Signal type identification method based on deep convolutional neural network

Abstract

The invention discloses a signal type identification method based on a deep convolutional neural network, and the method comprises the steps: carrying out the extension and enhancement of narrowband signal features for the features of narrowband signals, and making a data set for model training; then, feature extraction and network model training are carried out on the signal feature map through a deep convolutional neural network model, and visual separable signal types are identified by using a weight file; compared with a traditional signal identification method, the method bypasses a complex demodulation process, reduces the dependence on demodulation hardware equipment, improves the robustness and applicability of signal identification, and is higher in identification efficiency. Visually separable narrow-band signals are represented in an image form through feature enhancement, and meanwhile, a multi-dimensional representation convolutional neural network model designed by a deep learning technology is introduced for modeling, so that the complexity and difficulty of a signal identification method are reduced.

Description

Technical Field

[0001] The present invention relates to the field of communication technologies, and in particular, to a signal type recognition method based on a deep convolutional neural network. Background Art

[0002] With the rapid development of wireless communication technologies, the communication signal environment has become increasingly complex. Communication signal type recognition is widely applied in many fields such as electronic countermeasures and radio spectrum management. In the application scenarios of non-cooperative communication, identifying various signal types and modulation methods remains the current research focus. Currently, signal type recognition methods mostly rely on traditional communication demodulation technologies. For different types of signals, a large amount of manpower is required to design different and complex demodulation processes. The recognition difficulty and complexity are high, the recognition efficiency is low, and there is a serious dependence on the performance of demodulation devices. The image recognition technology based on deep learning has developed maturely, and the deep learning technology is gradually applied to the field of signal recognition. However, the existing signal recognition technologies are based on the principles of communication signals, with complex algorithm designs, low recognition efficiency, high requirements for device performance, and great dependence in the relevant demodulation processes. Summary of the Invention

[0003] The main object of the present invention is to provide a signal type recognition method based on a deep convolutional neural network, aiming to solve the problems of complex algorithm design and low recognition efficiency of the existing signal recognition technologies based on deep learning.

[0004] To achieve the above object, the present invention proposes a signal type recognition method based on a deep convolutional neural network. The signal type recognition method based on a deep convolutional neural network includes: Performing feature enhancement and extension processing on a narrowband signal to obtain a signal feature map, and constructing a data set; Training the data set through a deep convolutional neural network model to obtain a weight file; Combining the deep convolutional neural network model and the weight file to perform regression prediction on the signal feature map, obtaining a class label with a length of N, and performing threshold screening on the class label values to obtain the corresponding signal type.

[0005] In an embodiment, the specific steps of performing feature enhancement and extension processing on the narrowband signal to obtain a signal feature map and constructing a data set are as follows: Creating a feature container matrix, mapping the feature images of the narrowband signal onto the feature container matrix in three different forms respectively, and performing bandwidth extension on the feature images of the narrowband signal to expand the sample data; Performing sample annotation according to the visually separable characteristics of the sample data to form label data; Combining the sample data and the label data to construct the data set.

[0006] In one embodiment, the specific steps of creating the feature container matrix and mapping the feature images of the narrowband signal to the feature container matrix in three different forms are as follows: Generate a first container matrix with a size of W×H and feature pixel values of 0, directly map the feature pixel values of the narrowband signal visual map to the central position of the first container matrix, and convert it into feature image 1; Generate a second container matrix with a size of W×H and feature pixel values of 0, map the narrowband signal visual map without interval replication and tiling to the second container matrix, and convert it into feature image 2; Generate a third container matrix with a size of W×H and feature pixel values of 0, map the narrowband signal visual map with equal interval replication and tiling to the third container matrix, and convert it into feature image 3.

[0007] In one embodiment, the specific steps of training the dataset through the deep convolutional neural network model to obtain the weight file are as follows: Use a three-input deep convolutional neural network model to train the dataset, and stop training after the loss converges to obtain the weight file.

