Equipment identification method and system, electronic equipment and storage medium

Abstract

The invention discloses an equipment identification method and system, electronic equipment and a storage medium, and the method comprises the steps: constructing a hybrid model based on a bidirectional long-short-term memory neural network and a convolutional neural network, and extracting the time sequence features of CSI data through the bidirectional long-short-term memory neural network; according to the method, the CSI data of the to-be-identified equipment is extracted, local features of the CSI data are extracted through a convolutional neural network, and in addition, model parameters are optimized through a cross distillation loss function, so that the model can adapt to the CSI data of the to-be-identified equipment in an incremental learning scene to perform real-time updating, and memory of existing equipment features is kept. According to the method, the accuracy and the anti-interference capability of equipment identification are remarkably improved, higher identification precision can be shown in complex wireless environments such as fast time-varying characteristics, multipath effects and signal interference, and the problem that a traditional model can only identify types of fixed equipment is effectively solved.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technologies, and in particular, to a device identification method and system, an electronic device, and a computer-readable storage medium. Background Art

[0002] With the development of wireless communication technologies and the commercialization of the fifth-generation mobile communication technology (5G), more and more devices access the Internet wirelessly, such as smart home sensors, smart cars, and smart wearable devices. These wireless devices have penetrated into our daily lives, and the interconnection and communication between devices need to be effectively identified to ensure the security and reliability of data transmission during the communication process. In a wireless communication environment, the wireless channel is an important part of the communication process. The signal is often affected by signal fading and multipath effects during the transmission in the wireless channel, making the wireless channel characteristics of devices in different positions unique. Traditional device identification methods usually rely on the extraction and analysis of physical layer characteristics, such as Channel State Information (CSI). CSI contains detailed characteristic information of the wireless channel, including the multipath effect and fading that the signal suffers during the propagation process. However, due to the fast time-varying nature and multipath effect of the wireless channel, the amplitude and phase of the signal change with time and space, which makes it difficult to extract stable and distinguishable CSI characteristics, resulting in poor accuracy of device identification. In addition, in an actual wireless communication environment, signal interference between multiple devices may cause the CSI information to be disrupted or distorted, which also affects the accuracy and robustness of device identification. In addition, with the development of technology, deep learning technology has been introduced into the field of device identification. By constructing a device identification model, the device type can be automatically identified, but existing models usually can only identify devices of a fixed type, and the generality is poor. Summary of the Invention

[0003] The present invention provides a device identification method and system, an electronic device, and a computer-readable storage medium, which can significantly improve the accuracy rate and anti-interference ability of device identification, can exhibit higher identification accuracy in the face of complex wireless environments such as fast time-varying nature, multipath effect, and signal interference, and effectively solves the problem that traditional models can only identify devices of a fixed category.

[0004] According to one aspect of the present invention, a device identification method is provided, including the following:

[0005] Obtain CSI data of devices with known types, and make data labels to generate a training data set;

[0006] Construct a device recognition model based on a bidirectional long short-term memory neural network and a convolutional neural network, and use the training data set to train the device recognition model until the model converges;

[0007] Collect the CSI data of the device to be recognized, input the CSI data of the device to be recognized into the trained device recognition model, and output the device type of the device to be recognized.

[0008] Furthermore, the process of obtaining the CSI data of the device includes the following:

[0009] Obtain the sequence of channel measurement values obtained by the sender device and the receiver device during multiple rounds of channel probing;

[0010] Use the least squares method to estimate the sequence of channel measurement values of the sender device and the receiver device respectively to obtain the channel response estimation values of both;

[0011] Use the bidirectional differential quantization algorithm to quantize the channel response estimation value into a bit sequence, and perform Gray code encoding on the bit sequence to obtain a CSI data sequence;

[0012] Slice and normalize the CSI data sequence to obtain multiple CSI data subsequences.

[0013] Furthermore, the device recognition model consists of a double-layer bidirectional long short-term memory neural network and an improved AlexNet network. Among them, the double-layer bidirectional long short-term memory neural network is used as the pre-network to extract temporal features from the CSI data, and the improved AlexNet network is used to extract local features from the temporal features and perform classification discrimination.

[0014] Furthermore, it also includes the following:

[0015] If the request content judged according to the CSI data of the device to be recognized is identity registration, then the CSI data of the device to be recognized is labeled and input into the device recognition model for incremental learning.

