A method for identifying wireless devices based on multi-dimensional feature fusion deep learning
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-27
- Publication Date
- 2026-08-14
AI Technical Summary
在传统已知识别方法中，原始I/Q信号包含的信息有限，并且易受到各类无用信息的干扰，专家特征在特征提取的能力上表现欠缺，难以挖掘出信号潜在的细微特征
[0012]本发明的有益效果在于：通过结合传统信号处理领域的专家特征提取（如幅度、相位、均值、标准差、最大值、最小值、平均绝对偏差、偏度、峰度、信息熵、能量、均方根、FFT系数13种统计特征）与基于1DCNN的深度学习特征，充分挖掘信号在时域、频域、时频域和统计层面的多维特性。使用SE注意力机制对多维度特征进行融合，建立特征权重分配模型，实现专家特征与深度特征的优化组合，避免特征冗余。利用深度神经网络模型，通过设定训练时间、模型性能等约束条件，完成信号识别分类器的迭代训练，得到高识别准确率的模型。实现对已知设备的高准确率、高效率识别，保障智能家居安全，同时降低计算复杂度。
Smart Images
Figure CN122571484A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wireless equipment signal recognition technology, and in particular to a wireless equipment recognition method based on multi-dimensional feature fusion deep learning. Background Technology
[0002] With the rapid iteration and widespread application of IoT technology, modern families are gradually building a smart living system centered on intelligent terminals. Through wireless communication protocols, diverse intelligent terminals such as access control systems, environmental monitoring devices, and video recording equipment are deeply networked, forming a home ecosystem with environmental perception and autonomous decision-making capabilities. This technological innovation not only restructures the traditional home management model but also significantly improves the comfort and intelligence of living spaces through scenario-based services. However, the massive amount of interactive data generated during system operation makes information security a significant technical bottleneck restricting the continued development of this field. The open nature of wireless networks allows nearby users to receive signals emitted by any wireless device and also send radio signals to devices. While this convenience enhances the user experience, it also raises numerous security risks, including illegal eavesdropping, information tampering, and device impersonation attacks. These risks not only threaten users' privacy and property security but may also affect the overall security of the home. In this context, effectively identifying known authorized devices and ensuring network security is particularly important.
[0003] In traditional wireless communication system security, access authentication typically relies on MAC addresses or keys. MAC address-based authentication requires manually adding devices to the access list, and unauthorized users can forge MAC addresses. Key-based authentication requires shared keys, and once the key is cracked or leaked, unauthorized users can easily access the network. Currently widely used encryption methods in WLANs, such as WEP and WPA-PSK, are also vulnerable to cracking. To address this, researchers have proposed a smart home security system based on radio frequency fingerprint authentication and anomaly identification of wireless signals, a security strategy based on physical layer feature analysis. This system deeply integrates algorithms related to Specific Emitter Identification (SEI) technology, constructing a smart home security barrier through multi-dimensional signal feature analysis. The core challenge of smart home intrusion detection lies in distinguishing legitimate devices from potential threats. Traditional methods often misclassify devices as abnormal due to insufficient device identification capabilities, resulting in a high false alarm rate. Furthermore, devices may be affected by environmental noise, signal interference, and other factors during use, posing numerous challenges to device identification. Therefore, the identification of known authorized devices provides a reliable baseline environment for identifying unknown devices mixed in with the environment, which is crucial for smart home intrusion detection. In traditional known recognition methods, the original I / Q signal contains limited information and is easily interfered with by various kinds of useless information. Expert features are lacking in feature extraction capabilities and are difficult to uncover the subtle potential features of the signal. Summary of the Invention
[0004] The technical problem to be solved by this invention is to provide a wireless device identification method based on multi-dimensional feature fusion deep learning that can accurately and efficiently identify authorized wireless devices.
