A fast radio signal recognition method based on color-enhanced image recognition

Abstract

The present invention discloses a method for rapid radio signal recognition based on color-enhanced image recognition. It includes a radio fixed station monitoring data acquisition and processing module, a signal pattern annotation and training module, a radio signal image preprocessing module, and a signal monitoring module. The radio fixed station monitoring data acquisition and processing module collects and analyzes radio data; the signal pattern annotation and training module is used to annotate and train signals, and uses an image recognition algorithm to perform deep neural network learning to acquire a signal recognition model; the radio signal image preprocessing module uses pseudo-color enhancement to generate corresponding waterfall graphs by frequency band; and the signal monitoring module performs signal recognition on the preprocessed signal waterfall and extracts and stores it. The beneficial effects of the present invention are: it can automatically extract the pattern characteristics of radio waves, improving the ability to recognize radio signals in complex environments.

Description

Technical Field

[0001] The present invention relates to the technical field related to electromagnetic spectrum monitoring, and in particular to a method for rapid radio signal recognition based on color-enhanced image recognition. Background Art

[0002] In recent years, with the rapid development of radio communication technology, from the initial shortwave communication to the current mobile communication, radio signals have become an indispensable part of life. However, with the increasing number of radio signals, radio signal monitoring has become increasingly challenging.

[0003] Current radio signal identification data is collected from the air using various spectrum acquisition devices. After processing, it is displayed in visual formats such as spectrograms, waterfall plots, and afterglow plots. Professionals analyze the signal's time-frequency characteristics and search for target signals. This technology requires very high operator expertise, and efficiency and accuracy decrease significantly with longer monitoring times or when there are numerous radio signals. While radio experts can currently select key signal characteristics, such as maximum power spectral density, signal envelope, and phase pulses, and then classify them based on fixed rules or machine learning, this approach is computationally very complex and is affected by the diversity of signal waveforms and multipath fading. This complex, multi-feature algorithm lacks universal applicability. Summary of the Invention

[0004] The present invention aims to overcome the above-mentioned deficiencies in the prior art and provides a method for rapid radio signal recognition based on color-enhanced image recognition, which can improve the radio signal recognition capability in complex environments.

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

[0006] A method for rapid radio signal recognition based on color-enhanced image recognition includes a radio fixed station monitoring data acquisition and processing module, a signal pattern annotation and training module, a radio signal image preprocessing module, and a signal monitoring module. The radio fixed station monitoring data acquisition and processing module is used to collect radio data and parse radio data frames; the signal pattern annotation and training module is used to annotate and train existing signals, annotate signals in waterfall charts, and acquire a signal recognition model through deep neural network learning using an image recognition algorithm; the radio signal image preprocessing module is used to generate corresponding waterfall charts by frequency band based on the collected monitoring data through image pseudo-color enhancement; and the signal monitoring module is used to perform signal recognition on the preprocessed signal waterfall by calling the trained signal recognition model, extract corresponding signals, and finally save these signals. The method specifically includes the following steps:

[0007] (1) The radio fixed station monitoring data acquisition and processing module issues frequency sweep monitoring tasks to the monitoring equipment through network monitoring. For the collected frequency sweep data, the de-framing algorithm is called to de-frame it and the de-framed data is saved in the form of a frequency sweep level matrix.

[0008] (2) The radio signal image preprocessing module reads the frequency sweep level matrix of the radio fixed station monitoring data acquisition and processing module, selects a period of time and frequency band range to combine into a two-dimensional spectrum waterfall diagram, and at the same time enhances the image display by pseudo-enhancement of image color, making it easier to detect characteristic signals;

[0009] (3) The signal image annotation training module uses the waterfall chart formed by historical monitoring tasks to annotate the signals in the waterfall chart through the signal annotation method, and forms the annotated waterfall chart into a training sample. The training is performed through the YOLO algorithm, and the training sample is divided into a training set and a validation set in a ratio of 2:1. The corresponding paths of the training set and the validation set are configured in the YOLO algorithm. The training is implemented by configuring the training yaml file and selecting the corresponding image segmentation function until the set parameters are met. Finally, a signal recognition model that can be used for image signal recognition reasoning is formed;

