Personnel perception wireless signal generation method based on decoupling representation

Abstract

The invention discloses a personnel perception wireless signal generation method based on decoupling characterization, which belongs to the technical field of wireless perception, and comprises the following steps: converting the format of an input wireless signal, and converting and storing all real wireless signals as a spectrogram to obtain an original data set; pairing the wireless signals subjected to format conversion; inputting the paired wireless signal group into an encoder of a neural network to obtain a feature vector; the feature vectors are decoupled; modifying the decoupled feature vector to obtain a feature vector of the target; and inputting the modified feature vector into a neural network decoder to obtain a generated wireless signal. The personnel perception wireless signal generation method based on decoupling representation provided by the invention can effectively solve the problems in the prior art that complex information characteristics in signals are not fully considered, information of different targets in original signals is difficult to extract and decouple independently, and new signals different from the original signals in category cannot be generated.

Description

Technical Field

[0001] The present invention relates to the field of wireless sensing technology, and in particular to a method for generating a personnel sensing wireless signal based on decoupling characterization. Background Art

[0002] Wireless sensing technology is to transmit wireless signals through wireless devices (such as Wi-Fi, millimeter wave radar, ultra-wideband radar, etc.). The signal is transmitted in free space through electromagnetic waves. During the transmission process, the wireless signal will interact with the surrounding environment (such as reflection, refraction, etc.). The interacting objects include the human targets that need to be detected. The signals after interaction are collected and processed and analyzed to achieve the purposes of motion recognition, positioning, and monitoring of the detected targets.

[0003] Compared with traditional visual technology, wireless sensing technology has strong privacy protection and passivity, that is, the content of the collected signals cannot be visually understood, and the monitored objects will not be restricted during the monitoring process, and they can be allowed to move freely without wearing detection equipment. Therefore, this technology has great application value in smart home, Internet of Things, biomedicine and other fields. With the rapid development of artificial intelligence (AI), it has shown significant effects in various fields. Therefore, the application of AI in the field of wireless sensing is also very promising. In recent years, related work has emerged in an endless stream, which has led to the rapid development of this cross-field. However, there is a strict requirement for AI-based methods, that is, a large amount of data is needed for model training to achieve the desired effect. However, in the field of wireless sensing, the cost of wireless signal acquisition is high, which will consume a lot of manpower, equipment, time, venue, etc. Therefore, the lack of data limits its reliable and widespread application.

[0004] Thanks to the rapid development of generative artificial intelligence in recent years, generating the above wireless signals through simulation based on AI has become an effective solution, which eliminates the need to collect a large amount of signal data through real equipment in a real environment, thereby reducing costs. However, the existing wireless signal generation method is to input the real collected wireless signal into a neural network to extract features, then add a certain amount of Gaussian noise to it, and then generate a new simulated wireless signal through the processed features. The diversity of the wireless signals generated is insufficient. For example, the real collected signals of person 1 knocking and person 2 waving can only generate the same signals of person 1 knocking and person 2 waving based on these data. The data set can only be expanded in terms of quantity but not in terms of category. For example, it is impossible to generate signals of person 1 waving and person 2 knocking, which greatly reduces the diversity of the generated signals and limits its application. The reason for this limitation is that existing methods regard wireless signal data samples as complete individuals, ignoring the various types of information contained in the collection scene. They directly extract features from the original signals and add noise to regenerate new signals. They are unable to decouple the different information and then modify them in a targeted manner to generate new signals with different categories from the original signals. For example, in the task of person gesture recognition, the collected signals mainly contain information about people and gestures, and their mutual coupling exists in the collected signals. Existing methods are unable to extract the features of people and gestures separately, and therefore are unable to modify a certain type of information to generate new signals with different categories from the original signals. Summary of the invention

[0005] The purpose of the present invention is to provide a method for generating a personnel perception wireless signal based on decoupling characterization to solve the problems existing in the background technology.