[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:
[0006] A method for identifying wireless devices based on multi-dimensional feature fusion deep learning, comprising the following steps:
[0007] Construct a wireless communication environment and collect target signal data;
[0008] The target signal is preprocessed and expert features are extracted to obtain an expert feature matrix;
[0009] The signal is subjected to feature extraction based on deep feature learning using a deep neural network to obtain a deep feature matrix;
[0010] The expert feature matrix and the deep feature matrix are fused to obtain a feature fusion matrix;
[0011] A sample set is constructed from the feature fusion matrix, a deep neural network classification model is trained, and the trained model is used for wireless device identification.
[0012] The beneficial effects of this invention are as follows: By combining expert feature extraction from traditional signal processing fields (such as amplitude, phase, mean, standard deviation, maximum value, minimum value, mean absolute deviation, skewness, kurtosis, information entropy, energy, root mean square, and FFT coefficients—13 statistical features) with deep learning features based on 1DCNN, the multidimensional characteristics of signals in the time domain, frequency domain, time-frequency domain, and statistical level are fully explored. The SE attention mechanism is used to fuse multidimensional features, establishing a feature weight allocation model to achieve an optimized combination of expert features and deep features, avoiding feature redundancy. Using a deep neural network model, by setting constraints such as training time and model performance, iterative training of the signal recognition classifier is completed, resulting in a model with high recognition accuracy. This achieves high accuracy and efficiency in recognizing known devices, ensuring smart home security while reducing computational complexity. Attached Figure Description
[0013] Figure 1 This is a schematic diagram illustrating the implementation principle of a wireless device identification method based on multi-dimensional feature fusion deep learning according to an embodiment of the present invention.
[0014] Figure 2 This is a schematic diagram of the 1DCNN deep feature extraction network structure in a wireless device identification method based on multi-dimensional feature fusion deep learning according to an embodiment of the present invention;
[0015] Figure 3 This is a schematic diagram of the SE attention mechanism module in a wireless device identification method based on multi-dimensional feature fusion deep learning according to an embodiment of the present invention;
[0016] Figure 4 This is a comparison of the recognition performance results of an embodiment of a wireless device recognition method based on multi-dimensional feature fusion deep learning according to an embodiment of the present invention; Detailed Implementation
[0017] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.
[0018] Please refer to Figure 1 This invention provides a method for identifying wireless devices based on multi-dimensional feature fusion deep learning, comprising the following steps:
[0019] A method for identifying wireless devices based on multi-dimensional feature fusion deep learning, comprising the following steps:
[0020] Construct a wireless communication environment and collect target signal data;
[0021] The target signal is preprocessed and expert features are extracted to obtain an expert feature matrix;
[0022] The signal is subjected to feature extraction based on deep feature learning using a deep neural network to obtain a deep feature matrix;
[0023] The expert feature matrix and the deep feature matrix are fused to obtain a feature fusion matrix;
[0024] A sample set is constructed from the feature fusion matrix, a deep neural network classification model is trained, and the trained model is used for wireless device identification.
[0025] Furthermore, the technical implementation method for preprocessing the target signal includes: preprocessing the target signal by normalizing the maximum value and the minimum value.
[0026] Furthermore, the specific implementation method for expert feature extraction of the signal includes:
[0027] Thirteen expert features are extracted from the signal, including amplitude, phase, mean, standard deviation, maximum value, minimum value, mean absolute deviation, skewness, kurtosis, information entropy, energy, root mean square, and FFT coefficients of the I / Q channels.
[0028] For I / Q signals (where I is the in-phase component and Q is the quadrature component), the extraction methods for the above signal features are as follows:
[0029] Amplitude:
[0030]
[0031] Phase:
[0032]
[0033] Let the sample sequence of the I / Q signal be... The mean is calculated as follows:
[0034]
[0035] Standard deviation:
[0036]
[0037] Mean Absolute Deviation (MAD):
[0038]
[0039] Skewness
[0040]
[0041] Kurt:
[0042]
[0043] Information entropy:
[0044]
[0045] Let the sample sequence be The energy calculation is as follows:
[0046]
[0047] Root mean square (RMS):
[0048]
[0049] FFT:
[0050]
[0051] As can be seen from the above description, since the feature information contained in the signal itself is not obvious enough, it is necessary to perform normalization preprocessing on the signal and then extract expert features to obtain the expert feature matrix.