[0016] Furthermore, the loss function of the device recognition model is:

[0017]

[0018] Among them, \(L(\omega)\) represents the cross-distillation loss function, \(L C (\omega)\) represents the cross-entropy loss function, represents the distillation loss function, \(F\) represents the total number of all old classes in the model, \(N\) represents the number of samples in one training, \(C\) represents the current number of classifications of the model, \(p ij represents whether the true label of the \(i\)-th sample is the \(j\)-th class, \(p ij takes 0 or 1, \(qij Denotes the predicted probability value that the i-th sample belongs to the j-th class, pdist ij and qdist ij respectively denote the deformations of p ij and q ij obtained by softening p ij and q ij respectively.

[0019] Furthermore, the process of incremental learning includes the following:

[0020] Construct an incremental learning training set based on the representative CSI data in the original training dataset and the CSI data of the device to be identified;

[0021] Use the incremental learning training set to train the model for parameter optimization;

[0022] Use a subset of the original training dataset to train the model to achieve fine-tuning of the model parameters;

[0023] Store the CSI data of the device to be identified in the storage space of the representative CSI data.

[0024] Furthermore, Dropout regularization is set both inside each layer of the bidirectional long short-term memory neural network and between two layers of the bidirectional long short-term memory neural network.

[0025] In addition, the present invention also provides a device identification system, including:

[0026] A training data construction module for obtaining the CSI data of devices with known types and making data labels to generate a training dataset;

[0027] A model construction and training module for constructing a device identification model based on a bidirectional long short-term memory neural network and a convolutional neural network, and using the training dataset to train the device identification model until the model converges;

[0028] A device online identification module for collecting the CSI data of the device to be identified, inputting the CSI data of the device to be identified into the trained device identification model, and outputting the device type of the device to be identified.

[0029] Furthermore, the training data construction module includes:

[0030] A channel measurement value acquisition unit for obtaining the sequence of channel measurement values obtained by the sender device and the receiver device during multiple rounds of channel probing;

[0031] A channel response estimation unit for respectively estimating the sequences of channel measurement values of the sender device and the receiver device by using the least squares method to obtain the channel response estimation values of both.

[0032] A quantization encoding unit, which is used to quantize the channel response estimation value into a bit sequence by using a bidirectional differential quantization algorithm, and perform Gray code encoding on the bit sequence to obtain a CSI data sequence;

[0033] A normalization processing unit, which is used to slice and normalize the CSI data sequence to obtain multiple CSI data subsequences.

[0034] Furthermore, the device recognition model is composed of a double-layer bidirectional long short-term memory neural network and an improved AlexNet network. Among them, the double-layer bidirectional long short-term memory neural network is used as a pre-network to extract temporal features from the CSI data, and the improved AlexNet network is used to extract local features from the temporal features and perform classification discrimination.

[0035] Furthermore, it further includes:

[0036] An incremental learning module, which is used to perform labeling on the CSI data of the device to be recognized and input it into the device recognition model for incremental learning when it is determined according to the CSI data of the device to be recognized that the requested content is identity registration.

[0037] Furthermore, the loss function of the device recognition model is:

[0038]

[0039] Among them, L(ω) represents the cross-distillation loss function, and L C (ω) represents the cross-entropy loss function, represents the distillation loss function, F represents the total number of all old classes in the model, N represents the number of samples in one training, C represents the current number of classifications of the model, and p ij represents whether the true label of the i-th sample is the j-th class, and p ij takes 0 or 1, and q ij represents the predicted probability value that the i-th sample belongs to the j-th class, pdist ij and qdist ij respectively represent the deformations of p ij and q ij and are obtained by softening p ij and q ij

[0040] Furthermore, the incremental learning module includes:

[0041] An incremental learning training set construction unit, which is used to construct an incremental learning training set based on the representative CSI data in the original training set and the CSI data of the device to be recognized;

[0042] A parameter optimization unit for training the model using an incremental learning training set to optimize parameters.

[0043] A parameter fine-tuning unit for training the model using a subset of the original training data set to fine-tune the model parameters.

[0044] A data storage unit for storing the CSI data of the device to be recognized in the storage space of the representative CSI data.

[0045] Furthermore, Dropout regularization is set both inside each layer of the bidirectional long short-term memory neural network and between two layers of the bidirectional long short-term memory neural network. In addition, the present invention also provides an electronic device, including a processor and a memory. A computer program is stored in the memory, and the processor is configured to execute the steps of the method described above by calling the computer program stored in the memory.

[0046] In addition, the present invention also provides a computer-readable storage medium for storing a computer program for device recognition based on wireless channel characteristics. The computer program executes the steps of the method described above when running on a computer.