[0006] To achieve the above object, the present invention provides a method for generating a human perception wireless signal based on decoupling characterization, comprising the following steps: S1. Convert the format of the input wireless signal, convert all real wireless signals into spectrograms and save them to obtain the original data set; S2, pair the converted wireless signals, and pair the wireless signals in pairs from the original data set in S1. The pairing rule is that the two signals are paired , the two signals have the same information (for example, in gesture recognition tasks, the signal The information is: Person 1 waves, signal The information is: Person 2 waves, then the gesture information between them is the same); S3, inputting the paired wireless signal group into the encoder of the neural network to obtain a feature vector; S4, decoupling the feature vector; S5, modifying the decoupled feature vector to obtain the target feature vector; S6. Input the modified feature vector into the neural network decoder to obtain the generated wireless signal.

[0007] Preferably, the content of S3 is as follows: S31, the signal in S2 Feed the encoder neural network in batches , each batch gets the feature vector ; S32, the feature vector in S31 According to the number of information The sub-vectors are obtained and .

[0008] Preferably, the content of S4 is as follows: S41, the feature vector in S32 Input decoder neural network , and the generated wireless signal is ; S42, the wireless signal generated in S41 The signal pair in S2 At the same time, input the loss function , calculate the generated signal and its corresponding real signal The specific calculation process is as follows: ; in, is the L1 norm; For expectations; express Obey the real signal data distribution; S43, the wireless signal generated in S41 The signal pair in S2 At the same time, input the loss function , calculate the generated signal and its corresponding real signal The specific calculation process is as follows: ; in, express Obey the real signal data distribution; S44, the loss function in S42 and the loss function in S43 Add together to get the loss function , the specific calculation process is as follows: ; S45, the feature vector in S32 Swap the kth subvector and perform the swap operation Get the exchange eigenvector ; S46, the exchange feature vector in S45 Input the decoder neural network separately Get the generated wireless signal ; S47, the wireless signal generated in S46 The signal pair in S2 At the same time, input the loss function , calculate the generated signal and its corresponding real signal similarity; S48, the wireless signal generated in S46 The signal pair in S2 At the same time, input the loss function , calculate the generated signal and its corresponding real signal similarity; S49, the loss function in S47 and the loss function in S48 Add together to get the loss function , the specific calculation process is as follows: ; S410, the loss function in S44 And the loss function in S49 Add together to get the decoupling stage loss function , the specific calculation process is as follows: ; in is a hyperparameter; S411, use the decoupling phase loss function in S410 Encoder Neural Network and decoder neural network Conduct training, iterate N cycles, and complete decoupling training; S412, the encoder neural network Parameters frozen.

[0009] Preferably, the S45 calculation formula is as follows: .

[0010] Preferably, the calculation process of S47 is as follows: .

[0011] Preferably, the calculation process of S48 is as follows: .

[0012] Preferably, the content of S5 is as follows: S51, select two signals from the original data set in S1 , the two signals do not have the same information (for example, in gesture recognition tasks, the signal The information is: Person 1 waves, signal The information is: Person 2 taps, then the person information and gesture information between them are different); S52, the signal in S51 is Feed the encoder neural network in batches , each batch gets the feature vector ; S53, the feature vector in S52 According to the number of information The sub-vectors are obtained and ; S54, the feature vector in S53 Swap the kth subvector and perform the swap operation Get the exchange eigenvector The specific operation process is as follows: .

[0013] Preferably, S6 has the following content: S61, the feature vector in S54 Input decoder neural network , and the generated wireless signal is ; S62, select two signals from the original data set in S1 , where the signal The information and the wireless signal generated in S61 Same, signal The information and the wireless signal generated in S61 Same (for example, in gesture recognition tasks, the signal and The information is: Person 1 knocks, signal and The information is: Person 2 waves, then their personnel information and gesture information are the same); S63, the wireless signal generated in S61 With the signal in S62 At the same time, input the loss function , calculate the generated signal and its corresponding real signal similarity; S64, the wireless signal generated in S61 With the signal in S62 At the same time, input the loss function , calculate the generated signal and its corresponding real signal similarity; S65. Change the loss function in S63 and the loss function in S64 Add together to get the loss function , the specific calculation process is as follows: ; S66, use the generation stage loss function in S65 Decoder Neural Network Train and iterate cycles to complete the generation training; S67, S66 The generated wireless signal obtained after the training is completed Then generate a signal for the final goal .

