Radio signal intelligent identification method based on incremental learning

Abstract

The invention provides a radio signal intelligent identification method based on incremental learning. The radio signal intelligent identification method comprises the following steps: S100, acquiring a data set; s200, constructing a deep neural network model capable of realizing incremental training; s300, performing incremental training on the deep neural network model by adopting the data set; and S400, applying the trained deep neural network model to radio signal identification. Through fusion of an incremental learning mechanism, rapid and efficient matching of the deep neural network model and the radio signal waveform is realized, and the radio signal waveform identification accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless communications and deep learning, and in particular to a method for intelligently identifying radio signals based on incremental learning. Background Art

[0002] Traditional radio signal waveform recognition methods based on mathematical models require accurate mathematical modeling of radio signals. However, the hardware devices in actual scenarios have non-ideal characteristics, and non-ideal hardware will damage the radio signal waveform. Radio signal waveform recognition methods based on mathematical models are difficult to characterize the impact of hardware non-ideal characteristics on the signal waveform, and their accuracy is limited in actual scenarios.

[0003] Existing radio signal waveform recognition methods based on deep learning need to meet the consistency of training data and test data. In complex and dynamic actual scenarios, the emergence of new environments and new waveforms will cause this method to face the mismatch problem between the deep neural network model and the electromagnetic space, and the efficiency of reconstructing the new adaptive model is low, which is difficult to meet the actual application needs of the target radio signal recognition project. Summary of the invention

[0004] The present invention aims to provide a radio signal intelligent recognition method based on incremental learning to solve the problems existing in the above-mentioned existing methods.

[0005] The present invention provides a method for intelligently identifying radio signals based on incremental learning, comprising the following steps:

[0006] S100, collect data set;

[0007] S200, building deep neural network models capable of incremental training;

[0008] S300, performing incremental training on the deep neural network model using the data set;

[0009] S400, uses the trained deep neural network model for radio signal recognition.

[0010] In some embodiments, in step S100, time domain sequences and frequency domain sequences of radio signals under various signal modulation modes are collected as data sets.

[0011] In some embodiments, the time domain sequence is an IQ sequence, and the frequency domain sequence is an FFT sequence.

[0012] In some embodiments, in step S200, constructing a deep neural network model capable of incremental training includes:

[0013] A deep neural network model that can adapt to incremental training is constructed by adding an output correction layer, and the output port of the deep neural network model during incremental training is increased according to the new type of modulation method.

[0014] In some embodiments, the deep neural network model capable of incremental training includes:

[0015] The convolution module is used to perform real convolution and complex convolution on the time domain sequence and frequency domain sequence respectively to obtain four different feature streams;

[0016] The feature fusion module is used to perform adaptive feature fusion on the four feature streams through the attention mechanism to obtain fused features;

[0017] A feature mapping module is used to map the fused features to the signal modulation mode through a residual module and two hidden layers;

[0018] Output layer, used to output feature mapping results;

[0019] The output correction layer is used to adaptively scale the feature mapping results output by the output layer, and then map the scaled values ​​to (0,1) through the activation function Sigmoid;

[0020] The number of neurons in the output layer and the output correction layer is the sum of the number of old modulation modes and new modulation modes in incremental training; each neuron in the output layer is fully connected to the neurons in the previous layer and is singly connected to the neurons in the output correction layer.

[0021] In some embodiments, the adaptive feature fusion of the four feature streams through the attention mechanism includes: first, performing convolution operations on the time domain features and frequency domain features in the four different feature streams, then calculating the global average of each channel of the feature map after the convolution operation, and calculating the adaptive weights through the Sigmoid function, and finally multiplying the features and the weights to complete efficient feature fusion.

[0022] In some embodiments, in step S300, when incremental training is performed on the deep neural network model using the data set:

[0023] The parameters of the deep neural network model in the current incremental stage are optimized based on the parameters of the deep neural network model in the previous incremental stage;

[0024] The parameters of the deep neural network model are jointly optimized using data from the old and new types of modulation.

