where x_{ij} represents the RSS signal at the pixel in the i-th row and j-th column.
[0033] The present application also includes data enhancement processing on the RSS signal data set. The data enhancement processing methods include: Gaussian noise perturbation, data smoothing, data rotation, translation, scaling, etc., which can also be appropriately applied to enhance data diversity.
[0034] Gaussian noise perturbation: Add Gaussian noise to the original signal to generate new data samples.
[0035]
[0036] Data smoothing: Use filtering (such as mean filtering, median filtering, or Gaussian filtering) to reduce random fluctuations in the data and retain the main features.
[0037] This application selects some RSS signal data sets as training samples for subsequent training of neural network models.
[0038] Step S30 constructs a deep learning neural network model.
[0039] The deep learning neural network model is based on the convolutional neural network (CNN) and self-attention mechanism, and includes input layer, convolution layer, attention layer, and fully connected layer.
[0040] The input layer receives the RSS signal data after preprocessing and data enhancement.
[0041] The dimension of the input data of the input layer is usually in the form of a multi-channel image, such as M×N×C (where C represents the number of signal links or feature channels).
[0042] The convolutional layer has multiple convolution kernels. In this embodiment, a 5x5 kernel with a stride of 1 is used, with three layers in total. Each layer is padded to maintain the image size. Adjacent layers use an activation function (ReLU) and a pooling layer (max pooling). The convolutional layer uses multiple convolution kernels to extract local spatial feature vectors of the monitored area. The extracted local feature vectors are output by the convolutional layer, effectively mitigating noise interference and improving computational efficiency in subsequent processing.
[0043] The attention layer (Self-Attention) is located after the convolutional layer, receives the local feature vector output by the convolutional layer, and outputs the global feature vector.
[0044] Specifically, the local feature vectors extracted by the convolutional layer are used as input tokens for the transformer, and information is fused using next token prediction. A self-attention mechanism is introduced into the transformer to compensate for the limited local receptive field of the convolutional layer. By calculating the correlation weights between each position, global information is captured, thereby improving detection accuracy. A multi-head attention mechanism can be used to further enhance the model's expressiveness.
[0045] The fully connected layer (FC), located after the attention layer, maps the global feature vector obtained by the attention layer to the specific output space of target positioning and outputs the probability distribution p_{predict} of the target position for subsequent target positioning and reconstruction.
[0046] Step S40 uses a reinforcement learning optimization algorithm to train the deep learning neural network model to obtain the reinforcement learning neural network model.
[0047] The deep learning neural network model is trained using a reinforcement learning optimization algorithm to obtain a reinforcement learning neural network model. The algorithm used in the reinforcement learning optimization algorithm is the Proximal Policy Optimization algorithm (also known as the PPO algorithm). The PPO algorithm adaptively adjusts the network parameters. The loss function and optimizer are two indispensable components in deep learning training. In the PPO algorithm, the optimizer is the reward function, which is used for PPO training. The reward function design includes calculating the reward function based on the probability distribution of the target position and the difference from the true distribution. The reward function is defined as the opposite of the reconstruction square error. The reward function adjusts the model parameters based on the feedback of the loss function. The loss function measures the difference between the model prediction value and the true value, where the loss function is the reconstruction square error: L_target=||p_{predict}-p_{truth}||2 The final output probability distribution of the target position is p_{predict}, and the probability distribution of the true target position is p_{truth}. This reward function takes into account detection accuracy, computational efficiency, and error control.
[0048] In one embodiment of the present application, the following further comprises: Before inputting the enhanced RSS signal dataset into the deep learning neural network model network in batches, the enhanced RSS signal dataset is normalized to improve the convergence speed of the model and stabilize the training process.
[0049] The deep learning neural network model training process of this application is as follows: the normalized RSS signal dataset (training sample) is passed through the convolution layer to extract local features, and then the global information is integrated through the attention layer, and finally the probability distribution of the target position is obtained through the fully connected layer.
[0050] This application uses a reinforcement learning optimization algorithm to train a deep learning neural network model. This algorithm employs policy gradient optimization, selecting actions based on the current policy and calculating rewards for the current state. The PPO algorithm is used to update the policy, enabling the network to better adapt to RSS data distributions in different environments. For the update formula and specific implementation, please refer to the relevant PPO papers and open source code.
[0051] Step S50 inputs the new RSS signal data set into the reinforcement learning neural network model to obtain the probability distribution of the target position.
[0052] In this step, the trained reinforcement learning neural network model is used to perform forward inference on the new RSS signal dataset. The output of the reinforcement learning neural network model is the probability distribution function of the target position at each position in the pixel grid.
[0053] Step S60 determines the target position according to the output probability distribution of the target position.