[0052] Furthermore, the specific implementation method of extracting features from the signal through a deep feature learning branch to obtain a deep feature matrix includes:
[0053] The deep feature learning branch is a one-dimensional convolutional neural network (1DCNN). The network structure includes a convolutional layer connected to a pooling layer, followed by another convolutional layer and a pooling layer. After flattening, the concatenation results in high-dimensional features, which are then output as low-dimensional deep features through a fully connected layer. The original signal is input into the one-dimensional convolutional neural network via two branches (I and Q) for processing. The features output from the two branches are concatenated across modally in a fusion layer to form a high-dimensional mixed feature vector, resulting in the deep feature matrix.
[0054] The following is one of the preferred embodiments. The specific implementation may not be limited to the specific parameter configuration of this embodiment: The network includes a 64-channel convolutional layer connected to a pooling layer, a 128-channel convolutional layer and a pooling layer, which are then flattened and spliced into 65536-dimensional high-dimensional features, and then output as 256-dimensional deep features through a fully connected layer.
[0055] As described above, using the above method to extract features from a signal can extract information from multiple dimensions of the signal, which can be used for subsequent device identification.
[0056] Furthermore, the specific implementation of feature fusion of the expert feature matrix and the deep feature matrix includes:
[0057] A feature weight allocation model is established based on the SE (Squeeze-and-Excitation) attention mechanism module to realize the weighted calculation of the expert feature matrix and the deep feature matrix, and obtain the fused feature matrix.
[0058] The specific implementation process can be as follows: first, the features are compressed into channel description vectors through global average pooling; then, the nonlinear relationship between channels is learned through a bottleneck structure consisting of two fully connected layers with a compression ratio of 16; and finally, the channel weight matrix is generated through the Sigmoid function.
[0059] As described above, the SE attention mechanism can effectively compress the feature matrix and improve the effectiveness of the data.
[0060] Furthermore, based on the dataset and feature fusion matrix, the training set, validation set, and test set are divided, and the deep neural network classifier is optimized to obtain a signal recognition model for wireless device identification.
[0061] Furthermore, the 1DCNN network structure is as follows: Figure 2 As shown, the network mainly consists of convolutional layers, pooling layers, flattening layers, and fully connected layers.
[0062] As described above, the 1DCNN module consists of cascaded multi-network layers. This module is a convolutional neural network architecture for I / Q signal processing, employing a parallel branch design. The I and Q signals are first input into their respective branches, undergoing a feature extraction process of "input layer - convolutional layer - pooling layer - convolutional layer - pooling layer". The convolutional layers use learnable convolutional kernels to capture local signal features (such as frequency components and waveform patterns), while the pooling layers reduce dimensionality through downsampling and enhance robustness. Subsequently, the Flatten layer flattens the multi-dimensional feature map into a one-dimensional vector, and the pooling layer fuses the features from both paths. Finally, the fully connected layer learns the nonlinear relationships between features and outputs the processing results (such as signal identification category and parameter estimates). This approach is suitable for tasks requiring deep feature extraction from I / Q signals, such as communication signal recognition and modulation classification. By progressively abstracting features through a multi-layered structure, it improves processing accuracy and effectiveness.
[0063] Furthermore, the SE attention mechanism module is as follows: Figure 3 As shown, it mainly consists of spatial feature compression and direct feature transfer.
