WO2023159340A1 - 基于深度学习的标签识别、装置、电子设备及存储介质 - Google Patents

基于深度学习的标签识别、装置、电子设备及存储介质 Download PDF

Info

Publication number
WO2023159340A1
WO2023159340A1 PCT/CN2022/077200 CN2022077200W WO2023159340A1 WO 2023159340 A1 WO2023159340 A1 WO 2023159340A1 CN 2022077200 W CN2022077200 W CN 2022077200W WO 2023159340 A1 WO2023159340 A1 WO 2023159340A1
Authority
WO
WIPO (PCT)
Prior art keywords
signal
tag
sample
label
deep learning
Prior art date
Application number
PCT/CN2022/077200
Other languages
English (en)
French (fr)
Inventor
姚俊梅
余加宝
谢瑞桃
伍楷舜
Original Assignee
深圳大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳大学 filed Critical 深圳大学
Priority to PCT/CN2022/077200 priority Critical patent/WO2023159340A1/zh
Publication of WO2023159340A1 publication Critical patent/WO2023159340A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06K—GRAPHICAL DATA READING; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
    • G06K7/00—Methods or arrangements for sensing record carriers, e.g. for reading patterns
    • G06K7/10—Methods or arrangements for sensing record carriers, e.g. for reading patterns by electromagnetic radiation, e.g. optical sensing; by corpuscular radiation