[0014] Preferably, the calculation process of S63 is as follows: .

[0015] Preferably, the calculation process of S64 is as follows: .

[0016] Therefore, the present invention adopts the above-mentioned method for generating a human-perceived wireless signal based on decoupling characterization, which has the following beneficial effects: (1) It is possible to extract effective features for generation from the original signal. Compared with the traditional signal generation method, this method generates the simulated original signal by encoding the signal and then decoding it, so that the encoded and extracted features contain the effective information in the original signal; (2) Use the decoupling method to generate signals. Compared with the traditional signal generation method, this method can decouple different information in the signal and modify the feature part in a targeted manner. Based on the modified features, a new signal with a different category from the original signal can be generated, thereby improving the diversity of the generated signal.

[0017] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic diagram of a flow chart of an embodiment of the present invention; Figure 2 This is a spectrum diagram obtained after the wireless signal of an embodiment of the present invention is converted; Figure 3 A flow chart for generating wireless signal decoupling proposed in an embodiment of the present invention; Figure 4 The cosine similarity between each feature sub-vector in the embodiment of the present invention changes with the iteration process; Figure 5 The figure is a schematic diagram of testing an embodiment of the present invention on the public wireless signal dataset Dop-NET dataset. DETAILED DESCRIPTION

[0019] The following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] See also Figure 1-Figure 4 , a method for generating a human-sensing wireless signal based on decoupling characterization, comprising the following steps: S1. Convert the format of the input wireless signal, convert all real wireless signals into spectrograms and save them to obtain the original data set.

[0021] S2, pair the converted wireless signals, and pair the wireless signals in pairs from the original data set in S1. The pairing rule is that the two signals are paired , the two signals have the same information (for example, in gesture recognition tasks, the signal The information is: Person 1 waves, signal The information is: Person 2 waves, then the gesture information between them is the same).

[0022] S3. Input the paired wireless signal group into the encoder of the neural network to obtain a feature vector.

[0023] S31, the signal in S2 Feed the encoder neural network in batches , each batch gets the feature vector ; S32, the feature vector in S31 According to the number of information The sub-vectors are obtained and .

[0024] S4. Decouple the feature vector.

[0025] S41, the feature vector in S32 Input decoder neural network , and the generated wireless signal is ; S42, the wireless signal generated in S41 The signal pair in S2 At the same time, input the loss function , calculate the generated signal and its corresponding real signal The specific calculation process is as follows: ; in, is the L1 norm; For expectations; express Obey the real signal data distribution; S43, the wireless signal generated in S41 The signal pair in S2 At the same time, input the loss function , calculate the generated signal and its corresponding real signal The specific calculation process is as follows: ; in, express Obey the real signal data distribution; S44, the loss function in S42 and the loss function in S43 Add together to get the loss function , the specific calculation process is as follows: ; S45, the feature vector in S32 Swap the kth subvector and perform the swap operation Get the exchange eigenvector , the specific process is as follows: ; S46, the exchange feature vector in S45 Input the decoder neural network separately Get the generated wireless signal ; S47, the wireless signal generated in S46 The signal pair in S2 At the same time, input the loss function , calculate the generated signal and its corresponding real signal The specific calculation process is as follows: ; S48, the wireless signal generated in S46 The signal pair in S2 At the same time, input the loss function , calculate the generated signal and its corresponding real signal The specific calculation process is as follows: ; S49, the loss function in S47 and the loss function in S48 Add together to get the loss function , the specific calculation process is as follows: ; S410, the loss function in S44 And the loss function in S49 Add together to get the decoupling stage loss function , the specific calculation process is as follows: ; in is a hyperparameter; S411, use the decoupling phase loss function in S410 Encoder Neural Network and decoder neural network Training is performed, iterating N cycles to complete the decoupling training. During the N-cycle iteration, the feature vector in S32 Each sub-vector in is decoupled, and the success of decoupling can be evaluated by calculating the cosine similarity between them. Figure 3 As shown, the cosine similarity between feature sub-vectors gradually decreases as the decoupling proceeds, verifying the effectiveness of the decoupling method.

