阅读说明：本技术 一种基于深度多流神经网络的信号自动分类识别方法 (Signal automatic classification and identification method based on deep multi-flow neural network ) 是由 王岩 于 2020-02-25 设计创作，主要内容包括：本申请公开了一种基于深度多流神经网络的信号自动分类识别方法,将深度多流神经网络的网络结构在水平方向上扩展,以实现深度多流神经网络提取更丰富的信号特征。在深度多流神经网络的每个流中叠加预设量的卷积单元,以提高深度多流神经网络的分类性能,防止网络训练困难和容易过拟合；对改进后的所述深度多流神经网络进行训练和验证,用于对信号进行分类识别。改进后的深度多流神经网络随着网络结构宽度的增加,分类效果提高,参数数量减少。在验证阶段仅消耗有限的计算资源即可在几毫秒内完成相关任务。在通信系统部署中,使用训练好的改进后的深度多流神经网络,等效于验证阶段,具有较低的计算复杂度和实时处理速度,提高了信号分类的效率。(The application discloses a signal automatic classification and identification method based on a deep multi-flow neural network, which expands a network structure of the deep multi-flow neural network in the horizontal direction so as to extract richer signal characteristics by the deep multi-flow neural network. Superposing a preset amount of convolution units in each stream of the deep multi-stream neural network to improve the classification performance of the deep multi-stream neural network and prevent the network from being difficult to train and easy to overfit; and training and verifying the improved deep multi-flow neural network for carrying out classification and identification on signals. The improved deep multi-flow neural network has the advantages that the classification effect is improved and the number of parameters is reduced along with the increase of the width of the network structure. The related tasks can be completed within a few milliseconds with only limited consumption of computing resources during the verification phase. In the deployment of a communication system, the trained improved deep multi-flow neural network is used, which is equivalent to a verification stage, so that the method has lower computational complexity and real-time processing speed, and the efficiency of signal classification is improved.)

1. A signal automatic classification and identification method based on a deep multi-flow neural network is characterized by comprising the following steps:

the network structure of the deep multi-flow neural network is expanded in the horizontal direction, so that the deep multi-flow neural network extracts richer signal characteristics, and the method is suitable for classification of signal data sets of various modulation schemes;

superposing a preset amount of convolution units in each stream of the deep multi-stream neural network to improve the classification performance of the deep multi-stream neural network and prevent the network from being difficult to train and easy to overfit;

and training and verifying the improved deep multi-flow neural network for carrying out classification and identification on signals.

2. The method for automatically classifying and identifying signals based on the deep multi-flow neural network as claimed in claim 1, wherein the structure of the deep multi-flow neural network is expressed as:

where ζ represents the multi-stream convolution output, η represents the convolution stream composed of 1 × 1 and 2 × 2 convolution kernels, P represents the number of streams, P ═ 1,2, …, P, μ represents the convolution stream composed of 1 × 1 and 3 × 3 convolution kernels, Q represents the number of streams, Q ═ 1,2, …, Q,two streams are included, one including a convolution kernel of 1 × 1 and the other including a convolution kernel of 1 × 1 and AveragePooling.

3. The method for automatically classifying and identifying the signals based on the deep multi-flow neural network according to claim 2, wherein the step of superposing a preset amount of convolution units in each flow of the deep multi-flow neural network comprises: after each convolutional layer, a non-linear function layer and a normalization layer are added.

4. The method for automatically classifying and identifying signals based on the deep multi-flow neural network as claimed in claim 1, wherein the sequence of the input signals input into the improved deep multi-flow neural network is divided into different lengths, and each length of the signal sequence is input into the improved deep multi-flow neural network, so as to obtain rich modulation identification features and improve the signal modulation classification performance.

5. The method for automatically classifying and identifying the signals based on the deep multi-flow neural network as claimed in claim 4, wherein the signal sequences with different lengths are as follows:

h_l(v)＝r(l+v)

where l denotes the start position of the signal sequence, l ═ 1,2, …, L-1, L denote the total length of the signal r (·), V is the signal length, V ═ 0,1, …, V denotes the maximum fixed signal length that can be assumed.