[0010] (4) The signal monitoring module takes the spectrum waterfall diagram obtained after preprocessing as input. The YOLO algorithm performs image segmentation on the spectrum waterfall diagram according to the trained model based on the category and position of the input image to obtain a concrete image signal. For the image signal, the image features are converted into waveform signal features and time / frequency signal features through pixel coordinates.

[0011] Long-term radio signal monitoring can form a level spectrum matrix, from which a spectrum waterfall plot can be generated. However, since the monitored signal-to-noise ratio is often around 30dB, and the level values ​​often range from -10dBuv to 100dBuv, commonly used levels are often concentrated in a certain range. Mapping the level grayscale values ​​to 8 bits for image enhancement processing will result in a significant loss of order information (the brightness information representing the three-color image). Therefore, corresponding enhancement is required based on the level signal value. Because artificial intelligence image recognition processing has unique advantages in waterfall plot recognition, image enhancement using a swept level matrix forms a spectrum waterfall plot. Using image recognition deep learning algorithms, it is possible to automatically extract radio wave pattern features, improving the ability to recognize radio signals in complex environments.

[0012] Preferably, in step (1), specifically: the radio fixed station monitoring data acquisition and processing module issues a frequency sweep monitoring task to a certain monitoring station through the radio networking platform, obtains the corresponding socket frequency sweep data stream, calls the ultrashort wave atomic service protocol to deframe the socket frequency sweep data stream, and obtains the corresponding task information. The deframed data is saved according to the frequency sweep level matrix for subsequent waterfall image generation; the obtained task information includes the following information:

[0013] <start_freq,end_freq,step>

[0014] Among them, start_freq represents the starting frequency of the task, end_freq represents the ending frequency of the task, and step represents the step size of this task.

[0015] Preferably, the socket frequency sweep data stream is processed according to the following process after deframing:

[0016] (11) Obtain the data after de-framing of the socket frequency sweep data stream. If it is the first frame of data received, determine whether its first frequency point is the task start frequency point. If it is not the task start frequency point, discard it until the data with the first frequency point of the frame being the task start frequency point is obtained, and record the time of its data frame;

[0017] (12) Recording the data obtained in step (11), and sequentially obtaining the level values ​​of subsequent frequency points until the level value of the task end frequency point is obtained, thereby forming a one-dimensional complex sequence of level values, representing the spectrum at a certain time point;

[0018] (13) Multiple one-dimensional complex sequences of consecutive time points are stacked in sequence to form a frequency sweep level matrix. In this matrix, the horizontal axis represents the frequency point, the vertical axis represents the time, and the element value corresponds to the level value:

[0019]

[0020] Where tm represents the time of a complete scanning cycle, f i tj Indicates t j Time frequency f i The level value of this matrix also contains the task information of this task:<start_freq,end_freq,step> ;

[0021] (14) The new socket frequency sweep data stream forms a frequency sweep level matrix according to this processing flow.

[0022] Preferably, in step (2), specifically: the radio signal image preprocessing module forms a spectrum waterfall diagram according to the sweep frequency level matrix; the radio signal image preprocessing module starts from the task start time and obtains a sub-matrix of size 512*512. If the number of frequency points is not an integer multiple of 512, the last sub-matrix is ​​a sub-matrix with a column number less than 512; the time series needs to meet 512 time points. If the number of rows does not reach 512, wait for the radio fixed station monitoring data acquisition and processing module to accumulate enough data; after the sub-matrix is ​​selected, these sub-matrices are deleted from the original sweep frequency level matrix to avoid processing these data again in the next processing; in the sub-matrix, the element level value is mapped according to the grayscale value to obtain the corresponding grayscale spectrum waterfall diagram. In order to better display the image features for easy recognition, the color enhancement method is continued to be used on the basis of the grayscale image to perform color enhancement to form a spectrum waterfall diagram for subsequent signal recognition.

