Communication modulation identification method based on multiple scales

Abstract

The invention provides a communication modulation identification method based on multiple scales, and the method mainly comprises the steps: signal collection, signal preprocessing, data conversion, signal splicing, 2D network feature extraction, 3D multi-scale feature extraction, and feature information processing, and finally achieves modulation classification. Compared with a traditional method, the method has the advantages that the recognition capability of a complex modulation mode is improved through fusion processing; according to the method, the convolutional neural network and the improved LSTM are used for feature extraction, key features can be automatically extracted from input signals, the tedious process of manual feature design is avoided, and the recognition efficiency is improved; through the multi-scale feature extraction method, the recognition rate of the method can be improved, the feature information of the signals can be extracted more comprehensively, it is guaranteed that edge information is not lost, and therefore high robustness is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal analysis, and in particular relates to a communication modulation identification method. Background Art

[0002] With the rapid development of wireless communication technology, modulation recognition in communication systems has become a significant research area. Modulation recognition is a key step in signal processing in wireless communication systems and is crucial for applications such as signal demodulation, spectrum management, and radio interference cancellation. In complex wireless communication environments, signals can be affected by a variety of factors, such as noise, fading, and multipath. Traditional modulation recognition methods often suffer from low accuracy and insufficient generalization capabilities in these complex environments.

[0003] Existing modulation recognition methods have shifted from traditional manual feature extraction or high-order cumulants to automatic feature extraction and recognition using deep learning methods, as shown in Zhang Wenqi. Research on Radio Signal Modulation Recognition Methods Based on Machine Learning. Hainan University, 2022. However, existing methods generally only utilize the feature information of a certain domain of the signal, ignoring the potential correlation between signals in different domains. Summary of the Invention

[0004] In order to overcome the problems existing in the prior art, the present invention proposes a multi-scale communication modulation recognition method. The basic idea is: adopt a method of preprocessing the time domain signal and converting it into the frequency domain and the square of the time domain, making full use of the characteristics that different domains of the signal carry different information; in addition, according to the designed neural network model, the signal is feature extracted using different convolutional layer receptive fields, and the extracted feature information is fused to obtain more comprehensive feature information of the signal.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] A multi-scale communication modulation recognition method, comprising:

[0007] Step S1: signal acquisition, i.e., signal acquisition is performed through a signal receiving device to obtain a time domain signal;

[0008] Step S2: signal preprocessing, i.e. performing threshold detection and filtering on the collected time domain signal to obtain a time domain filtered signal;

[0009] Step S3: Data transformation, i.e., transforming the time domain filtered signal to obtain frequency domain data and time domain squared data;

[0010] Step S4: signal splicing, that is, splicing the collected time domain filtered signal, the processed frequency domain data and the processed time domain squared data to obtain the network input signal;

[0011] Step S5: 2D network feature extraction, that is, inputting the network input signal into the 2D convolutional neural network to extract initial feature information;

[0012] Step S6: 3D multi-scale feature extraction, that is, the initial feature information is further processed through a 3D multi-scale convolutional neural network to obtain small-scale feature information and large-scale feature information;

[0013] Step S7: Feature information processing, that is, concatenating the obtained small-scale feature information and large-scale feature information, flattening them, and then putting them into the LSTM network to extract time-related features and obtain the classification feature vector;

[0014] Step S8: Modulation classification, that is, sending the classification feature vector to the fully connected layer, and then classifying it through the Softmax layer, and outputting the network's probability distribution for different modulation types.

[0015] Further,

[0016] In step S1, the collected signal is stored in the form of time domain data to obtain a time domain signal x(t);

[0017] In step S2, the time domain signal x(t) is first subjected to energy detection, and the noise signal in the signal acquisition device area is subjected to energy detection to obtain an energy threshold T0. The time domain signal x(t) is then compared with the threshold T0. If the energy of the time domain signal x(t) is greater than the threshold T0, the signal in this part is the valid signal part, and the valid signal part is retained, and the invalid signal is filtered out. The upper limit f1 and lower limit f2 of the time domain signal transmission frequency are used as the passband frequency of the bandpass filter, and the valid signal part is bandpass filtered to obtain the filtered time domain filtered signal x'(t).