[0025] In some embodiments, the loss function for incremental training in step S300 includes a distillation loss for identifying old types of modulation and a poor loss for identifying new types of modulation.

[0026] In some embodiments, each incremental stage of incremental training in step S300 requires a container update so that a portion of the radio signal samples of the current incremental stage stored in the container is used as data of the old modulation method for incremental training in the next incremental stage; wherein the radio signal samples need to be selected based on the quantity and quality of the radio signal samples.

[0027] In some embodiments, when selecting radio signal samples, a certain number of radio signal samples are randomly selected in each modulation mode and put into a container for incremental training in the next incremental stage.

[0028] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0029] The present invention proposes a radio signal intelligent recognition method based on incremental learning. First, the time domain, frequency domain and other features of the radio signal are fully utilized, and the signal feature stream is obtained through real and complex domain convolution operations to realize multi-domain feature refinement extraction, and the feature stream is adaptively fused through the attention mechanism to improve the recognition performance; then, by integrating the incremental learning mechanism, the deep neural network model and the radio signal waveform are matched quickly and efficiently, and the radio signal waveform recognition accuracy is improved. During incremental learning, the parameters of the deep neural network model are jointly optimized by using the old class data and the new class data, and the knowledge learned by the old model is transferred to the new model in a fine-tuning manner to realize rapid model convergence, while retaining the knowledge of the old class data, effectively preventing the problem of catastrophic forgetting; and a container mechanism is introduced to balance the quantity and quality of signal samples, increase sample diversity and improve the recognition performance of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 Schematic diagram of a radio signal intelligent recognition method based on incremental learning in an embodiment of the present invention.

[0031] Figure 2 Schematic diagram of the structure of a deep neural network model capable of incremental training in an embodiment of the present invention.

[0032] Figure 3 The figure is a schematic diagram showing a performance comparison between the radio signal intelligent recognition method based on incremental learning in an embodiment of the present invention and the existing method.

[0033] Figure 4 The figure is a schematic diagram of performance comparison of different loss functions in the radio signal intelligent recognition method based on incremental learning in an embodiment of the present invention.

[0034] Figure 5 The figure is a schematic diagram showing the performance comparison of different sample selection strategies in the radio signal intelligent recognition method based on incremental learning in an embodiment of the present invention.

[0035] Figure 6 The figure is a schematic diagram of performance comparison of the radio signal intelligent recognition method based on incremental learning in an embodiment of the present invention using different sample selection strategies under different signal-to-noise ratios.

[0036] Figure 7 The figure is a schematic diagram of performance comparison of the radio signal intelligent recognition method based on incremental learning in an embodiment of the present invention using different container capacities. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0038] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] like Figure 1 As shown, the embodiment of the present invention proposes a method for intelligently identifying radio signals based on incremental learning, comprising the following steps:

[0040] S100, collect data set;

[0041] S200, building a deep neural network (DNN) model capable of incremental training;

[0042] S300, performing incremental training on the deep neural network model using the data set;

[0043] S400, uses the trained deep neural network model for radio signal recognition.

[0044] In some embodiments, the detection is mainly aimed at non-cooperative target transmitters, which involve the transmission of communication signals and radar signals. Therefore, in step S100, the time domain sequence and frequency domain sequence of the radio signal under each signal modulation mode are collected as a data set. The signal modulation modes considered in the embodiments of the present invention include binary phase shift keying (BPSK), quaternary phase shift keying (QPSK), octal phase shift keying (8PSK), continuous phase frequency shift keying (CPFSK), Gaussian frequency shift keying (GFSK), double-sideband modulation (DSB), single-sideband modulation (SSB), linear frequency modulation (LFM), 16 quadrature amplitude modulation (16QAM), 64QAM, pulse compression modulation (PCM), wideband frequency modulation (WBFM) and fourth generation pulse amplitude modulation (4Pulse Amplitude Modulation). Modulation, PAM4), a total of 13 signal modulation modes. The simulated channel environment is a Rice channel with a Rice factor of 4, and the signal-to-noise ratio range is -5dB to 15dB. Considering the relative motion between the transmitter and the receiver, the maximum Doppler shift can be set to 20kHz.

