CN109344772B - Ultrashort wave specific signal reconnaissance method based on spectrogram and deep convolutional network - Google Patents
Ultrashort wave specific signal reconnaissance method based on spectrogram and deep convolutional network Download PDFInfo
- Publication number
- CN109344772B CN109344772B CN201811159919.XA CN201811159919A CN109344772B CN 109344772 B CN109344772 B CN 109344772B CN 201811159919 A CN201811159919 A CN 201811159919A CN 109344772 B CN109344772 B CN 109344772B
- Authority
- CN
- China
- Prior art keywords
- feature
- signal
- deep convolutional
- time
- neural network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
Abstract
The invention belongs to the technical field of radio signal identification, and particularly relates to an ultrashort wave specific signal reconnaissance method based on a spectrogram and a deep convolutional network, which comprises the following steps: carrying out short-time Fourier transform on specific signals in a sample library to obtain a signal time-frequency map, wherein the specific signals are signals containing frame synchronization codes in a signal transmission data frame structure; training a deep convolutional neural network model by using a time-frequency map, and predicting a position target by using feature maps of different scales through a feature pyramid in the training process; and carrying out target detection and identification on a specific signal in the ultra-short wave communication by using the trained deep convolutional neural network model. The invention solves the problems of poor detection and identification efficiency and the like under the conditions of low signal-to-noise ratio and strong channel interference in the existing method, realizes ultrashort wave specific signal detection, time-frequency positioning and classification identification, improves the signal identification rate, has stable performance and high operation efficiency, provides ideas for subsequent related researches in the field, and has strong practical application value.
Description
Technical Field
The invention belongs to the technical field of radio signal identification, and particularly relates to an ultrashort wave specific signal reconnaissance method based on a spectrogram and a deep convolutional network.
Background
The signal detection technology is widely applied to radio detection, electronic countermeasure, software radio and the like, and the detection and identification technology of the ultrashort wave specific signal is one of the important components, and becomes a research hotspot in the field of signal analysis and processing. The ultra-short wave communication refers to communication for information transmission using radio waves in a 30MHz to 300MHz band. However, due to the propagation mode of the ultrashort wave communication, the transmitted signals have the phenomena of fading, interference, aliasing and the like, so that the research on the detection and identification of the ultrashort wave signals becomes a difficult problem. The specific signal is a signal containing a frame synchronization code (header frame) in data transmitted by the signal, and is mostly present in a short-wave, ultra-short-wave, satellite and other time division multiplex communication systems.
With the wide application of image processing technology in the field of communication signal processing, some signal detection algorithms combining image processing and signal time-frequency distribution emerge in recent years. For example, the automatic detection and identification of the Morse signal are carried out by utilizing image processing methods such as image enhancement, image segmentation, morphological denoising and the like; detecting satellite multiple access signals by using a time-frequency spectrogram generating method and combining with a morphological processing method; or, a method combining time frequency analysis and a statistical model is used for establishing a passive detection model of the three time frequency analysis methods for the passive detection of the underwater sound target signal; or, extracting time-frequency image characteristics of Link4A and Link11 data chain signals for detection by a mode matching and clustering related method; or, the edge extraction is carried out on the time-frequency spectrogram of the signal through digital image processing, and whether the frequency hopping signal exists or not is judged according to the result; or, the time-frequency spectrogram of the actual frequency hopping signal is subjected to image segmentation to separate background noise, and the detection of the frequency hopping signal is realized through template matching. Because the ultrashort wave specific signal has obvious visual characteristics on a time-frequency spectrogram, the detection problem of the specific signal can be researched from the angle of image target detection. However, from the current research results, many existing methods have poor detection and identification effects under the conditions of low signal-to-noise ratio and strong channel interference, and the extracted features cannot well characterize signals affected by the strong channel interference.
Disclosure of Invention
Therefore, the invention provides an ultrashort wave specific signal reconnaissance method based on a spectrogram and a deep convolutional network, which realizes ultrashort wave specific signal detection and classification and identification and improves the signal identification rate.
According to the design scheme provided by the invention, the ultrashort wave specific signal reconnaissance method based on the spectrogram and the deep convolutional network comprises the following contents:
carrying out short-time Fourier transform on specific signals in a sample library to obtain a signal time-frequency map, wherein the specific signals are signals containing frame synchronization codes in a signal transmission data frame structure;
training a deep convolutional neural network model by using a time-frequency map, and predicting a position target by using feature maps of different scales through a feature pyramid in the training process;
and carrying out target detection and identification on a specific signal in the ultra-short wave communication by using the trained deep convolutional neural network model.
In the above, the signal transmission data frame structure includes a frame synchronization code and a data frame, wherein the frame synchronization code is a regular data frame in the signal transmission data frame structure.
As described above, the deep convolutional neural network uses conv1 to conv5 of VGG16 as a base network, and convolution layers conv6 to conv11 are additionally added, wherein the convolution layer size in conv6 to conv11 decreases layer by layer.