[0018] Further,

[0019] The energy threshold T0 in step S2 is:

[0020]

[0021] Where N is the number of samples, M represents the collected noise signal, represents the N-point Fourier transform of M, and the square of the absolute value represents the vector norm, which is used to calculate the energy of the noise signal M sampled at N points;

[0022] Step S3 includes:

[0023] S3.1: Frequency domain transformation, that is, Fourier transform the time domain filtered signal x'(t) to obtain its frequency domain data f(t)

[0024] S3.2: Time domain square transform, that is, square the time domain filtered signal x'(t) to obtain the squared time domain square data x 2 (t).

[0025] Further,

[0026] In step S4, the network input signal S(t) is obtained by splicing;

[0027] S(t)=[x'(t),f(t),x 2 (t)].

[0028] Further,

[0029] In step S5, the convolution kernel of the 2D convolutional neural network is used to extract local features through a sliding window to obtain the initial feature information F1. The convolution layer parameters of the 2D convolutional neural network are {kernel=k=3, padding=p=1, stride=s=1}, and the convolution kernel size is a long sequence K of 1×3. Convolution is used to extract features from S(t), and the features obtained are:

[0030]

[0031] Where K[m] is the weight of the convolution kernel, i and j represent the index of the row and column after convolution respectively.

[0032] Corresponding output size H out and W out , whose values ​​are:

[0033]

[0034] Where s is the convolution stride size, which represents the step size that the convolution kernel moves on the input, H0 represents the length of S(t), and W0 represents the width of S(t).

[0035] Further,

[0036] In step S6, the 3D multi-scale convolutional neural network has two sub-networks, and the convolution kernel sizes of the two sub-networks are 1×3 and 1×5 respectively. The initial feature information F1 can be converted into small-scale feature information F2 and large-scale feature information F3 through these two sub-networks.

[0037] Further,

[0038] In step S6, the number of channels of the convolution kernel of the 3D multi-scale convolutional neural network is c=64, and the small-scale feature information F2 is extracted using the 3D multi-scale convolutional neural network. Its convolution layer parameters are {kernel=k=3, padding=p=1, stride=s=1}, and the convolution kernel size is a long sequence K of 1×3; the initial feature information F1 is extracted using convolution, and the obtained features are:

[0039]

[0040] Where K[m] is the weight of the convolution kernel, i and j represent the index of the row and column after convolution respectively.

[0041] Corresponding output size H out and W out , whose values ​​are:

[0042]

[0043] Where s is the convolution stride size, which indicates the step size of the convolution kernel moving on the input, H0 indicates the length of the initial feature information F1, and W0 indicates the width of the initial feature information F1;

[0044] The large-scale feature information F3 is extracted using a 3D multi-scale convolutional neural network. The convolution layer parameters are {kernel = k = 5, padding = p = 1, stride = s = 1}, and the convolution kernel size is a long sequence K of 1 × 5. The initial feature information F1 is extracted using convolution, and the features obtained are:

[0045]

[0046] Where K[m] is the weight of the convolution kernel, i and j represent the index of the row and column after convolution respectively.

[0047] Corresponding output size H out and W out , whose values ​​are:

[0048]

[0049] Where s is the convolution stride size, which indicates the step size of the convolution kernel moving on the input, H0 indicates the length of the initial feature information F1, and W0 indicates the width of the initial feature information F1.

[0050] Further,

[0051] In step S7, the small-scale feature information F2 and the large-scale feature information F3 after convolution calculation are spliced ​​to obtain multi-scale feature information F4, and then the multi-scale feature information F4 is tiled through the Fallten layer to obtain a tiled feature vector F5. Subsequently, the tiled feature vector F5 is input into the LSTM network, and the classification feature vector F6 is obtained through the functions of the forget gate, input gate and output gate of the LSTM.

[0052] Further,

[0053] In step S7, the flattened feature vector F5 is truncated to the length of the LSTM hidden unit γ = 128 and used as the input Q of the LSTM network t , Q t Put it into the forget gate of LSTM, the forget gate formula is:

[0054] L t =σ(W f ×[H t-1 ,Q t ]+b f )

[0055] Among them L t is the information controlled by the forget gate, H t-1 is the hidden information of the LSTM network, σ is the Sigmod formula, W f is the weight of the forget gate, b f is the bias,

[0056] For the input gate in LSTM, it is expressed as:

[0057] L i =σ(W i ×[H t-1 ,Q t ]+b i )

[0058] Among them L i is the information of input gate control, W i is the weight of the input gate, b i is the bias,

[0059] According to the last output and this input, we can get:

[0060] C ′ =tanh(W C ×[H t-1 ,Q t ]+b C )