[0064] As described above, the initial input feature cube represents the multi-channel feature map output by the convolutional layer. The number of channels and spatial size are determined by the preceding network. The first branch performs spatial feature compression using global average pooling, averaging the spatial dimension of each channel to compress the two-dimensional spatial information into a single channel's global statistics, extracting the overall channel response to obtain a one-dimensional vector. Next, channel feature learning is performed by a fully connected layer or a multilayer perceptron (MLP) to learn the non-linear dependencies between channels and uncover the correlations between features from different channels. Finally, channel attention information is output, using activation functions such as the sigmoid function to generate channel weights between 0 and 1, representing the importance of each channel. The second branch is direct feature propagation, where the input feature cube is directly propagated without attention processing. Finally, attention-weighted fusion multiplies the channel attention information with the corresponding channels of the directly propagated feature cube, enhancing important channel features and suppressing secondary channels. The output is a new feature cube incorporating channel attention, making the network focus more on key channel information and improving feature representation capabilities.
[0065] The wireless device identification method based on multi-dimensional feature fusion deep learning described above is applicable to scenarios requiring the detection and identification of wireless devices. The following detailed embodiments illustrate this method:
[0066] Example 1
[0067] Please refer to Figure 1-3 This embodiment of a wireless device identification method based on multi-dimensional feature fusion deep learning includes the following steps:
[0068] S1. Construct a wireless communication environment and collect target signal data;
[0069] S2. Based on the preprocessing of the target signal, expert feature extraction is performed, specifically including:
[0070] S21. Preprocess the target signal by normalizing the maximum value and the minimum value;
[0071] S22. Extract the amplitude and phase of the I / Q channels of the signal, specifically:
[0072] For I / Q signals (where I is the in-phase component and Q is the quadrature component), the amplitude is calculated as follows:
[0073]
[0074] Phase describes the angular position of a signal in the complex plane. Phase information is crucial for analyzing signal delay, modulation methods, and other parameters. Its calculation formula is:
[0075]
[0076] S23. Extract the mean, standard deviation, and mean absolute deviation of the I / Q channels of the signal. Specifically:
[0077] Let the sample sequence of the I / Q signal be... The mean is calculated as follows:
[0078]
[0079] The real and imaginary parts of the standard deviation (Std_real) are used to measure the dispersion of the real and imaginary parts of an I / Q signal. A larger standard deviation indicates more dispersed data. Standard deviation:
[0080]
[0081] The mean absolute deviation (MAD) is the average of the absolute deviations of each data point from the mean. If the mean of the signal is known, then the MAD is:
[0082]
[0083] S24. Extract skewness and kurtosis expert features from the I / Q sequences, specifically:
[0084] The real and imaginary parts of skewness are used to describe the asymmetry of the data distribution. If the skewness is positive, it means that the right side of the data distribution (the side with larger values) has a longer tail; if the skewness is negative, it means that the left side (the side with smaller values) has a longer tail.
[0085] If the mean and standard deviation of the sequence are known, then Skew is calculated as follows:
[0086]
[0087] Kurt is used to measure the sharpness of the peaks in a data distribution compared to a normal distribution. A kurtosis greater than 3 indicates that the data distribution is more peaked and has heavier tails than a normal distribution; less than 3 indicates that it is flatter. If the mean and standard deviation of the sequence are known, the kurtosis is calculated as follows:
[0088]
[0089] S25. Extract entropy and energy expert features from the I / Q sequence, specifically:
[0090] In information theory, entropy is a measure of uncertainty. For the real and imaginary parts of an I / Q signal, entropy can measure the complexity or information content of the signal. Information entropy:
[0091]
[0092] Signal energy calculation is typically related to the square of the signal. For the real and imaginary parts, the energy reflects the energy distribution of the signal in these two parts. Let the sample sequence be... The energy calculation is as follows:
[0093]
[0094] S26. Extract the root mean square of the I / Q sequence, specifically:
[0095] The root mean square (RMS) is an effective measure of signal amplitude, which is related to the signal's energy. For both the real and imaginary parts of the signal, the RMS reflects the average energy intensity of those parts. If the energy En of the sequence is known, then the RMS is:
[0096]
[0097] S27. Perform FFT on the real and imaginary parts of the I / Q sequence, specifically:
[0098] Let the sample sequence be Then, after FFT transformation, it becomes:
[0099]
[0100] S3. Construct a 1DCNN network. See [link to specific structure] for details. Figure 2 Specifically, it includes:
[0101] S31. The 1DCNN network structure mainly consists of convolutional layers, pooling layers, Flatten layers, and fully connected layers;
[0102] S32. The specific operation includes first adopting a parallel branch design for the I / Q signals. The I and Q signals are first input into their respective branches, and then go through the feature extraction process of "input layer - convolutional layer - pooling layer - convolutional layer - pooling layer". The convolutional layer uses learnable convolutional kernels to capture local features of the signal, and the pooling layer reduces dimensionality and enhances robustness through downsampling.