Preferably, feature capture boxes with corresponding sizes are designed for feature maps with different scales in the deep convolutional neural network, and the target category and the real target frame in the capture box are predicted by extracting the feature map features corresponding to the feature capture boxes; and taking the position deviation between the real target frame and the feature capture box and the target class probability contained in each feature capture box as loss values for neural network training.
Preferably, for the feature capture boxes matched with each real target frame, the intersection ratio of the feature capture boxes is obtained, and the feature capture box with the largest intersection ratio is selected as the best prediction result.
Preferably, in the deep convolutional neural network, by fusing different scale feature maps, the detail features of the A-scale feature map and the semantic features of the B-scale feature map are combined, and a mixed feature map is formed by combining according to channels, wherein the A scale is larger than the B scale.
Preferably, feature capture boxes with corresponding sizes are designed for feature graphs with different scales in the deep convolutional neural network, real target frames are clustered, scale clustering is carried out by using the intersection and parallel ratio of the real target frames and the feature capture boxes, and the length-width ratio parameters of the feature capture boxes are obtained; and predicting the object type and the real frame of the object in the grabbing box by extracting the feature graph features corresponding to the feature grabbing box.
Further, the ratio of the time of existence of the signal to the bandwidth is used as the aspect ratio parameter of the feature capture box.
Further, clustering is carried out on the real target frame through a K-means clustering method.
In the deep convolutional neural network, 7 additional convolutional layers conv 10-conv 16 are added to obtain a series of feature maps with different sizes; and training and detecting and identifying the characteristic graphs with different sizes after batch regularization treatment.
The invention has the beneficial effects that:
the method comprises the steps of firstly analyzing the visual characteristics of a specific signal on a time-frequency spectrogram, and then detecting and identifying the signal by using a deep convolutional neural network model, so that the problems of poor detection and identification efficiency and the like under the conditions of low signal-to-noise ratio and strong channel interference in the conventional method are solved, the identification of the ultrashort wave specific signal is realized, and the signal identification rate is improved; and according to the simulation experiment result, the effectiveness of the deep convolutional neural network applied to specific signal detection and identification is further verified, the performance is stable, the operation is efficient, an idea is provided for subsequent related researches in the field, and the deep convolutional neural network has a high practical application value.
Description of the drawings:
FIG. 1 is a schematic diagram of an ultrashort wave specific signal identification process in an embodiment;
FIG. 2 is a diagram illustrating a frame structure for transmitting data according to an embodiment;
FIG. 3 is a diagram illustrating a traffic channel header frame in an embodiment;
FIG. 4 is a time-frequency spectrum diagram of a traffic channel transmit signal in an embodiment;
FIG. 5 is a time-frequency spectrum diagram of a LINK11-CLEW signal in the example;
FIG. 6 is a time-frequency spectrum of FM-LINK11 signal in the example;
FIG. 7 is a diagram of an exemplary SSD network model architecture;
FIG. 8 is a schematic view of a feature capture box design in an embodiment;
FIG. 9 is a schematic diagram of an embodiment of a feature capture box prediction network;
FIG. 10 is a diagram illustrating an exemplary deep convolutional neural network model architecture;
FIG. 11 is a schematic diagram of a hybrid signature diagram in an embodiment;
FIG. 12 is a diagram illustrating the visual characteristics of different types of specific signals in time-frequency spectrum in an embodiment;
FIG. 13 is a schematic diagram of a time-frequency spectrum diagram of a finally generated specific signal in the embodiment;
FIG. 14 is a graph comparing model performance at different signal-to-noise ratios in the examples;
FIG. 15 is a diagram illustrating the detection effect of models on specific signals under different SNR in the example.
The specific implementation mode is as follows:
the present invention will be described in further detail below with reference to the accompanying drawings and technical solutions, and embodiments of the present invention will be described in detail by way of preferred examples, but the embodiments of the present invention are not limited thereto.
An image target detection and identification method based on a Deep learning technology becomes mainstream, breaks through the traditional image processing and machine learning algorithm processing process, and can simultaneously perform target identification and target positioning on an image by using a Deep Convolution Network (DCNN). To this end, referring to fig. 1, an embodiment of the present invention provides an ultrashort wave specific signal detection method based on a spectrogram and a deep convolutional network, including the following contents:
101. carrying out short-time Fourier transform on specific signals in a sample library to obtain a signal time-frequency map, wherein the specific signals are signals containing frame synchronization codes in a signal transmission data frame structure;
102. training a deep convolutional neural network model by using a time-frequency map, and predicting a position target by using feature maps of different scales through a feature pyramid in the training process;
103. and carrying out target detection and identification on a specific signal in the ultra-short wave communication by using the trained deep convolutional neural network model.
Aiming at the problems of specific signal detection and identification in ultra-short wave communication, the time-frequency spectrogram and the deep convolutional neural network are combined, the idea of the traditional signal detection method is broken through, signal detection, time-frequency positioning and classification identification can be realized at the same time, the effectiveness and the reliability of signal identification are improved, and the method has a high practical application value.