[0061] Among them, C ′ is the current input unit state, W Cis the weight of the unit state, b C is the bias,

[0062] Then pass the previous unit state C t-1 and the current input cell state C ′ , multiplied and added with the results of the forget gate and input gate respectively, to obtain the current cell state C t ,

[0063] C t =L t ×C t-1 +L i ×C ′

[0064] Then get the output according to the output gate of LSTM. The formula is as follows:

[0065] o t =σ(W o ×[H t-1 ,Q t ]+b o )

[0066] h t =o t tanh(C t )

[0067] Among them, t Indicates the information determined by the output gate, W o represents the weight of the output gate, b o is the bias, h t is the output of the LSTM module, namely the classification feature vector F6.

[0068] Further,

[0069] The model of the signal receiving device in step S1 is USRP_B210; the model of the computer equipment used in steps S2, S3, S4, S5, S6, S7, and S8 is: the GPU model is GeForce GTX1050 Ti, the CPU model is Intel(R) Core(TM) i7-8750H CPU @ 2.20GHz, and the software used is MATLAB R2022a; when sending and receiving signals, the environment is an open room with a length, width, and height of 20 meters, 3 meters, and 3 meters, respectively. Except for the computer and the two USRP devices, there are no other facilities or equipment. The two USRP devices, i.e., the sending device and the receiving device, are 3 meters apart.

[0070] In step S1, the length of the x(t) signal is 1×1000, which is a CLOVER-2000-BPSM time domain signal.

[0071] Compared with the prior art, the present invention has the following beneficial effects:

[0072] 1. Multi-class feature fusion: This invention applies multiple preprocessing methods to the signal in the time domain, including time domain waveform, frequency domain data, and time domain squared data, to obtain multiple signal representations. By fusing these features, the ability to recognize complex modulation schemes is improved.

[0073] 2. Automated multi-scale feature extraction: We designed convolutional neural sub-networks with different receptive fields and used convolutional neural networks and improved LSTM for feature extraction. This method can automatically extract key features from the input signal, avoiding the tedious process of manually designing features and improving recognition efficiency.

[0074] 3. Strong robustness: The multi-scale feature extraction method can not only improve the recognition rate of the method, but also more comprehensively extract the characteristic information of the signal, ensuring that edge information is not lost. Therefore, the present invention has strong robustness and can maintain high recognition accuracy in various environments, especially under low signal-to-noise ratio conditions, and can still stably extract effective features. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 Flow chart of the steps of the method of the present invention;

[0076] Figure 2 Schematic diagram of time domain signal of shortwave protocol CLOVER-2000-BPSM in an embodiment of the present invention;

[0077] Figure 3 Schematic diagram of the frequency domain signal of the shortwave protocol CLOVER-2000-BPSM in an embodiment of the present invention;

[0078] Figure 4 Schematic diagram of the time domain square signal of the shortwave protocol CLOVER-2000-BPSM in an embodiment of the present invention;

[0079] Figure 5 Schematic diagram of the convolutional neural network in the present invention;

[0080] Figure 6 This is a schematic diagram of the subcarrier modulation identification results in the shortwave protocol CLOVER-2000;

[0081] Figure 7 This is a time domain diagram of one of the signals in the public dataset RML2016.a;

[0082] Figure 8 This figure is a comparison chart with the results of other experiments using public datasets. DETAILED DESCRIPTION

[0083] In order to make the objectives, technical methods and advantages of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0084] In the hardware environment used in this embodiment, the signal transmitting device model is USRP_B210, which will be referred to as the transmitting device in subsequent embodiments. The signal receiving device model is also USRP_B210, which will be referred to as the receiving device in subsequent embodiments. The computer equipment used is: the GPU model is GeForce GTX1050 Ti, the CPU model is Intel(R) Core(TM) i7-8750H CPU @ 2.20GHz, and the software used is MATLAB R2022a. When sending and receiving signals, the environment is an open room with a length, width and height of 20 meters, 3 meters, and 3 meters respectively. Aside from the computer and two USRP devices, there are no other facilities or equipment. The two USRPs, namely the transmitting device and the receiving device, are 3 meters apart.

[0085] Figure 1 The figure is a flow chart of the steps of the method of the present invention.

[0086] The multi-scale communication modulation recognition method described in this embodiment includes the following steps:

[0087] S1: Signal acquisition, that is, signal acquisition through signal acquisition equipment.