[0047] The present invention has the following beneficial effects:

[0048] The device recognition method of the present invention constructs a hybrid model based on a bidirectional long short-term memory neural network and a convolutional neural network. First, the bidirectional long short-term memory neural network is used to extract the temporal features of the CSI data, and then the convolutional neural network is used to extract the local features of the CSI data, significantly improving the accuracy and anti-interference ability of device recognition. It can show higher recognition accuracy in the face of complex wireless environments such as fast time-variation, multipath effect, and signal interference, and effectively solves the problem that traditional models can only recognize fixed device categories.

[0049] In addition, the device recognition system of the present invention also has the above advantages.

[0050] In addition to the purposes, features, and advantages described above, the present invention has other purposes, features, and advantages. The present invention will be further described in detail below with reference to the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention. In the drawings:

[0052] Figure 1 is a schematic flowchart of the device recognition method of the preferred embodiment of this application;

[0053] Figure 2Yes Figure 1 It is a schematic diagram of the sub - process of step S1 in

[0054] Figure 3 It is a schematic diagram of the network structure of the bidirectional long - short - term memory neural network in the preferred embodiment of the present application;

[0055] Figure 4 It is a schematic diagram of the network structure of the double - layer bidirectional long - short - term memory neural network in the preferred embodiment of the present application;

[0056] Figure 5 It is a schematic diagram of the network structure of the improved AlexNet network in the preferred embodiment of the present application;

[0057] Figure 6 It is another schematic diagram of the process of the device recognition method in the preferred embodiment of the present application;

[0058] Figure 7 Yes Figure 6 It is a schematic diagram of the sub - process of step S4 in

[0059] Figure 8 It is a schematic diagram of the logical process of the incremental learning mechanism in the preferred embodiment of the present application;

[0060] Figure 9 It is a schematic diagram of the module structure of the device recognition system in another embodiment of the present application. Specific embodiments

[0061] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The present application will be described in detail below with reference to the drawings and in combination with the embodiments.

[0062] Referring to Figure 1 , the preferred embodiment of the present application provides a device recognition method, including the following:

[0063] Step S1: Obtain the CSI data of the devices with known types, and make data labels to generate a training data set;

[0064] Step S2: Construct a device recognition model based on the bidirectional long - short - term memory neural network and the convolutional neural network, and use the training data set to train the device recognition model until the model converges;

[0065] Step S3: Collect the CSI data of the device to be recognized, input the CSI data of the device to be recognized into the trained device recognition model, and output the device type of the device to be recognized.

[0066] It can be understood that for the device recognition method of this embodiment, first, the CSI data of devices with known types is collected, and data tags are made to generate sample data, and a training data set is constructed. Then, a device recognition model is constructed based on a bidirectional long short-term memory neural network and a convolutional neural network, and the training data set is used to train the device recognition model until the model converges. Finally, the CSI data of the device to be recognized is collected, and the CSI data of the device to be recognized is input into the trained device recognition model, and the device type of the device to be recognized is output, so as to realize online device recognition. The present invention constructs a hybrid model based on a bidirectional long short-term memory neural network and a convolutional neural network. First, the bidirectional long short-term memory neural network is used to extract the temporal features of the CSI data, and then the convolutional neural network is used to extract the local features of the CSI data, significantly improving the accuracy and anti-interference ability of device recognition. It can show higher recognition accuracy in the face of complex wireless environments such as fast time-varying, multipath effects, and signal interference, and effectively solves the problem that traditional models can only recognize fixed device categories.

[0067] It can be understood that as Figure 2 shown, in the step S1, the process of obtaining the CSI data of the device includes the following contents:

[0068] Step S11: Obtain the sequence of channel measurement values obtained by the sender device and the receiver device in multiple rounds of channel sounding;

[0069] Step S12: Use the least squares method to estimate the sequence of channel measurement values of the sender device and the receiver device respectively to obtain the channel response estimation values of both;

[0070] Step S13: Use the bidirectional differential quantization algorithm to quantize the channel response estimation value into a bit sequence, and perform Gray code encoding on the bit sequence to obtain a CSI data sequence;

[0071] Step S14: Perform slicing and normalization processing on the CSI data sequence to obtain multiple CSI data subsequences.