Referring to fig. 2, in another embodiment of the present invention, the signaling data frame structure includes a frame synchronization code and a data frame, wherein the frame synchronization code is a regular data frame in the signaling data frame structure. The frame synchronization code (also called header frame) refers to a specific regular data frame in the signaling frame structure data, and generally plays roles of frame synchronization, symbol synchronization, carrier synchronization, and the like in the front part of the data frame. Such special signals are commonly present in the short-wave, ultrashort-wave, microwave frequency bands, and are typically generated by time division multiplex communication systems. The time-frequency spectrogram is obtained by a short-time Fourier transform (STFT) method of the signals, reflects the change condition of the energy density of the signals along with time and frequency, and can obtain the time-frequency distribution characteristics of the signals by observing the time-frequency spectrogram. The specific signal can present unique visual characteristics on a time-frequency spectrogram, and the specific signals with different frame structures and different modulation modes can also generate different visual characteristics. The iridium satellite communication system and LINK11 will be described below as an example.
The iridium communication system is a global mobile communication system, and provides main services such as mobile phone, paging, and low-speed data transmission for users, and the frame structure of the traffic channel is shown in fig. 3, the length of the header frame of the traffic channel is 54 bits, and the header frame contains a plurality of repeated "1" bits. The repeated code words are subjected to DQPSK modulation to generate unique frequency characteristics. The time-frequency spectrum diagram of the traffic channel transmitted signal in fig. 4 shows that the header frame portion of the signal exhibits different visual characteristics than the following data frame portion, which can be a feature for identifying the signal.
The LINK11 datalink is a tactical data communication. The time-frequency spectrum of the conventional LINK11 signal is shown in fig. 5. As can be seen, LINK11 is a multiplexed signal with two audio header frames and 16 audio data frames. The LINK11 data chain is transmitted in a complex modulation mode. In the ultrashort wave band, the LINK11_ clean signal is modulated onto the carrier wave in the FM mode. The time-frequency spectrum of the LINK11_ CLEW signal after FM (FM _ LINK11 for short) is shown in FIG. 6. The FM _ LINK11 signal header frame also exhibits different visual characteristics than the data frame.
The short-time Fourier transform STFT is a basic method for time-frequency analysis, the visual characteristics of a specific signal are analyzed by applying the STFT method in the embodiment of the invention, so that the problem of cross terms caused by WVD and other transforms can be effectively avoided, the wavelet transform has no overlarge calculated amount, the time-frequency characteristics of the signal can be clearly reflected particularly under the condition of strong aliasing interference, and the method is very suitable for specific signal analysis. Therefore, the embodiment of the invention finds that the specific signal has obvious visual characteristics on the time-frequency spectrogram, the specific signals with different frame structures and different modulation modes can present different visual characteristics on the time-frequency spectrogram, the visual characteristics of the specific signal on the time-frequency spectrogram can be clearly reflected by applying the STFT method, and the visual characteristics can become the characteristics for identifying the specific signal; this visual characteristic is extracted by applying the STFT method to a specific signal, thereby realizing specific signal recognition.
The SSD, called Single Shot multi box Detector, is one of the main detection frameworks for the target detection algorithm at present, and has an obvious speed advantage compared to fast RCNN and an obvious performance advantage compared to YOLO. SSD has the following main features: the idea of converting detection into regression is inherited from the YOLO, and multi-target detection and identification can be completed only by observing the image once; based on an anchor mechanism in fast RCNN, a similar prior box is proposed; adding a detection mode of a characteristic pyramid (Central Feature Hierarchy), predicting the target at each position by using characteristic graphs with different scales, and comparing the accuracy with that of fast RCNN; and the higher detection and identification precision can be ensured for the image with lower resolution. Therefore, the embodiment of the invention uses the SSD network to extract the visual characteristics on the time-frequency spectrogram of the specific signal, overcomes the conditions of low signal-to-noise ratio and strong interference on the ultrashort wave channel, and realizes the detection and identification of the ultrashort wave specific signal. In another embodiment of the present invention, the SSD network model structure is shown in fig. 7, and can be mainly divided into the following parts: the very beginning of the SSD model is a base network (base network) used to extract depth features of the entire input picture, while the SSD network uses conv1 to conv5 of VGG16 as the base network. After the basic network structure, additional convolutional layers conv6 to conv11 are added, and the scale sizes of the convolutional layers are gradually reduced layer by layer, so that the extraction of multi-scale features is facilitated. Preferably, feature capture boxes with corresponding sizes are designed for feature maps with different scales in the deep convolutional neural network, and the target category and the real target frame in the capture box are predicted by extracting the feature map features corresponding to the feature capture boxes; and taking the position deviation between the real target frame and the feature capture box and the target class probability contained in each feature capture box as loss values for neural network training. As shown in fig. 8, different sized feature capture boxes (default boxes) are designed for different scales of depth feature maps (conv4_3, conv7, conv8_2, conv9_2, conv10_2, conv11_2), and the depth feature maps of different scales are divided into 8 × 8 or 4 × 4 grids, where a feature capture box is a series of fixed-sized boxes on each grid, that is, a series of frames formed by dashed lines in fig. 8.