[0088] The receiving device is a receiver used in a wireless communication system, capable of capturing the communication signal from the transmitter in real time. The parameters for sending and receiving signals, as well as the parameters for the sending and receiving devices, are set according to the shortwave protocols CLVER-2000, 3GALE, and 2GALE. The collected signal is stored in the form of time domain data to obtain the signal time domain data x(t), where x(t) represents the original time domain signal, and the signal length is 1×1000, as shown in the following example. Figure 2 Shown is the CLOVER-2000-BPSM time domain signal.

[0089] S2: Signal preprocessing, that is, threshold detection and filtering of the collected time domain signal.

[0090] First, the time-domain signal x(t) is energy-checked, and the noise signal within the signal acquisition device area is also energy-checked to obtain an energy threshold, T0. The time-domain signal x(t) is then compared with threshold T0. If the energy of the time-domain signal x(t) exceeds threshold T0, the signal in that portion is considered valid and retained, while the invalid signal is filtered out. Communication signals typically transmit within a specific frequency band. Therefore, the upper limit f1 and lower limit f2 of the time-domain signal transmission frequency are used as the passband frequencies of the bandpass filter. The valid signal portion is bandpass filtered to obtain the filtered time-domain filtered signal x'(t) to eliminate the effects of interference and noise.

[0091] Perform energy detection on the noise signal in the receiving device area and obtain the energy threshold T0 as:

[0092]

[0093] Where N is the number of samples, M represents the collected noise signal, It represents the N-point Fourier transform of M, and the square of the absolute value represents the vector norm, which is used to calculate the energy of the noise signal M sampled at N points.

[0094] S3: Data transformation, that is, transforming the time domain filtered signal x'(t) to obtain frequency domain data and time domain square data.

[0095] Step S3 includes:

[0096] S3.1: Frequency domain transformation, that is, the time domain filtered signal x'(t) is subjected to Fourier transformation to obtain its frequency domain data f(t). The purpose of frequency domain data is mainly to distinguish different frequencies of the signal, such as Figure 3 is the CLOVER-2000-BPSM time domain signal;

[0097] S3.2: Time domain square transform, that is, square the time domain filtered signal x'(t) to obtain the squared time domain square data x 2 (t), the purpose of time domain squaring is mainly to reduce the influence of high-order modulation. For the same type of modulation, high-order modulation and low-order modulation are easily confused. The main reason is that the same type of high-order modulation often contains low-order modulation, which will lead to a decrease in recognition rate. After squaring it, the high-order signal has a reduced order, so that the network can better extract feature information and identify the modulation mode of the signal. The results are shown in the figure. Figure 4 As shown, this is the time domain signal of CLOVER-2000-BPSM.

[0098] S4: Signal splicing.

[0099] The collected time domain filtered signal, the processed frequency domain data and the processed time domain squared data are spliced ​​together to obtain the network input signal S(t);

[0100] S(t)=[x'(t),f(t),x 2 (t)].

[0101] The network input signal provides the basis for subsequent feature extraction.

[0102] S5: 2D network feature extraction, that is, inputting the network input signal into the 2D convolutional neural network to extract the initial feature information. The deep learning network in the present invention is a convolutional neural network, such as Figure 5 As shown, it includes: 2D convolutional neural network and 3D multi-scale convolutional neural network.

[0103] The network input signal S(t) is input into the 2D convolutional neural network for convolution processing, that is, Figure 5 The convolution kernel of the 2D convolutional neural network extracts local features through a sliding window to obtain the initial feature information F1.

[0104] A 2D convolutional neural network is used for feature extraction. The convolution layer parameters are {kernel = k = 3, padding = p = 1, stride = s = 1}, and the convolution kernel size is a long sequence K of 1 × 3. Convolution is used to extract features from S(t), and the features obtained are:

[0105]

[0106] Where K[m] is the weight of the convolution kernel, i and j represent the index of the row and column after convolution respectively.

[0107] Corresponding output size H out and W out , whose values ​​are:

[0108]

[0109] Where s is the convolution stride size, which represents the step size that the convolution kernel moves on the input. H0 represents the length of S(t) and W0 represents the width of S(t).

[0110] S6: 3D multi-scale feature extraction.

[0111] The initial feature information F1 is then passed through a 3D multi-scale convolutional neural network. The number of channels of the convolution kernel of the 3D multi-scale convolutional neural network is c = 64, and feature information is extracted from multiple scales and dimensions. The 3D multi-scale convolutional neural network in the present invention has two sub-networks. The convolution kernel sizes of the two sub-networks are 1×3 and 1×5 respectively. The initial feature information F1 passes through these two sub-networks, and small-scale feature information F2 and large-scale feature information F3 can be obtained respectively. That is, Figure 5 The specific parameters of the medium-scale convolutional network are shown in Table 1.

[0112] Table 1

[0113]

[0114] By increasing the number of channels of the convolution kernel, the network can establish associations between different signal representations, further improving the accuracy and comprehensiveness of feature extraction.

[0115] For small-scale feature information F2:

[0116] A 3D multi-scale convolutional neural network is used for feature extraction. The convolution layer parameters are {kernel = k = 3, padding = p = 1, stride = s = 1}, and the convolution kernel size is a long sequence K of 1 × 3. Convolution is used to extract features from the initial feature information F1, and the features obtained are:

[0117]

[0118] Where K[m] is the weight of the convolution kernel, i and j represent the index of the row and column after convolution respectively.

[0119] Corresponding output size H out and W out , whose values ​​are:

[0120]

[0121]

[0122] Where s is the convolution stride size, which represents the step size of the convolution kernel moving on the input. H0 represents the length of the initial feature information F1, and W0 represents the width of the initial feature information F1.

[0123] As for the large-scale feature information F3:

[0124] A 3D multi-scale convolutional neural network is used for feature extraction. The convolution layer parameters are {kernel = k = 5, padding = p = 1, stride = s = 1}, and the convolution kernel size is a long sequence K of 1 × 5. Convolution is used to extract features from the initial feature information F1, and the features obtained are:

[0125]

[0126] Where K[m] is the weight of the convolution kernel, i and j represent the index of the row and column after convolution respectively.

[0127] Corresponding output size H out and W out , whose values ​​are:

[0128]

[0129] Where s is the convolution stride size, which represents the step size of the convolution kernel moving on the input. H0 represents the length of the initial feature information F1, and W0 represents the width of the initial feature information F1.

[0130] S7: Feature information processing, that is, concatenating the obtained small-scale feature information F2 and large-scale feature information F3, flattening them, and then putting them into the LSTM network to extract time-related features.

[0131] The small-scale feature information F2 and the large-scale feature information F3, after convolution, are concatenated to obtain multi-scale feature information F4. This is then tiled through a Fallten layer to produce a tiled feature vector F5. This tiled feature vector F5 is then input into an LSTM network. The LSTM (Long Short-Term Memory) network, leveraging its time series modeling capabilities and the contextual connections between neurons, captures the temporal characteristics of the signal and extracts dynamic information closely related to the signal modulation method. Based on the LSTM input, tiled feature vector F5, the classification feature vector F6 is generated through the LSTM's forget gate, input gate, and output gate.

[0132] Flatten is used to tile the feature map of c = 64 channels, and the resulting tiled feature vector F5 is:

[0133] F3=[F i1 ,F i2 ,...,F ic ]

[0134] Among them, F i1 Represents the feature information of the first channel, F i2 Represents the feature information of the second channel, and so on, F ic Represents the feature information of the c=64th channel.

[0135] LSTM processes the sequence by partially processing the sequence length. For the input gate in LSTM, the sequence length is matched with the number of hidden units of LSTM and used as the input of LSTM. After the complete LSTM model, another sequence with the same length is added to update the information of the entire sequence. The flattened feature vector F5 is truncated to the length of the number of hidden units of LSTM γ = 128 and used as the input Q of LSTM network. t , that is, the input at time t is Q. t Put it into the forget gate of LSTM, which determines which past information is forgotten. The forget gate formula is:

[0136] L t =σ(W f ×[H t-1 ,Q t ]+b f )

[0137] Among them L t is the information controlled by the forget gate, H t-1 is the hidden information of the LSTM network, σ is the Sigmod formula, W f is the weight of the forget gate, b f For bias.

[0138] For the input gate in LSTM, it can be expressed as:

[0139] L i =σ(W i ×[H t-1 ,Q t ]+b i )

[0140] Among them L i is the information of input gate control, W i is the weight of the input gate, b i For bias.

[0141] According to the last output and this input, we can get:

[0142] C ′ =tanh(W C ×[H t-1 ,Q t ]+b C )

[0143] Among them, C ′ is the current input unit state, W C is the weight of the unit state, b C For bias.

[0144] Then pass the previous unit state C t-1 and the current input cell state C′ , multiplied and added with the results of the forget gate and input gate respectively, to obtain the current cell state C t .

[0145] C t =L t ×C t-1 +L i ×C ′

[0146] Then get the output according to the output gate of LSTM. The formula is as follows:

[0147] o t =σ(W o ×[H t-1 ,Q t ]+b o )

[0148] h t =o t tanh(C t )

[0149] Among them, t Indicates the information determined by the output gate, W o represents the weight of the output gate, b o is the bias. t is the output of the LSTM module, namely the classification feature vector F6.

[0150] The weight matrices of the input, forget, and output gates are typically assigned values ​​using random initialization. In this embodiment, the biases of the input and output gates are set to 0, while the bias of the forget gate is set to 1. Backpropagation is used to calculate the gradients, and after each iteration, the model dynamically adjusts the weights and biases based on the training data and loss value. This is one of the advantages of using LSTM, as it can automatically update the weights.

[0151] S8: Modulation classification, that is, sending the classification feature vector F6 to the fully connected layer FC, how to classify it through the Softmax layer, and output the network's probability distribution for different modulation types. The Softmax layer can map the feature information to the classification space of each modulation method, and calculate the possibility of each modulation type based on the pattern of the feature information. By maximizing the output probability, the modulation method of the input signal is accurately identified. The classification result is as follows Figure 6 As shown in the figure, it can be seen that for the three shortwave protocols, using the method of the present invention, the training curve at 0dB can achieve a recognition rate close to 99.99%, and also has very good performance on the verification set, that is, the method can achieve relatively accurate recognition of the three shortwave protocols.

[0152] Example 2

[0153] In order to better verify the method of the present invention, the following experimental conditions were constructed in this example:

[0154] In the hardware environment used in this embodiment, the signal transmitting device is a USRP_B210, i.e., the transmitting device, and the signal receiving device is also a USRP_B210, i.e., the receiving device. The two computers used both have a GeForce GTX1050 Ti GPU and an Intel(R) Core(TM) i7-8750H CPU @ 2.20GHz. The software used is MATLAB R2022a. Signal transmission and reception are performed in an open room with a length, width, and height of 20 meters, 3 meters, and 3 meters, respectively. Aside from the two computers and the two USRPs, there are no other facilities or equipment. The two USRPs, i.e., the transmitting and receiving devices, are 3 meters apart.

[0155] Using the public dataset RML2016.a, we loaded the signals contained in the dataset using MATLAB software. The loaded signals were sent through the USRP device using the transmitting device according to the modulation method. After passing through the real channel, the signals were saved by the receiving device. The public dataset RML2016.10a includes eleven modulation methods, one of which is the BPSK signal waveform. Figure 7 shown.

[0156] The received signal is processed according to steps S1, S2 and S3 to obtain the spliced ​​signal S(t), and then the signal S(t) is obtained according to steps S4, S5, S6, S7 and S8. Figure 8 The recognition rate corresponding to the method of the present invention.

[0157] Keeping the environment and other hardware information unchanged, use other deep learning methods, such as Figure 8 As shown in , S3, S4, S5, S6, S7, and S8 are replaced by the convolutional neural network VGG method, the residual network ResNet method, the mobile neural network MobileNet method, and the long short-term memory network LSTM method. The final comparison results are as follows Figure 8 As shown, the recognition rate of the method of the present invention is significantly improved compared with common methods such as convolutional neural networks, residual networks, mobile neural networks, long short-term memory networks, etc. In the environment of -20dB to 20dB, the recognition rate can be improved by more than 2%. As the signal-to-noise ratio decreases, the influence of noise is too large and eventually it tends to be unrecognizable. This is because when the noise is too large, the characteristic information of the noise is more obvious than the characteristic information of the signal, resulting in the inability to normally distinguish which modulation type the signal belongs to. This is also a problem that cannot be solved in the current modulation classification research.

[0158] The present invention preprocesses signal time domain data to obtain three different types of signal expressions; splices the three preprocessed signals as network input to obtain more comprehensive signal feature information; uses a 2D convolutional neural network to perform primary feature extraction to obtain signal feature information; uses a 3D multi-scale convolutional neural network to further extract the primary feature information to obtain feature information of deeper dimensions; flattens the obtained feature map; uses an LSTM network to extract temporal connections, and finally uses a fully connected network to identify the modulation mode.