[0054] The argmax method is used to determine the target position. The argmax method directly selects the pixel with the largest probability value as the most likely location of the target. The formula of the argmax method is:
[0055] in is the final estimated target location coordinate, and f(x,y)=p_{predict} is the final output probability distribution of the target location. This method directly selects the pixel with the largest probability value as the most likely location of the target.
[0056] Step S70 maps the probability distribution of the target location to a heat map.
[0057] Heatmap Generation: Map the distribution of f(x,y) into a heatmap to visually display the distribution of targets within the monitoring area. Heatmaps can be used to detect the distribution trend of targets within the area and possible multi-target situations.
[0058] The order of the above steps S10 to S70 is not fixed, especially step S60 and step S70 have no order.
[0059] like Figure 2 As shown, the present application discloses a wireless tomography target detection system based on deep learning, the system 200 includes: The data acquisition unit 210 deploys a wireless sensor node link network in the monitoring area and obtains an RSS signal data set on the wireless sensor node link network; A model construction and training unit 220 constructs a deep learning neural network model, trains the deep learning neural network model using a reinforcement learning optimization algorithm, and obtains a reinforcement learning neural network model; The model calculation unit 230 inputs the new RSS signal data set into the reinforcement learning neural network model to obtain the probability distribution of the target position; The result calculation unit 240 determines the target position according to the output probability distribution of the target position.
[0060] The technical features of the wireless tomography target detection system correspond one-to-one to the above-mentioned wireless tomography target detection method and will not be repeated here.
[0061] The present application also discloses an electronic device, comprising: a processor; and a memory arranged to store computer executable instructions, wherein the executable instructions, when executed, enable the processor to perform the above method. The electronic device can also perform Figure 1The wireless tomography target detection method based on deep learning is not described in detail in the embodiments of this application.
[0062] The foregoing is merely an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should all be included within the scope of the claims of the present application.
Claims
1. A wireless tomography target detection method based on deep learning, characterized in that: Step S10, deploying a wireless sensor node link network in the monitoring area; Step S20, obtaining an RSS signal data set on the wireless sensor node link network; Step S30, constructing a deep learning neural network model; Step S40, using a reinforcement learning optimization algorithm to train a deep learning neural network model to obtain a reinforcement learning neural network model; Step S50, inputting the new RSS signal data set into the reinforcement learning neural network model to obtain the probability distribution of the target position; Step S60 determines the target position according to the output probability distribution of the target position.
2. The wireless tomography target detection method according to claim 1, wherein: The method further includes step S70 of mapping the probability distribution of the target position to a heat map.
3. The wireless tomography target detection method according to claim 1, wherein: Step S20 further includes: The wireless sensor node link network is discretized into a pixel grid. Each pixel records the change of the RSS signal. Each wireless sensor node link constitutes a multidimensional matrix. The RSS signal data set is represented as a multidimensional matrix.
4. The wireless tomography target detection method according to claim 1, wherein: The reinforcement learning optimization algorithm adopts the PPO algorithm, and the PPO algorithm adopts a reward function, which calculates the reward function according to the difference between the probability distribution of the acquired target position and the probability distribution of the actual target position.
5. The wireless tomography target detection method according to claim 4, wherein: The reward function is the inverse of the squared reconstruction error, and the loss function is the squared reconstruction error: L_target=||p_{predict}-p_{truth}||2 The probability distribution of the final output target position is p_{predict}, and the probability distribution of the true target position is p_{truth}.
6. The wireless tomography target detection method according to claim 1, wherein: Step S60 further includes: determining the target position using an argmax method, where the argmax method directly selects a pixel point with a maximum probability value as the most probable position of the target.
7. The wireless tomography target detection method according to claim 1, wherein: Also includes: The RSS signal data set is subjected to data enhancement processing, wherein the data enhancement processing method includes: Gaussian noise perturbation, data smoothing processing, data rotation, translation, and scaling.
8. The wireless tomography target detection method according to claim 1, wherein: The deep learning neural network model includes an input layer, multi-layer convolutional layers, an attention layer and a fully connected layer connected in sequence, wherein activation functions and pooling layer structures are used between the convolutional layers, and the attention layer adopts a multi-head attention mechanism.
9. A wireless tomography target detection system based on deep learning, characterized in that: include: The data acquisition unit deploys a wireless sensor node link network in the monitoring area and obtains the RSS signal data set on the wireless sensor node link network; The model construction training unit constructs a deep learning neural network model, uses a reinforcement learning optimization algorithm to train the deep learning neural network model, and obtains the reinforcement learning neural network model; a model calculation unit, inputting the new RSS signal data set into the reinforcement learning neural network model to obtain a probability distribution of the target position; The result calculation unit determines the target position according to the output probability distribution of the target position.
10. An electronic device comprising: processor; and a memory arranged to store computer-executable instructions, which, when executed, cause the processor to perform the method of claims 1 to 8.