[0008] The technical solution of the present invention first expands and enhances the narrowband signal features according to the characteristics of the narrowband signal to produce a dataset for model training; then, through the designed deep convolutional neural network model, feature extraction and network model training are performed on the signal feature map, and the visually separable signal types are identified using the weight file; compared with the traditional signal recognition method, this method bypasses the complex demodulation process, reduces the input cost of manual research, reduces the dependence on demodulation hardware equipment, improves the robustness and applicability of signal recognition, and has higher recognition efficiency; the visually separable narrowband signals are characterized in the form of images through feature enhancement, and at the same time, a multi-dimensional representation convolutional neural network model designed by introducing deep learning technology is used for modeling, reducing the complexity and difficulty of the signal recognition method; compared with the single feature input method, the multi-dimensional representation dimension improves the accuracy of signal recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 It is the overall process schematic diagram of the signal type recognition method based on the deep convolutional neural network of the present invention; Figure 2 It is the process schematic diagram of constructing a customized dataset for the signal type recognition method based on the deep convolutional neural network of the present invention; Figure 3 It is the process schematic diagram of training the deep convolutional neural network model for the signal type recognition method based on the deep convolutional neural network of the present invention; Figure 4This is a schematic diagram of the structure of the deep convolutional neural network model for the signal type recognition method based on the deep convolutional neural network of the present invention. Detailed implementation manners

[0010] To make the objectives, 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 with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Components of the embodiments of the present invention usually described and illustrated in the accompanying drawings here can be arranged and designed in various different configurations.

[0011] 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 claimed present invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0012] It should be noted that like reference numerals and letters denote like items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0013] In the description of the present invention, it should be understood that the orientation or positional relationships indicated by the terms "upper", "lower", "inner", "outer", "left", "right", etc. are based on the orientation or positional relationships shown in the accompanying drawings, or the orientation or positional relationships in which the inventive product is customarily placed during use, or the orientation or positional relationships commonly understood by those skilled in the art. These are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as limiting the present invention.

[0014] In addition, the terms "first", "second", etc. are only used for descriptive distinction and cannot be understood as indicating or implying relative importance.

[0015] In the description of the present invention, it should also be noted that unless otherwise clearly defined and limited, terms such as "arrangement" and "connection" should be understood in a broad sense. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, and it can be the communication inside two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0016] Since the existing deep learning-based signal recognition technology is based on the principles of communication signals, its algorithm design is complex, the recognition efficiency is low, and it has high requirements for and relies heavily on the performance of the device during the demodulation-related process.

[0017] To solve the above problems, the present invention proposes a signal type recognition method based on a deep convolutional neural network. The following will describe the specific embodiments of the present invention in detail with reference to the accompanying drawings.

[0018] As Figures 1-4 shown, the signal type recognition method based on a deep convolutional neural network includes the following steps: Perform feature enhancement and expansion processing on the narrowband signal to obtain a signal feature map, and construct a data set; Train the data set through a deep convolutional neural network model to obtain a weight file; Combine the deep convolutional neural network model and the weight file to perform regression prediction on the signal feature map, obtain a class label with a length of N, and perform threshold screening on the class label values to obtain the corresponding signal type.

[0019] In one embodiment, the specific steps of performing feature enhancement and expansion processing on the narrowband signal to obtain a signal feature map and constructing a data set are as follows: Create a feature container matrix, map the feature images of the narrowband signal onto the feature container matrix in three different forms respectively, and perform bandwidth expansion on the feature images of the narrowband signal to expand the sample data; Perform sample annotation according to the visually separable characteristics of the sample data to form label data; Combine the sample data and the label data to construct the data set.