[0159] The present invention has a simple preprocessing method, automatically extracts feature information, and can fully utilize information from different signal domains; specifically, it has the following beneficial effects:

[0160] 1. Multi-class feature fusion: This invention applies multiple preprocessing methods to the signal in the time domain, including time domain waveform, frequency domain data, and time domain squared data, to obtain multiple signal representations. By fusing these features, the ability to recognize complex modulation schemes is improved.

[0161] 2. Automated multi-scale feature extraction: We designed convolutional neural sub-networks with different receptive fields and used convolutional neural networks and improved LSTM for feature extraction. This method can automatically extract key features from the input signal, avoiding the tedious process of manually designing features and improving recognition efficiency.

[0162] 3. Strong robustness: The multi-scale feature extraction method can not only improve the recognition rate of the method, but also more comprehensively extract the characteristic information of the signal, ensuring that edge information is not lost. Therefore, the present invention has strong robustness and can maintain high recognition accuracy in various environments, especially under low signal-to-noise ratio conditions, and can still stably extract effective features.

Claims

1. A communication modulation recognition method based on multi-scale, comprising: Step S1: signal acquisition, i.e., acquiring a signal through a signal receiving device to obtain a time domain signal; Step S2: signal preprocessing, i.e., threshold detection and filtering of the collected time domain signal to obtain a time domain filtered signal; Step S3: data transformation, i.e., transforming the time domain filtered signal to obtain frequency domain data and time domain squared data; Step S4: signal splicing, that is, splicing the collected time domain filtered signal, the processed frequency domain data and the processed time domain square data to obtain a network input signal; Step S5: 2D network feature extraction, that is, inputting the network input signal into the 2D convolutional neural network to extract initial feature information; Step S6: 3D multi-scale feature extraction, that is, the initial feature information is further passed through the 3D multi-scale convolutional neural network to obtain small-scale feature information and large-scale feature information; Step S7: feature information processing, that is, concatenating the obtained small-scale feature information and large-scale feature information, flattening them, and then putting them into the LSTM network to extract time-related features to obtain a classification feature vector; Step S8: Modulation classification, that is, sending the classification feature vector to the fully connected layer, and then classifying it through the Softmax layer, and outputting the network's probability distribution for different modulation types.

2. A communication modulation identification method based on multi-scale as claimed in claim 1, characterized in that: In step S1, the collected signal is stored in the form of time domain data to obtain a time domain signal x(t); In step S2, the time domain signal x(t) is first subjected to energy detection, and the noise signal in the signal acquisition device area is subjected to energy detection to obtain an energy threshold value T0, and then the time domain signal x(t) is compared with the threshold value T0. If the energy of the time domain signal x(t) is greater than the threshold value T0, the signal in this part is a valid signal part, and the valid signal part is retained, and the invalid signal is filtered out; the upper limit f1 and the lower limit f2 of the time domain signal transmission frequency are used as the passband frequency of the bandpass filter, and the valid signal part is subjected to bandpass filtering to obtain the filtered time domain filtered signal x'(t).

3. A communication modulation identification method based on multi-scale as claimed in claim 2, characterized in that: The energy threshold T0 in step S2 is: Where N is the number of samples, M represents the collected noise signal, represents the N-point Fourier transform of M, and the square of the absolute value represents the vector norm, which is used to calculate the energy of the noise signal M sampled as N points; Step S3 includes: S3.1: Frequency domain transformation, that is, Fourier transform the time domain filtered signal x'(t) to obtain its frequency domain data f(t) S3.2: Time domain square transformation, that is, square the time domain filtered signal x'(t) to obtain the squared time domain square data x 2 (t).

4. A communication modulation identification method based on multi-scale as claimed in claim 3, characterized in that: In step S4, the network input signal S(t) is obtained by splicing; S(t)=[x'(t),f(t),x 2 (t)]。 5. A communication modulation identification method based on multi-scale as claimed in claim 4, characterized in that: In step S5, the convolution kernel of the 2D convolutional neural network is used to extract local features through a sliding window to obtain initial feature information F1. The convolution layer parameters of the 2D convolutional neural network are {kernel=k=3, padding=p=1, stride=s=1}, and the convolution kernel size is a long sequence K of 1×3. Convolution is used to extract features from S(t), and the features obtained are: Where K[m] is the weight of the convolution kernel, i and j represent the indexes of the rows and columns after convolution, respectively. Corresponding output size H out and W out , whose values ​​are: Where s is the convolution stride size, which represents the step size that the convolution kernel moves on the input, H0 represents the length of S(t), and W0 represents the width of S(t).