6. The method for automatically classifying and identifying signals based on the deep multi-flow neural network as claimed in claim 1, wherein the improved deep multi-flow neural network extracts various characteristics of the signals through different forms of convolution kernels within the same receptive field range to obtain stronger nonlinear representation and ensure that the final classification result is more accurate.

7. The method for automatically classifying and identifying signals based on the deep multi-flow neural network as claimed in claim 2, wherein the advanced classification representation is improved, a preset number of 2 × 2 or 3 × 3 convolution kernels are connected in series, more nonlinear features are combined, the network has better nonlinear fitting capability, and the output representation is shown as

Wherein o is_cOutput data size, z, representing the current c-layer_cRepresenting the input data size of the current c-layer, g represents the size of the convolution kernel, where g takes 1,2 or 3, respectively, corresponding to 1 × 1,2 × 2, 3 × 3 convolution kernels, s represents the step size, andindicating a rounded down symbol.

8. The method for automatically classifying and identifying the signals based on the deep multi-flow neural network as claimed in claim 1, wherein the improved deep multi-flow neural network adopts a stochastic gradient descent method in a training process, an initial learning rate is set to 0.001, momentum is set to 0.9, and a loss function adopts a classification cross entropy and is set to 128 in batches in a verification process.

Technical Field

The application relates to the technical field of signal identification, in particular to a signal automatic classification and identification method based on a deep multi-flow neural network.

Background

In recent years, AMC for wireless communication has received increasing attention, complex wireless channel conditions may distort signals and degrade performance of communication systems.

There are two major modulation classification algorithms, classification methods Based on likelihood criteria (L ikelihood-Based, L B) and classification methods Based on Feature-Based (FB). L B method is optimal to reduce the likelihood of classification errors.

Disclosure of Invention

In order to solve the technical problems, the following technical scheme is provided:

in a first aspect, an embodiment of the present application provides a method for automatically classifying and identifying a signal based on a deep multi-flow neural network, where the method includes: the network structure of the deep multi-flow neural network is expanded in the horizontal direction, so that the deep multi-flow neural network extracts richer signal characteristics, and the method is suitable for classification of signal data sets of various modulation schemes; superposing a preset amount of convolution units in each stream of the deep multi-stream neural network to improve the classification performance of the deep multi-stream neural network and prevent the network from being difficult to train and easy to overfit; and training and verifying the improved deep multi-flow neural network for carrying out classification and identification on signals.

By adopting the implementation mode, the improved deep multi-flow neural network has the advantages that the classification effect is improved and the number of parameters is reduced along with the increase of the width of the network structure. The related tasks can be completed within a few milliseconds with only limited consumption of computing resources during the verification phase. In the deployment of a communication system, the trained improved deep multi-flow neural network is used, which is equivalent to a verification stage, so that the method has lower computational complexity and real-time processing speed, and the efficiency of signal classification is improved.

With reference to the first aspect, in a first possible implementation manner of the first aspect, the structure of the deep multiflow neural network is expressed as:

where ζ represents the multi-stream convolution output, η represents the convolution stream composed of 1 × 1 and 2 × 2 convolution kernels, P represents the number of streams, P ═ 1,2, …, P, μ represents the convolution stream composed of 1 × 1 and 3 × 3 convolution kernels, Q represents the number of streams, Q ═ 1,2, …, Q,two streams are included, one including a convolution kernel of 1 × 1 and the other including a convolution kernel of 1 × 1 and AveragePooling.

With reference to the first possible implementation manner of the first aspect, in a second possible implementation manner of the first aspect, the superimposing, in each stream of the deep multi-stream neural network, a preset amount of convolution units includes: after each convolutional layer, a non-linear function layer and a normalization layer are added.

With reference to the first aspect, in a third possible implementation manner of the first aspect, a sequence of an input signal input to the improved deep multi-stream neural network is divided into different lengths, and each length of the signal sequence is input to the improved deep multi-stream neural network, so as to obtain rich modulation identification features and improve signal modulation classification performance.