[0072] It can be understood that the acquisition of CSI data includes four stages, namely, the channel measurement stage, the channel estimation stage, the quantization and coding stage, and the normalization processing stage. Specifically, in the channel measurement stage, the sender device Alice and the receiver device Bob send known pilot signal frames to each other. Within the coherence time Δt, the signal frames sent by Alice and Bob reach each other through the same multipath effect and channel fading, and both parties use the received signal frames to estimate the channel within the time Δt. Taking the i-th round of channel sounding as an example, at time t i , Alice sends a pilot signal frame x(t i ) to Bob, then the signal received by Bob is:

[0073]

[0074] After Bob receives the pilot signal frame sent by Alice, at time t i +Δt, Bob sends the same signal frame to Alice. Then the signal received by Alice is as follows:

[0075]

[0076] where h AB (t i , τ) and h BA (t′ i , τ) represent the channel impulse responses from Alice to Bob and from Bob to Alice respectively. t′ i = t i +Δt, n A (t′ i ) and n B (t i ) represent Gaussian white noise, and y A (t′ i ) and y B (t i ) represent the channel measurement values obtained by Alice and Bob in the i-th round of channel sounding. Next, Alice and Bob send pilot signal frames to each other multiple times to obtain sufficient channel sounding values. After multiple rounds of channel sounding between the two parties, the sequences of channel measurement values obtained by Alice and Bob are:

[0077] y A = [y A (t′1), y A (t′2),..., y A (t′ N )]

[0078] y B = [y B (t1), y B (t2),..., y B (t N )].

[0079] Then, after obtaining the sequences of channel measurement values, enter the channel estimation stage. Use the LS algorithm to perform LS estimation on the sequences of channel measurement values obtained by Alice and Bob respectively to obtain the estimated channel response values, which can be expressed as:

[0080]

[0081] Next, the channel response estimate is quantized into a bit sequence by the Bidirectional Differential Quantization (BDQ) algorithm. For example, for the data value h A (k), where k = 1, 2, 3,..., N, the following rules are used for quantization:

[0082]

[0083] where and respectively represent the 2 bits obtained after quantizing h A (n). After quantizing and extracting the CSI data, a bit string of the CSI data can be obtained. These bit strings are encoded in groups of 4 bits using the Gray code to obtain the CSI data sequence. Encoding the bit string is beneficial to improving the stability of bit string extraction.

[0084] Finally, the CSI data sequence is sliced, divided into multiple equal-length CSI data subsequences of length 256. If the length of some subsequences is not enough, 0 values are filled. Then, the CSI data subsequences are normalized to limit the range of the subsequence data values within the interval [0, 1]. The normalization method is as follows:

[0085]

[0086] where x norm represents the data after normalization, x represents the data before normalization, x min and x max respectively represent the minimum and maximum values of the data in this subsequence. It can be understood that slicing and normalizing the CSI data sequence is beneficial to accelerating the training speed of the model, enabling the model to find the global optimal solution faster, improving the generalization ability of the model, reducing the differences in features on different datasets, and improving the accuracy of model classification. In addition, the size of a single CSI data matrix after normalization is (256, 1). In addition, after obtaining the CSI data of each device, corresponding data labels are made to generate sample data, and a training dataset is constructed. The training dataset is divided into a training set, a validation set, and a test set according to a preset ratio.

[0087] It can be understood that in step S2, a device recognition model is constructed based on a bidirectional long short-term memory neural network and a convolutional neural network, and the device recognition model is trained using the training dataset constructed in step S1 until the model converges. Among them, the device recognition model consists of a double-layer bidirectional long short-term memory neural network and an improved AlexNet network. Among them, the double-layer bidirectional long short-term memory neural network is used as a pre-network to extract temporal features from CSI data, and the improved AlexNet network is used to extract local features from the temporal features and perform classification and discrimination.

[0088] It can be understood that the present invention uses a bidirectional long short-term memory neural network to extract the temporal features of CSI data. The bidirectional long short-term memory neural network can learn both the front and back directions of the data sequence at the same time, and can capture context information more comprehensively when learning sequence features, which is beneficial to improving the accuracy of classification and recognition. Among them, the structure of the bidirectional long short-term memory neural network is as Figure 3 shown. During the training process of the bidirectional long short-term memory neural network, the parameters are updated through the outputs in the forward and reverse directions of the model, so that each memory cell state contains both past information and future information, so that the model can better utilize the features in the temporal dimension for comprehensive learning. And, the bidirectional long short-term memory neural network is trained in both the forward and reverse directions, and the learned features are more comprehensive, reducing the risk of model overfitting. The output h t of the bidirectional long short-term memory neural network is composed of the output of the forward long short-term memory neural network and the output of the reverse long short-term memory neural network in series. The specific calculation method is as follows:

[0089]

[0090] Among them, represents the output of the forward long short-term memory neural network, represents the output of the reverse long short-term memory neural network, represents the exclusive OR operator. Therefore, through the bidirectional long short-term memory neural network, CSI data can be processed in the forward order and the reverse order to extract rich CSI temporal features.