[0045] In some embodiments, the time domain sequence and frequency domain sequence of the radio signal under each signal modulation mode are specifically as follows:

[0046] The time domain sequence of the radio signal is represented as:

[0047] r(t)=x(t)*h(t)+n(t)

[0048] Among them, x(t) represents the modulated signal transmitted by the transmitter, h(t) represents the impulse response of the channel, and n(t) represents the additive Gaussian noise. After receiving the signal, the receiver converts the signal into a discrete form r[n], which is usually expressed as two sets of orthogonal components, namely:

[0049] r[n]=r I[n]+jr Q [n]

[0050] Where n represents the sampling number, j represents the imaginary unit and j 2 =-1, r I [n] represents the in-phase (I) component, r Q [n] represents the orthogonal (quadrature, Q) component, and thus the time domain sequence of the signal is the IQ sequence.

[0051] The frequency domain sequence of a signal is usually obtained by performing a discrete Fourier transform (DFT) on the signal, which can be expressed as:

[0052]

[0053] Where N is the length of DFT, The DFT algorithm has a fast implementation method, namely, Fast Fourier Transform (FFT). FFT can effectively reduce the computational complexity of frequency domain transformation. Therefore, the frequency domain sequence of the signal is the FFT sequence.

[0054] The present invention uses time domain sequences and frequency domain sequences as data sets, which has the advantages that the IQ sequence is a native signal directly received by the receiver and does not require additional processing; and the FFT algorithm currently has a mature and complete implementation strategy.

[0055] In some embodiments, in step S200, a deep neural network model capable of incremental training is constructed, including: constructing a deep neural network model capable of adapting to incremental training by adding an output correction layer, and adding an output port of the deep neural network model during incremental training according to a new type of modulation method.

[0056] like Figure 2 As shown, the deep neural network model capable of incremental training includes:

[0057] The convolution module is used to perform real convolution and complex convolution on the time domain sequence and frequency domain sequence respectively to obtain four different feature streams;

[0058] The feature fusion module is used to perform adaptive feature fusion on the four feature streams through the attention mechanism to obtain fused features;

[0059] A feature mapping module is used to map the fused features to the signal modulation mode through a residual module and two hidden layers;

[0060] Output layer, used to output feature mapping results;

[0061] The output correction layer is used to adaptively scale the feature mapping results output by the output layer, and then map the scaled values ​​to (0,1) through the activation function Sigmoid;

[0062] The number of neurons in the output layer and the output correction layer is the sum of the number of old modulation modes and new modulation modes in incremental training; each neuron in the output layer is fully connected to the neurons in the previous layer and is singly connected to the neurons in the output correction layer.

[0063] The construction process of the deep neural network model capable of incremental training is as follows:

[0064] S201, build the initial stage of the deep neural network model

[0065] like Figure 2 In the white area shown, the deep neural network model in the initial stage first performs FFT transformation on the received time domain sequence to obtain the frequency domain sequence, then performs real convolution and complex convolution on the time domain sequence and the frequency domain sequence respectively to obtain four different feature streams, and then performs adaptive feature fusion on the four feature streams through the attention mechanism to obtain fusion features to improve recognition performance; then the fusion features are mapped to the signal modulation mode through a residual module and two hidden layers. Among them, the attention mechanism is inspired by the convolution attention mechanism and the channel attention mechanism. The adaptive feature fusion of the four feature streams through the attention mechanism is specifically: first, convolution operation is performed on the time domain features and frequency domain features in the four different feature streams, and then the global average of each channel of the feature map after the convolution operation is calculated, and the adaptive weight is calculated through the Sigmoid function, and finally the feature is multiplied by the weight to complete efficient feature fusion.