Firstly, different scale parameters are designed for feature maps with different scales, and if m feature maps exist, the scale parameters are as follows:
where m is the number of feature maps, minimum SminIs 0.2, maximum value SmaxIs 0.9. The aspect ratio parameters are:
is characterized in that the width of the grabbing box is
Has a height of
Further, the target category and the real border of the target in the box are predicted by extracting the features of the depth feature map corresponding to the feature capture boxes. As shown in fig. 9, in each depth feature map with different scales, the deviation (offsets) of the position between the real target frame and the feature capture box and the class probability (scores) of the target contained in each feature capture box are obtained, and these probabilities and deviations are used as loss values for training. The best prediction results were screened by Non-maximum inhibition (Non-maximum suppression), as shown in fig. 8. And calculating the intersection ratio (jaccard overlap) of some feature capture boxes matched with each real target frame, and selecting the feature capture box with the maximum intersection ratio as the best prediction result. The objective function of SSD training may be designed as:
obviously, the loss value of the SSD model is divided into two parts, namely confidence loss and location loss, where N is the number of feature capture boxes matched to the real target frame; and the α parameter is used to adjust the ratio between confidence loss and location loss, and α is 1 by default. Confidence loss in SSDs is typically softmax loss:
wherein
Represents the firstThe i feature capture boxes are matched to the jth real target frame with the category p,
representing the confidence coefficient that the ith feature capture box is matched with the p-th class target;
while the location loss in SSD is used to measure the predicted performance of the bounding box, using the typical smooth L1 loss:
wherein
Representing the deviation between the jth real target frame and the ith feature capture box, wherein m belongs to { cx, cy, w, h }, (cx, cy) represents the coordinate of the center point of the frame, and (w, h) represents the width and height of the frame.
Wherein
In summary, for ultrashort wave specific signal detection and identification, in the embodiment, multi-scale information is added to the SSD deep convolutional neural network model, and visual characteristics on a time-frequency spectrogram of a specific signal are extracted from feature maps of different layers of the deep neural network to perform regression prediction on a target, so that more judgments are made, and speed and performance are considered at the same time. And the convolution layer is used end to end, so that the influence of noise and interference on a time-frequency spectrogram can be effectively reduced.
In the field of image detection and identification, the speed and the performance of an SSD network model can achieve satisfactory effects. However, the SSD model still does not solve the influence of the contradiction between the positioning and the recognition on the detection result, the large-size characteristic diagram has a good positioning capability and a poor recognition capability, and the small-size characteristic diagram is opposite, so that the characterization capability of the small target is not strong enough, and an ideal effect is not achieved when the SSD model is used to detect some specific signals with a narrow bandwidth and a short occurrence time in the ultra-short wave band. In another embodiment of the present invention, as shown in fig. 10, an SSD network model is improved according to the characteristics of the ultrashort wave specific signal, and the detection and identification performance is improved on the premise of maintaining the speed advantage, wherein the improvement content is as follows:
(1) and fusing the characteristic diagrams, combining the detail characteristics of the A-scale characteristic diagram and the semantic characteristics of the B-scale characteristic diagram by fusing different-scale characteristic diagrams, and combining according to channels to form a mixed characteristic diagram, wherein the A scale is larger than the B scale. The detail features (positioning) of the large-size feature map and the semantic features (recognition) of the small-size feature map are combined to solve the contradiction between the positioning and the recognition existing in the SSD network. As shown in fig. 11, Conv4_3, Conv7 and Conv8_2 are combined, the dimensions of the three layers are reduced to 256 by convolution with 1x1, then Conv7 and Conv8_2 are bilinearly interpolated to the same size as Conv4_3, and finally the three layers are combined by channels to form a mixed feature map.
(2) The method for improving and solving the length-width ratio parameter of the feature capture box comprises the steps of designing the feature capture boxes with corresponding sizes for feature graphs with different scales in a deep convolutional neural network, clustering real target frames, carrying out scale clustering by using the intersection and parallel ratio of the real target frames and the feature capture boxes, and obtaining the length-width ratio parameter of the feature capture boxes; and predicting the object type and the real frame of the object in the grabbing box by extracting the feature graph features corresponding to the feature grabbing box. Because the detection network is used for fitting the real target frame by finely adjusting the feature capture box, and the prior information of the feature capture box has very obvious influence on the prediction result, the method replaces manual setting of the length-width ratio parameter of the feature capture box, and clustering the real target frame of the data set by using a clustering method such as K-means and the like, and uses the intersection and the comparison of the real target frame and the feature capture box as a clustering scale to cluster the length-width ratio parameter of the feature capture box. For an ultrashort wave specific signal, and further using the ratio of the time of existence to the bandwidth of the signal as the aspect ratio parameter of the feature capture box, and the aspect ratio parameter of the specific signal may greatly exceed the conventional aspect ratio range, the detection effect of the improved method on the specific signal can be very obvious.