[0091] Among them, the network structure diagram of the double-layer long short-term memory neural network is as Figure 4 shown. The output matrix of the first-layer bidirectional long short-term memory neural network is: The output matrix of the second-layer bidirectional long short-term memory neural network is: It can be understood that the present invention uses a double-layer long short-term memory neural network, which further improves the extraction accuracy of CSI data temporal features and is beneficial to improving the accuracy of device classification and recognition.

[0092] Optionally, considering that the present model uses a bidirectional long short-term memory neural network and a large amount of data is input during the training process, in order to enhance the generalization ability of the model, Dropout regularization is adopted for setting inside each layer of the bidirectional long short-term memory neural network and between two layers of the bidirectional long short-term memory neural network, that is, a Dropout layer is added to the input data, and a Dropout layer is added in the middle of the double-layer network.

[0093] In addition, a convolutional neural network is further used to perform feature extraction and classification on the output information of the double-layer bidirectional long short-term memory neural network. Through the convolution operation of the convolutional neural network, local features in the CSI time series features can be further mined. The present invention improves on the existing AlexNet network. Among them, the network structure of the improved AlexNet network is as Figure 5 shown. Specifically, from the model parameters of the bidirectional long short-term memory neural network, it can be known that the dimension of its output feature matrix is (256, 256×2). Therefore, the size of the convolutional kernel of the first convolutional layer of the improved AlexNet network is set to 39×39, the number of convolutional kernels is 96, no padding is performed, and the convolutional stride is set to 4. The size of the processed output matrix can be obtained as 96×55×55, and then max pooling is performed. The size of the pooling layer is 3×3, and the pooling stride is 2. Therefore, the size of the matrix finally output by the first convolutional layer is 96×27×27, and the parameters of the following four convolutional layers are the same as those of the existing AlexNet network. After 5 layers of convolution operations, the size of the data matrix becomes 256×6×6. This data matrix is input into a three-layer fully connected layer, and a Dropout layer is added between the fully connected layers to prevent overfitting. Finally, the result output by the model is classified and discriminated through the softmax function to identify the wireless channel feature information of the device, and 5 types of device identity types can be output.

[0094] It can be understood that in step S3, in the online recognition stage, the CSI data of the device to be recognized is collected and the request content of the device to be recognized is judged. If the request content of the device to be recognized is to apply for identity recognition, the CSI data of the device to be recognized is input into the trained device recognition model, and the device type of the device to be recognized can be output, thereby realizing device online recognition. Among them, the device recognition model will output a set of recognition probabilities. If the maximum value in the probability set is greater than or equal to the set threshold, it is determined that the identity of the device to be recognized is the identity label corresponding to the maximum value, otherwise it is determined that the device to be recognized is an illegal device.

[0095] In addition, as Figure 6 shown, the device recognition further includes the following content:

[0096] Step S4: If it is determined that the requested content is identity registration based on the CSI data of the device to be recognized, the CSI data of the device to be recognized is tagged and then input into the device recognition model for incremental learning.

[0097] It can be understood that as the devices to be recognized emerge continuously, the model needs to be continuously updated to adapt to the addition of the devices to be recognized. Therefore, the present invention sets up an incremental learning mechanism for model update. Among them, as Figure 7 shown, the process of incremental learning includes the following:

[0098] Step S41: Construct an incremental learning training set based on the representative CSI data in the original training dataset and the CSI data of the device to be recognized;

[0099] Step S42: Use the incremental learning training set to train the model to optimize the parameters;

[0100] Step S43: Use a subset of the original training dataset to train the model to achieve fine-tuning of the model parameters;

[0101] Step S44: Store the CSI data of the device to be recognized in the storage space of the representative CSI data.

[0102] Specifically, as Figure 8 shown, first construct an incremental learning training set based on the CSI data of the device to be recognized and the representative CSI data of the old devices in the representative sample storage space to ensure the effective integration of the data of the new and old devices in training; then, use the incremental learning training set to train the model and optimize the parameters through the loss function to ensure that the model maintains the recognition ability of the old devices while learning the features of the device to be recognized; next, use a subset of the original training dataset to train the model to ensure that the number of samples in each category is the same, and perform more refined parameter adjustment by reducing the learning rate to avoid overfitting and bias, ensuring the stability and accuracy of the model during the update process; finally, store the CSI data of the device to be recognized in the representative sample storage space and make room through the sample removal method to continuously expand the recognition range of the model in subsequent incremental learning.

[0103] It can be understood that traditional models often need to be retrained or manually adjusted when identifying devices to be identified, and it is impossible to continuously learn the identity characteristics of the devices to be identified based on the original model, resulting in insufficient flexibility and scalability. However, the present invention introduces an incremental learning mechanism, allowing the model to continuously learn the characteristics of the devices to be identified while maintaining the ability to identify old devices, overcoming the limitation that traditional methods can only identify fixed device categories, thereby enhancing the scalability and adaptability of the model in multi-device identification, enabling efficient device identification in an environment where the device categories and quantities change dynamically, ensuring that the model does not need to retrain the entire model when facing devices to be identified, and being able to quickly adapt to the changing wireless communication environment.

[0104] Optionally, the loss function of the device identification model is:

[0105]

[0106] where L(ω) represents the cross-distillation loss function, L C (ω) represents the cross-entropy loss function, represents the distillation loss function, F represents the total number of all old classes in the model, and old classes refer to the device categories that the model has been able to identify after previous training. N represents the number of samples in one training, C represents the current number of classifications of the model, p ij represents whether the true label of the i-th sample is the j-th class, p ij takes 0 or 1, q ij represents the predicted probability value that the i-th sample belongs to the j-th class, pdist ij and qdist ij respectively represent the deformations of p ij and q ij obtained by softening p ij and q ij . Additionally, the specific softening calculation process is:

[0107]

[0108] where T represents the distillation parameter, which is used to control the softening effect of the output class probability values. When T = 1, the probability value distribution is the same as that of the Softmax function; when T < 1, the probability value distribution is sharper than the original distribution, that is, the gap between classes is larger, highlighting some classes with larger probabilities; when T > 1, the probability value distribution is more moderate than the original distribution, that is, the gap between classes is smaller, making the probability distributions of all classes relatively similar.

[0109] It can be understood that the present invention uses a cross-distillation loss function to optimize the model parameters, which combines a distillation loss function and a multi-class cross-entropy loss function. The distillation loss function can preserve the old-class classification ability of the model, that is, the ability of the model to identify old devices, while the multi-class cross-entropy loss function can update the learning of the model's classification ability for all classes, enabling the model to continuously learn the features of the devices to be identified while retaining the ability to identify old devices, greatly improving the accuracy and robustness of device identity recognition.

[0110] In addition, as Figure 9 shown, another embodiment of the present invention further provides a device recognition system, preferably adopting the device recognition method as described above, including:

[0111] A training data construction module, configured to obtain CSI data of devices with known types and make data labels to generate a training data set;

[0112] A model construction and training module, configured to construct a device recognition model based on a bidirectional long short-term memory neural network and a convolutional neural network, and use the training data set to train the device recognition model until the model converges;

[0113] A device online recognition module, configured to collect CSI data of the device to be recognized, input the CSI data of the device to be recognized into the trained device recognition model, and output the device type of the device to be recognized.

[0114] It can be understood that for the device recognition system of this embodiment, first, the CSI data of devices with known types is collected, and data labels are made to generate sample data, and a training data set is constructed. Then, a device recognition model is constructed based on a bidirectional long short-term memory neural network and a convolutional neural network, and the training data set is used to train the device recognition model until the model converges. Finally, the CSI data of the device to be recognized is collected, the CSI data of the device to be recognized is input into the trained device recognition model, and the device type of the device to be recognized is output, thereby realizing online device recognition. The present invention constructs a hybrid model based on a bidirectional long short-term memory neural network and a convolutional neural network. First, the bidirectional long short-term memory neural network is used to extract the temporal features of the CSI data, and then the convolutional neural network is used to extract the local features of the CSI data, significantly improving the accuracy and anti-interference ability of device recognition, and can show higher recognition accuracy in the face of complex wireless environments such as fast time-variation, multipath effect, and signal interference, and effectively solves the problem that traditional models can only recognize fixed device categories.

[0115] In addition, the device recognition system further includes:

[0116] An incremental learning module, which is used to perform incremental learning by tokenizing the CSI data of the device to be recognized and inputting it into the device recognition model when it is determined according to the CSI data of the device to be recognized that the requested content is identity registration.

[0117] Among them, the loss function of the device recognition model is:

[0118]

[0119] Among them, L(ω) represents the cross-distillation loss function, and L C (ω) represents the cross-entropy loss function. represents the distillation loss function, F represents the total number of all old classes in the model, N represents the number of samples in one training, C represents the current number of classifications of the model, and p ij represents whether the true label of the i-th sample is the j-th class, and p ij takes 0 or 1, and q ij represents the predicted probability value that the i-th sample belongs to the j-th class, pdist ij and qdist ij respectively represent the deformations of p ij and q ij and are obtained by softening p ij and q ij .

[0120] Among them, the incremental learning module includes:

[0121] An incremental learning training set construction unit, which is used to construct an incremental learning training set based on the representative CSI data in the original training dataset and the CSI data of the device to be recognized;

[0122] A parameter optimization unit, which is used to train the model with the incremental learning training set to optimize the parameters;

[0123] A parameter fine-tuning unit, which is used to train the model with a subset of the original training dataset to fine-tune the model parameters;

[0124] A data storage unit, which is used to store the CSI data of the device to be recognized in the storage space of the representative CSI data.

[0125] Among them, Dropout regularization is set both inside each layer of the bidirectional long short-term memory neural network and between two layers of the bidirectional long short-term memory neural network.

[0126] It can be understood that each module in the embodiments of this system corresponds to each step in the embodiments of the above method. Therefore, the specific working processes of each module will not be elaborated here, and reference can be made to the embodiments of the above method for correspondence.

[0127] In addition, another embodiment of the present invention further provides an electronic device, including a processor and a memory. A computer program is stored in the memory. The processor is configured to execute the steps of the method as described above by invoking the computer program stored in the memory.

[0128] In addition, another embodiment of the present invention further provides a computer-readable storage medium for storing a computer program for device identification based on wireless channel characteristics. The computer program, when running on a computer, executes the steps of the method as described above.

[0129] The forms of common computer-readable storage media generally include: floppy disks, flexible disks, hard disks, magnetic tapes, any other magnetic media, CD-ROMs, any other optical media, punch cards, paper tapes, any other physical media with a pattern of holes, random access memories (RAMs), programmable read-only memories (PROMs), erasable programmable read-only memories (EPROMs), flash erasable programmable read-only memories (FLASH-EPROMs), any other memory chips or cartridges, or any other media readable by a computer. The instructions can further be transmitted or received by a transmission medium. The term transmission medium can include any tangible or intangible medium that can be used to store, encode, or carry instructions for execution by a machine, and includes digital or analog communication signals or an intangible medium that facilitates the communication of the above instructions. The transmission medium includes coaxial cables, copper wires, and optical fibers, which include the wires of a bus used to transmit a computer data signal.

[0130] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program code. The solutions in the embodiments of the present application can be implemented in various computer languages. For example, object-oriented programming languages such as Java and interpreted scripting languages such as JavaScript.

[0131] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one or more of the flows Figure 1 one or more of the flows and / or blocks Figure 1 or means for implementing the functions specified in one or more of the blocks.

[0132] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means for implementing the functions specified in one or more of the flows Figure 1 one or more of the flows and / or blocks Figure 1 or means for implementing the functions specified in one or more of the blocks.

[0133] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more of the flows Figure 1 one or more of the flows and / or blocks Figure 1 or means for implementing the functions specified in one or more of the blocks.

[0134] Although the preferred embodiments of the present application have been described, those skilled in the art can make additional changes and modifications once they learn the basic creative concepts. Therefore, the appended claims are intended to be construed to include the preferred embodiments as well as all changes and modifications falling within the scope of the present application.

[0135] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application is also intended to include these changes and modifications.

[0136] The above description is only for the preferred embodiments of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A device identification method, characterized in that: Includes the following: Obtain CSI data of devices of known types and create data labels to generate training datasets; A device recognition model is constructed based on a bidirectional long short-term memory neural network and a convolutional neural network, and the device recognition model is trained using a training data set until the model converges; Collect the CSI data of the device to be identified, input the CSI data of the device to be identified into the trained device identification model, and output the device type of the device to be identified.

2. The device identification method according to claim 1, characterized in that: The process of obtaining CSI data for a device of known type includes the following: Obtaining a sequence of channel measurement values ​​obtained by a sending device and a receiving device during multiple rounds of channel detection; The least square method is used to estimate the channel measurement value sequences of the transmitting device and the receiving device respectively to obtain the channel response estimation values ​​of the two devices; The channel response estimation value is quantized into a bit sequence using a bidirectional differential quantization algorithm, and the bit sequence is encoded by Gray code to obtain a CSI data sequence; The CSI data sequence is sliced ​​and normalized to obtain multiple CSI data subsequences.

3. The device identification method according to claim 1, characterized in that: The device identification model consists of a double-layer bidirectional long short-term memory neural network and an improved AlexNet network, wherein the double-layer bidirectional long short-term memory neural network is used as a front network to extract time series features from CSI data, and the improved AlexNet network is used to extract local features from the time series features and perform classification and discrimination.