[0066] S202, adding an output correction layer to the initial deep neural network model

[0067] Due to the imbalance between the amount of new and old data in the dataset, the deep neural network model in the initial stage cannot be directly applied to incremental training. In order to accelerate the incremental training of new modulation methods, the amount of new modulation method data is usually much larger than the amount of old modulation method data. When calculating the loss, the loss of new modulation method data dominates. The model mainly updates parameters in the direction of minimizing the data loss of new modulation methods, which will cause a deviation in the recognition accuracy of old and new modulation methods. This is an inherent problem in incremental training.

[0068] In order to alleviate the deviation of model performance caused by data imbalance, the embodiment of the present invention adds an output correction layer after the deep neural network model in the initial stage, such as Figure 2The gray area shown. First, remove the Softmax function of the output layer of the deep neural network model in the initial stage, and then add the output correction layer. If there is no output correction layer, the output value of the new type of modulation port will be much larger than the old type of modulation port, because the data imbalance makes the model more inclined to identify the input signal as a new type of modulation. In order to solve this problem, the output correction layer first adaptively scales the feature mapping results output by the output layer, and then further maps the scaled value to (0,1) through the activation function Sigmoid. The output correction layer can be expressed as:

[0069] y i =Sigmoid(w i *x i +b i )

[0070] Among them, x i represents the value of the i-th port of the output layer, w i and b i are the parameters of the i-th port of the output rectification layer, both of which are trainable.

[0071] S203, adding an output port of the deep neural network model according to the new type of modulation method

[0072] In order to adapt to the recognition task of new types of modulation, the structure of the deep neural network model needs to be further modified. For the convenience of description, the number of old and new modulation modes in the mth incremental stage is denoted as S m and T m First, we need to add T in the output layer m Correspondingly, T also needs to be increased in the output correction layer. m neurons and use Sigmoid activation function, such as Figure 2 The black points shown. In this way, both the output layer and the output rectification layer have S m +T m neurons, among which S m belongs to the old type of modulation, T m Each neuron in the output layer is fully connected to the neurons in the previous layer and has a single connection to the neurons in the output rectification layer.

[0073] In some embodiments, in step S300, when the data set is used to perform incremental training on the deep neural network model, there are two main design ideas:

[0074] (1) Since part of the deep neural network model in the current incremental stage (referred to as the new model) is the structure of the deep neural network model in the previous incremental stage (referred to as the old model), the knowledge learned by the old model can be transferred to the new model. Therefore, the parameters of the deep neural network model in the current incremental stage are optimized based on the parameters of the deep neural network model in the previous incremental stage, which can help the new model converge quickly.

[0075] (2) The parameters of the deep neural network model are optimized jointly using the data of the old and new modulation methods, so that the knowledge of the old modulation methods can be retained while learning the knowledge of the new modulation methods. This can effectively prevent the problem of catastrophic forgetting.

[0076] Note m =[Θ m,1 ,Θ m,2 ], where Θ m,1 represents the parameters before the output port of the deep neural network model is increased in the mth incremental stage, Θ m,2 represents the newly added parameters. Since the parameters before the addition of the mth incremental stage are the same as the parameters after convergence in the m-1th incremental stage, we can get Θ m,1 =Θ m-1 To train Θ m , where Θ m,1 and θ m,2 The training is performed alternately until both parts converge. In this embodiment, an epoch is defined as training once Θ m,1 or Θ m,2 The process, the epoch number is recorded as k, Θ m,1 and θ m,2 The initial value at the kth increment stage is recorded as and In the mth stage, multiple epochs are needed to train the parameters. m,1 and θ m,2 The convergence values ​​of and