[0020] In one embodiment, the specific steps of creating a feature container matrix and mapping the feature images of the narrowband signal onto the feature container matrix in three different forms respectively are as follows: Generate a first container matrix with a size of W×H and a feature pixel value of 0, directly map the feature pixel values of the narrowband signal visual map to the center position of the first container matrix, and convert it into feature image 1; Generate a second container matrix with a size of W×H and a feature pixel value of 0, map the narrowband signal visual map by non-interval replication and tiling onto the second container matrix, and convert it into feature image 2; Generate a third container matrix with a size of W×H and a feature pixel value of 0, map the narrowband signal visual map by equally spaced replication and tiling onto the third container matrix, and convert it into feature image 3.

[0021] In one embodiment, the specific steps of training the dataset through the deep convolutional neural network model to obtain the weight file are as follows: training the dataset using a three-input deep convolutional neural network model, stopping the training after the loss converges, and obtaining the weight file.

[0022] The signal type recognition method based on the deep convolutional neural network of the present invention consists of two parts: data processing, and dataset production, model training and prediction; specifically, the method first expands and enhances the narrowband signal features according to the characteristics of the narrowband signal to produce a dataset for model training; then, extracts features from the signal feature map and trains the network model through the designed deep convolutional neural network model, and uses the weight file to identify visually separable signal types; compared with the traditional signal recognition method, this method bypasses the complex demodulation process, reduces the input cost of manual research, reduces the dependence on demodulation hardware devices, improves the robustness and applicability of signal recognition, and has higher recognition efficiency; represents the visually separable narrowband signals in the form of images through feature enhancement, and at the same time introduces a multi-dimensional representation convolutional neural network model designed by deep learning technology for modeling, reducing the complexity and difficulty of the signal recognition method; compared with the single feature input method, the multi-dimensional representation dimension improves the accuracy of signal recognition.

[0023] As Figure 1 shown, in this embodiment, a dataset is constructed through feature images, the dataset is input into the deep convolutional neural network to obtain the weight file, and then the feature matrix is regressively predicted through the deep convolutional neural network model and the weight file to obtain a class label of length N, and the largest class label value is selected. If the class label value is greater than the threshold, the regression prediction is correct; otherwise, it is judged as an unknown signal, thus completing the recognition of the visually separable predicted signal.

[0024] As Figure 2As shown in the figure, in order to comprehensively represent the visual features of narrowband signals, in this embodiment, a feature enhancement method is used to increase the multiple representations of features. Specifically, a feature container matrix with a size of W×H and feature pixel values of 0 is first created, and then the feature images of the narrowband signals are mapped onto the feature container matrix in three different forms to obtain three types of feature matrices with a size of W×H, so as to achieve the visual feature enhancement of narrowband signals, namely, a single narrowband feature image 1, a non-gap narrowband tiled feature image 2, and an equal-gap narrowband tiled feature image 3. In order to meet the robustness of recognition for the same signal under different bandwidths, in this embodiment, the bandwidth of the feature image is extended to achieve the expansion of sample data; and label data is defined according to different narrowband signal types, that is, sample annotation is performed according to the visually separable characteristics of the sample data to form label data, that is, one-dimensional category labels, and then the expanded sample data together constitute a sample data set, so as to represent the visually separable narrowband signals in the form of images through feature enhancement and realize the construction of a customized data set.

[0025] As Figure 3 shown, in this embodiment, the combined feature map after feature extension and sample expansion is input into a three-input deep convolutional neural network to train the data set until the training stops after the loss converges, and a weight file is obtained.