6. A communication modulation identification method based on multi-scale as claimed in claim 5, characterized in that: In step S6, the 3D multi-scale convolutional neural network has two sub-networks, and the convolution kernel sizes of the two sub-networks are 1×3 and 1×5 respectively. The initial feature information F1 can obtain small-scale feature information F2 and large-scale feature information F3 respectively through these two sub-networks.

7. A communication modulation identification method based on multi-scale as claimed in claim 6, characterized in that: In step S6, the number of channels of the convolution kernel of the 3D multi-scale convolutional neural network is c=64, and the small-scale feature information F2 is extracted using the 3D multi-scale convolutional neural network, and its convolution layer parameters are {kernel=k=3, padding=p=1, stride=s=1}, and the convolution kernel size is a long sequence K of 1×3; the initial feature information F1 is extracted using convolution, and the features obtained are: Where K[m] is the weight of the convolution kernel, i and j represent the indexes of the rows and columns after convolution, respectively. Corresponding output size H out and W out , whose values ​​are: Where s is the convolution stride size, which indicates the step length of the convolution kernel moving on the input, H0 indicates the length of the initial feature information F1, and W0 indicates the width of the initial feature information F1; The large-scale feature information F3 is extracted using a 3D multi-scale convolutional neural network, with convolution layer parameters of {kernel=k=5, padding=p=1, stride=s=1}, and a convolution kernel size of 1×5 long sequence K; the initial feature information F1 is extracted using convolution, and the features obtained are: Where K[m] is the weight of the convolution kernel, i and j represent the indexes of the rows and columns after convolution, respectively. Corresponding output size H out and W out , whose values ​​are: Where s is the convolution stride size, which indicates the step length of the convolution kernel moving on the input, H0 indicates the length of the initial feature information F1, and W0 indicates the width of the initial feature information F1.

8. A communication modulation identification method based on multi-scale as claimed in claim 7, characterized in that: In step S7, the small-scale feature information F2 and the large-scale feature information F3 after convolution calculation are concatenated to obtain multi-scale feature information F4, and then the multi-scale feature information F4 is tiled through the Fallten layer to obtain a tiled feature vector F5. Subsequently, the tiled feature vector F5 is input into the LSTM network, and the classification feature vector F6 is obtained through the function of the forget gate, input gate and output gate of the LSTM.

9. A communication modulation identification method based on multi-scale as claimed in claim 8, characterized in that: In step S7, the flattened feature vector F5 is truncated to the length of the number of LSTM hidden units γ = 128 as the input Q of the LSTM network t , Q t Put it into the forget gate of LSTM, the forget gate formula is: L t =σ(W f ×[H t-1 ,Q t ]+b f ) Where L t is the information controlled by the forget gate, H t-1 is the hidden information of the LSTM network, σ is the Sigmod formula, W f is the weight of the forget gate, b f is the bias, For the input gate in LSTM, it is expressed as: L i =σ(W i ×[H t-1 ,Q t ]+b i ) Where L i is the information of input gate control, W i is the weight of the input gate, b i is the bias, Based on the last output and this input, we can get: C ′ =tanh(W C ×[H t-1 ,Q t ]+b C ) Among them, C ′ is the current input unit state, W C is the weight of the unit state, b C is the bias, Then pass the last unit state C t-1 and the current input cell state C ′ , multiplied and added with the results of the forget gate and the input gate, respectively, to obtain the current cell state C t , C t =L t ×C t-1 +L i ×C ′ Then get the output according to the output gate of LSTM, the formula is as follows: o t =σ(W o ×[H t-1 ,Q t ]+b o ) h t =o t ·tanh(C t ) Among them, t Indicates the information determined by the output gate, W o represents the weight of the output gate, b o is the bias, h t is the output of the LSTM module, i.e., the classification feature vector F6.

10. The multi-scale communication modulation identification method according to claim 1, characterized in that: The model of the signal receiving device in step S1 is USRP_B210; the models of the computer devices used in steps S2, S3, S4, S5, S6, S7, and S8 are: the model of GPU is GeForce GTX1050 Ti, the model of CPU is Intel(R)Core(TM)i7-8750H CPU@2.20GHz, and the software used is MATLAB R2022a; when sending and receiving signals, the environment is an open room with a length, width, and height of 20 meters, 3 meters, and 3 meters respectively; the two USRPs, i.e., the sending device and the receiving device, are 3 meters apart; In step S1, the length of the x(t) signal is 1×1000, which is a CLOVER-2000-BPSM time domain signal.

Images