(3) The pre-trained model parameters are improved. In general, a method of transfer learning is applied to pre-train a network model when a target detection network is trained. For the SSD network model, it needs to be pre-trained using the parameters of the VGG16 model. Embodiments pre-train the network model herein by improving the pre-trained model parameters, replacing the VGG16 model parameters with the SSD's model parameters.
(4) Carrying out batch regularization treatment, and adding 6 additional convolution layers conv 10-conv 16 to obtain a series of characteristic graphs with different sizes; and training or learning after batch regularization processing is carried out on the feature maps with different sizes. After the mixed feature map is formed, adding some extra convolutional layers (conv10, conv11, conv12, conv13, conv14, conv15 and conv16) to obtain a series of feature maps with different sizes, and performing batch regularization on the feature maps, and inputting the feature maps into a prediction network for detection. The batch regularization processing can accelerate the convergence speed in the training process and avoid falling into the condition of local optimum.
To further verify the effectiveness of the technical scheme of the present invention, the following further explanation is made through specific simulation experimental data:
a simulation experiment is carried out aiming at ultra-short wave specific signal broadband spectrum detection, four representative specific signals including iridium DQPSK, MIL-STD SOQPSK, LINK4A and FM-LINK11 are selected as data samples, wherein the sampling rates are all 1MHz, and the specific information is shown in Table 1.
TABLE 1 data set information
The time-frequency spectrogram is generated in an MATLAB environment, different types of specific signals have different visual characteristics on the time-frequency spectrogram, and the different visual characteristics are mainly reflected in the shape, size and amplitude of a header frame, such as (a) an iridium DQPSK spectrogram, (b) an MIL-STD SOQPSK spectrogram, (c) a LINK4A spectrogram and (d) an FM-LINK11 spectrogram in FIG. 12.
In order to balance the time resolution and the frequency resolution of the time-frequency spectrogram, and synthesize basic parameters of the ultrashort wave specific signals to obtain more remarkable visual characteristics, the four specific signals are firstly superposed according to different carrier frequencies, and then 1000-point short-time Fourier transformation is carried out to obtain the time-frequency spectrogram with the size of 1000 x 1000, wherein the frequency resolution is 1 KHz. There are four non-overlapping signals in the time-frequency spectrogram, each of which has random horizontal offset (delay) and vertical offset (frequency offset), and in order to meet the situation of the ultrashort wave actual channel, various actual ultrashort wave interferences are also superimposed in the spectrogram. Fig. 13 is a time-frequency spectrum diagram of four types of specific signals finally generated, wherein the four types of specific signals and the superimposed actual ultrashort wave channel interference can be seen.
Model training and testing are completed by calling a TensorFlow deep learning library issued by Google under an Anaconda3 platform, wherein the programming language is Python language, and the image processor is 6GB NVidia GeForce GTX 1080. In the range of 0dB to 18dB of signal-to-noise ratio (ES/n0), 400 specific signal spectra were generated every 2dB, with 4000 training and test sets. The learning rate was 0.001, the learning attenuation rate was 0.99, the regularization coefficient was 0.0004, the batch normalized attenuation was 0.99, the number of training times was 50000, and 20 experiments were performed.
The detection performance of the SSD model and the original SSD model on the ultrashort wave specific signals is improved under different signal-to-noise ratios (Es/n0), as shown in FIG. 14, wherein the average detection rate represents the average detection and identification probability of the different models on four types of ultrashort wave specific signals under the condition that the false alarm rate is zero. As can be seen from fig. 14, as the signal-to-noise ratio (Es/n0) increases, the improved SSD network model and the average detection rate of the SSD network model for a specific signal are both improved, wherein the performance of the improved network model of the present invention is obviously better than that of the SSD network model. When the signal-to-noise ratio is larger than 4dB, the average recognition rate of the method in the improved network model of the invention basically reaches 100%, and when the signal-to-noise ratio reaches 0dB, the average recognition rate can still reach more than 95%. And the SSD model is poor in detection mainly because of poor detection capability for small target signals. The improved SSD network model in the scheme of the invention has good anti-noise capability aiming at the method for improving the network model by the ultrashort wave specific signal, and can be suitable for the detection and identification of the specific signal under the low signal-to-noise ratio.
The specific operations for optimizing the SSD network model for ultrashort wave specific signals are shown in table 3. The method comprises the steps of obtaining a characteristic capture box, designing a characteristic capture box by using a K-means clustering method, designing a Batch Norm by using an additional characteristic diagram, and obtaining an average detection recognition probability by testing all test sets with signal-to-noise ratios ranging from 0dB to 18dB by using an average detection rate.
