Communication signal detection and identification method based on improved YOLOv5
By improving the YOLOv5 model, using lightweight structure and feature fusion module, combined with feature attention mechanism, the problem of multi-parameter signal recognition in non-cooperative communication scenarios is solved, real-time identification and efficient deployment of communication signals are realized, and suitable for a variety of terminal devices. .
Patent Information
- Application Number
- CN202410512596.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-26
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2044-04-26
AI Technical Summary
The existing communication signal detection and recognition algorithms are difficult to effectively identify the multi-parameter characteristics of signals in non-cooperative communication scenarios, and the traditional deep learning model is complex in computing, hardware resource dependence, and application deployment is difficult.
The YOLOv5 model is improved, and a lightweight structure and targeted scale feature fusion module are adopted. Combined with the feature attention module, real-time identification of communication signals is achieved through signal time-frequency diagrams, reducing the calculation amount, and convenient deployment in various terminal devices.
Real-time multi-parameter recognition of communication signals is realized, the recognition performance is improved, the calculation is small, suitable for deployment of various platforms, and the application difficulty is reduced.
Smart Images

Figure CN119051798A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of signal detection and recognition, and more specifically to a communication signal detection and recognition method based on improved YOLOv5. Background Art
[0002] Over the past decades, with the rapid development of mobile communication technology, especially the full deployment of 5G in recent years, the Internet of Things has begun to rise, various communication devices have emerged, wireless spectrum resources have become increasingly scarce, and the wireless electromagnetic space has become increasingly complex and changeable. In order to ensure the legal and reasonable use of wireless spectrum resources and maintain the wireless communication environment, spectrum monitoring has become an important research topic in the field of wireless communications.
[0003] Spectrum monitoring is mainly responsible for listening, measuring, analyzing and judging radio communication signals, so as to grasp the use of radio electromagnetic space spectrum, the existence of signal sources, etc., and realize the management and scheduling of radio. In the process of spectrum monitoring, we often face the situation of non-cooperative communication, that is, the monitoring party as the receiving end cannot confirm the prior information of the communication signal with the transmitting end in advance, and the communication signal is a blind signal. The detection and identification of blind signals is the beginning of blind signal processing, and it is also the most important step in spectrum monitoring under non-cooperative communication. Therefore, this paper mainly studies the detection and identification of communication signals under non-cooperative communication.
[0004] Detecting and identifying signals under non-cooperative communication is an important part of spectrum monitoring. Its main goal is to estimate and identify parameters such as the existence, carrier frequency, bandwidth, duration, and modulation method of communication signals, laying the foundation for the subsequent identification and decryption of parameters such as channel coding, source coding, source identity, and communication information.
[0005] Communication signal detection and recognition algorithms have been a hot research topic with the popularization of mobile communication technology, but traditional detection and recognition algorithms often have the following defects: traditional algorithms are extremely dependent on the extraction of key features of communication signals, including instantaneous amplitude, power spectrum, and high-order statistics, and the processing process is complex and computationally intensive.
[0006] With the rapid development of artificial intelligence (AI) technology, especially deep learning, the detection and recognition of communication signals combined with deep learning technology has become a new research direction. The detection and recognition of communication signals is essentially still the extraction and induction of signal features, and classification is completed based on this. Deep learning, by simulating the cognitive structure of the human brain to establish a learning model, obtains powerful feature extraction capabilities, and interprets various input data, has significant advantages in classification problems. Using the powerful feature learning ability of deep learning, designing an efficient signal detection and recognition network model, and realizing a communication signal detection and recognition system under non-cooperative communication has extremely high practical significance.
[0007] Traditional signal detection and recognition algorithms based on deep learning generally only detect single parameter features of communication signals. When detecting multiple signal parameters, multiple algorithms are often required to be combined, which cannot be widely used. In particular, when faced with parameters such as signal duration, carrier frequency, and bandwidth, it is difficult for deep learning models for target classification tasks to achieve classification functions, because the number of these parameter categories is almost impossible to enumerate. Although deep learning has significant advantages in feature learning and fitting classification, the performance of deep learning often depends on complex network models, and the application reasoning of complex network models ultimately depends on local hardware resources, which is not convenient for application deployment. Summary of the invention
[0008] Technical problem: The purpose of the present invention is to provide a communication signal detection and recognition method based on improved YOLOv5. The method proposes a signal detection and recognition scheme that can realize a series of functions such as signal acquisition, processing, training, real-time detection and recognition, and is easy to migrate and deploy on various platforms. The present invention has high recognition performance in real-time detection and recognition of multiple parameters of communication signals, and only uses one model with small calculation amount and low difficulty of application and deployment in various terminal devices.