[0026] S412, the encoder neural network Parameters frozen.

[0027] S5. Modify the decoupled feature vector to obtain the target feature vector.

[0028] S51, select two signals from the original data set in S1 , the two signals do not have the same information (for example, in gesture recognition tasks, the signal The information is: Person 1 waves, signal The information is: Person 2 taps, then the person information and gesture information between them are different); S52, the signal in S51 is Feed the encoder neural network in batches , each batch gets the feature vector ; S53, the feature vector in S52 According to the number of information The sub-vectors are obtained and ; S54, the feature vector in S53 Swap the kth subvector and perform the swap operation Get the exchange eigenvector The specific operation process is as follows: .

[0029] S6. Input the modified feature vector into the neural network decoder to obtain the generated wireless signal.

[0030] S61, the feature vector in S54 Input decoder neural network , and the generated wireless signal is ; S62, select two signals from the original data set in S1 , where the signal The information and the wireless signal generated in S61 Same, signal The information and the wireless signal generated in S61 Same (for example, in gesture recognition tasks, the signal and The information is: Person 1 knocks, signal and The information is: Person 2 waves, then their personnel information and gesture information are the same); S63, the wireless signal generated in S61 With the signal in S62 At the same time, input the loss function , calculate the generated signal and its corresponding real signal The specific calculation process is as follows: ; S64, the wireless signal generated in S61 With the signal in S62 At the same time, input the loss function , calculate the generated signal and its corresponding real signal The specific calculation process is as follows: ; S65. Change the loss function in S63 and the loss function in S64 Add together to get the loss function , the specific calculation process is as follows: ; S66, use the generation stage loss function in S65 Decoder Neural Network Train and iterate cycles to complete the generation training; S67, S66 The generated wireless signal obtained after the training is completed Then generate a signal for the final goal .

[0031] The test was conducted on the public wireless signal dataset Dop-NET dataset. The specific results are as follows: Figure 5 As shown, the wireless signal of person 1 tapping and person 2 waving is decoupled to regenerate the wireless signal of person 1 waving.

[0032] Therefore, the present invention adopts the above-mentioned method for generating personnel perception wireless signals based on decoupling representation. This method can effectively solve the problem that the existing technology does not fully consider the complex information characteristics in the signal, it is difficult to extract and decouple the information of different objects in the original signal separately, and it is impossible to generate a new signal different from the original signal category.

[0033] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for generating a human perception wireless signal based on decoupling characterization, characterized in that: The following steps are involved: S1. Convert the format of the input wireless signal, convert all real wireless signals into spectrograms and save them to obtain the original data set; S2, pair the converted wireless signals, and pair the wireless signals in pairs from the original data set in S1. The pairing rule is that the two signals are paired , the two signals have a common information; S3, inputting the paired wireless signal group into the encoder of the neural network to obtain a feature vector; S4, decoupling the feature vector; S5, modifying the decoupled feature vector to obtain the target feature vector; S6. Input the modified feature vector into the neural network decoder to obtain the generated wireless signal.

2. According to claim 1, a method for generating a human perception wireless signal based on decoupling characterization is characterized in that: The S3 content is as follows: S31, the signal in S2 Feed the encoder neural network in batches , each batch gets the feature vector ; S32, the feature vector in S31 According to the number of information The sub-vectors are obtained and .