4. The device identification method according to claim 1, characterized in that: Also included are the following: If it is determined that the request content is identity registration based on the CSI data of the device to be identified, the CSI data of the device to be identified is labeled and then input into the device identification model for incremental learning.

5. The device identification method according to claim 4, characterized in that: The loss function of the device identification model is: Among them, L(ω) represents the cross distillation loss function, L C (ω) represents the cross entropy loss function, represents the distillation loss function, F represents the total number of all old classes in the model, N represents the number of samples trained at one time, C represents the current number of classifications of the model, and p ij Indicates whether the true label of the i-th sample is the j-th category, p ij Takes 0 or 1, q ij Indicates the predicted probability value that the i-th sample belongs to the j-th class, pdist ij and qdist ij Respectively represent p ij and q ij The deformation of p ij and q ij Soften to obtain.

6. The device identification method according to claim 4, characterized in that: The incremental learning process includes the following: Construct an incremental learning training set based on the representative CSI data in the original training data set and the CSI data of the device to be identified; The model is trained using the incremental learning training set for parameter optimization; Use a subset of the original training data set to train the model to fine-tune the model parameters; The CSI data of the device to be identified is stored in the storage space of the representative CSI data.

7. The device identification method according to claim 3, characterized in that: Dropout regularization is set inside each layer of bidirectional long short-term memory neural network and between two layers of bidirectional long short-term memory neural network.

8. A device identification system, characterized in that: include: A training data building module is used to obtain CSI data of devices of known types and make data labels to generate training data sets; Model building and training module, used to build a device recognition model based on bidirectional long short-term memory neural network and convolutional neural network, and train the device recognition model using training data set until the model converges; The device online identification module is used to collect the CSI data of the device to be identified, input the CSI data of the device to be identified into the trained device identification model, and output the device type of the device to be identified.

9. The device identification system according to claim 8, characterized in that: The training data building module includes: A channel measurement value acquisition unit, used to acquire a channel measurement value sequence obtained by a sending device and a receiving device during multiple rounds of channel detection; A channel response estimation unit, used to estimate the channel measurement value sequences of the transmitting device and the receiving device respectively by using the least square method to obtain channel response estimation values ​​of both; A quantization coding unit, used to quantize the channel response estimation value into a bit sequence by using a bidirectional differential quantization algorithm, and perform Gray code encoding on the bit sequence to obtain a CSI data sequence; The normalization processing unit is used to slice and normalize the CSI data sequence to obtain multiple CSI data subsequences.

10. The device identification system according to claim 8, characterized in that: The device identification model consists of a double-layer bidirectional long short-term memory neural network and an improved AlexNet network, wherein the double-layer bidirectional long short-term memory neural network is used as a front network to extract time series features from CSI data, and the improved AlexNet network is used to extract local features from the time series features and perform classification and discrimination.

11. The device identification system according to claim 8, characterized in that: Also includes: The incremental learning module is used to label the CSI data of the device to be identified and input it into the device identification model for incremental learning when it is determined that the request content is identity registration based on the CSI data of the device to be identified.

12. The device identification system according to claim 11, characterized in that: The loss function of the device identification model is: Among them, L(ω) represents the cross distillation loss function, L C (ω) represents the cross entropy loss function, represents the distillation loss function, F represents the total number of all old classes in the model, N represents the number of samples trained at one time, C represents the current number of classifications of the model, and p ij Indicates whether the true label of the i-th sample is the j-th category, p ij Takes 0 or 1, q ij Indicates the predicted probability value that the i-th sample belongs to the j-th class, pdist ij and qdist ij Respectively represent p ij and q ij The deformation of p ij and q ij Soften to obtain.

13. The device identification system according to claim 11, characterized in that: The incremental learning module includes: An incremental learning training set construction unit, used to construct an incremental learning training set based on representative CSI data in an original training data set and CSI data of a device to be identified; A parameter optimization unit, used to train the model using the incremental learning training set to perform parameter optimization; A parameter fine-tuning unit, used to train the model using a subset of the original training data set to achieve model parameter fine-tuning; The data storage unit is used to store the CSI data of the device to be identified in the storage space of the representative CSI data.

14. The device identification system according to claim 10, characterized in that: Dropout regularization is set inside each layer of bidirectional long short-term memory neural network and between two layers of bidirectional long short-term memory neural network.

15. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores a computer program, and the processor executes the steps of the method according to any one of claims 1 to 7 by calling the computer program stored in the memory.

16. A computer-readable storage medium for storing a computer program for device identification based on wireless channel characteristics, characterized in that: When the computer program is run on a computer, the steps of the method according to any one of claims 1 to 7 are executed.

Images