TABLE 3 SSD model improved performance comparisons
| Model (model) | Feature map fusion | K-means | Pre-trained VGG16 | Pre-training SSD | Batch Norm | |
| 1 | ╳ | ╳ | √ | ╳ | ╳ | 0.89 |
| 2 | ╳ | √ | ╳ | √ | ╳ | 0.96 |
| 3 | √ | ╳ | √ | ╳ | ╳ | 0.91 |
| 4 | √ | √ | √ | ╳ | ╳ | 0.96 |
| 5 | √ | √ | ╳ | √ | ╳ | 0.98 |
| 6 | √ | √ | ╳ | √ | √ | 0.99 |
As can be seen from table 3, the model in the scheme of the present invention is obtained by gradually improving the SSD model, and the performance of the model is improved by 11.1% compared with the SSD model. The method for fusing the feature maps is improved by 2.2%, the method for designing the feature capture box by K-means is improved by 5.5%, the parameters of the improved pre-training model are improved by 2%, the batch regularization treatment on the feature maps is improved by 1%, and the effectiveness of improving the SSD model for the ultrashort wave specific signals is verified. Meanwhile, the method for grabbing the box by the K-means design features has the most improvement effect as can be seen from the table 3, because the aspect ratio of the MIL-STD SOQPSK signal is 7.1 as seen from the table 1, which greatly exceeds the conventional aspect ratio
The detection effect is influenced, so that for specific signal detection, the effect is improved most by using the method of designing the feature capture box by using the K-means, and the fact that the prior information of the feature capture box is particularly important for the detection of specific signals is verified.
To analyze the effect of the fused profile method on small targets in more detail, the test compares the detection effect of model 2 and model 5 on four specific signals in table 3 below under different signal-to-noise ratios, as shown in fig. 15. In FIG. 13, the Iridium DQPSK signal and the LINK4A signal are small target signals, while the MIL-STD SOQPSK signal and the FM-LINK11 signal are relatively large target signals. As can be seen from fig. 15, model 5 has better detection effect on small target signals than model 2; further verifies that the method for fusing the characteristic graphs can improve the detection performance of small target signals. In conclusion, for the ultrashort wave specific signal identification, the SSD network model is effectively improved by optimization, the detection capability of the SSD network model on small target signals is improved, and meanwhile, the method has better anti-noise interference capability and is an effective ultrashort wave specific signal detection and identification method.
According to the invention, by analyzing and researching visual characteristics on the ultra-short wave specific signal time-frequency spectrogram, the large difference of different types of specific signal time-frequency spectrograms is found, the image target detection and identification method is applied to the field of signal detection and identification, and a deep convolutional neural network model is introduced, so that the traditional detection and identification algorithm processing mode is broken, and signal detection, time-frequency positioning and classification identification can be simultaneously carried out. Experimental results show that the method can effectively reduce the interference influence on an ultrashort wave channel, realize ultrashort wave specific signal detection and identification under low signal-to-noise ratio, and improve the detection performance by optimizing network structures and other methods. In the specific simulation experiment of the invention, only four types of specific signals are selected, and the type number of the specific signals can be increased to continue the experimental study.
The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other. The device disclosed by the embodiment corresponds to the method disclosed by the embodiment, so that the description is simple, and the relevant points can be referred to the method part for description.
The elements of the various examples and method steps described in connection with the embodiments disclosed herein may be embodied in electronic hardware, computer software, or combinations of both, and the components and steps of the examples have been described in a functional generic sense in the foregoing description for clarity of hardware and software interchangeability. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the solution. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.
Those skilled in the art will appreciate that all or part of the steps of the above methods may be implemented by instructing the relevant hardware through a program, which may be stored in a computer-readable storage medium, such as: read-only memory, magnetic or optical disk, and the like. Alternatively, all or part of the steps of the foregoing embodiments may also be implemented by using one or more integrated circuits, and accordingly, each module/unit in the foregoing embodiments may be implemented in the form of hardware, and may also be implemented in the form of a software functional module. The present invention is not limited to any specific form of combination of hardware and software.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims (6)
1. An ultrashort wave specific signal reconnaissance method based on a spectrogram and a deep convolutional network is characterized by comprising the following contents:
carrying out short-time Fourier transform on specific signals in a sample library to obtain a signal time-frequency spectrogram, wherein the specific signals are signals containing frame synchronization codes in a signal transmission data frame structure;
training a deep convolution neural network model by using a time-frequency spectrogram, and predicting a position target by using feature maps of different scales through a feature pyramid in the training process;
carrying out target detection and identification on a specific signal in the ultra-short wave communication by using the trained deep convolutional neural network model;
the signal transmission data frame structure comprises a frame synchronization code and a data frame, wherein the frame synchronization code is a regular data frame in the signal transmission data frame structure; the deep convolutional neural network uses conv1 to conv5 of VGG16 as a basic network, and convolution layers conv6 to conv11 are additionally added, wherein the sizes of the convolution layers in conv6 to conv11 are gradually decreased layer by layer; designing feature capture boxes with corresponding sizes for feature maps with different scales in the deep convolutional neural network, and predicting the object type and the real frame of the object in the capture boxes by extracting the feature map features corresponding to the feature capture boxes; taking the position deviation between the real target frame and the feature capture boxes and the target class probability contained in each feature capture box as loss values for neural network training;
supposing that m characteristic graphs are provided, the scale parameters are as follows:
wherein m is the number of feature maps and the minimum value is SminMaximum value of SmaxUsing the ratio of the time of existence of the signal to the bandwidth as the aspect ratio parameter a of the feature capture boxrCharacterized by a width of the grasping box of
Has a height of
In the deep convolutional neural network, through fusing different scale feature maps, combining the detail features of the A scale feature map and the semantic features of the B scale feature map, and combining according to channels to form a mixed feature map, wherein the A scale is larger than the B scale.