3. According to the method for generating a human perception wireless signal based on decoupling characterization according to claim 2, it is characterized in that: S4 content is as follows: S41, the feature vector in S32 Input decoder neural network , and the generated wireless signal is ; S42, the wireless signal generated in S41 The signal pair in S2 At the same time, input the loss function , calculate the generated signal and its corresponding real signal The specific calculation process is as follows: ； in, is the L1 norm; For expectations; express Obey the real signal data distribution; S43, the wireless signal generated in S41 The signal pair in S2 At the same time, input the loss function , calculate the generated signal and its corresponding real signal The specific calculation process is as follows: ； in, express Obey the real signal data distribution; S44, the loss function in S42 and the loss function in S43 Add together to get the loss function , the specific calculation process is as follows: ； S45, the feature vector in S32 Swap the kth subvector and perform the swap operation Get the exchange eigenvector ; S46, the exchange feature vector in S45 Input decoder neural network respectively Get the generated wireless signal ; S47, the wireless signal generated in S46 The signal pair in S2 At the same time, input the loss function , calculate the generated signal and its corresponding real signal similarity; S48, the wireless signal generated in S46 The signal pair in S2 At the same time, input the loss function , calculate the generated signal and its corresponding real signal similarity; S49, the loss function in S47 and the loss function in S48 Add together to get the loss function , the specific calculation process is as follows: ； S410, the loss function in S44 And the loss function in S49 Add together to get the decoupling stage loss function , the specific calculation process is as follows: ； in is a hyperparameter; S411, use the decoupling phase loss function in S410 Encoder Neural Network and decoder neural network Conduct training, iterate N cycles, and complete decoupling training; S412, the encoder neural network Parameters frozen.

4. According to claim 3, a method for generating a human perception wireless signal based on decoupling characterization is characterized in that: The S45 calculation formula is as follows: 。 5. The method for generating a human perception wireless signal based on decoupling characterization according to claim 4, characterized in that: The S47 calculation process is as follows: 。 6. The method for generating a human perception wireless signal based on decoupling characterization according to claim 5, characterized in that: The S48 calculation process is as follows: 。 7. A method for generating a human perception wireless signal based on decoupling characterization according to claim 6, characterized in that: S5 content is as follows: S51, select two signals from the original data set in S1 , the two signals do not have the same information; S52, the signal in S51 is Feed the encoder neural network in batches , each batch gets the feature vector ; S53, the feature vector in S52 According to the number of information The sub-vectors are obtained and ; S54, the feature vector in S53 Swap the kth subvector and perform the swap operation Get the exchange eigenvector , the operation process is as follows: 。 8. The method for generating a human perception wireless signal based on decoupling characterization according to claim 7, characterized in that: S6 content is as follows: S61, the feature vector in S54 Input decoder neural network , and the generated wireless signal is ; S62, select two signals from the original data set in S1 , where the signal The information and the wireless signal generated in S61 Same, signal The information and the wireless signal generated in S61 same; S63, the wireless signal generated in S61 With the signal in S62 At the same time, input the loss function , calculate the generated signal and its corresponding real signal similarity; S64, the wireless signal generated in S61 With the signal in S62 At the same time, input the loss function , calculate the generated signal and its corresponding real signal similarity; S65. Change the loss function in S63 and the loss function in S64 Add together to get the loss function , the specific calculation process is as follows: ； S66, use the generation stage loss function in S65 Decoder Neural Network Train and iterate cycles to complete the generation training; S67, S66 The generated wireless signal obtained after the training is completed Then generate a signal for the final goal .

9. The method for generating a human perception wireless signal based on decoupling characterization according to claim 8, characterized in that: The S63 calculation process is as follows: 。 10. The method for generating a human perception wireless signal based on decoupling characterization according to claim 8, characterized in that: The S64 calculation process is as follows: 。

Images