With reference to the third possible implementation manner of the first aspect, in a fourth possible implementation manner of the first aspect, the signal sequences with different lengths are:

h_l(v)＝r(l+v)

where l denotes the start position of the signal sequence, l ═ 1,2, …, L-1, L denote the total length of the signal r (·), V is the signal length, V ═ 0,1, …, V denotes the maximum fixed signal length that can be assumed.

With reference to the first aspect, in a fifth possible implementation manner of the first aspect, the improved deep multi-flow neural network extracts various features of a signal through convolution kernels in different forms within the same receptive field range to obtain a stronger nonlinear representation, so that a final classification result is ensured to be more accurate.

With reference to the first possible implementation manner of the first aspect, in a sixth possible implementation manner of the first aspect, the advanced classification representation is improved, a preset number of 2 × 2 or 3 × 3 convolution kernels are connected in series, more nonlinear features are combined, the network has better nonlinear fitting capability, and the output is represented as

Wherein o is_cOutput data size, z, representing the current c-layer_cRepresenting the input data size of the current c-layer, g represents the size of the convolution kernel, where g takes 1,2 or 3, respectively, corresponding to 1 × 1,2 × 2, 3 × 3 convolution kernels, s represents the step size, andindicating a downward roundingThe symbol of (2).

With reference to the first aspect, in a seventh possible implementation manner of the first aspect, in a training process of the improved deep multi-flow neural network, the optimizer adopts a random gradient descent method, an initial learning rate is set to 0.001, momentum is set to 0.9, in a verification process, a loss function adopts a classification cross entropy, and a batch is set to 128.

Drawings

Fig. 1 is a schematic flowchart of a method for automatically classifying and identifying signals based on a deep multi-flow neural network according to an embodiment of the present application;

fig. 2 is a schematic structural diagram of a convolutional neural network provided in an embodiment of the present application;

fig. 3 is a schematic structural diagram of a deep multi-flow neural network provided in an embodiment of the present application;

fig. 4 is a schematic diagram of a channel model according to an embodiment of the present application;

fig. 5 is a schematic diagram of modulation classification accuracy of a network involved in different flow structures according to an embodiment of the present application;

fig. 6 is a schematic diagram for comparing classification effects of related networks with other neural network methods under different overlapping units according to an embodiment of the present application.

Detailed Description

The present invention will be described with reference to the accompanying drawings and embodiments.

Fig. 1 is a schematic flowchart of a method for automatically classifying and identifying signals based on a deep multi-flow neural network according to an embodiment of the present application, and referring to fig. 1, the method includes:

s101, expanding a network structure of the deep multi-flow neural network in the horizontal direction to extract richer signal characteristics by the deep multi-flow neural network, and being suitable for classification of signal data sets of various modulation schemes.

The Deep learning Method (Deep L earning Method, D L M) has shown classification capability in various tasks, such as image recognition, automatic speech recognition and machine translation, the main reason is that multiple hidden layers can learn high-level representations hidden in the dataset, another reason is that better classification results are non-linear logic functions used in the network layers D L M can better handle the problems encountered by L B and FB above, and automatically obtain better classification thresholds D L M's algorithm design primarily considers the choice of network structure in depth and width.

D L M currently has numerous applications in the wireless communication field, such as radio frequency signal processing, radio resource allocation, radio control, MIMO detection, channel estimation and Internet of Things (IoT) signal detection with the commercialization of 5G, D L M application faces many opportunities and challenges AMC typically uses D L M to employ either a Convolutional Neural Network (CNN) or a Recurrent Neural Network (RNN) structure.

In the general form of deep CNN network structures, the features of local data blocks of different receptive fields are extracted, convolution kernels of different sizes are passed, the CNN network outputs results by activating functions and then performs a pooling operation, taking a commonly used nonlinear function Re L U as an example, after convolution, the features are obtained using the activating functions

f＝max(w^Tα,0)

Where w represents the weight matrix, T represents the transpose of the matrix, α represents the input data max (-) when w^Tα>When 0 is w^Tα when w^Tα<0-CNN network structures are composed of alternating convolutional and pooling layers the convolutional layers form a feature map, e.g., Re L U, sigmoid, tanh, etc., by using a linear combination of nonlinear activation functions.

In fact, the convolution operation in CNN can also be regarded as the inner product of two vectors, which is a generalized linear model. If the extracted features are linear or low degree of non-linearity, the features obtained by convolution operations are more desirable. If the extracted features are highly non-linear, such as data sets containing various signal modulation schemes, it is difficult for the convolution kernel method to obtain the desired features. In this case, the conventional CNN method will attempt to extract various potential features by overcomplete convolution kernels, i.e., initialize a large number of convolution kernels to extract as many features as possible to cover the extracted desired signal features. It results in a complex network structure and a huge parameter space.

In view of these problems encountered with conventional CNN methods, a stronger nonlinear function approximator, referred to as a Network In Network (NiN), is used to improve the extraction capability of local data features, the nonlinear approximator may have a variety of options, and a multilayer perceptron (M L P) may be used as a micro-network that uses a more nonlinear structure to approximate data features, the M L P layer may be considered as a micro-multilayer network contained in a conventional CNN layer, as shown in FIG. 2^Tα,0) may be modified to:

…

where b represents the deviation, k represents the data in the k layer of the network, m, n represents the input data in m rows and n columns, and j represents the weight of the corresponding data.

M L P is suitable as a micro-network because it allows a network of such CNN architecture to be highly integrated in this way it can be trained with the BP algorithm and increases the number of hidden layers by itself, which is consistent with the theory of feature reuse.

The convolution in CNN is basically a multi-dimensional feature map, which has the same number of dimensions as the convolution kernel, it uses a multi-dimensional convolution kernel for feature extraction, if a 1x1 convolution kernel is used, the obtained values are independent of surrounding data points, this operation achieves a linear combination of multiple feature maps and achieves a variation in feature map dimensions, a cascade of multiple 1 × 1 convolution kernels can accomplish a non-linear combination of multi-dimensional feature maps, in conjunction with activation functions, the NiN structure can be achieved by M L P, at the same time, the number of convolution kernels can be adjusted in dimension by 1 × 1 convolution kernel operations, thus reducing parameters, in a network of microstructures embedded in M L P, the network can learn complex and useful cross-feature maps, since the integration between different feature maps is obtained by different convolution kernels, each layer corresponding to a convolution layer with a 1 × 1 convolution kernel.

And S102, superposing a preset amount of convolution units in each stream of the deep multi-stream neural network to improve the classification performance of the deep multi-stream neural network and prevent the network from being difficult to train and easy to overfit.

In the deep multi-stream neural network form adopted by the application, in the same receptive field range, the deep multi-stream neural network can extract various characteristics of signals through convolution kernels in different forms, so that stronger nonlinear representation is obtained. Richer features also mean that the final classification result is more accurate. In the network structure, it is mainly composed of four forms of multi-stream convolutional layers. The structure of the deep multi-flow neural network in this embodiment is expressed as follows:

where ζ represents the multi-stream convolution output, η represents the convolution stream composed of 1 × 1 and 2 × 2 convolution kernels, P represents the number of streams, P ═ 1,2, …, P, μ represents the convolution stream composed of 1 × 1 and 3 × 3 convolution kernels, Q represents the number of streams, Q ═ 1,2, …, Q,two streams are included, one including a convolution kernel of 1 × 1 and the other including a convolution kernel of 1 × 1 and AveragePooling.

The network structure behaves as a sparse connection in the feature dimension and performs convolution and re-aggregation in multiple dimensions to collect signal features with strong correlation. Each size convolution outputs only a portion of the plurality of signal features, thereby improving classification performance of the network.

In the superposition convolution unit according to the embodiment, when the convolution layer has a large number of input features, the convolution operation directly performed on the input generates a huge amount of calculation, if the input is firstly reduced in dimension, the calculation complexity is greatly saved after the number of the features is reduced, that is, the 1 × 1 convolution kernel is used in the multi-stream convolution layer, as long as the number of the features in the final output is unchanged, the reduction of the middle size is equivalent to the compression effect, and the final training result is not influenced, in order to further improve the high-level classification representation, a plurality of 2 × 2 or 3 × 3 convolution kernels are connected in series, so that more nonlinear features can be combined, the network has better nonlinear fitting capability, and the output can be represented as a high-level classification representation

Wherein o is_cOutput data size, z, representing the current c-layer_cRepresenting the input data size of the current c-layer, g represents the size of the convolution kernel, where g takes 1,2 or 3, respectively, corresponding to 1 × 1,2 × 2, 3 × 3 convolution kernels.s represents the step size, andindicating a rounded down symbol. Through multilayer convolution processing, richer nonlinear characteristics can be combined to better fit the distribution of original signal data, so that the classification effect of signal modulation is improved.

Other settings of the deep multi-flow neural network are as follows: adding a non-linear function layer and a normalization layer after each convolution layer further improves the net classification capability and they also prevent overfitting problems.

And S103, training and verifying the improved deep multi-flow neural network for signal classification and identification.

During training, the optimizer employs a random gradient descent method, and the initial learning rate is set to 0.001 and momentum is set to 0.9. During validation, the loss function uses class cross entropy with the batch set to 128.

The signal preprocessing section includes: input data in neural networks is typically fixed in length. In order to further exploit the modulation characteristics to improve the classification effect, the sequence of input signals may be divided into different lengths. Each length of the signal sequence is input into the deep multi-stream neural network, so that rich modulation identification characteristics are obtained, and better signal modulation classification is realized. The signal sequences of different lengths are as follows

h_l(v)＝r(l+v)

Wherein l represents the starting position of the signal sequence, l is 1,2, …, L-1, L represents the total length of the signal r (·), V is the signal length, V is 0,1, …, V represents the maximum fixed signal length that can be taken.

In an embodiment of the present application, a NiN structure first performs a general convolution operation, which follows a 1 × convolution kernel.1 convolution kernel × convolution kernel reduces dimensionality and reduces computational complexity.a deep multi-stream neural network performs several 1 × convolution operations, which are equivalent to performing fully concatenated calculations on all signal characteristics.a NiN structure is simply a multi-layered convolution on the same scale, with no added pool layer in between.a deep multi-stream neural network passes through a 1 × convolution kernel in the first few layers and passes through a convergence layer to better extract signal features, as shown in fig. 3. Conv stands for convolution operations, where 1 ×,2 × and 3 × stand for the size of the convolution kernel.the number after the convolution kernel represents the size of the receiving domain.the sensing fields of the three 1 × convolution kernels after the input layer are 32, 32 and 64, respectively.32 corresponding to the branching networks containing 1 × convolution kernel branches are 64, the sensing fields corresponding to the flow containing avgegeotropic Poeing are ×,2 is represented by a solid line containing 2 convolution kernel, 2 is represented by a maximum value of a box containing a virtual wire box representing the convolution kernel 3. the result of the convolution kernel containing avflag 3. the average of the convolution kernel 733, 3. the convolution kernel containing the result, 3. the result of the virtual box containing avphylline 3. the virtual wire box containing avpling 3. the convolution kernel representing the average convolution kernel.

In the present embodiment, the form of the channel model representation to which the communication signal relates is substantially similar to a general communication model, but differs from the basic form in detail, and can be described in the following form: wireless channels are subject to multipath, doppler, and Additive White Gaussian Noise (AWGN). Channel model as shown in fig. 4, the received signal can be represented as

Where g (t) is the transmitted signal, n (t) is AWGN, h (t, η) represents the multipath channel, d_i(t) is the attenuation of the ith multipath signal,representing a convolution, η_i(t) is the ith multipath delay N is the number of multipath signals and all paths have similar Doppler scaling factor σ η_i(t)≈η_i- σ t. The transmitted signal may be digital (e.g., quadrature amplitude modulation) or analog (e.g., frequency shift keying).

To further illustrate the structural advantages exhibited by the present invention in the testing phase of the protocol implementation, experimental verification was performed and the proposed method was evaluated using a signal data set generated using the following parameters. Ten modulation schemes are considered, such as binary phase-shift keying (BPSK), quadrature phase-shift keying (QPSK) and 8PSK,16 quadrature amplitude modulation (16 QAM), 64QAM, 4frequency-shift keying (4 FSK), 8FSK, Pulse Amplitude Modulation (PAM), double-sideBand modulation (DSB), Frequency Modulation (FM). The signal lengths are set to 32, 64, 128 and 256. The number of training vectors is 500 for each modulation scheme, and there are 500 validation vectors in one modulation scheme. The signal-to-noise ratio considered is from-20 dB to 20 dB. The rayleigh distribution is used for the temporal fading model. The additive noise is assumed to be band-limited, zero-mean, white gaussian noise. In the frequency selective fading simulation, the number of sine waves was set to 6, and the noise source seeded by the random number generator was set to 0x 1337. The standard deviation of the drift process for the sampling rate is 0.01 Hz per sample and the maximum sampling rate offset is 50 Hz. The filter uses raised cosine pulse shaping with a roll-off coefficient of 0.3.

In fig. 5, S1 denotes the number of streams containing 3 633 convolution kernels, S2 denotes the number of streams containing 2 × convolution kernels, from-15 dB to 0dB, when S1 is 2, S2 is 3, S1 is 3, S2 is 2, the classification effect is similar, and there is the best classification result, these two stream formats are 1, S2 are 1% higher than S1, S2 is 4% higher than S1, S2 is 3.5% higher than S1, S2 is 3, and is approximately 1.5% higher than the other stream formats, at a lower signal-to-noise ratio, increasing the number of streams of two convolution kernels can increase the classification effect, when the convolution kernels of each stream are sufficient to extract modulation characteristics in the signal data set, the classification effect does not increase significantly, from 0dB to 15dB, when S1 is 3, S5 is 1, S583 is 24, S583 is 2, S593 is equal to S863, S593 is equal to S593, S593 is equal to four cases.

In fig. 6, R1 represents the number of superimposed units with a convolution kernel of 3 ×, R2 represents the number of superimposed units with a convolution kernel of 2 × 2, while a deep multiflow neural network is compared with Resnet, densenert and ResNeXT, rennet and Densenet represent the structural forms of typical longitudinal deep networks, ResNeXT is mainly a horizontal deep network structure, in the range of-20 dB to-10 dB, R1 is 3, R2 is 3, R1 is 4, R2 has the best classification effect, the former is slightly better than the latter by about 1%, except for R1, R2 is 1, the two structural forms have a classification effect which is nearly 7% higher than the other cases, which is 21% lower than the two structural forms, in this range of signal to noise ratios, their classification effect is higher than the net and densnet 20%, the classification effect is higher than the other forms by about 7% than the normal network structures, which are higher than the net10% than the net10, the R8, the performance of the wide-range of the networks is higher than the net10% of the common network structures, the R6858, the network structures which are similar to the higher than the netnet 8, the network structures of the high R8, the high R8.

Table 1 compares the parameter size of the method of the present invention with other network models, where the best classification effect of S1-2, S2-3 and length-256 is selected. As the number of overlay units increases, the parameter size increases significantly. The parameter size of R1 ═ 4, R2 ═ 4 is almost 2 for R1, twice for R2 ═ 2, and 20 ten thousand more than that of R1 ═ 3 and R2 ═ 3. The parameter sizes of ResNet, densneet and ResNeXT are 18 times, 3 times and 6 times of R1-4, R2-4, respectively.

TABLE 1 comparison of parameter quantities for different network methods

It should be noted that, in this document, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other identical elements in a process, method, article, or apparatus that comprises the element.

Of course, the above description is not limited to the above examples, and technical features that are not described in this application may be implemented by or using the prior art, and are not described herein again; the above embodiments and drawings are only for illustrating the technical solutions of the present application and not for limiting the present application, and the present application is only described in detail with reference to the preferred embodiments instead, it should be understood by those skilled in the art that changes, modifications, additions or substitutions within the spirit and scope of the present application may be made by those skilled in the art without departing from the spirit of the present application, and the scope of the claims of the present application should also be covered.