[0077] In the first epoch of the mth incremental phase (k=1), the knowledge from the (m-1)th incremental phase is transferred to initialize Θ m,1 ,Right now at the same time, Randomly initialize. Then, from the dataset D m Randomly select Radio signal samples form a mini-batch of data in and Represents D in the kth epoch of the mth incremental stage m,old and D m,new The loss function designed in this embodiment is:

[0078]

[0079] The first term on the right side of the equal sign represents the distillation loss for identifying the old type of modulation, which can be expressed as:

[0080]

[0081] The second term on the right side of the equal sign represents the poorer loss of identifying new types of modulation, which can be expressed as:

[0082]

[0083] Among them, X i represents the i-th modulation signal in the data set, Y i is the modulation mode number corresponding to it. i =y, The value is 1 if yes, otherwise it is 0. It is represented as the value of the yth output correction layer port in the kth epoch of the mth incremental stage. Represents the output value of the converged model in the (m-1)th incremental stage.

[0084] Afterwards freeze, Update in the direction of minimizing loss, that is:

[0085]

[0086] Among them, η represents the learning rate, Represents the loss function The partial derivative of .

[0087] In the second epoch (k=2), let Then recalculate the loss and update it in the direction of minimizing the loss It can be expressed as:

[0088]

[0089] When k>2, update according to whether k is odd or even. Incrementally train the two parts of parameters until convergence, denoted as and And remember

[0090] It is worth noting that the loss function used in this embodiment is different from the existing distillation loss. The existing distillation loss is expressed as:

[0091]

[0092] By comparison, it can be found that the existing loss function ignores a basic fact: the model trained in the previous stage has not learned the knowledge about the new type of modulation. Using the old model to calculate the loss of identifying the new type of modulation may cause serious errors and lead to deviations in the training direction. On the contrary, the loss function proposed in the embodiment of the present invention calculates the old and new type of modulation signals respectively, which can accurately reflect the real loss, thereby improving the performance of the incremental training stage.

[0093] In some embodiments, in order to alleviate the problem of catastrophic forgetting in incremental training, that is, the deep neural network model will forget the learned knowledge when learning new knowledge, the container mechanism is widely used in incremental training. Due to the limited data storage capacity, only part of the radio signal samples in the current incremental stage can be saved as old data for the next incremental stage. That is to say, each incremental stage of incremental training requires a container update, so that part of the radio signal samples in the current incremental stage stored in the container is used as data of the old modulation mode for incremental training in the next incremental stage; wherein, the radio signal samples need to be selected according to the quantity and quality of the radio signal samples to ensure the performance of each incremental stage. At the same time, the total amount of radio signal samples stored in the container should not exceed the upper limit of the container capacity.

[0094] It is easy to ensure that the number of stored radio signal samples does not exceed the upper limit of the container capacity, but it is difficult to ensure the quality of the radio signal samples. In the embodiment of the present invention, a random selection method is adopted, that is, when selecting radio signal samples, a certain number of radio signal samples are randomly selected in each modulation mode and put into the container for incremental training in the next incremental stage. This method believes that the diversity of data has a greater impact on recognition performance.

[0095] In step S400, when the trained deep neural network model is used for radio signal recognition, the receiver inputs the received radio signal into the trained deep neural network model to obtain the calculated probability result of modulation perception, and the index with the highest calculated probability result is used as the modulation mode corresponding to the received radio signal, which is expressed as:

[0096]

[0097] in, is the output value of the i-th port of the output correction layer.

[0098] The practicality of the present invention is illustrated below with the help of simulation examples and drawings. First, the data set parameters and system model parameter settings are provided, and then the simulation results are provided to evaluate the performance of the present invention. At the same time, the method of the present invention is compared with the existing methods, and the influence of different sampling points on the perception accuracy of the present invention is analyzed.