2. The method of claim 1, wherein for each feature capture box matched with each real target frame, the cross-over ratio of the feature capture boxes is obtained, and the feature capture box with the largest cross-over ratio is selected as the best prediction result.
3. The method for detecting ultrashort wave specific signals based on spectrogram and deep convolutional network as claimed in claim 1, wherein feature capture boxes with corresponding sizes are designed for different scale feature maps in the deep convolutional neural network, real target frames are clustered, and scale clustering is performed by using the intersection ratio of the real target frames and the feature capture boxes to obtain the aspect ratio parameters of the feature capture boxes; and predicting the object type and the real frame of the object in the grabbing box by extracting the feature graph features corresponding to the feature grabbing box.
4. The method of claim 3, wherein the ratio of the time of existence of the signal to the bandwidth is used as the aspect ratio parameter of the feature capture box.
5. The spectrogram and deep convolutional network-based ultrashort wave specific signal reconnaissance method as claimed in claim 3, wherein the real target borders are clustered by a K-means clustering method.
6. The method for detecting ultrashort wave specific signals based on spectrogram and deep convolutional network as claimed in claim 1, wherein 7 additional convolutional layers conv 10-conv 16 are added in the deep convolutional neural network to obtain a series of feature maps with different sizes; and training and detecting and identifying the characteristic graphs with different sizes after batch regularization treatment.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811159919.XA CN109344772B (en) | 2018-09-30 | 2018-09-30 | Ultrashort wave specific signal reconnaissance method based on spectrogram and deep convolutional network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811159919.XA CN109344772B (en) | 2018-09-30 | 2018-09-30 | Ultrashort wave specific signal reconnaissance method based on spectrogram and deep convolutional network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109344772A CN109344772A (en) | 2019-02-15 |
| CN109344772B true CN109344772B (en) | 2021-01-26 |
Family
ID=65307994
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811159919.XA Active CN109344772B (en) | 2018-09-30 | 2018-09-30 | Ultrashort wave specific signal reconnaissance method based on spectrogram and deep convolutional network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109344772B (en) |
Families Citing this family (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110048977B (en) * | 2019-03-14 | 2022-03-01 | 中国人民解放军战略支援部队信息工程大学 | Short wave signal system identification method and device based on gray level co-occurrence matrix texture feature detection |
| CN110059555B (en) * | 2019-03-14 | 2021-04-20 | 中国人民解放军战略支援部队信息工程大学 | Signal modulation classification identification method and device based on multi-resolution spectrogram watershed image segmentation and decision tree |
| CN111079665A (en) * | 2019-12-20 | 2020-04-28 | 长沙深之瞳信息科技有限公司 | Morse code automatic identification method based on Bi-LSTM neural network |
| CN111325290B (en) * | 2020-03-20 | 2023-06-06 | 西安邮电大学 | Traditional Chinese painting image classification method based on multi-view fusion multi-example learning |
| CN111551920A (en) * | 2020-04-16 | 2020-08-18 | 重庆大学 | Three-dimensional target real-time measurement system and method based on target detection and binocular matching |
| CN111541511B (en) * | 2020-04-20 | 2022-08-16 | 中国人民解放军海军工程大学 | Communication interference signal identification method based on target detection in complex electromagnetic environment |
| CN111814953B (en) * | 2020-06-16 | 2024-02-13 | 上海瀚讯信息技术股份有限公司 | Positioning method of deep convolution neural network model based on channel pruning |
| CN111558542A (en) * | 2020-06-24 | 2020-08-21 | 重庆视智迪科技有限公司 | Ceramic tile surface quality online detection sorting system and method |
| CN111680689B (en) * | 2020-08-11 | 2021-03-23 | 武汉精立电子技术有限公司 | Target detection method, system and storage medium based on deep learning |
| CN112464837B (en) * | 2020-12-03 | 2023-04-07 | 中国人民解放军战略支援部队信息工程大学 | Shallow sea underwater acoustic communication signal modulation identification method and system based on small data samples |
| CN112751629A (en) * | 2021-01-15 | 2021-05-04 | 中国人民解放军战略支援部队信息工程大学 | Broadband specific signal detection method based on time-frequency image processing |
| CN113312996B (en) * | 2021-05-19 | 2023-04-18 | 哈尔滨工程大学 | Detection and identification method for aliasing short-wave communication signals |
| CN113376610B (en) * | 2021-06-22 | 2023-06-30 | 西安电子科技大学 | Narrow-band radar target detection method based on signal structure information |
| CN113808197A (en) * | 2021-09-17 | 2021-12-17 | 山西大学 | Automatic workpiece grabbing system and method based on machine learning |
| CN114205821B (en) * | 2021-11-30 | 2023-08-08 | 广州万城万充新能源科技有限公司 | Wireless radio frequency anomaly detection method based on depth prediction coding neural network |