[0103] S33. The multidimensional feature map is flattened into a one-dimensional vector by the Flatten layer, and then the two features are fused by the convergence layer. Finally, the non-linear relationship between the features is learned by the fully connected layer, and the processing result is output.
[0104] S4. Construct the SE attention mechanism module; see [link to module for details]. Figure 3 Specifically, it includes:
[0105] S41. The SE attention mechanism module mainly consists of spatial feature compression and direct feature transfer;
[0106] S42. Input Features Compress spatial dimensions using global average pooling:
[0107]
[0108] Where H and W are the height and width of the feature map, respectively. : The global statistics of the c-th channel, representing the "global response" of the spatial characteristics of that channel;
[0109] S43. Channel Feature Learning: This method learns the dependencies between channels through nonlinear transformations and generates channel attention weights. Two fully connected layers are used to learn the nonlinear relationships between channels. First, the compressed feature vectors are passed through a dimensionality-reducing fully connected layer, then through a ReLU activation function, then through an up-dimensional fully connected layer, and finally through a Sigmoid activation function to obtain the weights (values between 0 and 1) for each channel.
[0110]
[0111] in, It is the compressed feature vector. It is a dimension reduction matrix (dimension is ). ), It is the ReLU activation function. It is an upgraded matrix (dimension is 1). ), It's the Sigmoid function. Ultimately, it yields the feature vector for each channel. ;
[0112] S44. Multiply the learned channel weights s with the original feature map channel by channel, that is, multiply the feature map of each channel by the corresponding weight value, so as to achieve feature recalibration.
[0113]
[0114] in, It is the weight of the c-th channel, which is a scalar; It is the c-th channel of the original feature map; It is the recalibrated feature map.
[0115] Ultimately, by enhancing the features of important channels and suppressing those of secondary channels, a new feature cube with channel attention is output, making the network pay more attention to key channel information and improving its feature representation ability.
[0116] S5, Signal Recognition Model Training
[0117] S51. Based on the dataset and feature fusion matrix, divide the training set, validation set, and test set into a 6:2:2 ratio.
[0118] S52. Construct a deep neural network model for learning and training to obtain a signal recognition model.
[0119] The network is trained using the training set and validation set, and its performance is tested using the test set. The model training is completed using the following parameters, and the specific parameter configuration is shown in Table 1:
[0120] Table 1 Hyperparameter settings during training
[0121] Optimizer Adam epoch 3000 batch_size 32 learning_rate 0.0001
[0122] S6. Deploy the trained model for wireless device identification, specifically including:
[0123] S61. Reference Figure 1 After obtaining the I / Q signals, they are normalized and then expert features are extracted to obtain the expert feature matrix.
[0124] S62. Obtain the depth feature matrix using the same depth feature extraction method as in step S2;
[0125] S63. Perform feature fusion using the same attention mechanism module as in step S3;
[0126] S64. Input the fused feature matrix into the signal recognition model, and the model output is the recognition result.
[0127] S65. To verify the effectiveness of this embodiment of the invention, the recognition results using only the expert feature matrix described in S2, the deep feature matrix described in S3, and the fusion feature matrix described in S4 as feature data are verified. In this embodiment, 8 Bluetooth devices and 2 WiFi devices are used for verification, and the results are as follows: Figure 4 As shown, the multi-dimensional feature fusion deep learning method proposed in this invention significantly improves the recognition accuracy of the device.