[0009] Technical solution: In order to achieve the above-mentioned purpose of the invention, the present invention provides a communication signal detection and recognition method based on improved YOLOv5, which can realize real-time recognition and extraction of communication signal existence, bandwidth, carrier frequency, duration, modulation mode and other characteristics through signal time-frequency diagram.
[0010] The method comprises the following steps:
[0011] Step 1, obtain a communication signal frequency spectrum data set: obtain communication signal simulation data through signal simulation, and then convert the signal simulation data into a signal frequency spectrum, obtain a communication signal frequency spectrum target simulation data set and a communication signal frequency spectrum data set measured by instruments and receiving equipment, pre-process the communication signal frequency spectrum, mark it according to the parameters of the communication signal duration, bandwidth, carrier frequency, existence, and modulation mode, and divide it into a training data set and a test data set;
[0012] Step 2, establish an improved YOLOv5 model, that is, use a lightweight structure in the YOLOv5 backbone network, and improve the YOLOv5 model by combining a targeted scale feature fusion module and a feature attention module. The targeted scale feature fusion module is improved based on the bidirectional feature pyramid BiFPN, and the feature attention module is based on the simple parameter-free attention module SimAM attention mechanism.
[0013] Step 3: Input the training data set into the improved YOLOv5 model for training to obtain a trained improved YOLOv5 model;
[0014] Step 4: Input the test data set into the trained improved YOLOv5 model to obtain the detection and recognition results.
[0015] This method specifically includes the following four parts:
[0016] In the first part, a receiving device is used to receive communication signals in real time through cables and antennas, and communication signal data is obtained through analog-to-digital conversion, signal processing, and signal acquisition steps. The communication signal data is converted into a signal time-frequency diagram and transmitted in real time;
[0017] The second part is to establish, train and deploy the improved YOLOv5 model, including: establishing the improved YOLOv5 model, that is, using a lightweight structure in the YOLOv5 backbone network, and improving it by combining the targeted scale feature fusion module and the feature attention module. The targeted scale feature fusion module is improved based on BiFPN, and the feature attention module is based on the SimAM attention mechanism. The activation function uses ReLU, Leaky ReLU, Swish, Mish, PReLU and other activation functions, and the loss function uses IOU, CIOU, Wise IOU and other loss functions; training the improved YOLOv5 model, that is, inputting the training data set into the improved YOLOv5 model for training to obtain the trained improved YOLOv5 model; deploying the improved YOLOv5 model, that is, converting the trained improved YOLOv5 model into a format suitable for deployment, using TensorFlow Lite, ONNX and other formats, and optimize the trained improved YOLOv5 model, using int8, FP16, FP32 and other quantization methods and branch reduction operations, and finally deploy the model to the target platform, using servers, edge devices, and mobile devices.
[0018] In the third part, the real-time transmitted signal time-frequency diagram is input into the deployed improved YOLOv5 model for real-time detection and recognition, using picture or video input;
[0019] In the fourth part, combined with the system operation time and the characteristics of the signal time-frequency diagram, the recognition results of the improved YOLOv5 model are converted into multiple real-time parameters of the communication signal, including converting the positioning box of the communication signal in the time-frequency diagram into parameters such as duration, bandwidth, carrier frequency, and existence, converting the target type of the signal into the type of modulation parameter, and organizing multiple parameters at the same time to form a detection and recognition report that is more in line with the reading habits of the public.
[0020] The lightweight structure is used in the backbone network of YOLOv5. The lightweight structure makes lightweight improvements to the standard Conv unit, CBL unit, Res unit unit, and CSP unit in YOLOv5, specifically including:
[0021] The standard Conv unit is split into lightweight Conv units. First, the input of the standard Conv unit is evenly split into two parts. One part is convolved according to the standard convolution process. The convolution kernel size remains unchanged, but the number of output channels is halved to obtain half of the normal output feature map. The other part first passes through the standard convolution with a convolution kernel size of 1×1, and then the convolution output is grouped and convolved. The convolution kernel size remains unchanged, the number of output channels is halved, and the number of groups is 1 / 4 of the number of input channels to obtain the other half of the output feature map. The output feature maps of the two parts are spliced by the splicing layer and the channels are rearranged to form a complete output feature map.
[0022] The standard Conv unit in the CBL unit is replaced by a lightweight Conv unit. At the same time, the BN layer and ReLU layer superimposed after the standard Conv unit are placed before the output feature map is concatenated. The outputs of the two parts of the convolution process are executed separately to form a lightweight CBL unit.
[0023] The convolution process of the standard Conv unit in the CBL unit is replaced by a grouped convolution process to form a grouped CBL unit;
[0024] The CBL unit in the Res unit is replaced by a lightweight CBL unit and a grouped CBL unit to form a lightweight Res unit: first, the two-way process of the Res unit is changed to an input average split, and then split into two parts; one part replaces the CBL units with convolution kernel sizes of 1×1 and 3×3 in the Res unit with improved CBL units with a convolution kernel of 1×1 and grouped CBL units with a convolution kernel of 3×3 to extract features; the other part uses a separate 1×1 CBL unit, which is responsible for the change of the output feature size dimension but not for feature extraction; finally, the addition layer in the Res unit is replaced by a splicing layer to complete the combination of the output feature maps of the split part, and the channel rearrangement is used to output the complete feature map;
[0025] The CBL unit and Res unit in the CSP unit are replaced by lightweight CBL unit and lightweight Res unit, and the structure is simplified to form a lightweight CSP unit: first, a CBL unit with a convolution kernel of 1×1 and a standard Conv layer with a convolution kernel of 1×1 in the CSP unit are merged into a lightweight CBL unit with a convolution kernel of 1×1, and the subsequent two operations share the output of the lightweight CBL unit; at the same time, the convolution unit structure composed of the convolution layer, splicing layer, BN layer, and ReLU layer at the output end of the CSP unit is deleted, and only the combined output features of the splicing layer are retained, and the subsequent processing is completed by channel rearrangement.
[0026] The targeted scale feature fusion module uses a targeted scale feature fusion module based on BiFPN; the targeted scale feature fusion module based on BiFPN blocks the path of fusion of small-scale features to large-scale features, and large-scale features are only fused with moderate-scale features; the targeted scale feature fusion module is mainly based on moderate scale, and adopts the bidirectional structure in BiFPN to fuse features upward and downward from the moderate scale respectively, so as to realize large-scale and small-scale feature extraction; at the same time, an additional path is added on the moderate scale to fuse the three scale features again, and extract the final moderate-scale feature.
[0027] The feature attention module based on the SimAM attention mechanism is used; the feature attention module based on the SimAM attention mechanism realizes the weight allocation of each neuron through separability and activation function sigmoid, completes the spatial and channel attention allocation, adds the fusion of features of different scales in YOLOv5, and extracts features of different scales in the targeted scale feature fusion module based on BiFPN.
[0028] The preprocessing of the communication signal time-frequency diagram includes: adding noise, changing brightness, cropping, translating, rotating and scaling the data set; and adjusting the duration range, bandwidth range, sampling frequency and power display range of the time-frequency diagram display.
[0029] The labeling according to the parameters of the communication signal, such as duration, bandwidth, carrier frequency, existence, and modulation mode, includes: labeling each image in the training set using a graphic labeling tool, and automatically generating a labeling text file according to rules using a code.
[0030] Beneficial effects: The present invention discloses a communication signal detection and recognition method based on improved YOLOv5, which can realize real-time recognition and extraction of multiple parameter features such as the existence, bandwidth, carrier frequency, duration and modulation mode of the communication signal through the signal time-frequency diagram. The method uses a lightweight improved YOLOv5, that is, a lightweight structure is used in the backbone network to reduce the amount of calculation, and the recognition ability is enhanced by combining a targeted scale feature fusion module and a feature attention module. The method proposes a signal detection and recognition scheme, which can realize a series of functions such as signal acquisition, processing, training, real-time detection and recognition, and is easy to migrate and deploy on various platforms as a whole. The present invention has high recognition performance in real-time detection and recognition of multiple parameters of communication signals, and only uses one model with a small amount of calculation and low difficulty in application and deployment in various terminal devices. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 This is a diagram of a communication signal detection and identification scheme of the present invention;
[0032] Figure 2This is a diagram showing the lightweight improvement of the standard Conv unit of the present invention;
[0033] Figure 3 This is a diagram of the lightweight improvement of the CBL unit of the present invention;
[0034] Figure 4 This is a diagram of the lightweight improvement of the Res unit of the present invention;
[0035] Figure 5 This is a diagram showing the lightweight improvement of the CSP unit of the present invention;
[0036] Figure 6 It is a schematic diagram of the targeted scale feature fusion module of the present invention;
[0037] Figure 7 Schematic diagram of the feature attention module based on the SimAM attention mechanism of the present invention;
[0038] Figure 8 This is a comparison chart of the computational complexity of the improved YOLOv5 and the standard YOLOv5. DETAILED DESCRIPTION
[0039] In order to make the purpose, technical solution and advantages of the embodiment of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly and completely described in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all the embodiments. Based on the embodiment of the present invention, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present invention. It should be noted that the terms used here are only for describing specific embodiments, and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "include" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof. Unless otherwise specified, the relative arrangement of the components and steps described in these embodiments, the numerical expressions and numerical values do not limit the scope of the present invention. At the same time, it should be understood that for ease of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship. The techniques, methods and devices known to ordinary technicians in the relevant fields may not be discussed in detail, but where appropriate, the techniques, methods and devices should be regarded as part of the authorization specification. In all examples shown and discussed herein, any specific value should be interpreted as being merely exemplary and not limiting. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters represent similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further discussed in subsequent drawings.
[0040] The use scenario of the present invention mainly refers to the detection and identification of blind signals in third-party non-cooperative communication scenarios. The specific implementation plan for unknown signals is as follows:
[0041] In the first part, the receiving device is used to receive the communication signal in real time through cables and antennas, and the communication signal data is obtained through analog-to-digital conversion, signal processing, signal acquisition and other steps. The communication signal data is converted into a signal time-frequency diagram and transmitted in real time.
[0042] The second part is to establish, train and deploy an improved YOLOv5 model, including: establishing an improved YOLOv5 model, that is, using a lightweight structure in the YOLOv5 backbone network, and improving it by combining a targeted scale feature fusion module and a feature attention module. The targeted scale feature fusion module is improved based on BiFPN, and the feature attention module is based on the SimAM attention mechanism. The activation function can use ReLU, Leaky ReLU, Swish, Mish, PReLU and other activation functions, and the loss function can use IOU, CIOU, Wise IOU and other loss functions; training the improved YOLOv5 model, that is, inputting the training data set into the improved YOLOv5 model for training to obtain a trained improved YOLOv5 model; deploying the improved YOLOv5 model, that is, converting the trained improved YOLOv5 model into a format suitable for deployment, which can be done using TensorFlow Lite, ONNX and other formats, and optimize the trained improved YOLOv5 model, which can adopt various quantization methods and branch reduction operations such as int8, FP16, FP32, and finally deploy the model to the target platform, which can adopt multiple platforms such as servers, edge devices, mobile devices, etc.
[0043] In the third part, the real-time transmitted signal time-frequency diagram is input into the deployed improved YOLOv5 model for real-time detection and recognition. Multiple input methods such as pictures and videos can be used.
[0044] In the fourth part, combined with the system operation time and the characteristics of the signal time-frequency diagram, the recognition results of the improved YOLOv5 model are converted into multiple real-time parameters of the communication signal, including converting the positioning box of the communication signal in the time-frequency diagram into parameters such as duration, bandwidth, carrier frequency, and existence, converting the target type of the signal into the type of modulation parameter, and organizing multiple parameters at the same time to form a detection and recognition report that is more in line with the public's reading habits.
[0045] Furthermore, the steps for establishing and training the improved YOLOv5 model in the second part are as follows:
[0046] Step 1, obtain a communication signal frequency spectrum data set: obtain communication signal simulation data through signal simulation, and then convert the signal simulation data into a signal frequency spectrum to obtain a communication signal frequency spectrum target simulation data set and a communication signal frequency spectrum data set measured by instruments and receiving equipment, pre-process the communication signal frequency spectrum, mark it according to the communication signal duration, bandwidth, carrier frequency, existence, modulation mode and other parameters, and divide it into a training data set and a test data set;
[0047] Step 2, establish an improved YOLOv5 model, that is, use a lightweight structure in the backbone network of YOLOv5, and improve it by combining a targeted scale feature fusion module and a feature attention module. The targeted scale feature fusion module is improved based on BiFPN, and the feature attention module is based on the simple parameter-free attention module SimAM attention mechanism. The activation function can use ReLU, Leaky ReLU, Swish, Mish, PReLU and other activation functions, and the loss function can use IOU, CIOU, Wise IOU and other loss functions;
[0048] Step 3: Input the training data set into the improved YOLOv5 model for training to obtain a trained improved YOLOv5 model;
[0049] Step 4: Input the test data set into the trained improved lightweight YOLOv5 model to obtain the detection and recognition results.
[0050] Furthermore, the main points of the improved YOLOv5 model in step 2 are as follows:
[0051] The lightweight structure is used in the backbone network of YOLOv5 to make lightweight improvements to the standard Conv unit, CBL unit, Res unit unit, and CSP unit in YOLOv5.
[0052] The standard Conv unit is lightly improved, such as Figure 2 As shown in the figure, the standard Conv unit is split into a lightweight Conv unit. First, the input of the standard Conv unit is evenly split into two parts. One part will be convolved according to the standard convolution process, the convolution kernel size remains unchanged, but the number of output channels is halved, and half of the normal output feature map is obtained; the other part will first pass the standard convolution with a convolution kernel size of 1×1, and then the convolution output will be grouped and convolved, the convolution kernel size remains unchanged, the number of output channels is halved, and the number of groups is 1 / 4 of the number of input channels, to obtain the other half of the output feature map; the output feature maps of the two parts are spliced by the splicing layer and the channels are rearranged to form a complete output feature map.
[0053] The CBL unit is improved to be lightweight, such as Figure 3 As shown in the figure, the standard Conv unit in the CBL unit is replaced by a lightweight Conv unit, and the BN layer and ReLU layer superimposed after the standard Conv unit are placed before the output feature map is concatenated. The outputs of the two parts of the convolution process are executed separately to form a lightweight CBL unit.
[0054] The Res unit is improved to be lightweight, such as Figure 4As shown in the figure, the CBL unit in the Res unit is replaced by a lightweight CBL unit and a grouped CBL unit to form a lightweight Res unit. First, the two-way process of the Res unit is changed to an average input split, which is split into two parts. One part replaces the CBL unit with a convolution kernel size of 1×1 and 3×3 in the Res unit with an improved CBL unit with a convolution kernel of 1×1 and a grouped CBL unit with a convolution kernel of 3×3 to extract features; the other part uses a separate 1×1 CBL unit, which is responsible for the change of the output feature size dimension, but not for feature extraction. Finally, the addition layer in the Res unit is replaced by a splicing layer to complete the combination of the output feature maps of the split part, and the channel rearrangement is used to output the complete feature map.
[0055] The grouped CBL unit is formed by replacing the convolution process of the standard Conv unit in the CBL unit with the grouped convolution process.
[0056] The CSP unit is lightened by improvements such as Figure 5 As shown in the figure, the CBL unit and Res unit in the CSP unit are replaced by lightweight CBL unit and lightweight Res unit, and the structure is simplified to form a lightweight CSP unit. First, the CBL unit with a convolution kernel of 1×1 and the standard Conv layer with a convolution kernel of 1×1 in the CSP unit are merged into a lightweight CBL unit with a convolution kernel of 1×1, and the subsequent two operations share the output of the lightweight CBL unit. At the same time, the convolution unit structure composed of the convolution layer, splicing layer, BN layer, and ReLU layer at the output end of the CSP unit is deleted, and only the combined output features of the splicing layer are retained, and the subsequent processing is completed by channel rearrangement.
[0057] The joint targeted scale feature fusion module is improved by replacing the FPN+PAN in YOLOv5 with a targeted scale feature fusion module based on BiFPN.
[0058] The BiFPN-based targeted scale feature fusion module, such as Figure 6 As shown in the figure, the path for the fusion of small-scale features to large-scale features is cut off, and large-scale features are only fused with moderate-scale features. The targeted scale feature fusion module is mainly based on moderate scale, and adopts the bidirectional structure in BiFPN to fuse features from moderate scale upward and downward respectively to achieve large-scale and small-scale feature extraction; at the same time, an additional path is added to the moderate scale to fuse the three scale features again, and the final moderate-scale feature is extracted from them.
[0059] The joint feature attention module is improved to be a feature attention module based on the SimAM attention mechanism. The feature attention module based on the SimAM attention mechanism is as follows: Figure 7 As shown in the figure, the weight allocation of each neuron is realized by combining separability with the activation function sigmoid, completing the spatial and channel attention allocation, adding the fusion of features of different scales in YOLOv5 and the feature extraction of different scales in the targeted scale feature fusion module based on BiFPN.
[0060] The present invention demonstrates two experimental examples to evaluate the recognition performance of a communication signal detection and recognition method based on an improved YOLOv5 and the application performance of a communication signal detection and recognition scheme based on an improved YOLOv5.
[0061] In Experimental Example 1, a signal simulation time-frequency diagram dataset is used, with a total of 11 types of communication signals with modulation modes, including BPSK, QPSK, QAM8, 16QAM, QAM32, 64QAM, AM, AM-DSB, AM-SSB, FM, and FSK. When conducting a signal existence test, all modulation signal samples are divided into one type of signal samples, namely, the target communication signal samples.
[0062] The parameters of the signal simulation time-frequency graph dataset are shown in Table 1. Since YOLOv5 image target detection will adaptively scale the target image, the image size of the signal simulation time-frequency graph dataset is not 640×640, but the size of all images remains the same. The signal simulation time-frequency graph dataset is mainly composed of low-frequency narrowband signal graphs, but the results are still of practical significance for high-frequency broadband signals. When the high-frequency broadband signal is processed into a time-frequency graph, compared with the overall time-frequency graph, the high-frequency broadband signal graph is a "low-frequency narrowband signal graph" in a relative sense. The image features are basically the same, and the relative frequency band information of the signal is stored in the time-frequency graph. The noise environment of the signal simulation time-frequency graph dataset is -20dB, -10dB, 0dB, 10dB, and 20dB according to the ratio of the overall average power of the signal sequence to the channel noise. All signals in a single time-frequency graph are in the same noise environment, and the ratio of the test set to the training set is 9:1.
[0063] The signal simulation time-frequency graph data set is divided into 6 subsets according to the noise environment, namely, the mixed noise environment data set, which contains 500 time-frequency graphs of noise environments of -20dB, -10dB, 0dB, 10dB, and 20dB, for a total of 2500 graphs; the noise environment is a -20dB data set, which contains 1000 -20dB time-frequency graphs; the noise environment is a -10dB data set, which contains 1000 -10dB time-frequency graphs; the noise environment is a 0dB data set, which contains 1000 0dB time-frequency graphs; the noise environment is a 10dB data set, which contains 1000 10dB time-frequency graphs; the noise environment is a 20dB data set, which contains 1000 20dB time-frequency graphs. There are 7500 time-frequency graphs in the time-frequency graph data set, and the number of signal samples exceeds 45,000.
[0064] System configuration: CPU+GPU is an edge terminal based on the Jetson Orin module, the operating system is Ubuntu20.04, the software environment is Python 3.8, CUDA 11.4, and the software framework is PyTorch V1.12+torchvisionv0.13.0.
[0065] like Figure 8 As shown in the figure, it is a comparison chart of the computational amount of the improved YOLOv5 of the present invention and the standard YOLOv5, which shows the computational gap between different versions of YOLOv5. FLOPs is the number of floating point operations, which is the computational amount used to measure the complexity of the model, and GFLOPs is 1 billion floating point operations. Figure 8 As shown in the figure, the improved YOLOv5 is adapted to different versions of YOLOv5, and different computational loads are displayed with different versions. Compared with the standard YOLOV5 versions, the computational load of the improved YOLOv5 versions is greatly reduced, and the difference in the computational load of the corresponding versions is at least 1.6 times and at most 5.7 times. Among them, the computational load of the improved YOLOv5s is only 5.1GFLOPs, which is close to the computational load of the standard YOLOv5n, while the lightweight YOLOv5n has only 1.6GFLOPs of computational load. The smaller computational load requirement is greatly conducive to the deployment and application of the improved YOLOv5 on different platforms.
[0066] As shown in Table 2, it is a simulation signal existence detection result table of the improved YOLOv5 of the present invention. In the case of third-party non-cooperative communication, the detection of the existence of the communication signal pays more attention to the detection omission of the target signal, that is, it pays more attention to the R value. The higher the R value, the less the target signal is missed. As shown in Table 2, when the noise environment is -20dB, due to the limitation of signal quality, there are more omissions and the average accuracy is also low. The mAP50-95 value is less than 30%, which obeys the general cognitive law. The higher P value is because the energy in the signal sequence is not evenly distributed, but mostly concentrated in the signal, while the noise obeys the Gaussian distribution. There is a high probability that the target signal exists at higher energy, rather than noise. With the improvement of signal quality and noise environment, the R value continues to increase, reaching 80% at -10dB; in the noise environment above 0dB, the R value is higher than 80%; in the noise environment of 20dB, the R value reaches the highest 97.4%, and there is almost no omission of the target signal. The average accuracy is close to 100% when the IOU threshold is 50%, proving that the improved YOLOv5 model can effectively detect the existence of the target signal. In the mixed noise environment, the R value also reached 82.8%, proving that the improved YOLOv5 model still maintains excellent performance in a complex noise environment and can effectively detect the target signal.
[0067] As shown in Table 3, this is the simulation signal existence detection result table of the improved YOLOv5 of the present invention. Due to the increase in the communication signal classification target, the improved YOLOv5s modulation mode recognition performance is reduced compared to the signal existence detection. As shown in Table 3, when the noise environment is -20dB, limited by the signal quality, the P value is 57.4%, the R value is 46.3%, and the mAP50-95 value is less than 25%. However, at -10dB, the P value is increased to about 75%, and the R value is increased to about 70%; when the noise environment is better than 0dB, the P value is higher than 85%, the R value is higher than 880%, the mAP50 value is not less than 90%, and the P value in the 10dB noise environment is an average accuracy of 93.1%. And in a mixed noise environment, the P value reaches 83.3%, the R value reaches 72.5%, the mAP50 value is 79.9%, and the mAP50-95 is higher than 55%, which can maintain the modulation mode recognition ability in a complex noise environment.
[0068] In Experimental Example 2, the measured signal time-frequency diagram test is adopted. The communication signal is generated in real time by SMW 200A, received and processed into a time-frequency diagram by LimeSDR in real time, and handed over to the improved YOLOv5 model for real-time recognition. As shown in Table 4, it is the parameter table of the measured signal time-frequency diagram data set of the present invention. The system configuration is consistent with that in Experimental Example 1.
[0069] As shown in Table 5, the communication signal detection and identification scheme based on the improved YOLOv5 proposed in the present invention can basically realize the identification of the third-party non-cooperative measured modulation signals, and the various detection values in the 8 modulation signal identification are close to 100%. With the subsequent expansion of the data set, the present invention will have a good recognition ability for the third-party non-cooperative communication signals.
[0070] The parts not described in detail in the present invention are all known technologies to those skilled in the art. The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.
[0071] Table 1 Parameter table of signal simulation time-frequency diagram data set of the present invention
[0072]
[0073] Table 2 The simulation signal existence detection results of the improved YOLOv5 of the present invention
[0074]
[0075] Table 3 The simulation signal modulation mode recognition result table of the improved YOLOv5 of the present invention
[0076]
[0077] Table 4 Parameters of the measured signal time-frequency diagram data set of the present invention
[0078]
[0079] Table 5 The measured signal modulation mode recognition results of the improved YOLOv5 of the present invention
[0080] type P R mAP50 mAP50-95 all 99.7% 99.8% 99.4% 97.9% 128QAM 100% 100% 99.5% 98.5% 64QAM 98.3% 98.7% 98.7% 97.8% 32QAM 99.9% 100% 99.5% 98.1% 8PSK 100% 100% 99.5% 97.7% QPSK 99.9% 100% 99.5% 97.9% BPSK 99.9% 100% 99.5% 97.2% ASK 99.6% 99.6% 99.5% 98.5% MSK 100% 100% 99.5% 97.4%
Claims
1. A communication signal detection and recognition method based on improved YOLOv5, characterized in that: The method comprises the following steps: Step 1, obtain a communication signal frequency spectrum data set: obtain communication signal simulation data through signal simulation, and then convert the signal simulation data into a signal frequency spectrum, obtain a communication signal frequency spectrum target simulation data set and a communication signal frequency spectrum data set measured by instruments and receiving equipment, pre-process the communication signal frequency spectrum, mark it according to the parameters of the communication signal duration, bandwidth, carrier frequency, existence, and modulation mode, and divide it into a training data set and a test data set; Step 2, establish an improved YOLOv5 model, that is, use a lightweight structure in the YOLOv5 backbone network, and improve the YOLOv5 model by combining a targeted scale feature fusion module and a feature attention module. The targeted scale feature fusion module is improved based on the bidirectional feature pyramid BiFPN, and the feature attention module is based on the simple parameter-free attention module SimAM attention mechanism. Step 3: Input the training data set into the improved YOLOv5 model for training to obtain a trained improved YOLOv5 model; Step 4: Input the test data set into the trained improved YOLOv5 model to obtain the detection and recognition results.
2. The communication signal detection and recognition method based on improved YOLOv5 according to claim 1, characterized in that: This method specifically includes the following four parts: In the first part, a receiving device is used to receive communication signals in real time through cables and antennas, and communication signal data is obtained through analog-to-digital conversion, signal processing, and signal acquisition steps. The communication signal data is converted into a signal time-frequency diagram and transmitted in real time; The second part is to establish, train and deploy the improved YOLOv5 model, including: establishing the improved YOLOv5 model, that is, using a lightweight structure in the YOLOv5 backbone network, and improving it by combining the targeted scale feature fusion module and the feature attention module. The targeted scale feature fusion module is improved based on BiFPN, and the feature attention module is based on the SimAM attention mechanism. The activation function uses ReLU, Leaky ReLU, Swish, Mish, PReLU and other activation functions, and the loss function uses IOU, CIOU, Wise IOU and other loss functions; training the improved YOLOv5 model, that is, inputting the training data set into the improved YOLOv5 model for training to obtain the trained improved YOLOv5 model; deploying the improved YOLOv5 model, that is, converting the trained improved YOLOv5 model into a format suitable for deployment, using TensorFlow Lite, ONNX and other formats, and optimize the trained improved YOLOv5 model, using int8, FP16, FP32 and other quantization methods and branch reduction operations, and finally deploy the model to the target platform, using servers, edge devices, and mobile devices. In the third part, the real-time transmitted signal time-frequency diagram is input into the deployed improved YOLOv5 model for real-time detection and recognition, using picture or video input; In the fourth part, combined with the system operation time and the characteristics of the signal time-frequency diagram, the recognition results of the improved YOLOv5 model are converted into multiple real-time parameters of the communication signal, including converting the positioning box of the communication signal in the time-frequency diagram into parameters such as duration, bandwidth, carrier frequency, and existence, converting the target type of the signal into the type of modulation parameter, and organizing multiple parameters at the same time to form a detection and recognition report that is more in line with the reading habits of the public.
3. The communication signal detection and recognition method based on improved YOLOv5 according to claim 2, characterized in that: The lightweight structure is used in the backbone network of YOLOv5. The lightweight structure makes lightweight improvements to the standard Conv unit, CBL unit, Res unit unit, and CSP unit in YOLOv5, specifically including: The standard Conv unit is split into lightweight Conv units. First, the input of the standard Conv unit is evenly split into two parts. One part is convolved according to the standard convolution process. The convolution kernel size remains unchanged, but the number of output channels is halved to obtain half of the normal output feature map. The other part first passes through the standard convolution with a convolution kernel size of 1×1, and then the convolution output is grouped and convolved. The convolution kernel size remains unchanged, the number of output channels is halved, and the number of groups is 1 / 4 of the number of input channels to obtain the other half of the output feature map. The output feature maps of the two parts are spliced by the splicing layer and the channels are rearranged to form a complete output feature map. The standard Conv unit in the CBL unit is replaced by a lightweight Conv unit. At the same time, the BN layer and ReLU layer superimposed after the standard Conv unit are placed before the output feature map is concatenated. The outputs of the two parts of the convolution process are executed separately to form a lightweight CBL unit. The convolution process of the standard Conv unit in the CBL unit is replaced by a grouped convolution process to form a grouped CBL unit; The CBL unit in the Res unit is replaced by a lightweight CBL unit and a grouped CBL unit to form a lightweight Resunit unit: first, the two-way process of the Res unit is changed to an input average split, and then split into two parts; one part replaces the CBL units with convolution kernel sizes of 1×1 and 3×3 in the Res unit with improved CBL units with a convolution kernel of 1×1 and grouped CBL units with a convolution kernel of 3×3 to extract features; the other part uses a separate 1×1 CBL unit, which is responsible for the change of the output feature size dimension but not for feature extraction; finally, the addition layer in the Res unit is replaced by a splicing layer to complete the combination of the output feature maps of the split part, and the channel rearrangement is used to output the complete feature map; The CBL unit and Res unit in the CSP unit are replaced by lightweight CBL unit and lightweight Res unit, and the structure is simplified to form a lightweight CSP unit: first, a CBL unit with a convolution kernel of 1×1 and a standard Conv layer with a convolution kernel of 1×1 in the CSP unit are merged into a lightweight CBL unit with a convolution kernel of 1×1, and the subsequent two operations share the output of the lightweight CBL unit; at the same time, the convolution unit structure composed of the convolution layer, splicing layer, BN layer, and ReLU layer at the output end of the CSP unit is deleted, and only the combined output features of the splicing layer are retained, and the subsequent processing is completed by channel rearrangement.
4. The communication signal detection and recognition method based on improved YOLOv5 according to claim 2, characterized in that: The targeted scale feature fusion module uses a targeted scale feature fusion module based on BiFPN; The targeted scale feature fusion module based on BiFPN blocks the path of fusion from small-scale features to large-scale features, and large-scale features are only fused with moderate-scale features; The targeted scale feature fusion module is mainly based on the moderate scale, and uses the bidirectional structure in BiFPN to fuse features from the moderate scale upward and downward respectively to achieve large-scale and small-scale feature extraction; at the same time, an additional path is added to the moderate scale to fuse the three scale features again and extract the final moderate-scale feature.
5. The communication signal detection and recognition method based on improved YOLOv5 according to claim 2, characterized in that: The feature attention module based on the SimAM attention mechanism is used; the feature attention module based on the SimAM attention mechanism realizes the weight allocation of each neuron through separability and activation function sigmoid, completes the spatial and channel attention allocation, adds the fusion of features of different scales in YOLOv5, and extracts features of different scales in the targeted scale feature fusion module based on BiFPN.
6. The communication signal detection and recognition method based on improved YOLOv5 according to claim 1, characterized in that: The preprocessing of the communication signal time-frequency diagram includes: adding noise, changing brightness, cropping, translating, rotating and scaling the data set; and adjusting the duration range, bandwidth range, sampling frequency and power display range of the time-frequency diagram display.
7. The communication signal detection and recognition method based on improved YOLOv5 according to claim 1, characterized in that: The labeling according to the parameters of the communication signal, such as duration, bandwidth, carrier frequency, existence, and modulation mode, includes: labeling each image in the training set using a graphic labeling tool, and automatically generating a labeling text file according to rules using a code.
Citation Information
Patent Citations
Automatic driving target detection algorithm based on improved Yolov5
CN116563825A
Signal detection method based on improved YOLOv5
CN116738225A
Gesture recognition method based on improved YOLOv5
CN117238024A
Cited By
Non-cooperative signal identification method based on time-frequency edge enhancement and physical characteristics
CN122634333A