[0026] As Figure 4As shown in the figure, during prediction in this embodiment, the prediction signal is first enhanced with visualization features, and the narrowband visualization features are regressively predicted by combining a deep convolutional neural network and the trained model weights to obtain a class label of length N. After screening the class label values by threshold size, the corresponding visualization signal type can be obtained. In addition, for data outside the sample label type, the model can also achieve a good regression effect. Specifically, the deep convolutional neural network is a three-input network that inputs three feature images with a size of 512*256*3. The deep convolutional neural network first uses a convolutional kernel with a size of 3*3 and 64 channels to extract features from the three feature images respectively, and then performs max-pooling and BN (Batch Normalization) operations on the feature matrices respectively to complete feature extraction. Then, the three obtained feature matrices are used with a convolutional kernel with a size of 3*3 and 128 channels to extract features, and max-pooling and BN operations are performed on the feature matrices respectively to complete feature extraction. Continuing, the three obtained feature matrices are used with a convolutional kernel with a size of 3*3 and 256 channels to extract features, and then channel dimension stacking, max-pooling, and BN operations are performed on the three feature matrices to complete feature extraction, obtaining a fused feature matrix. The obtained fused feature matrix is used with a convolutional kernel with a size of 3*3 and 1024 channels to extract features, and max-pooling and BN operations are performed on the feature matrix to extract features. The obtained feature matrix is used with a convolutional kernel with a size of 3*3 and 2048 channels to extract features, and max-pooling and BN operations are performed on the feature matrix to extract features. The obtained feature matrix is used with a convolutional kernel with a size of 3*3 and 4096 channels to extract features, and max-pooling and BN operations are performed on the feature matrix to extract features. Finally, the feature matrix is flattened and fully connected to a feature vector of length 1, regularized by Dropout, and non-linearly processed using the Sigmoid activation function, and the class label is output as required. It should be noted that to prevent the "cold start" problem, the LeakyReLu activation function is selected for non-linear transformation after each convolution operation.

[0027] In summary, the signal type recognition method based on a deep convolutional neural network according to the present invention combines a deep convolutional neural network and a customized data set, and directly realizes the type recognition of visually separable narrowband signals based on the visual images of the signals, that is, first enhances the features of the visualization images of the narrowband signals, and then performs prediction and classification recognition through the deep convolutional neural network model, which can improve the efficiency and accuracy of signal recognition in the case of a large narrowband bandwidth span and a low signal-to-noise ratio.

[0028] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the technical principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A signal type recognition method based on deep convolutional neural network, characterized in that: The signal type recognition method based on deep convolutional neural network comprises the following steps: Perform feature enhancement and expansion processing on narrowband signals to obtain signal feature maps and construct data sets; The data set is trained by a deep convolutional neural network model to obtain a weight file; The signal feature graph is regressed and predicted in combination with the deep convolutional neural network model and the weight file to obtain a category label with a length of N, and the category label value is threshold-screened to obtain the corresponding signal type.

2. The signal type recognition method based on deep convolutional neural network according to claim 1 is characterized in that: The specific steps of performing feature enhancement and expansion processing on the narrowband signal to obtain a signal feature graph and constructing a data set are: Creating a feature container matrix, mapping the feature image of the narrowband signal onto the feature container matrix in three different forms respectively, and performing bandwidth expansion on the feature image of the narrowband signal to expand sample data; Annotate the samples according to the visually separable characteristics of the sample data to form label data; The sample data and the labeled data are combined to construct the data set.

3. The signal type recognition method based on deep convolutional neural network according to claim 2 is characterized in that: The specific steps of creating a feature container matrix and mapping the feature image of the narrowband signal onto the feature container matrix in three different forms are as follows: Generate a first container matrix with a size of W×H and a feature pixel value of 0, directly map the feature pixel value of the narrowband signal visual image to the center position of the first container matrix, and convert it into a feature image 1; Generate a second container matrix with a size of W×H and a feature pixel value of 0, copy and tile the narrowband signal visual image onto the second container matrix without intervals, and convert it into a feature image 2; Generate a third container matrix with a size of W×H and a characteristic pixel value of 0, and evenly copy and tile the narrowband signal visual image onto the third container matrix and convert it into a characteristic image 3.

4. The signal type recognition method based on deep convolutional neural network according to claim 1 is characterized in that: The specific steps of training the data set through a deep convolutional neural network model to obtain the weight file are: training the data set using a three-input deep convolutional neural network model, stopping the training after the loss converges, and obtaining the weight file.

Images