Definitions

  • the invention belongs to the technical field of physical layer identification of wireless equipment, and in particular relates to a tag identification method, device, electronic equipment and storage medium based on deep learning.
  • Radio Frequency Identification RFID
  • RFID Radio Frequency Identification
  • the purpose of the embodiments of this specification is to provide a tag recognition method, device, electronic device and storage medium based on deep learning. In order to solve the problem that the authenticity of the identity of the label cannot be identified when the attacker forges a label with the same ID as the real label.
  • the present application provides a label recognition method based on deep learning.
  • the method includes: performing signal collection on multiple label samples to obtain signal samples; wherein the signal samples include the signal of the label sample; the signal of the label sample includes the signal of the label sample Random number RN16 signal; signal preprocessing is performed on the signal sample to obtain a data set; the data set includes the random number RN16 signal of the label sample in the signal sample, and the RN16 signal is saved in the data set in the form of an input vector of the deep learning model; using the data Set training to obtain a deep learning model; where the deep learning model is used to perform feature learning on the RN16 signal of the tag to obtain the feature information of the RN16 signal of the tag, and the deep learning model is used to identify the tag.
  • signal acquisition is performed on multiple tag samples, and obtaining signal samples includes: constructing a reader conforming to the Gen2 protocol by using USRP and software radio SDR; using the constructed reader to perform signal acquisition on multiple tag samples, Get a signal sample.
  • the communication channel between the reader and different tag samples in the plurality of tag samples is the same.
  • the signal preprocessing is performed on the signal samples to obtain the data set including: using the time domain characteristics of the signal of the reader and the time domain characteristics of the tags to extract the signals of the tag samples in the signal samples from the signal samples; Extract the RN16 signal of the label sample from the signal of the label sample through the sliding window; where the length of the sliding window is greater than the length of the RN16 signal; convert the RN16 signal of the label sample extracted from the signal sample into an input that conforms to the deep learning model Data in the form of vectors, and store the converted data in the dataset.
  • the deep learning model is a convolutional neural network (CNN) model; wherein, the input vector of the CNN model is the RN16 signal in the form of a four-dimensional vector, and the output feature of the CNN model is the feature information of the RN16 signal.
  • CNN convolutional neural network
  • using the data set to train the deep learning model includes: vectorizing the RN16 signal in the data set to obtain a sample vector; where the input dimension of a single sample vector in the sample vector is the number of I/Q channels ⁇ sample Length; use the vectorized sample vector to train the initialized CNN model and repeatedly debug to obtain a stable CNN model, and use the stable CNN model as the final CNN model for identifying tags.
  • the feature information of the RN16 signal of the tag is used to identify the tag, including: if the characteristics of the RN16 signal of the tag to be tested communicate with the reader and the RN16 signal of the tag learned by the deep learning model If the features are inconsistent, it is determined that the identity of the tag to be tested is illegal; otherwise, if they match, it is determined that the identity of the tag to be tested is true.
  • the present application provides a deep learning-based label recognition device, which includes: a signal acquisition module for collecting signals from multiple label samples to obtain signal samples; wherein the signal samples include signals of label samples; label samples
  • the signal includes the random number RN16 signal of the label sample;
  • the signal preprocessing module is used to perform signal preprocessing on the signal sample to obtain a data set;
  • the data set includes the random number RN16 signal of the label sample in the signal sample, and the RN16 signal is in depth
  • the input vector form of the learning model is stored in the data set;
  • the signal recognition module is used to train the deep learning model by using the data set; wherein, the deep learning model is used to perform feature learning on the RN16 signal of the label to obtain the characteristic information of the RN16 signal of the label,
  • a deep learning model is used to identify tags.
  • the present application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and operable on the processor.
  • the processor executes the program, the deep learning-based label recognition as in the first aspect is realized. method.
  • the present application provides a readable storage medium on which a computer program is stored, and when the program is executed by a processor, the deep learning-based label recognition method as in the first aspect is implemented.
  • this solution uses a large number of signal samples to train to obtain a deep learning model that can output the characteristic information of the RN16 signal, and identifies the label based on the learned characteristic information of the RN16 signal. Since the RN16 signal The characteristics of the tag are closely related to the hardware characteristics of the tag. The RN16 signals generated by tags with the same ID but different hardware characteristics are different and fluctuate. Subsequently, if the characteristics of the RN16 signal of a tag read by the reader do not match the characteristics of the RN16 signal learned by the deep learning model, it is determined that the tag currently communicating with the reader may be forged and unreal, and vice versa. is true.
  • the method shown in the present invention is compatible with current label industry standards and can be smoothly deployed on existing equipment.
  • deep learning is based on a multi-layer nonlinear neural network, combined with a large amount of training data, automatically extracts features and abstracts them layer by layer to directly obtain features from the data, reducing the workload of designing feature extractors for each problem, and based on deep learning.
  • the classification method shows stronger performance for the classification of UHF labels.
  • Fig. 1 is the schematic diagram of the RFID system provided by the application.
  • Fig. 2 is the sequence diagram of the communication between the reader and the tag provided by the present application.
  • Fig. 3 is a schematic flow chart of the label recognition method based on deep learning provided by the present application.
  • FIG. 4 is a schematic diagram of the internal structure of the reader provided by the present application.
  • Figure 5- Figure 6 is a schematic diagram of the time domain diagram and frequency domain diagram of the signal of the reader provided by the present application.
  • Figures 7-8 are schematic diagrams of the time-domain diagram and frequency-domain diagram of the signal of the tag provided by the present application.
  • FIG. 9 is a schematic diagram of identifying RN16 signals using a sliding window provided by the present application.
  • Fig. 10 is a schematic diagram of the CNN model provided by the present application.
  • FIG. 11 is a schematic diagram of a deep learning-based label recognition device provided by the present application.
  • FIG. 12 is a schematic structural diagram of an electronic device provided by the present application.
  • RFID equipment has become an important part of many systems, such as electronic passports, supply chain logistics RFID applications, and airport luggage management.
  • the authenticity and privacy of tags are very important.
  • people have done a lot of research to improve the security of tags.
  • the relevant research mainly adopts two technical solutions: 1. Design a secure identification and authentication protocol based on an encryption algorithm. 2. Based on the identification method of the physical layer, the fingerprint related to the device is obtained by analyzing the communication signal.
  • a more secure authentication protocol based on encryption algorithm is provided.
  • a lightweight algorithm based on block cipher principles has been proposed, or] the proposed scheme is that when the tag is successfully read each time, the tag saves a new hash value, and the modified hash value is determined by hash of random text; or prevent forgery by sharing a key between the tag and the back-end database, and changing the tag's response with each query to prevent tracking. Or by using a public key to ensure security during tag communication.
  • an RFID tag can be easily programmed to act as a counterfeit tag, or an emulation device can be used to be programmed to behave like other tags, and the authentication system cannot detect the counterfeit tag because it uses application layer data, not the tag
  • the identification method based on the physical layer.
  • the identification method based on the physical layer uses small differences in hardware to obtain fingerprints related to devices by analyzing communication signals. For example, since the frequency range of the RFID tag is 860MHz to 960MHz, the minimum power response of the tag can be measured at multiple frequencies through the MPRMF identification method for tag identification; or the time domain of the signal of the tag can be collected by a specially customized reader and spectral features, where time-domain features include Time Interval Error (TIE) and Average Baseband Power (Average baseband power, PB), TIE measures the distance between each active edge of the clock and the ideal position, and PB measures the label The average baseband power of the signal; or GenePrint utilizes the internal similarity between the preamble signal pulses of the tag to extract hardware features as fingerprints, including covariance-based pulse features and signal internal features based on power spectral density.
  • TIE Time Interval Error
  • PB Average Baseband Power
  • GenePrint utilizes the internal similarity between the preamble signal pulses of the tag to extract hardware features as fingerprint
  • the identification method based on the physical layer distinguishes different devices through the hardware attributes of the devices themselves, rather than authenticity identification through the ID of the tag itself.
  • Existing physical layer identification methods have some limitations. For example, MPRMF physical layer identification features are sensitive to the propagation distance of the signal, and require specific environments and specialized equipment to extract corresponding features. Another example is that the TIE feature entropy value is relatively low so that the uniqueness of the feature is limited. Secondly, the extraction of this feature requires specially customized and expensive equipment. Another example is that the algorithm complexity of pulse feature based on covariance and signal internal feature extraction based on power spectral density is not low. Secondly, this feature has low robustness to the environment and the recognition distance is short.
  • the present invention proposes a tag recognition method based on deep learning, which can also be called the physical layer recognition method of UHF tags based on deep learning, and uses a large number of signal samples to train to obtain features that can output RN16 signals
  • the deep learning model of information identifies tags based on the learned feature information of the RN16 signal. If the features of the RN16 signal of a tag read by the reader do not match the features of the RN16 signal learned by the deep learning model, then It is determined that the tag currently communicating with the reader may be forged and unreal, and vice versa, it is real. This method is compatible with current labeling industry standards and can be deployed smoothly on existing equipment.
  • deep learning is based on a multi-layer nonlinear neural network, combined with a large amount of training data, automatically extracts features and abstracts them layer by layer to directly obtain features from the data, reducing the workload of designing feature extractors for each problem, and based on deep learning.
  • the classification method shows stronger performance for the classification of UHF labels.
  • the RFID system may include a tag (Tag) 101 and a universal software radio peripheral (Universal Software Radio Peripheral, USRP) 102. Further, a server (Server) may also be included.
  • the USRP102 performs non-contact two-way data communication with the tag 101 through RFID technology, and reads and writes the tag.
  • the tag 101 may be called an RFID tag, and the tag 101 may be understood as a passive or semi-active IoT tag.
  • the tag 101 can be attached to the object to identify the target object.
  • the tag stores the information of the object, and each tag has a globally unique electronic product code (Electronic Product Code, EPC).
  • EPC Electronic Product Code
  • the tags are divided into passive tags and semi-active tags. There are two types of tags, which work differently.
  • the working mode of the passive tag is: when the tag enters the effective identification range of the USRP102, the tag receives the radio frequency signal sent by the USRP102, obtains energy by virtue of the induced current and sends out the information stored in the chip (this tag is a passive tag).
  • the working method of the semi-active tag is: the tag obtains energy through solar energy and other means, and actively sends the stored information through the obtained energy. USRP102 receives and decodes the information and sends it to the server for data processing.
  • the label of the present invention can be regarded as a terminal, and the terminal management method of the present invention can be understood as a label management method or a label management method, etc., without limitation.
  • USRP102 can be included in a reader or a reader (Reader), and USRP102 can have a transmitting (Transmitter) antenna (abbreviated as TX) and a receiving (Receiver) antenna (abbreviated as RX) for realizing wireless radio frequency communication, and USRP102 can use It is used to read the information stored in the specified tag through the radio frequency signal. If it is an inventory operation, the reader can take an inventory of the tag information within its management range. If it is a read operation, the reader reads the data in the storage area of the tag. Optionally, in some occasions where the information stored in the tag needs to be rewritten, the reader can also have a writing function. If it is a writing operation, the reader will write the data into the storage area of the tag.
  • TX transmitting
  • RX receiving
  • USRP102 can use It is used to read the information stored in the specified tag through the radio frequency signal. If it is an inventory operation, the reader can take an inventory of the tag information within its management range. If it is a
  • the reader can also perform an invalidation operation on the label.
  • USRP102 may also be called a reader.
  • the USRP102 can be deployed independently, or deployed/integrated in other devices, such as a terminal, and the USRP102 is regarded as a terminal at this time.
  • the main application scenarios of RFID systems include warehouse management, inventory, logistics, etc.
  • the goods can be embedded or affixed with labels.
  • the labels correspond to the relevant information of the goods.
  • USRP102 can use the above method to communicate with the labels, obtain the relevant information of the goods from the labels, and report them to the central information system , so that the management personnel can quickly query the relevant information of the goods in the central information system, improve the speed and accuracy of the delivery of goods, and if there are abnormal situations such as loss of goods, the management personnel can also be informed and deal with them in the first time.
  • some places with huge assets or valuable items such as libraries, art galleries, and museums can embed or label items.
  • USRP102 can use the above method to communicate with tags and obtain Item related information, and reported to the central information system, so that in the case of abnormal changes in the storage information of books or valuables, the administrator will be reminded to deal with it as soon as possible.
  • the operation instructions issued by the USRP 102 to the tags may include one or more of an inventory operation, a read operation, a write operation, and a failure operation. It should be understood that the present invention does not limit the type of operation and the name of the operation. In addition to the inventory operation, read operation, write operation and invalidation operation, other operations can also be included. In addition, these operations can also be named by other names , without restriction.
  • the inventory operation can be used to indicate the inventory of the tags.
  • the inventory is the operation of obtaining the tag ID.
  • the inventory can also be understood as an inventory, and the two have equivalent concepts.
  • the read operation can be used to read the information stored in the tag.
  • the write operation can be used to indicate to write information into the tag's internal storage.
  • the invalidate operation can be used to indicate invalidated tags, and invalidated tags cannot be inventoried and read/written.
  • the process includes: (1) The reader (such as USRP102 in FIG. 1 ) sends a Select command to the tag (such as tag 101 in FIG. 1 ). For example, when the reader receives the inventory command sent by the server (or the inventory command can be sent by the server to the reader through the middleware), a selection command is generated, which includes the range of the label (such as some specific ranges of EPC codes) ). After the tag receives the selection command, it judges whether it belongs to the range of the tag indicated by the selection command, and if the tag belongs to the range of the tag indicated, it returns information after receiving the query command subsequently.
  • the reader sends a Query command to the tags within the specified range.
  • the tag finds itself belonging to the range of tags in the selection command, return a random number to the reader.
  • the tag may return a random number to the reader through competition, for example, the random number may be a 16-bit random number (16-bit Random Number, RN16).
  • the reader After the reader receives the random number sent from the tag, it sends an acknowledgment (Acknowledgment, ACK) command to the tag, and the ACK command includes the received random number (such as RN16).
  • ACK acknowledgment
  • the tag receives the response command sent by the reader and verifies that the random number is correct, it returns its EPC code to the reader.
  • the information/signaling message transmitted between the reader and the tag can be carried in the signal, for example, the above query command, random number RN16, ACK command, EPC code, etc. can be carried in the signal for transmission, which carries
  • the signal of the inquiry command may be called the inquiry signal
  • the signal carrying the random number RN16 may be called the RN16 signal
  • the signal carrying the ACK command may be called the ACK signal
  • the signal carrying the EPC code may be called the EPC code signal, etc.
  • the amplitude and frequency of different signals are different.
  • RN16 is a signal with a fixed pattern.
  • the random number RN16 generated by tag 1 is 1100 1011 0111 1010.
  • the hardware characteristics of the tag will cause some small fluctuations in the RN16 signal.
  • the present invention collects RN16 signals as raw data, performs feature analysis on RN16 signals generated by tags with different hardware characteristics through a deep learning model, and performs tag identification based on RN16 signals generated by tags.
  • this application provides a tag recognition method based on deep learning. This method is suitable for tag recognition and solves the problem of incorrect identification of the tag's real identity caused by malicious/criminal counterfeit tag IDs. As shown in Figure 3, the method includes:
  • S310 Collect signals from multiple tag samples to obtain signal samples.
  • the plurality of label samples may include labels of different design models, for example, labels of different models may be obtained from two or more manufacturers (it is understandable that these manufacturers are all legitimate manufacturers) .
  • all acquired tags may be used as tag samples, and signal samples are obtained by performing signal collection on all tags.
  • a large number of acquired tags may be divided into multiple groups (for example, three groups), and each group of tags is used as a tag sample, and signals of the group of tags are collected to obtain signal samples.
  • Each set of labels includes labels of different design models. It should be noted that when signal collection is performed on tag samples, the communication channel between the reader and the tag samples is fixed. In order to verify the universal applicability of the method shown in the present invention, a large number of random numbers corresponding to each tag sample are collected. RN16 signals, such as collecting 5000 RN16 signals for each tag, etc.
  • the signal sample includes the signal generated during the communication between the reader and the tag.
  • the signal sample can include the signal of the reader, the signal of the tag, such as the RN16 signal generated by the tag, etc.
  • the signal sample can also include other signals. , such as a DC carrier signal.
  • a reader conforming to the Gen2 protocol can be constructed by using USRP and software defined radio (SDR), wherein the SDR uses the GNU Radio software radio development kit. Signal collection is performed on multiple tag samples through the constructed reader to obtain signal samples. In this way, through the GU-RFID platform designed by the hardware device USRP and the software GNU Radio, the signal of the UHF RFID tag (or UHF tag for short) is collected, and the reader test and signal modulation are carried out using the GU-RFID platform .
  • SDR software defined radio
  • the reader of the GU-RFID platform is composed of a computer (such as a laptop), a USRP device (or called a USRP source) and two circularly polarized antennas, and the computer and the USRP device pass through Network connection or Universal Serial Bus (Universal Serial Bus, USB) interface connection.
  • These two circularly polarized antennas are connected to the USRP device.
  • One antenna is responsible for sending reader commands (such as query commands, response commands, etc.) to tags through RFID technology, and the other is responsible for receiving backscattered signals from tags (such as RN16 signals, EPC code, etc.).
  • the backscatter signal of the tag in the present invention may be called a tag response, and the signal sent by the reader may be called a reader command.
  • the reader includes a USRP source, a matched filter (Matched Filter), a gate module (or called a tag response gate (Tag Response Gate) module), tag decoder (Tag Decoder), Gen2 logic module (or called reader logic (Reader Logic) module) and USRP receiver module, where the URSP receiver module includes a transmitting antenna (or TX for short) and a receiving antenna (or RX for short). Further, as shown in FIG. 4 , the reader may further include an amplifier (Amplifier) module.
  • An amplifier Analogifier
  • the reader may include a receiving chain and a sending chain, wherein the receiving chain is used to receive backscattered signals of tags (such as RN16 signals, EPC codes, etc.).
  • the sending chain is used to send reader commands (such as query commands, response commands, etc.) to tags.
  • FIG. 4 is only an exemplary drawing, and the reader in this application includes but is not limited to the components shown in FIG. 4 , and also includes other components.
  • the reader is used to collect the signal of the tag sample, and the signal sample obtained may include:
  • the logic module of the reader in the sending chain generates a query command, which is used to query the signal (such as RN16 signal) generated by the tag. Then, the query command is amplified through the configurable amplifier module, and the amplified query command is sent through TX, for example, to the tag through TX.
  • the query command in this application may be called an on-off keying command, and the query command may be a command generated by an on-off keying method (or an amplitude keying (Amplitude Shift Keying, ASK) adjustment method and a switch monitoring method).
  • the tag After the tag receives the query command (or switch key control command) sent by the reader, it returns the backscatter signal (such as RN16 signal) of the tag, and the RX of the reader receives the signal returned by the tag, and Store the received signal into the Linux kernel. Further, the receiving chain of the reader extracts signal samples of the received signal from the Linux kernel. The sample first passes through the matched filter module, which is responsible for filtering the received signal with a square pulse of half the symbol period, so as to maximize the signal-to-noise ratio (Signal Noise Ratio, SNR) of the signal transmitted by the tag.
  • SNR Signal-to-noise ratio
  • the gate structure is responsible for identifying the query command of the reader, and identifies the query command of the reader by tracking the amplitude of the received signal, so as to only process the returned signal samples of the tags queried by the reader, which Computational resources are also reduced, as the tag decoding process is only performed when a valid tag response is expected, reducing overall system latency.
  • the signal samples processed by the gate module are used as the input of the next module for the next step of processing.
  • the tag decoding module is responsible for frame synchronization, channel estimation and detection of tag responses, and channel synchronization is achieved by combining the received signal with the known preamble Code correlation is completed, and finally the decoder demodulates the signal of the tag response through amplitude shift keying (ASK) to obtain signal samples.
  • ASK amplitude shift keying
  • the USRP-based reader can communicate with the tag sample, and the signal in the communication process between the reader and the tag can be collected as a signal sample.
  • the RN16 signal of the label is included in the dataset.
  • the RN16 signal is included in the data set in the form of an input vector of the deep learning model.
  • S320 may include the following steps S3201 and S3202:
  • the original signal/signal sample received by the USRP (or USRP source) in the reader includes the DC carrier signal, the reader signal (or reader command) and the tag signal (or tag response).
  • the extracted signal has nothing to do with the data, but is related to the hardware characteristics of the tag itself, it is first necessary to extract/separate the RN16 data packet from the original signal/signal sample.
  • a sliding window can be designed, and the sliding window can be used to traverse the entire signal during the communication process between the reader and the tag. Combined with the time-frequency characteristics of each signal, the RN16 signal is extracted from it.
  • extracting the RN16 signal may include the following two steps: Step 1, by observing the time domain diagram and frequency domain diagram of the signal of the tag and the signal of the reader, as shown in Figures 5 and 6, the time domain diagram and frequency domain diagram of the signal of the reader Figures 7 and 8 are the time-domain diagrams and frequency-domain diagrams of the tag's signal. It can be found that there are obvious differences between the reader's signal and the frequency-domain diagram of the tag's signal, and then the fast Fourier transform can be used to detect the frequency domain of the window. Whether the signal energy of the signal conforms to the signal mode of the tag, if so, it is determined that the signal in the current sliding window is the signal of the tag, otherwise, it is the signal of the reader.
  • Step 2 since the RN16 signal of the tag and the ID signal of the tag (or the EPC code of the tag) have the same pattern in the frequency domain, that is, have the same frequency domain characteristics, it is necessary to further distinguish the RN16 signal of the tag and the ID signal of the tag .
  • the time length of the set sliding window is slightly longer than the length of the RN16 signal, which means that the RN16 signal will fall into the sliding window.
  • Figure 9 shows the operation of the label sliding window. It can be seen that the front end and the back end of the sliding window are DC carrier signals. If this condition is met, it is a candidate window. The signal located in the candidate window is the RN16 signal, which can be extracted from the candidate window. The entire RN16 signal is output; on the contrary, those that do not meet this condition are non-candidate windows, and the signal located in the non-candidate window is the ID signal of the tag (or called the EPC code of the tag).
  • N example represents the number of vectors.
  • the deep learning model is used to perform feature learning on the RN16 signal of the label to obtain feature information (such as time domain characteristics and/or frequency domain) of the RN16 signal of the label characteristics, etc.), the characteristic information of the RN16 signal of the tag is used to identify the tag.
  • the deep learning model can be a convolutional neural network (Convolutional Neural Network, CNN) model, and the CNN model includes a certain number of convolutional layers and fully connected layers.
  • This CNN model features the raw IQ samples of each radio signal, which have been normalized.
  • the present invention does not perform any expert feature extraction or other preprocessing on the original radio signal, but directly uses the network to learn the characteristics of the original time series samples on the high-dimensional data.
  • FIG. 10 it is a schematic diagram of the CNN model, as shown in Figure 10, the CNN model uses three convolutional layers and four fully connected layers, except that the last output classification layer uses the Softmax (normalized exponential function) function , other layers use the Rectified Linear Unit (ReLU) function as the activation function.
  • the three convolutional layers contain 128, 64, and 32 filters respectively, and the four fully connected layers contain 512, 256, 128, and 50 neurons respectively.
  • the regularization method is adopted, and the dropout technology is applied in the fully connected layer, which is a very effective method to improve the generalization ability and reduce over-fitting.
  • the input dimension of a single sample vector is the number of I/Q channels ⁇ sample length, wherein the number of I/Q channels is 2, and the sample length is 1200 samples.
  • the cross-entropy loss function and Adaptive Momentum Estimation (Adaptive Momentum Estimation, Adam) optimizer are used for optimization training, and finally the CNN model shown in Figure 10 is obtained.
  • the process of using the data set to train the model and repeatedly debug it to finally obtain the convolutional neural network capable of outputting the characteristic information of the RN16 signal can refer to the existing relevant model training and debugging process, and will not be described in detail.
  • the feature information of the RN16 signal of the tag used to identify the tag may include: if the feature of the RN16 signal of a certain tag to be tested communicates with the reader and the RN16 signal of the tag learned by the deep learning model Inconsistent/inconsistent with the characteristics of the tag to be tested, it is determined that the identity of the tag to be tested is false and illegal. It may be a tag that is counterfeited or forged by the attacker with the same ID as the real tag. The identity of the tag under test is real.
  • a large number of signal samples are used to train a deep learning model that can output the characteristic information of the RN16 signal, and the label is identified based on the learned characteristic information of the RN16 signal. If the characteristics of the RN16 signal do not match the characteristics of the RN16 signal learned by the deep learning model, it is determined that the tag currently communicating with the reader may be forged and unreal, otherwise, it is real.
  • the method shown in the present invention is compatible with current label industry standards and can be smoothly deployed on existing equipment.
  • deep learning is based on a multi-layer nonlinear neural network, combined with a large amount of training data, automatically extracts features and abstracts them layer by layer to directly obtain features from the data, reducing the workload of designing feature extractors for each problem, and based on deep learning.
  • the classification method shows stronger performance for the classification of UHF labels.
  • the deep learning-based tag recognition device 110 may include: a signal collection module 1110 , a signal preprocessing module 1120 and a signal recognition module 1130 .
  • the signal collection module 1110 is used to collect signals from multiple label samples to obtain signal samples; wherein the signal samples include the signals of the label samples; the signal of the label samples includes the random number RN16 signal of the label samples.
  • the signal preprocessing module 1120 is used to perform signal preprocessing on the signal sample to obtain a data set; wherein the data set includes the random number RN16 signal of the label sample in the signal sample, and the RN16 signal is stored in the data set in the form of an input vector of the deep learning model .
  • the signal identification module 1130 is used to obtain a deep learning model by using data set training; wherein, the deep learning model is used to perform feature learning on the RN16 signal of the label to obtain the feature information of the RN16 signal of the label, and the deep learning model is used to identify the label .
  • the function of the signal acquisition module 1110 is to use the Gnuradio software radio to emulate the USRP N210 device as a reader conforming to the Gen2 protocol standard to collect the signal of the tag.
  • the signal collection module 1110 constructs a reader conforming to the Gen2 protocol by using the USRP and the software defined radio SDR; and uses the constructed reader to collect signals from multiple tag samples to obtain signal samples.
  • the communication channel between the reader and different tag samples in the plurality of tag samples is the same.
  • the signal preprocessing module 1120 locates the RN16 signal corresponding to the tag and extracts the corresponding RN16 signal.
  • the collected signal is stored in the form of a 32-bit floating point number and needs to be converted into a format that can be learned by deep learning.
  • the signal preprocessing module 1120 uses the time-domain characteristics of the signal of the reader and the time-domain characteristics of the tag to extract the signal of the tag sample in the signal sample from the signal sample; extract the tag sample from the signal of the tag sample through a sliding window RN16 signal; wherein, the length of the sliding window is greater than the length of the RN16 signal; the RN16 signal of the label sample extracted from the signal sample is converted into data in the form of an input vector conforming to the deep learning model, and the converted data is stored in the data set.
  • the deep learning model is a convolutional neural network (CNN) model
  • the input vector of the CNN model is the RN16 signal in the form of a four-dimensional vector
  • the output feature of the CNN model is the feature information of the RN16 signal.
  • CNN convolutional neural network
  • the signal identification module 1130 performs deep learning model training and model performance improvement according to the collected signals, and identifies and classifies labels by this.
  • the signal identification module 1130 vectorizes the RN16 signal in the data set to obtain a sample vector; the input dimension of a single sample vector in the sample vector is the number of I/Q channels ⁇ sample length; the vectorized sample vector is used to initialize
  • the CNN model is trained and repeatedly debugged to obtain a stable CNN model, and the stable CNN model is used as the final CNN model for identifying tags.
  • the feature information of the RN16 signal of the tag is used to identify the tag includes: if the feature of the RN16 signal of the tag to be tested communicated with the reader is inconsistent with the feature of the RN16 signal of the tag learned by the deep learning model , it is determined that the identity of the tag to be tested is illegal, otherwise, if it matches, it is determined that the identity of the tag to be tested is real.
  • a deep learning-based label recognition device provided by the present invention can execute the embodiments of the above methods, and its implementation principles and technical effects are similar, and will not be repeated here.
  • FIG. 12 is a schematic structural diagram of an electronic device provided by an embodiment of the present invention. As shown in FIG. 12 , a schematic structural diagram of an electronic device 1200 suitable for implementing the present invention is shown.
  • an electronic device 1200 includes a central processing unit (CPU) 1201, which can operate according to a program stored in a read-only memory (ROM) 1202 or a program loaded from a storage section 1208 into a random access memory (RAM) 1203 Instead, various appropriate actions and processes are performed.
  • ROM read-only memory
  • RAM random access memory
  • various programs and data necessary for the operation of the device 1200 are also stored.
  • the CPU 1201, ROM 1202, and RAM 1203 are connected to each other through a bus 1204.
  • An input/output (I/O) interface 1205 is also connected to the bus 1204 .
  • the following components are connected to the I/O interface 1205: an input section 1206 including a keyboard, a mouse, etc.; an output section 1207 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker; a storage section 1208 including a hard disk, etc. and a communication section 1209 including a network interface card such as a LAN card, a modem, or the like.
  • the communication section 1209 performs communication processing via a network such as the Internet.
  • Drive 1210 is also connected to I/O interface 1206 as needed.
  • a removable medium 1211 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is mounted on the drive 1210 as necessary so that a computer program read therefrom is installed into the storage section 1208 as necessary.
  • the process described above with reference to FIG. 3 may be implemented as a computer software program.
  • embodiments of the present disclosure include a computer program product including a computer program tangibly contained on a machine-readable medium, the computer program including program code for executing the above deep learning-based tag recognition method.
  • the computer program may be downloaded and installed from a network via communication portion 1209 and/or installed from removable media 1211 .
  • each block in a flowchart or block diagram may represent a module, program segment, or part of code that includes one or more logical functions for implementing specified executable instructions.
  • the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or they may sometimes be executed in the reverse order, depending upon the functionality involved.
  • each block of the block diagrams and/or flowchart illustrations, and combinations of blocks in the block diagrams and/or flowchart illustrations can be implemented by a dedicated hardware-based system that performs the specified functions or operations , or may be implemented by a combination of dedicated hardware and computer instructions.
  • the units or modules involved in the embodiments described in the present application may be implemented by means of software or by means of hardware.
  • the described units or modules may also be provided in a processor.
  • the names of these units or modules do not constitute limitations on the units or modules themselves in some cases.
  • a typical implementing device is a computer.
  • the computer can be, for example, a personal computer, a notebook computer, a mobile phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or any of these devices combination of devices.
  • the present invention also provides a storage medium, which may be the storage medium included in the aforementioned device in the above embodiment; or may be a storage medium that exists independently and is not assembled into the device.
  • the storage medium stores one or more programs, and the aforementioned programs are used by one or more processors to execute the tag recognition method based on the deep learning model described in the present invention.
  • Storage media includes permanent and non-permanent, removable and non-removable media.
  • Information storage can be realized by any method or technology.
  • Information may be computer readable instructions, data structures, modules of a program, or other data.
  • Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Flash memory or other memory technology, Compact Disc Read-Only Memory (CD-ROM), Digital Versatile Disc (DVD) or other optical storage, Magnetic tape cartridge, tape magnetic disk storage or other magnetic storage device or any other non-transmission medium that can be used to store information that can be accessed by a computing device.
  • computer-readable media excludes transitory computer-readable media, such as modulated data signals and carrier waves.

Landscapes

  • Physics & Mathematics (AREA)
  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Electromagnetism (AREA)
  • General Health & Medical Sciences (AREA)
  • Toxicology (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Near-Field Transmission Systems (AREA)

Abstract

一种基于深度学习的标签识别方法，涉及无线设备的物理层识别技术领域，解决无法准确识别标签的真实身份的问题。该方法包括：对多个标签样本进行信号采集，获得信号样本；其中信号样本包括标签样本的信号；标签样本的信号包括标签样本的随机数RN16信号；对信号样本进行信号预处理，得到数据集；其中数据集中包括信号样本中标签样本的随机数RN16信号，且RN16信号以深度学习模型的输入向量形式保存在数据集中；利用数据集训练得到深度学习模型；其中，深度学习模型用于对标签的RN16信号进行特征学习得到标签的RN16信号的特征信息，深度学习模型用于对标签进行身份识别。

Description

基于深度学习的标签识别、装置、电子设备及存储介质 技术领域
本发明属于无线设备的物理层识别技术领域，特别涉及一种基于深度学习的标签识别方法、装置、电子设备及存储介质。
背景技术
随着物联网的普及，无线射频识别(Radio Frequency Identification，RFID)系统成为促进各种应用自动化管理的平台，其最基本的功能是标签识别。
然而攻击者很容易假冒或者伪造一个与真实标签身份(Identification，ID)相同的标签，对于假冒或者伪造的标签，无法识别标签身份的真实性。
发明内容
本说明书实施例的目的是提供一种基于深度学习的标签识别方法、装置、电子设备及存储介质。以解决现有在攻击者伪造一个与真实标签的ID相同的标签而导致无法识别标签身份的真实性的问题。
为解决上述技术问题，本申请实施例通过以下方式实现的：
第一方面，本申请提供一种基于深度学习的标签识别方法，方法包括：对多个标签样本进行信号采集，获得信号样本；其中信号样本包括标签样本的信号；标签样本的信号包括标签样本的随机数RN16信号；对信号样本进行信号预处理，得到数据集；其中数据集中包括信号样本中标签样本的随机数RN16信号，且RN16信号以深度学习模型的输入向量形式保存在数据集中；利用数据集训练得到深度学习模型；其中，深度学习模型用于对标签的RN16信号进行特征学习得到标签的RN16信号的特征信息，深度学习模型用于对标签进行身份识别。
一种可能的设计中，对多个标签样本进行信号采集，获得信号样本包括：利用USRP和软件无线电SDR构建了符合Gen2协议的阅读器；利用构建的阅读器对多个标签样本进行信号采集，获得信号样本。
一种可能的设计中，阅读器与多个标签样本中不同标签样本之间的通信信道相同。
一种可能的设计中，对信号样本进行信号预处理，得到数据集包括：利用阅读器的信号的时域特性以及标签的时域特性，从信号样本中提取出信号样本中标 签样本的信号；通过滑动窗口的从标签样本的信号中提取出标签样本的RN16信号；其中，滑动窗口的长度大于RN16信号的长度；将信号样本中提取出的标签样本的RN16信号转换成符合深度学习模型的输入向量形式的数据，将转换后的数据存在数据集中。
一种可能的设计中，深度学习模型为卷积神经网络CNN模型；其中，CNN模型的输入向量为四维向量形式的RN16信号，CNN模型的输出特征为RN16信号的特征信息。
一种可能的设计中，利用数据集训练得到深度学习模型包括：将数据集中的RN16信号进行向量化处理，得到样本向量；其中样本向量中的单个样本向量输入维度是I/Q信道数×样本长度；利用向量化处理后的样本向量对初始化CNN模型进行训练以及反复调试得到稳定的CNN模型，将稳定的CNN模型作为最终用于对标签进行身份识别的CNN模型。
一种可能的设计中，标签的RN16信号的特征信息用于对标签进行身份识别包括：若与阅读器通信的待测标签的RN16信号的特征与经深度学习模型学习得到的该标签的RN16信号的特征不一致，则确定待测标签的身份是不合法的，反之，若符合，则确定该待测标签的身份是真实的。
第二方面，本申请提供一种基于深度学习的标签识别装置，装置包括：信号采集模块，用于对多个标签样本进行信号采集，获得信号样本；其中信号样本包括标签样本的信号；标签样本的信号包括标签样本的随机数RN16信号；信号预处理模块，用于对信号样本进行信号预处理，得到数据集；其中数据集中包括信号样本中标签样本的随机数RN16信号，且RN16信号以深度学习模型的输入向量形式保存在数据集中；信号识别模块，用于利用数据集训练得到深度学习模型；其中，深度学习模型用于对标签的RN16信号进行特征学习得到标签的RN16信号的特征信息，深度学习模型用于对标签进行身份识别。
第三方面，本申请提供一种电子设备，包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，处理器执行程序时实现如第一方面的基于深度学习的标签识别方法。
第四方面，本申请提供一种可读存储介质，其上存储有计算机程序，该程序被处理器执行时实现如第一方面的基于深度学习的标签识别方法。
由以上本说明书实施例提供的技术方案可见，该方案利用大量的信号样本训练得到能够输出RN16信号的特征信息的深度学习模型，基于学习到的RN16信号的特征信息对标签进行识别，由于RN16信号的特征与标签的硬件特性有紧密关系，ID相同但硬件特性不同的标签所产生的RN16信号是不同的，是有波动的。后续如果阅读器读取的某个标签的RN16信号的特征与经深度学习模型学习得到的RN16信号的特征不符，则确定当前与阅读器通信的标签可能是被伪造的，是不真实，反之，则是真实的。此外，本发明所示方法兼容当前的标签工业标准，可以顺利部署到现有的设备上。此外，深度学习基于多层非线性神经网络，结合大量训练数据，自动抽取特征并逐层抽象直接从数据中获取特征，减少了为每个问题设计特征提取器的工作量，并且基于深度学习的分类方法对超高频标签的分类表现了更强的性能。
附图说明
为了更清楚地说明本说明书实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本说明书中记载的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1为本申请提供的RFID系统示意图；
图2为本申请提供的阅读器与标签之间的通信时序图；
图3为本申请提供的基于深度学习的标签识别方法的流程示意图；
图4为本申请提供的阅读器内部结构示意图；
图5-图6为本申请提供的阅读器的信号的时域图和频域图的示意图；
图7-图8为本申请提供的标签的信号的时域图和频域图的示意图；
图9为本申请提供的利用滑动窗口识别RN16信号的示意图；
图10为本申请提供的CNN模型示意图；
图11为本申请提供的基于深度学习的标签识别装置示意图；
图12为本申请提供的电子设备的结构示意图。
具体实施方式
为了使本技术领域的人员更好地理解本说明书中的技术方案，下面将结合本 说明书实施例中的附图，对本说明书实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本说明书一部分实施例，而不是全部的实施例。基于本说明书中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都应当属于本说明书保护的范围。
以下描述中，为了说明而不是为了限定，提出了诸如特定系统结构、技术之类的具体细节，以便透彻理解本申请实施例。然而，本领域的技术人员应当清楚，在没有这些具体细节的其它实施例中也可以实现本申请。在其它情况中，省略对众所周知的系统、装置、电路以及方法的详细说明，以免不必要的细节妨碍本申请的描述。
在不背离本申请的范围或精神的情况下，可对本申请说明书的具体实施方式做多种改进和变化，这对本领域技术人员而言是显而易见的。由本申请的说明书得到的其他实施方式对技术人员而言是显而易见得的。本申请说明书和实施例仅是示例性的。
关于本文中所使用的“包含”、“包括”、“具有”、“含有”等等，均为开放性的用语，即意指包含但不限于。本申请中的“份”如无特别说明，均按质量份计。
相关技术中，RFID设备成为许多系统的重要组成部分，如电子护照、供应链物流RFID的应用及机场行李管理等，标签的真实性和私密性非常重要。近年来人们已经做了许多研究来提升标签的安全性，相关的研究主要采用两种技术方案：一、基于加密算法，设计安全的识别和认证协议。二、基于物理层的识别方法，通过分析通信信号获得与设备相关的指纹。下面对这两种相关技术进行介绍
一、基于加密算法，设计安全的识别和认证协议。
主要通过修改现有的超高频RFID系统协议规范，提供基于加密算法更安全的认证协议。比如近年来提出的利用一种基于分组密码准则的轻量级算法，或者]提出的方案是当标签每次被成功读取时，标签保存新的哈希值，修改后的哈希值是由随机文本的哈希组成的；或者通过标签与后端数据库之间共享密钥的方式来防止伪造，并且利用每次查询来改变标签的响应来防止跟踪。或者通过使用一个公共的密钥来保证标签通信过程中安全性。
但是，基于加密算法，设计安全的识别和认证协议存在弊端：对于传统的设计基于加密的识别和认证协议，由于标签的低成本以及尺寸小，标签的计算能力 和资源十分有限，这一限制使得传统的加密和安全协议的实现效率低下。其次即使实现了传统的或者轻量级的基于加密的协议，不恰当的实现和非标准化的算法也会导致攻击者拥有足够的资源破坏协议来找到底层数据。比如一个RFID标签可以被轻易编程为一个伪造标签，或者一个仿真设备可以用来被编程像其他标签一样工作，而身份验证系统无法检测到伪造标签，因为其使用的是应用层数据，而不是标签的一个或者多个独特的物理属性。同时由于改变了标签的认证协议，与现有的协议不兼容，很难得到大规模应用以及部署。
二、基于物理层的识别方法。
基于物理层的识别方法是利用硬件的微小差异，通过分析通信信号获得与设备相关的指纹，其目的是通过设备的硬件特征来区分不同的设备，而不是通过设备的id来进行真实性识别。比如，由于RFID标签的频率范围在860MHz到960MHz，通过MPRMF识别方法实现在多个频率下测量标签的最小功率响应，用于进行标签识别；或者利用专门定制的阅读器收集标签的信号的时域和谱特征，其中时域特征包括时间间隔误差(Time Interval Error，TIE)和平均基带功率(Average baseband power，PB)，TIE测量时钟的每个活动边缘与理想位置之间的距离，PB测量标签的信号的平均基带功率；或者GenePrint利用了标签的前导码信号脉冲之间的内部相似性来提取硬件特征作为指纹，包括基于协方差的脉冲特征和基于功率谱密度的信号内部特征。
但是，基于物理层的识别方法是通过设备本身的硬件属性来区分不同的设备，而不是通过标签本身的ID来进行真实性识别。现存的物理层识别方法存在一些限制，比如MPRMF物理层识别特征对信号的传播距离敏感，并且需要在特定环境和专门的设备来提取相应的特征。又比如TIE特征熵值比较低以至于限制了特征的唯一性，其次该特征的提取需要专门定制的昂贵的设备。再比如基于协方差的脉冲特征和基于功率谱密度的信号内部特征提取的算法复杂度不低，其次该特征对环境的鲁棒性低，识别距离较短。
基于此，本发明提出一种基于深度学习的标签识别方法，该识别方法还可以称为基于深度学习的超高频标签的物理层识别方法，利用大量的信号样本训练得到能够输出RN16信号的特征信息的深度学习模型，基于学习到的RN16信号的特征信息对标签进行识别，后续如果阅读器读取的某个标签的RN16信号的特征 与经深度学习模型学习得到的RN16信号的特征不符，则确定当前与阅读器通信的标签可能是被伪造的，是不真实，反之，则是真实的。此方法兼容当前的标签工业标准，可以顺利部署到现有的设备上。此外，深度学习基于多层非线性神经网络，结合大量训练数据，自动抽取特征并逐层抽象直接从数据中获取特征，减少了为每个问题设计特征提取器的工作量，并且基于深度学习的分类方法对超高频标签的分类表现了更强的性能。
下面结合附图和实施例对本发明进一步详细说明。
参照图1，其示出了适用于本申请实施例提供的方法的RFID系统示意图。如图1所示，RFID系统可以包括标签(Tag)101、通用软件无线电外设(Universal Software Radio Peripheral，USRP)102。进一步的，还可以包括服务器(Server)。USRP102通过RFID技术与标签101进行非接触双向数据通信、对标签进行读写。
其中，标签101可以称为RFID标签，标签101可以理解为无源或半有源物联网标签。标签101可以附着在物体上标识目标对象，标签中存储该物体的信息，且每个标签都有一个全球唯一的电子产品代码(Electronic Product Code，EPC)，标签分为无源标签和半有源标签两类，二者的工作方式不同。无源标签的工作方式为：当标签进入USRP102有效识别范围内时，标签接收USRP102发出的射频信号，凭借感应电流获得能量发出存储在芯片中的信息(此标签为无源标签)。半有源标签的工作方式为：由标签通过太阳能等方式获取能量，通过获取到的能量主动发送存储的信息。USRP102接收信息并解码后送至服务器进行数据处理。应理解，本发明的标签可以看做是一个终端，本发明的终端管理方法可以理解标签管理方法或对标签的管理方法等，不予限制。
USRP102可以包括在读写器或者阅读器(Reader)中，USRP102可以具备用于实现无线射频通信的发射(Transmitter)天线(简称为TX)以及接收(Receiver)天线(简称为RX)，USRP102可以用于通过射频信号来阅读指定的标签中存储的信息。若为盘点操作，阅读器可以盘点其管理范围内的标签的信息。若为读操作，阅读器则读取该标签存储区中的数据。可选的，在一些需要改写标签内存储的信息的场合下，阅读器还可具有写的功能，若为写操作，则阅读器将数据写入标签的存储区中。可选的，在一些需要失效标签的场合下，阅读器还可以对标签执行失效操作。本发明中，USRP102也可以称为阅读器。USRP102可以独立部 署，或者部署在/集成在其他设备中，如终端中，此时该USRP102看做一个终端。
RFID系统的主要应用场景包括仓库管理、盘点、物流等。例如，以物流运输为例，可以给货品嵌入或贴上标签，该标签对应货品的相关信息，USRP102可以采用上述方式与标签进行数据通信，从标签获取货品的相关信息，并上报给中央信息系统，以使得管理人员可以在中央信息系统中迅速查询货品的相关信息，提高货品交接速度和准确率，如果有货品丢失等异常情况，管理人员也可以第一时间获知并处理。又例如，以固定资产管理为例，图书馆、艺术馆及博物馆等资产庞大或者物品贵重的一些场所，可以给物品嵌入或贴上标签，USRP102可以采用上述方式与标签进行数据通信，从标签获取物品的相关信息，并上报给中央信息系统，以便于在书籍或者贵重物品的存放信息有异常变动的情况下，第一时间提醒管理员进行处理。
本发明实施例中，USRP102对标签下发的操作指令可以包括盘点操作、读操作、写操作以及失效操作中的一项或者多项。应理解，本发明不限制操作操作的类型、操作操作的命名，除盘点操作、读操作、写操作以及失效操作之外，还可以包括其他操作操作，此外，这些操作操作还可以命名为其他名称，不予限制。
盘点操作可以用于指示对标签进行盘点，盘点是获取标签标识的操作，盘点还可以理解为盘存，二者具有等效概念。读操作可以用于读取标签内存储的信息。写操作可以用于指示向标签的内存储中写入信息。失效操作可以用于指示失效标签，被失效的标签不能被盘点、被读/写。
下面参照图2，以USRUP102包括在阅读器中为例，描述阅读器对标签执行读操作的过程。该过程包括：(1)阅读器(如图1中的USRP102)向标签(如图1中的标签101)发送选择(Select)命令。例如，当阅读器收到服务器发送的盘点命令(或者，该盘点命令可由服务器通过中间件向阅读器发送)，生成选择命令，该选择命令中包括标签的范围(如某些特定范围的EPC码)。标签接收选择命令后，判断自己是否属于该选择命令指示的标签的范围，若该标签属于指示的标签的范围，则在后续接收查询命令后返回信息。若该标签不属于应判断的标签范围，则在接收查询命令后不必返回信息。(2)阅读器向指定范围内的标签发送查询(Query)命令。(3)当标签发现自己属于选择命令中的标签范围时，向阅读器返回一个随机数。例如，标签可以通过竞争的方式向阅读器返回随机数， 例如，随机数可以是16比特的随机数(16-bit Random Number，RN16)。(4)当阅读器收到来自标签发送的随机数后，向标签发送应答(Acknowledgement，ACK)命令，该ACK命令中包含了接收到的随机数(比如RN16)。当标签收到阅读器发送的应答命令，并验证该随机数正确后，向阅读器返回其EPC码。
本申请实施例中，阅读器与标签之间传输的信息/信令消息可以携带在信号中，比如上述查询命令、随机数RN16、ACK命令、EPC码等都可以携带在信号中传输，其中携带查询命令的信号可以称为查询信号，携带随机数RN16的信号可以称为RN16信号，携带ACK命令的信号可以称为ACK信号，携带EPC码的信号可以称为EPC码信号等。如图2所示，不同信号的幅度和频率是不同的。
其中，根据协议，每一轮标签的ID(比如EPC码)的读取过程都如上，都以阅读器的查询命令开始，标签在接收到查询的命令后会发送一个随机数RN16。其中RN16是一个固定模式的信号，比如标签1生成的随机数RN16为1100 1011 0111 1010。可选的，标签的硬件特性会导致RN16信号产生某些微小波动，比如假设具有相同EPC码的标签的硬件特性不同，则在该标签的进行识别的过程中，这些具有相同EPC码的标签生成的RN16信号的特征(比如时域特性和/或频域特性)可能是不同的。因此，本发明采集RN16信号作为原始数据，通过深度学习模型对具有不同硬件特性的标签生成的RN16信号进行特征分析，根据标签生成的RN16信号进行标签识别。
参照图3为本申请提供的一种基于深度学习的标签识别方法，该方法适用于标签识别，解决恶意者/不法分子假冒标签的ID带来的标签真实身份识别有误的问题。如图3所示，该方法包括：
S310、对多个标签样本进行信号采集，获得信号样本。
其中，多个标签样本可以包括不同设计模型的标签，比如可以获取来自两个或者两个以上的制造厂商(可以理解的是，这些制造厂商都是合法的制造厂商)的多种不同型号的标签。一种可能的设计中，可以将获取到的所有标签作为标签样本，对所有标签进行进行信号采集获得信号样本。又一种可能的设计中，可以将获取到的大量的标签分为多组(比如分为三组)，针对每一组标签，作为标签样本，对该组标签进行信号采集获得信号样本。其中每组标签包括不同设计模型的标签。需要说明的是，对标签样本进行信号采集时，阅读器与标签样本之间的 通信信道是固定的，为验证本发明所示方法的普遍适用性，收集每个标签样本对应的大量的随机数RN16信号，比如收集每个标签的5000个RN16信号等。
其中信号样本包括阅读器与标签通信过程中产生的信号，比如信号样本可以包括阅读器的信号、标签的信号，比如该标签产生的RN16信号等，除此之外，信号样本还可以包括其他信号，比如直流载波信号。
示例性，可以利用USRP和软件无线电(software defined radio，SDR)构建了符合Gen2协议的阅读器，其中SDR使用的是GNU Radio软件无线电开发包。通过构建的阅读器对多个标签样本进行信号采集，获得信号样本。如此，通过硬件设备USRP和软件GNU Radio设计的GU-RFID平台，采集超高频RFID标签(或者简称为超高频标签)的信号，利用GU-RFID平台开展阅读器的测试与信号的调制工作。
具体的，参照图4，GU-RFID平台的阅读器由一台计算机(比如笔记本电脑)、一台USRP设备(或者称为USRP源)和两个圆极化天线组成，该计算机与USRP设备通过网络连接或者通用串行总线(Universal Serial Bus，USB)接口连接。这两个圆极化天线连接在USRP设备上，一个天线负责通过RFID技术向标签发送阅读器命令(比如查询命令、响应命令等)，另一个负责接收标签的反向散射信号(比如RN16信号、EPC码等等)。应理解，本发明中标签的反向散射信号可以称为标签响应，阅读器发出的信号可以称为读取器命令。
继续参照图4中本申请的阅读器内部的结构图，如图4所示，该阅读器包括USRP源、匹配滤波器(Matched Filter)、门模块(或者称为标签响应门(Tag Response Gate)模块)、标签解码器(Tag Decoder)、Gen2逻辑模块(或者称为阅读器逻辑(Reader Logic)模块)和USRP接收器模块，其中URSP接收器模块包括发送天线(或者简称为TX)和接收天线(或者简称为RX)。进一步的，如图4所示，该阅读器还可以包括放大器(Amplifier)模块。
需要说明的是，在图4所示架构图中，阅读器可以包括接收链和发送链，其中接收链用于接收标签的反向散射信号(比如RN16信号、EPC码等等)。发送链用于向标签发送阅读器命令(比如查询命令、响应命令等)。此外，图4仅为示例性附图，本申请中的阅读器包括但不限于图4所示部件，还包括其他部件。
具体的，利用阅读器对标签样本进行信号采集，获得信号样本可以包括：
(1)发送链中的阅读器逻辑模块生成查询命令，该查询命令用于查询标签生成的信号(比如RN16信号)。然后将查询命令经过可配置的放大器模块进行放大处理，将放大处理后的查询命令通过TX发送出去，比如通过TX向标签发送。应理解，本申请的查询命令可以称为开关键控命令，该查询命令可以是通过开关键控方式(或者振幅键控(Amplitude Shift Keying，ASK)调整方式和开关监控方式)生成的命令。
(2)相应的，标签接收到阅读器发出的查询命令(或者称为开关键控命令)后，返回标签的反向散射信号(比如RN16信号)，阅读器的RX接收标签返回的信号，并将接收到的信号存储到Linux内核中。进一步的，阅读器的接收链从Linux内核中提取到接收信号的信号样本。样本首先通过匹配滤波器模块，该滤波器负责用半符号周期的方形脉冲对接收到的信号进行滤波，以此达到最大限度地提高标签传输的信号的信噪比(Signal Noise Ratio，SNR)。然后过滤后的信号进入到门模块，门结构负责识别阅读器的查询命令，通过跟踪接收信号的幅度来识别阅读器的查询命令，从而只处理阅读器所查询的标签的返回的信号样本，这也同时减少了计算资源，因为只有在预期有效标签响应时才执行标签的解码流程，从而减少了整体的系统延迟。经过门模块处理的信号样本作为下一个模块的输入进行下一步的处理，标签解码模块负责帧同步、信道估计和检测标签的响应，而信道的同步是通过将接收到的信号与已知的前导码相关联完成的，最后解码器经过幅移键控(amplitude shift keying，ASK)解调标签响应的信号，得到信号样本。
如此，可以通过基于USRP的阅读器与标签样本进行通信，采集阅读器与标签通信过程的信号，作为信号样本。
S320、对获取到的信号样本进行信号预处理，得到数据集。
其中，数据集中包括标签的RN16信号。且该RN16信号以深度学习模型的输入向量的形式包括在数据集中。
具体的，S320可以包括下述步骤S3201及S3202：
S3201、对获取到的信号样本进行定位并提取出信号样本中的RN16信号。
由于阅读器中的USRP(或者称为USRP源)接收到的原始信号/信号样本包括直流载波信号、阅读器的信号(或者称为阅读器命令)和标签的信号(或者称为标签响应)。为了实现提取的信号与数据无关，而是与标签本身的硬件特性有 关，首先需要从原始信号/信号样本中提取/分离RN16数据包。
结合图2所示的阅读器与标签之间的通信时序图可知，阅读器的信号与标签的信号发生在不同时序，且不同信号具有不同特性，比如具有不同的频域特性。为了精确地将RN16信号从原始信号/信号样本中提取出来，可以设计一个滑动窗口，利用该滑动窗口来遍历阅读器与标签通信过程中的整个信号。结合各信号的时频特性从中提取出RN16信号。
比如提取RN16信号可以包括以下两个步骤：步骤一，通过观察标签的信号和阅读器的信号的时域图和频域图，如图5、6为阅读器的信号的时域图和频域图，图7、8为标签的信号的时域图和频域图，可以发现阅读器的信号和标签的信号的频域图存在明显的差异，进而可以利用快速傅里叶变换检测该窗口中的信号能量是否符合标签的信号模式，若符合，则确定当前滑动窗中的信号为标签的信号，反之，则是阅读器的信号。步骤二，由于标签的RN16信号与标签的ID信号(或者标签的EPC码)在频域上有相同的模式，即具有相同的频域特性，则需要进一步区分标签的RN16信号和标签的ID信号。比如设置的滑动窗口的时间长度为略大于RN16信号的长度，这意味着RN16信号会落入滑动窗口中。如图9为标签滑动窗口的操作，可以看出的是滑动窗口的前端与后端是直流载波信号，满足这个条件为候选窗口，位于候选窗口中的信号为RN16信号，可以从候选窗口中提取出整个RN16信号；反之，不满足这个条件的为非候选窗口，位于非候选窗口中的信号为标签的ID信号(或者称为标签的EPC码)。
应理解，为了获得更精确的RN16信号，需要设置更小的滑动窗口来寻找标签的脉冲图形，精确地定位标签的信号的瞬态点。
S3202、根据信号样本中提取出的RN16信号构建数据集。
由于阅读器中的USRP(或者称为USRP源)采集的信号样本存储为32位的浮点数，为了构建适用于卷积网络模型训练的向量集。则将采集到的信号样本中的RN16信号(即原始I/Q数据流)转换成卷积神经网络常用的四维向量形式N example×Dim channel×Dim IQ×Dim value。
其中Dim IQ＝2分别存储I/Q两个通道的数据，Dim value＝1200意味着每个数据向量包括1200个样本点，取1200个点作为RN16信号长度，这是为了减少逻辑数据对标签识别分类的影响，Dim channel＝1代表单色样本类型，N example代 表向量数量。对信号样本中RN16信号进行向量化样本数据流处理得到用四维向量形式表示的向量后，将向量化样本进行序列化处理得到一个序列，将得到的序列存储到文件中，以便后续用于训练得到卷积神经网络模型。
S330、利用S320中获取到的数据集训练得到深度学习模型；其中，该深度学习模型用于对标签的RN16信号进行特征学习得到标签的RN16信号的特征信息(比如时域特性和/或频域特性等)，该标签的RN16信号的特征信息用于对标签进行身份识别。
其中，该深度学习模型可以为卷积神经网络(Convolutional Neural Network，CNN)模型，该CNN模型包括一定数量的卷积层和全连接层。该CNN模型的特征是每个无线电信号的原始IQ样本，这些原始IQ样本已经进行归一化处理。本发明不对原始的无线电信号进行任何的专家特征提取或其他预处理，而是直接利用网络在高维数据上学习原始时间序列样本的特征。
具体的，参照图10，为CNN模型示意图，如图10所示，该CNN模型采用三层卷积层和四个全连接层，除了最后一个输出分类层采用Softmax(归一化指数函数)函数，其他层都采用线性整流函数(Rectified Linear Unit，ReLU)函数作为激活函数。其中三个卷积层分别包含128、64、32个过滤器，四个全连接层分别包含了512、256、128、50个神经元。为了防止过拟合的出现，采用了正则化的方法，在全连接层应用了弃权(dropout)技术，这是一种非常有效提高泛化能力，降低过拟合的方法。
示例性的，数据集经向量化后，单个样本向量输入维度是I/Q信道数×样本长度，其中I/Q通道数是2，样本长度取1200个样本数。经过大规模训练与反复调试，利用交叉熵损失函数和自适应矩估计(Adaptive Momentum Estimation，Adam)优化器进行优化训练，最终得到如图10所示的CNN模型。其中利用数据集对模型训练与反复调试最终得到能够输出RN16信号的特征信息的卷积神经网络的过程可以参照现有有关模型训练及调试的过程，不予详述。
应理解，该标签的RN16信号的特征信息用于对标签进行身份识别可以包括：若与阅读器通信的某个待测标签的RN16信号的特征与经深度学习模型学习得到的该标签的RN16信号的特征不一致/不符，则确定该待测标签的身份是虚假的，不合法的，有可能是被攻击者假冒或者伪造的一个与真实标签的ID相同的 标签，反之，若符合，则确定该待测标签的身份是真实的。
基于图4所示方法，利用大量的信号样本训练得到能够输出RN16信号的特征信息的深度学习模型，基于学习到的RN16信号的特征信息对标签进行识别，后续如果阅读器读取的某个标签的RN16信号的特征与经深度学习模型学习得到的RN16信号的特征不符，则确定当前与阅读器通信的标签可能是被伪造的，是不真实，反之，则是真实的。此外，本发明所示方法兼容当前的标签工业标准，可以顺利部署到现有的设备上。此外，深度学习基于多层非线性神经网络，结合大量训练数据，自动抽取特征并逐层抽象直接从数据中获取特征，减少了为每个问题设计特征提取器的工作量，并且基于深度学习的分类方法对超高频标签的分类表现了更强的性能。
参照图11，其示出了根据本发明一个实施例描述的基于深度学习的标签识别装置110的结构示意图。如图11所示，基于深度学习的标签识别装置110，可以包括：信号采集模块1110、信号预处理模块1120与信号识别模块1130。
其中，信号采集模块1110，用于对多个标签样本进行信号采集，获得信号样本；其中信号样本包括标签样本的信号；标签样本的信号包括标签样本的随机数RN16信号。
信号预处理模块1120，用于对信号样本进行信号预处理，得到数据集；其中数据集中包括信号样本中标签样本的随机数RN16信号，且RN16信号以深度学习模型的输入向量形式保存在数据集中。
信号识别模块1130，用于利用数据集训练得到深度学习模型；其中，深度学习模型用于对标签的RN16信号进行特征学习得到标签的RN16信号的特征信息，深度学习模型用于对标签进行身份识别。
可选的，信号采集模块1110的作用是利用Gnuradio软件无线电将USRP N210设备仿真为符合Gen2协议标准的阅读器采集标签的信号。比如信号采集模块1110利用USRP和软件无线电SDR构建了符合Gen2协议的阅读器；利用构建的阅读器对多个标签样本进行信号采集，获得信号样本。
可选的，阅读器与多个标签样本中不同标签样本之间的通信信道相同。
可选的，信号预处理模块1120定位标签相应的RN16信号并提取相应的RN16信号，采集到的信号是以32位浮点数的形式存储，需要转化为深度学习 能够学习的格式。比如信号预处理模块1120利用阅读器的信号的时域特性以及标签的时域特性，从信号样本中提取出信号样本中标签样本的信号；通过滑动窗口的从标签样本的信号中提取出标签样本的RN16信号；其中，滑动窗口的长度大于RN16信号的长度；将信号样本中提取出的标签样本的RN16信号转换成符合深度学习模型的输入向量形式的数据，将转换后的数据存在数据集中。
可选的，深度学习模型为卷积神经网络CNN模型；CNN模型的输入向量为四维向量形式的RN16信号，CNN模型的输出特征为RN16信号的特征信息。
可选的，信号识别模块1130根据采集的信号进行深度学习模型的训练以及模型的性能提升，并以此来识别和分类标签。比如信号识别模块1130将数据集中的RN16信号进行向量化处理，得到样本向量；其中样本向量中的单个样本向量输入维度是I/Q信道数×样本长度；利用向量化处理后的样本向量对初始化CNN模型进行训练以及反复调试得到稳定的CNN模型，将稳定的CNN模型作为最终用于对标签进行身份识别的CNN模型。
可选的，标签的RN16信号的特征信息用于对标签进行身份识别包括：若与阅读器通信的待测标签的RN16信号的特征与经深度学习模型学习得到的该标签的RN16信号的特征不一致，则确定待测标签的身份是不合法的，反之，若符合，则确定该待测标签的身份是真实的。
本发明提供的一种基于深度学习的标签识别装置，可以执行上述方法的实施例，其实现原理和技术效果类似，在此不再赘述。
图12为本发明实施例提供的一种电子设备的结构示意图。如图12所示，示出了适于用来实现本发明的电子设备1200的结构示意图。
如图12所示，电子设备1200包括中央处理单元(CPU)1201，其可以根据存储在只读存储器(ROM)1202中的程序或者从存储部分1208加载到随机访问存储器(RAM)1203中的程序而执行各种适当的动作和处理。在RAM 1203中，还存储有设备1200操作所需的各种程序和数据。CPU 1201、ROM 1202以及RAM 1203通过总线1204彼此相连。输入/输出(I/O)接口1205也连接至总线1204。
以下部件连接至I/O接口1205：包括键盘、鼠标等的输入部分1206；包括诸如阴极射线管(CRT)、液晶显示器(LCD)等以及扬声器等的输出部分1207；包括硬盘等的存储部分1208；以及包括诸如LAN卡、调制解调器等的网络接口 卡的通信部分1209。通信部分1209经由诸如因特网的网络执行通信处理。驱动器1210也根据需要连接至I/O接口1206。可拆卸介质1211，诸如磁盘、光盘、磁光盘、半导体存储器等等，根据需要安装在驱动器1210上，以便于从其上读出的计算机程序根据需要被安装入存储部分1208。
特别地，根据本公开的实施例，上文参考图3描述的过程可以被实现为计算机软件程序。例如，本公开的实施例包括一种计算机程序产品，其包括有形地包含在机器可读介质上的计算机程序，计算机程序包含用于执行上述基于深度学习的标签识别方法的程序代码。在这样的实施例中，该计算机程序可以通过通信部分1209从网络上被下载和安装，和/或从可拆卸介质1211被安装。
附图中的流程图和框图，图示了按照本发明各种实施例的系统、方法和计算机程序产品的可能实现的体系架构、功能和操作。在这点上，流程图或框图中的每个方框可以代表一个模块、程序段、或代码的一部分，前述模块、程序段、或代码的一部分包含一个或多个用于实现规定的逻辑功能的可执行指令。也应当注意，在有些作为替换的实现中，方框中所标注的功能也可以以不同于附图中所标注的顺序发生。例如，两个接连地表示的方框实际上可以基本并行地执行，它们有时也可以按相反的顺序执行，这依所涉及的功能而定。也要注意的是，框图和/或流程图中的每个方框、以及框图和/或流程图中的方框的组合，可以用执行规定的功能或操作的专用的基于硬件的系统来实现，或者可以用专用硬件与计算机指令的组合来实现。
描述于本申请实施例中所涉及到的单元或模块可以通过软件的方式实现，也可以通过硬件的方式来实现。所描述的单元或模块也可以设置在处理器中。这些单元或模块的名称在某种情况下并不构成对该单元或模块本身的限定。
上述实施例阐明的系统、装置、模块或单元，具体可以由计算机芯片或实体实现，或者由具有某种功能的产品来实现。一种典型的实现设备为计算机。具体的，计算机例如可以为个人计算机、笔记本电脑、行动电话、智能电话、个人数字助理、媒体播放器、导航设备、电子邮件设备、游戏控制台、平板计算机、可穿戴设备或者这些设备中的任何设备的组合。
作为另一方面，本发明还提供了一种存储介质，该存储介质可以是上述实施例中前述装置中所包含的存储介质；也可以是单独存在，未装配入设备中的存储 介质。存储介质存储有一个或者一个以上程序，前述程序被一个或者一个以上的处理器用来执行描述于本发明的基于深度学习模型的标签识别方法。
存储介质包括永久性和非永久性、可移动和非可移动媒体可以由任何方法或技术来实现信息存储。信息可以是计算机可读指令、数据结构、程序的模块或其他数据。计算机的存储介质的例子包括，但不限于相变内存(PRAM)、静态随机存取存储器(SRAM)、动态随机存取存储器(DRAM)、其他类型的随机存取存储器(RAM)、只读存储器(ROM)、电可擦除可编程只读存储器(EEPROM)、快闪记忆体或其他内存技术、只读光盘只读存储器(CD-ROM)、数字多功能光盘(DVD)或其他光学存储、磁盒式磁带，磁带磁磁盘存储或其他磁性存储设备或任何其他非传输介质，可用于存储可以被计算设备访问的信息。按照本文中的界定，计算机可读介质不包括暂存电脑可读媒体(transitory media)，如调制的数据信号和载波。
需要说明的是，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、商品或者设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、商品或者设备所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括要素的过程、方法、商品或者设备中还存在另外的相同要素。
本说明书中的各个实施例均采用递进的方式描述，各个实施例之间相同相似的部分互相参见即可，每个实施例重点说明的都是与其他实施例的不同之处。尤其，对于系统实施例而言，由于其基本相似于方法实施例，所以描述的比较简单，相关之处参见方法实施例的部分说明即可。

Claims (10)

  1. 一种基于深度学习的标签识别方法，其特征在于，所述方法包括：
    对多个标签样本进行信号采集，获得信号样本；其中所述信号样本包括标签样本的信号；所述标签样本的信号包括所述标签样本的随机数RN16信号；
    对所述信号样本进行信号预处理，得到数据集；其中所述数据集中包括所述信号样本中标签样本的随机数RN16信号，且所述RN16信号以深度学习模型的输入向量形式保存在所述数据集中；
    利用所述数据集训练得到深度学习模型；其中，所述深度学习模型用于对标签的RN16信号进行特征学习得到标签的RN16信号的特征信息，所述标签的RN16信号的特征信息用于对所述标签进行身份识别。
  2. 根据权利要求1所述的方法，其特征在于，所述对多个标签样本进行信号采集，获得信号样本包括：
    利用USRP和软件无线电SDR构建了符合Gen2协议的阅读器；
    利用构建的所述阅读器对多个标签样本进行信号采集，获得信号样本。
  3. 根据权利要求2所述的方法，其特征在于，
    所述阅读器与所述多个标签样本中不同标签样本之间的通信信道相同。
  4. 根据权利要求1所述的方法，其特征在于，所述对所述信号样本进行信号预处理，得到数据集包括：
    利用阅读器的信号的时域特性以及标签的时域特性，从所述信号样本中提取出所述信号样本中标签样本的信号；
    通过滑动窗口的从所述标签样本的信号中提取出标签样本的RN16信号；其中，所述滑动窗口的长度大于所述RN16信号的长度；
    将所述信号样本中提取出的标签样本的RN16信号转换成符合所述深度学习模型的输入向量形式的数据，将转换后的数据存在所述数据集中。
  5. 根据权利要求1所述的方法，其特征在于，
    所述深度学习模型为卷积神经网络CNN模型；
    其中，所述CNN模型的输入向量为四维向量形式的RN16信号，所述CNN模型的输出特征为所述RN16信号的特征信息。
  6. 根据权利要求5所述的方法，其特征在于，所述利用所述数据集训练得到深度学习模型包括：
    将所述数据集中的RN16信号进行向量化处理，得到样本向量；其中样本向量中的单个样本向量输入维度是I/Q信道数×样本长度；
    利用向量化处理后的样本向量对初始化CNN模型进行训练以及反复调试得到稳定的CNN模型，将所述稳定的CNN模型作为最终用于对标签进行身份识别的CNN模型。
  7. 根据权利要求1所述的方法，其特征在于，所述标签的RN16信号的特征信息用于对所述标签进行身份识别包括：
    若与阅读器通信的待测标签的RN16信号的特征与经所述深度学习模型学习得到的该标签的RN16信号的特征不一致，则确定所述待测标签的身份是不合法的，反之，若符合，则确定该待测标签的身份是真实的。
  8. 一种基于深度学习的标签识别装置，其特征在于，所述装置包括：
    信号采集模块，用于对多个标签样本进行信号采集，获得信号样本；其中所述信号样本包括标签样本的信号；所述标签样本的信号包括所述标签样本的随机数RN16信号；
    信号预处理模块，用于对所述信号样本进行信号预处理，得到数据集；其中所述数据集中包括所述信号样本中标签样本的随机数RN16信号，且所述RN16信号以深度学习模型的输入向量形式保存在所述数据集中；
    信号识别模块，用于利用所述数据集训练得到深度学习模型；其中，所述深度学习模型用于对标签的RN16信号进行特征学习得到标签的RN16信号的特征信息，所述深度学习模型用于对标签进行身份识别。
  9. 一种电子设备，包括存储器、处理器及存储在存储器上并可在处理器上 运行的计算机程序，其特征在于，所述处理器执行所述程序时实现如权利要求1-7中任一所述的基于深度学习的标签识别方法。
  10. 一种可读存储介质，其上存储有计算机程序，其特征在于，该程序被处理器执行时实现如权利要求1-7中任一所述的基于深度学习的标签识别方法。
PCT/CN2022/077200 2022-02-22 2022-02-22 基于深度学习的标签识别、装置、电子设备及存储介质 WO2023159340A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2022/077200 WO2023159340A1 (zh) 2022-02-22 2022-02-22 基于深度学习的标签识别、装置、电子设备及存储介质

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2022/077200 WO2023159340A1 (zh) 2022-02-22 2022-02-22 基于深度学习的标签识别、装置、电子设备及存储介质

Publications (1)

Publication Number Publication Date
WO2023159340A1 true WO2023159340A1 (zh) 2023-08-31

Family

ID=87764296

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/077200 WO2023159340A1 (zh) 2022-02-22 2022-02-22 基于深度学习的标签识别、装置、电子设备及存储介质

Country Status (1)

Country Link
WO (1) WO2023159340A1 (zh)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120176219A1 (en) * 2008-12-12 2012-07-12 Massimo Fiorucci Access identification and control device
CN106022185A (zh) * 2016-05-09 2016-10-12 天津理工大学 一种基于连续检测机制适用于EPC Gen2协议的防冲突算法
CN106452780A (zh) * 2016-09-08 2017-02-22 中国科学院信息工程研究所 一种适用于mimo rfid系统的身份认证方法
CN110503354A (zh) * 2019-07-02 2019-11-26 北京交通大学 一种基于深度学习的rfid标签位置估计方法
CN112116742A (zh) * 2020-08-07 2020-12-22 西安交通大学 融合用户多源发声特征的身份认证方法、存储介质及设备
CN113887502A (zh) * 2021-10-21 2022-01-04 西安交通大学 一种通信辐射源时频特征提取与个体识别方法及系统

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120176219A1 (en) * 2008-12-12 2012-07-12 Massimo Fiorucci Access identification and control device
CN106022185A (zh) * 2016-05-09 2016-10-12 天津理工大学 一种基于连续检测机制适用于EPC Gen2协议的防冲突算法
CN106452780A (zh) * 2016-09-08 2017-02-22 中国科学院信息工程研究所 一种适用于mimo rfid系统的身份认证方法
CN110503354A (zh) * 2019-07-02 2019-11-26 北京交通大学 一种基于深度学习的rfid标签位置估计方法
CN112116742A (zh) * 2020-08-07 2020-12-22 西安交通大学 融合用户多源发声特征的身份认证方法、存储介质及设备
CN113887502A (zh) * 2021-10-21 2022-01-04 西安交通大学 一种通信辐射源时频特征提取与个体识别方法及系统

Similar Documents

Publication Publication Date Title
US11391819B2 (en) Object verification using radar images
Han et al. GenePrint: Generic and accurate physical-layer identification for UHF RFID tags
US20210142337A1 (en) Product Authenticity Verification Based on Environment
US20060181394A1 (en) Radio frequency fingerprinting to detect fraudulent radio frequency identification tags
Zhang et al. CSI-based human activity recognition with graph few-shot learning
CN102968606A (zh) Rfid双向认证方法
Parada et al. Gesture detection using passive RFID tags to enable people-centric IoT applications
CN102945528A (zh) 一种基于手机(或移动终端)和rfid电子标签的商品防伪方法
EP3740894A1 (en) A method and system for automatically annotating and identifying a living being or an object with an identifier providing a subject identification
Liu et al. Pinpointing anomaly RFID tags: Situation and opportunities
WO2023159340A1 (zh) 基于深度学习的标签识别、装置、电子设备及存储介质
US10929741B2 (en) Cross-registration for unclonable chipless RFID tags
Zanetti et al. Types and origins of fingerprints
Li et al. Recovery of collided RFID tags with frequency drift on physical layer
CN114662510A (zh) 基于深度学习的标签识别、装置、电子设备及存储介质
WO2022222576A1 (zh) 一种场景识别方法及电子设备
CN205334511U (zh) 一种通用身份标识识别装置
Maher Abdulzahra Novel anti-collision algorithm in RFID tag identification process
Goyal et al. Security system for vehicle using number plate detection and rfid
Hinders et al. Classification of RFID tags with wavelet fingerprinting
Peng et al. A support vector machine classification-based signal detection method in ultrahigh-frequency radio frequency identification systems
CN106056339A (zh) 一种利用电磁波穿透性的物品鉴定方法
Arulmozhi et al. ALRC: A novel adaptive linear regression based classification for grade based student learning using radio frequency identification
Cui System efficiency of collision recover binary tree algorithm in RFID
Zhao et al. Leveraging Topic Model for CSI Based Human Activity Recognition

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22927637

Country of ref document: EP

Kind code of ref document: A1