[0099] The modulation modes considered in this example include BPSK, QPSK, 8PSK, CPFSK, GFSK, DSB, SSB, WBFM, QAM16, QAM64, LFM, PCM and PAM4. The baseband bandwidth is 100MHz, the carrier frequency is 2.4GHz, and the upsampling factor is 8. The channel environment is a Rice channel with a Rice factor of 4, and the simulated signal-to-noise ratio range is -5dB to 15dB. Considering the relative motion between the transmitter and the receiver, the simulated maximum Doppler frequency deviation is 20KHz.

[0100] The parameters of the deep neural network model in this example are set as follows. Figure 2 As shown in the overall structure, in the convolution module, a complex convolution consists of two real convolution layers, where the size of the real convolution kernel is (1, 3), Relu is used as the activation function, and the output channels are 16. The convolution kernel size of the channel dimension convolution in the feature fusion module is (1, 3), and the number of output channels of the 1×1 convolution layer is 16. In the above convolution operations, except for the 1×1 convolution, the rest use zero padding operations to keep the dimension of the feature map constant. After passing through the feature fusion module, the feature vector obtained by the above convolution layer is converted into a one-dimensional vector. The length of the hidden layer in the fully connected layer is 2, the number of neurons is 128 and 64 respectively, Selu is used as the activation function, the number of neurons and the type of modulation mode of the output layer and the output correction layer are consistent, the output layer has no activation function, and the activation function of the output correction layer is Sigmoid. The optimizer used is Adam, the learning rate is 0.001, and the attenuation coefficients of the first-order and second-order moment estimation are 0.9 and 0.999 respectively. At each signal-to-noise ratio, this example collects 2000 frames of radio signals under each modulation mode, and each frame has 128 sampling points. In the dataset, the ratio of the training set to the test set is 7:3.

[0101] Figure 3 The recognition performance curves of different algorithms are shown, with the full-data and no-data methods as upper and lower bounds, respectively. Figure 3It can be seen that the performance curve of the method proposed in the present invention has the same trend as full data learning and is very close, and is higher than other algorithms. As the number of modulation modes increases, the recognition accuracy first increases and then decreases. This is because if the new type of modulation mode is easier to identify, that is, the recognition accuracy of the new type of modulation mode is higher than that of the old type of modulation mode, the average recognition accuracy will increase. On the contrary, if the accuracy of the new type of modulation mode is lower than that of the old type of modulation mode, the average recognition accuracy will be reduced. Specifically, under the set simulation environment, the accuracy of the method proposed in the present invention is less than 5% higher than the upper bound. At the same time, the accuracy of the method proposed in the present invention is about 9% higher than that of the iCaRL algorithm and 7% higher than that of the BiC algorithm.

[0102] Figure 4 The performance curves of different loss functions are shown. As described in the above embodiments, the present invention adopts a new loss function. Figure 4 It can be seen that the algorithm using the new loss function outperforms the cross entropy loss and knowledge distillation loss functions in overall performance, which proves the advantage of the proposed loss function. In particular, as the number of modulation methods increases, the performance gap becomes more and more obvious. This is because as the number of modulation methods increases, the recognition task becomes more and more difficult, and failure to distinguish the data will lead to deviations in the direction of model optimization.

[0103] Figure 5 The recognition performance curves of different selection strategies are shown. Figure 6 The recognition performance curves of different selection strategies under different signal-to-noise ratios are shown, including random selection, clustering algorithm, and selection strategy based on signal-to-noise ratio. Figure 5 In the above analysis, the overall accuracy of random selection is higher than that of the other two selection strategies. Specifically, clustering algorithms emphasize the clustering of features rather than their diversity. In actual scenarios, the high-speed dynamic radio environment leads to the dispersion of signal sample features. In this case, signal diversity is more important than feature clustering. The method based on signal-to-noise ratio also has the same problem. Figure 6 It can be verified in the results that this method performs well only in the high signal-to-noise ratio area.

[0104] Figure 7 The performance impact of different container storage sizes is shown, where the storage size is 2, 1, 0.5, 0.3, and 0.1 times the size of the 29400 training samples. Again, full data learning and no data learning are used as upper and lower bounds. Figure 7It can be seen that as the number of stored samples increases, the performance of the modulation recognition algorithm based on IL will also increase, proving that storing more historical samples can effectively solve the problem of catastrophic forgetting. It is worth noting that even with only 2940 frames (0.1 times) of samples stored, the accuracy of identifying 13 modulation modes still exceeds 70%, which is much higher than the data-free learning method.

[0105] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A radio signal intelligent recognition method based on incremental learning, characterized in that: The steps include: S100, collect data set; S200, building deep neural network models capable of incremental training; S300, performing incremental training on the deep neural network model using the data set; S400, uses the trained deep neural network model for radio signal recognition.

2. The method for intelligent identification of radio signals based on incremental learning according to claim 1, characterized in that: In step S100, time domain sequences and frequency domain sequences of radio signals under various signal modulation modes are collected as data sets.

3. The method for intelligent identification of radio signals based on incremental learning according to claim 2, characterized in that: The time domain sequence is an IQ sequence, and the frequency domain sequence is an FFT sequence.

4. The method for intelligent identification of radio signals based on incremental learning according to claim 2, characterized in that: In step S200, a deep neural network model capable of incremental training is constructed, including: A deep neural network model that can adapt to incremental training is constructed by adding an output correction layer, and the output port of the deep neural network model during incremental training is increased according to the new type of modulation method.

5. The method for intelligent identification of radio signals based on incremental learning according to claim 4, characterized in that: The deep neural network model capable of incremental training includes: The convolution module is used to perform real convolution and complex convolution on the time domain sequence and frequency domain sequence respectively to obtain four different feature streams; The feature fusion module is used to perform adaptive feature fusion on the four feature streams through the attention mechanism to obtain fused features; A feature mapping module is used to map the fused features to the signal modulation mode through a residual module and two hidden layers; Output layer, used to output feature mapping results; The output correction layer is used to adaptively scale the feature mapping results output by the output layer, and then map the scaled values ​​to (0,1) through the activation function Sigmoid; The number of neurons in the output layer and the output correction layer is the sum of the number of old modulation modes and new modulation modes in incremental training; each neuron in the output layer is fully connected to the neurons in the previous layer and is singly connected to the neurons in the output correction layer.

6. The method for intelligent identification of radio signals based on incremental learning according to claim 5, characterized in that: The adaptive feature fusion of the four feature streams through the attention mechanism includes: firstly performing convolution operation on the time domain features and frequency domain features in the four different feature streams, then calculating the global average of each channel of the feature map after the convolution operation, and calculating the adaptive weight through the Sigmoid function, and finally multiplying the features and the weights to complete efficient feature fusion.

7. The method for intelligent identification of radio signals based on incremental learning according to claim 5, characterized in that: In step S300, when the deep neural network model is incrementally trained using the data set: The parameters of the deep neural network model in the current incremental stage are optimized based on the parameters of the deep neural network model in the previous incremental stage; The parameters of the deep neural network model are jointly optimized using data from the old and new types of modulation.

8. The method for intelligent identification of radio signals based on incremental learning according to claim 7, characterized in that: The loss function for incremental training in step S300 includes a distillation loss for identifying old types of modulation and a poor loss for identifying new types of modulation.

9. The method for intelligent identification of radio signals based on incremental learning according to claim 7, characterized in that: Each incremental stage of incremental training in step S300 requires a container update so that a portion of the radio signal samples of the current incremental stage stored in the container is used as data of the old modulation method for incremental training in the next incremental stage; wherein the radio signal samples need to be selected based on the quantity and quality of the radio signal samples.

10. The method for intelligent identification of radio signals based on incremental learning according to claim 9, characterized in that: When selecting radio signal samples, a certain number of radio signal samples are randomly selected in each modulation mode and put into a container for incremental training in the next incremental stage.

Images