| CN114609604B (en) * | 2022-03-25 | 2023-06-09 | 电子科技大学 | Unmanned aerial vehicle cluster target detection and target contour and cluster scale estimation method |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103220055A (en) * | 2013-05-10 | 2013-07-24 | 厦门大学 | Multi-fractal gradient characteristic fingerprint identification method of wireless transmitter signal |
| CN107122738A (en) * | 2017-04-26 | 2017-09-01 | 成都蓝色起源科技有限公司 | Automatic Communication Signals Recognition based on deep learning model and its realize system |
| CN107221005A (en) * | 2017-05-04 | 2017-09-29 | 美的集团股份有限公司 | Object detecting method and device |
| CN108282427A (en) * | 2017-12-15 | 2018-07-13 | 西安电子科技大学 | Radio signal recognition recognition methods based on multiple dimensioned light weight network |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9204448B2 (en) * | 2013-03-26 | 2015-12-01 | Empire Technology Department Llc | Predictive spectral allocation in mobile networks |
| US10133987B2 (en) * | 2014-07-14 | 2018-11-20 | Federated Wireless, Inc. | Categorizing radio signals while obfuscating characteristics of federal radio signals |
-
2018
- 2018-09-30 CN CN201811159919.XA patent/CN109344772B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103220055A (en) * | 2013-05-10 | 2013-07-24 | 厦门大学 | Multi-fractal gradient characteristic fingerprint identification method of wireless transmitter signal |
| CN107122738A (en) * | 2017-04-26 | 2017-09-01 | 成都蓝色起源科技有限公司 | Automatic Communication Signals Recognition based on deep learning model and its realize system |
| CN107221005A (en) * | 2017-05-04 | 2017-09-29 | 美的集团股份有限公司 | Object detecting method and device |
| CN108282427A (en) * | 2017-12-15 | 2018-07-13 | 西安电子科技大学 | Radio signal recognition recognition methods based on multiple dimensioned light weight network |
Non-Patent Citations (3)
| Title |
|---|
| "ＢＮ对ＶＧＧ神经网络的影响研究";陈强普 等;《合肥工业大学学 报（自然科学版）》;20180131;第41卷(第1期);第35-38页 * |
| "FSSD: Feature Fusion Single Shot Multibox Detector";Zuo-Xin Li 等;《arXiv.org》;20171231;第1-10页 * |
| "目标检测网络SSD的区域候选框的设置问题研究";翁昕;《万方数据 知识服务平台》;20180929;第19-66页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109344772A (en) | 2019-02-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109344772B (en) | Ultrashort wave specific signal reconnaissance method based on spectrogram and deep convolutional network | |
| CN109274621B (en) | Communication protocol signal identification method based on depth residual error network | |
| CN109379311B (en) | Ultra-short wave specific signal identification method based on convolutional neural network | |
| CN104135327B (en) | frequency spectrum sensing method based on support vector machine | |
| CN110428842A (en) | Speech model training method, device, equipment and computer readable storage medium | |
| CN114881093B (en) | Signal classification and identification method | |
| CN114358064B (en) | Interference detection device and method based on deep support vector data description | |
| CN116866129A (en) | Wireless communication signal detection method | |
| CN110932807A (en) | Spectrum sensing method of MIMO (multiple input multiple output) system under non-Gaussian noise | |
| CN110991378B (en) | Individual identification method and device for power amplifier | |
| CN117233706A (en) | Radar active interference identification method based on multilayer channel attention mechanism | |
| CN113191224A (en) | Unmanned aerial vehicle signal extraction and identification method and system | |
| Chen et al. | Indoor location method of interference source based on deep learning of spectrum fingerprint features in Smart Cyber-Physical systems | |
| CN116318249A (en) | Short wave frequency hopping signal sorting method based on improved YOLOv5 | |
| CN111753803A (en) | Unmanned aerial vehicle graph signal identification method based on fuzzy clustering and high-order cumulant | |
| CN110048977A (en) | The recognition methods of short-wave signal system and device based on the detection of gray level co-occurrence matrixes textural characteristics | |
| Le et al. | Performance analysis of convolutional neural networks with different window functions for automatic modulation classification | |
| CN115932808A (en) | Passive sonar intelligent detection method based on multi-feature fusion | |
| CN114936570A (en) | Interference signal intelligent identification method based on lightweight CNN network | |
| Tekbıyık et al. | Real-world considerations for deep learning in wireless signal identification based on spectral correlation function | |
| CN116827486B (en) | Blind detection system and method for short-wave communication signals | |
| Fu et al. | The evaluation method of speech interference effect based on deep learning | |
| US20240106683A1 (en) | Co-channel signal classification using deep learning | |
| CN114970618B (en) | Sign language identification method and system based on millimeter wave radar and with environment robustness | |
| CN115598615B (en) | Power spectrum information geometric radar target detection method and device based on sub-band filtering |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |