Radio frequency fingerprint recognition method based on multi-task and multi-feature fusion

Abstract

The invention discloses a radio frequency fingerprint recognition method based on multi-task and multi-feature fusion, comprising: data preprocessing: normalizing three signals, namely, power spectrum, time spectrum and cross-power spectrum, and performing data enhancement operation; signal feature extraction: extracting three signal features, namely, power spectrum feature, time spectrum feature and cross-power spectrum feature, as input of a neural network model; network training: training the neural network model using the three signal features, namely, power spectrum feature, time spectrum feature and cross-power spectrum feature; classification and recognition: performing radio frequency fingerprint recognition on a test set signal using the trained neural network model to obtain a classification and recognition result; the multi-feature fusion and multi-task method proposed by the invention is improved over the single-feature and single-task method, and is an effective radio frequency fingerprint recognition method.

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing, and in particular to a radio frequency fingerprint recognition method based on multi-task and multi-feature fusion. Background Art

[0002] According to statistics, the total number of IoT devices connected to the network in 2020 has exceeded 21.7 billion, and is still on an upward trend, and is expected to reach 41.2 billion in 2025. While the massive number of IoT devices connected to the network has brought great convenience, it has also brought severe challenges to information security. Among them, the access authentication of IoT devices is a key link in security measures. It is the first barrier to information security and an important foundation for building a secure and trusted network. Traditional IoT communication security means mainly rely on encryption technology, such as the common challenge-response based protocol. Such protocols often rely on pre-shared public keys. Due to the limited computing power and power of IoT devices, it is difficult to directly transplant and apply them.

[0003] Radio Frequency Fingerprint (RFF) is a signal feature information that does not disappear with changes in signal transmission scenarios due to inevitable minor errors in electronic components during the manufacturing process of the device. This type of feature information is small and unique, just like a human fingerprint, it is unique and cannot be copied. Radio Frequency Fingerprint Identification (RFFI) is an emerging IoT security technology that uses the unique and unalterable errors generated by the hardware manufacturing of wireless devices to achieve device identification and authentication. RF fingerprint technology has the characteristics of not causing unnecessary overhead to the device and being lightweight, and can be used as an ideal technology for IoT device access authentication.

[0004] Existing RF fingerprint recognition research can be mainly divided into feature engineering-based methods and deep learning-based methods. Feature engineering-based methods require manual extraction of statistical information or physical parameters such as signal modulation information, transform domain information, basic parameter information, image information, etc. as features, and then use classifiers to identify devices. This method is highly dependent on feature extraction algorithms, requires high knowledge and experience in signal processing, and some hardware features are interrelated and difficult to extract manually. In recent years, deep learning has ushered in rapid development and has gradually been applied to the field of RF fingerprint recognition. IoT device recognition methods based on deep learning usually rely on classification neural networks to process the original signal or its transformed form and directly infer the device identity, eliminating the complicated feature engineering and attracting widespread attention. Some scholars have used I / Q data, differential constellation trajectory diagrams, short-time Fourier transform (STFT) to obtain time-frequency spectrum, power spectrum, bispectrum, and cyclic spectrum as inputs to the neural network and have achieved good results, but the use of a single signal feature still has the problem of low recognition accuracy in the experimental process. Summary of the invention

[0005] In order to solve the problems existing in the prior art, the purpose of the present invention is to provide a radio frequency fingerprint recognition method based on multi-task and multi-feature fusion. The multi-feature fusion multi-task method proposed in the present invention is improved compared with the single-feature single-task method and is an effective radio frequency fingerprint recognition method.

[0006] To achieve the above object, the technical solution adopted by the present invention is: a radio frequency fingerprint recognition method based on multi-task multi-feature fusion, comprising the following steps:

[0007] Step 1: Data preprocessing: normalize the power spectrum, time spectrum and cross power spectrum and perform data enhancement operations;

[0008] Step 2, signal feature extraction: extract three signal features, namely power spectrum feature, time-frequency spectrum feature and cross-power spectrum feature, as input of the neural network model;

[0009] Step 3, network training: use power spectrum features, time-frequency spectrum features and cross-power spectrum features to train the neural network model;

[0010] Step 4: Classification and identification: Use the trained neural network model to perform RF fingerprint identification on the test set signal to obtain the classification and identification results.

[0011] As a further improvement of the present invention, in step 2, the method for extracting the power spectrum feature is specifically as follows:

[0012] set up is a random signal sequence, divided into L segments, the length of each segment is M, the length of the overlap between two adjacent segments is Mk, and the signal of the i-th segment is:

[0013] x i (n)=x[n+(i-1)k],0≤n≤M-1,1≤i≤L

[0014] Wherein, 0<k<M, L and M satisfy (L-1)k+M≤N.

[0015] The power spectrum of the i-th segment is:

[0016]

[0017]

[0018] In the formula, U is the normalization factor, ω(n) is the window function, and the power spectrum of the L-segment signal is averaged to obtain the power spectrum of the entire signal x(n) based on the Welch algorithm:

[0019]

[0020] As a further improvement of the present invention, in step 2, the method for extracting the time-frequency spectrum feature is specifically as follows:

[0021] The discrete-time STFT is expressed as:

[0022]

[0023] In the formula, S k,m are the elements of the STFT complex matrix S, k = 1, 2, ..., N, m = 1, 2, ..., M, M is the number of columns, N is the length of the window function ω[n], R is the number of hops, s[n] is the signal being analyzed, and its time spectrum Expressed on the dB scale as:

[0024]

[0025] Here, || returns the magnitude.

[0026] As a further improvement of the present invention, in step 2, the method for extracting the cross-power spectrum feature is specifically as follows:

[0027] Assume that the symbol sequences of two adjacent preamble codes are:

[0028] y s1 (n) = x s1 (n)*h(n)+v s1 (n),y s2 (n) = x s2 (n)*h(n)+vs2 (n)

[0029] In the formula, y(n) is the received signal, x(n) is the transmitted signal, h(n) is the system impulse response, v(n) is the additive Gaussian white noise, v(n) can be eliminated by the cross-correlation function, and the length of each symbol sequence is L s ,y s1 (n) and y s2 The cross-correlation function of (n) is:

[0030]

[0031] Where m = -L s +1,-L s +2,...,0,...,L s -2,L s -1, where v(n) is additive white Gaussian noise. The useful signal is uncorrelated with the noise signal, and the correlation coefficient is 0, so:

[0032]

[0033] The cross-power spectral density can be obtained by Fourier transforming the cross-correlation function.

[0034] As a further improvement of the present invention, the step 3 is specifically as follows:

[0035] The power spectrum feature, time-frequency spectrum feature and cross-power spectrum feature are used as input to train the multi-feature fusion network model. The multi-feature fusion network model completes the classification task through the Softmax classifier. The output of the Softmax classifier is a set of probability values ​​p, and its mathematical expression is:

[0036]

[0037] In the formula, δ i is the output of the last layer of the network model, i is the index of the category, and the total number of categories is C; the output probability satisfies the following conditions:

[0038]

[0039] The classification result finally output by the multi-feature fusion network model is the signal transmitting device corresponding to the index with the largest probability value;

[0040] The network model of multi-task learning is obtained by training the device classification task as the main task of the multi-feature fusion network model and the classification of signal noise content as the second task, that is, the neural network model.

[0041] The present invention proposes a multi-task radio frequency fingerprint recognition method based on the fusion of three signal features: power spectrum based on Welch algorithm, time-frequency spectrum based on STFT transformation, and cross-power spectrum, which improves the problem of low accuracy of single signal feature recognition method. In addition, based on the multi-feature fusion method, the present invention further proposes a multi-task learning method, which uses noise information as prior knowledge to help neural network learning for the first time, with device classification as the main task and signal noise content classification as the second task of the network to assist the main task training.

[0042] The beneficial effects of the present invention are:

[0043] The radio frequency fingerprint recognition method of the present invention combines the three features of power spectrum based on Welch algorithm, time-frequency spectrum obtained by short-time Fourier transform, and cross-power spectrum to effectively classify the transmitting device. At the same time, in order to solve the problem that the signal fluctuation caused by noise drowns the device features, a multi-task learning method using signal noise information as prior knowledge to help network training is proposed, that is, a multi-task learning method with device classification as the main task and classification of signal noise content as the second task of the network. Experiments have proved that the multi-feature fusion method is superior to the single-feature method in both single-task and multi-task learning scenarios; the multi-task learning method using multi-feature fusion can further improve the recognition accuracy on the basis of the multi-feature fusion method. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a diagram of a radio frequency fingerprint recognition model based on multi-feature fusion in an embodiment of the present invention;

[0045] Figure 2 A neural network structure diagram based on multi-feature fusion in an embodiment of the present invention;

[0046] Figure 3 This is a multi-task learning network model diagram based on multi-feature fusion in an embodiment of the present invention;

[0047] Figure 4 This is a comparison chart of the recognition effects of four methods in the embodiments of the present invention;

[0048] Figure 5 A schematic diagram of a confusion matrix of a multi-feature fusion MTL method in an embodiment of the present invention;

[0049] Figure 6 4 is a comparison chart of the recognition effects of the three methods in the embodiments of the present invention. DETAILED DESCRIPTION

[0050] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0051] Example

[0052] A radio frequency fingerprint recognition method based on multi-task and multi-feature fusion, the radio frequency fingerprint recognition system model based on multi-feature fusion of the present invention is as follows Figure 1 As shown. The overall process of the system is as follows:

[0053] 1) Data preprocessing. Normalize the signal and perform data enhancement operations. By adding noise data, the robustness of the model can be effectively increased;

[0054] 2) Signal feature extraction. Three signal features are extracted as input of the neural network;

[0055] 3) Network training. Use three signal features to train the network model;

[0056] 4) Classification and recognition: Use the trained network to perform RF fingerprint recognition on the test set signal to obtain the classification and recognition results.

[0057] Signal feature extraction:

[0058] The multi-feature fusion method proposed in this embodiment is the fusion of the following three features.

[0059] Power spectrum characteristics:

[0060] The power spectrum is the power of a signal per unit frequency, which describes the power distribution of the signal in the frequency domain. In order to obtain a better power spectrum curve, this embodiment uses the power spectrum based on the Welch algorithm as the input of the network. The Welch algorithm is optimized based on the Bartlett method in terms of variance performance and spectral distortion.

[0061] set up is a random signal sequence, divided into L segments, the length of each segment is M, the length of the overlap between two adjacent segments is Mk, and the signal of the i-th segment is:

[0062] x i (n)=x[n+(i-1)k],0≤n≤M-1,1≤i≤L

[0063] Wherein, 0<k<M, L and M satisfy (L-1)k+M≤N.

[0064] The power spectrum of the i-th segment is:

[0065]

[0066]

[0067] Where U is the normalization factor and ω(n) is the window function. The power spectrum of the L-segment signal is averaged to obtain the power spectrum of the entire signal x(n) based on the Welch algorithm:

[0068]

[0069] Time-frequency spectrum characteristics:

[0070] Fourier transform can only reflect the characteristics of the signal in the frequency domain, but cannot reflect the signal characteristics in the time domain. At present, short-time Fourier transform is a widely used time-frequency analysis algorithm. STFT divides the original signal into several small segments, and performs Fourier transform on each segment to obtain a set of local "spectrum" information about the original signal. Discrete-time STFT can be expressed mathematically as:

[0071]

[0072] In the formula, S k,m are the elements of the STFT complex matrix S, k = 1, 2, ..., N, m = 1, 2, ..., M, M is the number of columns, N is the length of the window function ω[n], R is the number of hops, s[n] is the signal being analyzed, and its time spectrum In dB scale it can be expressed as:

[0073]

[0074] Where || returns the amplitude, and the time-spectrum spectrum can effectively represent the change of frequency over time, as well as some signal parameters such as bandwidth and symbol duration.

[0075] Cross power spectrum characteristics:

[0076] The cross power spectrum is the Fourier transform of the cross-correlation function of two signal sequences, which describes the statistical correlation between the two signals in the frequency domain as a signal feature. Assume that the symbol sequences of two adjacent preamble codes of the signal are:

[0077] y s1 (n) = x s1 (n)*h(n)+v s1 (n),y s2 (n) = x s2 (n)*h(n)+v s2 (n)

[0078] In the formula, y(n) is the received signal, x(n) is the transmitted signal, h(n) is the system impulse response, and v(n) is the additive white Gaussian noise. v(n) can be eliminated by the cross-correlation function. Assume that the length of each symbol sequence is L s ,y s1 (n) and y s2 The cross-correlation function of (n) is:

[0079]

[0080] Where m = -L s +1,-L s +2,...,0,...,L s -2,L s -1, where v(n) is additive white Gaussian noise. The useful signal is uncorrelated with the noise signal, and the correlation coefficient is 0, so:

[0081]

[0082] Therefore, the cross-correlation function between two symbol sequences can greatly eliminate the influence of additive noise. The cross-power spectrum density can be obtained by Fourier transforming the cross-correlation function.

[0083] Neural Network Model:

[0084] Network model for multi-feature fusion:

[0085] This embodiment designs a multi-feature fusion neural network model with reference to the ResNet34 network. The model sets three input terminals with input sizes of 1×1024, 7×2047, and 1×256×63, which correspond to the power spectrum, cross-power spectrum, and time-frequency spectrum obtained by STFT transformation based on the Welch algorithm, respectively. The structure of the network is shown in the figure below. Figure 2 As shown in the figure, it consists of 35 convolutional layers, 3 fully connected layers, 3 maximum pooling layers, and 1 average pooling layer. In the figure, / represents the step size of convolution or pooling. The curve next to the residual block represents the residual edge, which can avoid the problem of gradient disappearance. The solid residual edge represents direct connection, and the dotted residual edge represents a 1×1 convolution to change the corresponding number of channels. The size of the convolution in the front part is 7, the step size is 2, and the padding size is 3. Compared with the fusion method of compressing features of the same size into multiple channels of a single input end, the three input ends retain their own dimensions and perform convolution to extract features separately, and then reconstruct and splice them into the same dimension to fuse input features of different sizes, thereby achieving the purpose of retaining their own features and avoiding the problem of partial feature information loss due to input size cropping. Among them, inputs 1 and 2 perform 1D convolution operations, and input 3 performs 2D convolution operations. Because the time-frequency spectrum after STFT transformation reflects the information of time and frequency, it is a 2D matrix, so a 2D convolution operation is required. Each residual block trains the fused features, and finally completes the classification task through the Softmax classifier. The output of the Softmax classifier is a set of probability values ​​p, and its mathematical expression is:

[0086]

[0087] In the formula, δ i is the output of the last layer of the network model, i is the index of the category, and the total number of categories is C. The output probability satisfies the following conditions:

[0088]

[0089] The classification result finally output by the network is the signal transmitting device corresponding to the index with the largest probability value.

[0090] Compared with the method that only uses a single signal feature, the RF fingerprint recognition method that integrates three signal features can represent device characteristics from different angles, learn more signal features, and effectively alleviate the problem that the information learned by the network from a single signal feature is incomplete.

[0091] Network model for multi-task learning:

[0092] Considering that the signal will change with the noise content, causing the input signal characteristics of the network to fluctuate, these fluctuations may even overwhelm the feature differences of the device itself, reducing the recognition accuracy. This embodiment further improves the multi-feature fusion method and proposes a multi-task learning method for multi-feature fusion. This method uses noise information labels to assist network training based on the multi-feature fusion method.

[0093] The multi-task learning method of multi-feature fusion proposed in this embodiment takes the device classification task as the main task and the classification of signal noise content as the second task of the network. The signal noise power after adding artificial Gaussian white noise is divided into 10 levels according to size, corresponding to the 10 labels 0-9. Because the noise information is added manually, the noise information can be used as prior knowledge to help network training during the training phase to optimize the performance of the network in different noise intensity scenarios, and the knowledge learned in each task in multi-task learning can help better learn other tasks.

[0094] Figure 3 is the network model structure of multi-task learning used in this embodiment. The model is divided into two parts, namely the shared layer and the specific task layer. There are two task branches. The shared layer is designed with a residual structure. Figure 2 Same as part 1. Figure 2 Compared with the model of Figure 2 Part 1 introduces the task 2 branch of the network, and the rest is the same. Figure 3 The task 2 branch in the figure corresponds to the classification task of the signal noise content, and the task 1 branch corresponds to the classification task of the device. The shared parameter part of the network model adopts a hard parameter sharing mechanism. The size of the input end is consistent with the above, and the Softmax classifier is used to complete the classification task. Because the underlying parameters are shared, it does not cause excessive increase in parameters. The hard parameter sharing mechanism also reduces the risk of overfitting in network training.

[0095] This multi-task learning approach allows the network to simultaneously improve the accuracy of device classification and the accuracy of signal noise content classification. Since the network learns the device feature information under a specific noise content during the training phase, rather than just learning the device feature information, using the classification of signal noise content as the second task of the network can alleviate the problem of reduced recognition accuracy due to signal fluctuations caused by noise submerging device features.

[0096] The present embodiment is further analyzed by experiments as follows:

[0097] Introduction to data set and experimental environment:

[0098] All neural network models in the experiment were run on a Windows system with an Intel i5-11400F 64-bit CPU and an NVDIA GeForce RTX 3060 GPU. The framework used was Pytorch. Adam was used as the optimizer for training, the initial learning rate was set to 0.001, the split ratio of the training set and the validation set was 9:1, the cross entropy function was used as the loss function, the batch-size was set to 32, and the GPU was used to accelerate the operation.

[0099] The data set used in the experiment comes from the open source data of IEEEDataPort, which is the signal preamble code part of 10 LoRa devices, that is, 8 preamble code symbol sequences, where the number of data packets corresponding to each device is 100, and the transmission configuration of LoRa is spreading factor SF=7, bandwidth B=125kHz, and carrier frequency 868MHz.

[0100] Data preprocessing:

[0101] Adding noise data can effectively increase the robustness of the model, so before transforming the signal data, we first perform an enhancement operation on it to generate more data for training. We replicate and expand the data set, and add 20-40dB of random noise to the normalized signal for training, so that the model can learn more noise signals.

[0102] Performance analysis of multi-feature fusion:

[0103] First, the multi-feature fusion method proposed in this embodiment is compared with the method using a single input feature. Experiments are conducted on the test set signal without adding noise and adding 20-40dB random noise. The recognition results are shown in Table 1, and the accuracy is the average of 5 results. It can be seen that the recognition accuracy of the multi-feature fusion method can reach 99.7% when no noise is added to the test set signal, and the recognition accuracy of the other three methods is also above 97%, all showing good performance. However, in the case of adding random noise, the recognition accuracy of the multi-feature fusion method designed in this embodiment is significantly better than that of the method using a single signal feature. Compared with the other three methods, the recognition accuracy can be increased by 14.2%, 8.2%, and 15.3%, respectively.

[0104] Table 1 Comparison of recognition accuracy of four methods

[0105]

[0106] Figure 4 We further compared the performance of the method based on multi-feature fusion and the method using a single signal feature in different noise environments. Figure 4 It can be seen that the four methods all show good performance in recognition rate when the signal-to-noise ratio is higher than 27dB, all exceeding 90%; in addition, the recognition accuracy of the method based on multi-feature fusion is improved compared with the method using only a single feature in all signal-to-noise ratio scenarios. When the signal-to-noise ratio is 20-25dB, the improvement is most obvious, with an increase of about 10%. At 22dB, the accuracy rate exceeds 90%.

[0107] Performance analysis of multi-task learning methods:

[0108] In order to verify the effectiveness of the proposed multi-task learning method, this embodiment uses a multi-task learning method that trains three signal features separately and a multi-task learning method that integrates multiple features to compare with their corresponding single-task learning methods. First, the test set signal is denoised. Figure 5 It is the confusion matrix of four multi-task learning methods when 20-40dB random noise is added to the test set signal. The confusion matrix records the prediction results of the classifier. The classification accuracy, recall rate and other indicators can be calculated through the confusion matrix.

[0109] exist Figure 5Based on this, Table 2 gives a comparison of the accuracy of the four methods. It can be seen from Table 2 that the method proposed in this embodiment has the highest accuracy, reaching 87.3%, which is 3.9% higher than the single task of multi-feature fusion. The multi-task learning method using time-frequency spectrum improves the accuracy of single-task learning by more than 5%, and the accuracy exceeds 80% when only one signal feature is used. It can be seen that although only the simplest hard parameter sharing mechanism is used, the multi-task learning method that introduces noise content as the second label has a certain improvement over the single-task learning method when all other conditions remain unchanged, which shows the effectiveness of the multi-task learning method using noise information.

[0110] It can also be seen from Table 2 that the multi-feature fusion method has the highest accuracy among the four methods in both single-task and multi-task learning scenarios, which is 8.2% and 6.4% higher than the time-frequency spectrum method with the highest accuracy among the single signal feature methods, further illustrating the effectiveness of using the multi-feature fusion method.

[0111] Table 2 Comparison of multi-task learning accuracy and single-task learning accuracy

[0112]

[0113] Performance comparison of this embodiment method with other methods:

[0114] In order to further verify the effectiveness of the method proposed in this embodiment, two methods proposed in recent years are compared and analyzed with the multi-task learning method of multi-feature fusion proposed in this embodiment: a method based on power spectrum combined with 1DCNN, and a method based on channel-independent time-frequency spectrum combined with CNN. The experimental results of the three methods without adding noise and adding 20-40dB random noise are shown in Table 3. It can be seen that the three methods all show very good performance when no noise is added to the test set signal, but after adding random noise, thanks to the richer input features of the network and stronger network performance, the method proposed in this embodiment is significantly better than the other two methods.

[0115] Table 3 Comparison of recognition accuracy of three methods

[0116]

[0117] Figure 6 By comparing the performances of the three methods in different noise environments, it can be seen that the performance of the method proposed in this embodiment is higher than that of the other two methods in different noise environments, which shows that the method based on multi-feature fusion proposed in this embodiment shows strong robustness and accuracy in noisy scenarios.

[0118] The above-mentioned embodiments only express the specific implementation of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A radio frequency fingerprint recognition method based on multi-task and multi-feature fusion, characterized in that: The following steps are involved: Step 1: Data preprocessing: normalize the power spectrum, time spectrum and cross power spectrum and perform data enhancement operations; Step 2, signal feature extraction: extract three signal features, namely power spectrum feature, time-frequency spectrum feature and cross-power spectrum feature, as input of the neural network model; The method for extracting the power spectrum feature is specifically as follows: set up is a random signal sequence, divided into L segments, the length of each segment is M, the length of the overlap between two adjacent segments is Mk, and the signal of the i-th segment is: x i (n)=x[n+(i-1)k],0≤n≤M-1,1≤i≤L Where, 0<k<M, L and M satisfy (L-1)k+M≤N; The power spectrum of the i-th segment is: In the formula, U is the normalization factor, ω(n) is the window function, and the power spectrum of the L-segment signal is averaged to obtain the power spectrum of the entire signal x(n) based on the Welch algorithm: The method for extracting the time-frequency spectrum features is specifically as follows: The discrete-time STFT is expressed as: In the formula, S k,m are the elements of the STFT complex matrix S, k = 1, 2, ..., N, m = 1, 2, ..., M, M is the number of columns, N is the length of the window function ω[n], R is the number of hops, s[n] is the signal being analyzed, and its spectrum S is expressed in dB scale as: In the formula, | | returns the amplitude; The method for extracting the cross power spectrum feature is specifically as follows: Assume that the symbol sequences of two adjacent preamble codes are: y s1 (n)=x s1 (n)*h(n)+v s1 (n),y s2 (n)=x s2 (n)*h(n)+v s2 (n) In the formula, y(n) is the received signal, x(n) is the transmitted signal, h(n) is the system impulse response, v(n) is the additive Gaussian white noise, v(n) can be eliminated by the cross-correlation function, and the length of each symbol sequence is L s ,y s1 (n) and y s2 The cross-correlation function of (n) is: Where m = -L s +1,-L s +2,...,0,...,L s -2,L s -1, where v(n) is additive white Gaussian noise. The useful signal is uncorrelated with the noise signal, and the correlation coefficient is 0, so: The cross-power spectrum density can be obtained by Fourier transforming the cross-correlation function; Step 3, network training: use power spectrum features, time-frequency spectrum features and cross-power spectrum features to train the neural network model; Step 4: Classification and identification: Use the trained neural network model to perform RF fingerprint identification on the test set signal to obtain the classification and identification results.

2. The radio frequency fingerprint recognition method based on multi-task and multi-feature fusion according to claim 1 is characterized in that: The step 3 is as follows: The power spectrum feature, time-frequency spectrum feature and cross-power spectrum feature are used as input to train the multi-feature fusion network model. The multi-feature fusion network model completes the classification task through the Softmax classifier. The output of the Softmax classifier is a set of probability values ​​p, and its mathematical expression is: In the formula, δ i is the output of the last layer of the network model, i is the index of the category, and the total number of categories is C; the output probability satisfies the following conditions: The classification result finally output by the multi-feature fusion network model is the signal transmitting device corresponding to the index with the largest probability value; The network model of multi-task learning is obtained by training the device classification task as the main task of the multi-feature fusion network model and the classification of signal noise content as the second task, that is, the neural network model.

Images