Claims
1. A method for identifying wireless devices based on multi-dimensional feature fusion deep learning, characterized in that, Including the following steps: Construct a wireless communication environment and collect target signal data; The target signal is preprocessed and expert features are extracted to obtain an expert feature matrix; A deep feature matrix is obtained by using a deep neural network to perform feature extraction based on deep feature learning on the signal. The expert feature matrix and the deep feature matrix are fused to obtain a feature fusion matrix; A sample set is constructed from the feature fusion matrix, a deep neural network classification model is trained, and the trained model is used to identify wireless devices.
2. The wireless device identification method based on multi-dimensional feature fusion deep learning as described in claim 1, characterized in that, The preprocessing of the target signal includes preprocessing the target signal by normalizing the maximum value and normalizing the minimum value.
3. The wireless device identification method based on multi-dimensional feature fusion deep learning as described in claim 1, characterized in that, The extraction of expert features based on prior knowledge includes calculating specific parameters of the signal based on the mathematical expression of the signal and its signal performance in different dimensions such as time domain, frequency domain, time-frequency domain, and modulation domain. These parameters may include, but are not limited to, amplitude, phase, mean, standard deviation, maximum value, minimum value, mean absolute deviation, skewness, kurtosis, information entropy, energy, root mean square, and FFT coefficients.
4. The wireless device identification method based on multi-dimensional feature fusion deep learning as described in claim 1, characterized in that, The deep neural network is a one-dimensional convolutional neural network.
5. The wireless device identification method based on multi-dimensional feature fusion deep learning as described in claim 1, characterized in that, The feature fusion of the expert feature matrix and the deep feature matrix includes establishing a feature weight allocation model based on the SE (Squeeze-and-Excitation) attention mechanism module, realizing the weighted calculation of the expert feature matrix and the deep feature matrix, and obtaining the fused feature matrix.
6. The wireless device identification method based on multi-dimensional feature fusion deep learning as described in claim 4, characterized in that, The one-dimensional convolutional neural network comprises a convolutional layer connected to a pooling layer, followed by another convolutional layer and a pooling layer. After flattening, the concatenation results in high-dimensional features, which are then output as low-dimensional depth features through a fully connected layer. The original signal is input into the one-dimensional convolutional neural network via two branches (I and Q) for processing. The features output from the two branches are concatenated across modalities in a fusion layer to form a high-dimensional hybrid feature vector, resulting in a depth feature matrix. The specific parameter configuration can be, but is not limited to, the following implementation: The network contains a 64-channel convolutional layer connected to a pooling layer, a 128-channel convolutional layer and a pooling layer, which are then flattened and spliced into 65536-dimensional high-dimensional features, and then output as 256-dimensional deep features through a fully connected layer.
7. The wireless device identification method based on multi-dimensional feature fusion deep learning as described in claim 5, characterized in that, The specific implementation method for feature fusion of expert feature matrix and deep feature matrix includes: For input features Compress spatial dimensions using global average pooling: Where H and W are the height and width of the feature matrix, respectively. is the global statistic for the c-th channel, representing the global response of the spatial characteristics of that channel; Channel feature learning learns the dependencies between channels through nonlinear transformations, generating channel attention weights. Two fully connected layers are used to learn the nonlinear relationships between channels: first, the compressed feature vectors are passed through a dimensionality-reducing fully connected layer, then through a ReLU activation function, followed by an dimensionality-increasing fully connected layer, and finally, a Sigmoid activation function is used to obtain the weights for each channel. in, It is the compressed feature vector. It is a dimension reduction matrix (dimension is ). ), () is the ReLU activation function. It is an upgraded matrix (dimension is 1). ), () is the Sigmoid function, which ultimately yields the feature vector for each channel. ; The learned channel weights s are multiplied channel by channel with the original feature map, that is, the feature map of each channel is multiplied by the corresponding weight value, thereby realizing the recalibration of features. in, It is the weight of the c-th channel, which is a scalar; It is the c-th channel of the original feature map; It is the recalibrated feature map.