[0023] Preferably, in the color enhancement, color enhancement is performed according to the radio level distribution characteristics, and the process is as follows:

[0024] (21) The sweep level matrix is ​​quantized in bit depth, and the quantization formula is as follows:

[0025] Graxy(x,y)=(Matrix(x,y)+20)*2 8 / 80

[0026] Where Graxy(x,y) represents the grayscale value of the pixel coordinate (x,y), Matrix(x,y) is the radio frequency sweep level matrix, and 80 represents the signal-to-noise ratio in radio detection;

[0027] (22) According to the characteristics of radio level, a grayscale enhancement algorithm is constructed to perform color enhancement. Based on step (21), a JET mapping function adapted to radio grayscale is constructed, and its mapping function is defined as WGLFC JET (R, G, B), where the mapping relationship between R, G, and B channels is as follows:

[0028]

[0029] (23) The RGB color space after color enhancement is transposed to the HSV color space to enhance the color, depth, and brightness. In order to enhance the color vividness of the image under the HSV model, the V channel is set to the maximum value of 1, and an enhancement coefficient is introduced to solve the problem of low image contrast caused by small frequency level fluctuations due to low device resolution.

[0030] Preferably, in step (3), the annotation is specifically as follows: the signal image annotation training module uses the labelme tool to annotate the color-enhanced spectrum waterfall diagram, annotates the signal according to each device, and monitors the corresponding signal recognition model for each device; performs signal annotation for each device, annotates one by one according to the service frequency band, and automatically generates annotation information of the image in the background by annotating the visual border; the annotation information refers to the json information of the signal border, identifies the outline of the signal by the spectrum diagram, and adds the corresponding signal label; after the annotation is completed, it is converted into a corresponding txt file, including the number of the annotated signal, the coordinates of the corresponding contour point of the annotated signal on the diagram, the information of the image where the annotated signal is located, and the Base64 code of the image, so as to facilitate segmentation and recognition according to the corresponding points during model training.

[0031] Preferably, in step (3), the training is specifically as follows: dividing and training the data set according to the labeled images, outputting the signal recognition model of each site, and naming the model using the MFID code of the device; in the training model, in order to enable the output layer to retain higher resolution features and reduce the probability of gradient disappearance, the ReLU nonlinear activation function is used to reduce the probability of neuron death during training:

[0032]

[0033] Here, x represents the pixel value of the image.

[0034] Preferably, in step (4), the image signal is obtained as follows: the signal monitoring module loads the corresponding model, takes the pre-processed spectrum waterfall diagram to be judged as input, and obtains a concrete signal image through the YOLO algorithm; first, for the spectrum waterfall diagram to be judged, obtain the monitoring station corresponding to its task, select the YOLO model trained on the monitoring station according to the MFID corresponding to the monitoring station, and perform feature extraction on the model with the processed two-dimensional spectrum waterfall diagram to extract relevant information of the image signal:

[0035] <tx,ty,tw,th,t0>

[0036] Wherein, they represent the coordinates of the upper left point of the pixel signal image (tx, ty), tw represents the width of the pixel signal image border, th represents the height of the pixel signal image border, and t0 represents the coordinates of the first pixel point in the row where the pixel signal image (tx, ty) is extracted.

[0037] Preferably, in step (4), the image signal conversion is specifically as follows:

[0038] (41) Signal start frequency: calculated by tx coordinate and image start coordinate, the calculation formula is as follows

[0039] freq_start=f0+(tx-t0)*step

[0040] Where freq_start represents the starting frequency of the signal, f0 represents the starting frequency corresponding to this spectrum waterfall diagram, tx represents the pixel coordinates of the upper left point of the pixel signal image border, t0 represents the coordinates of the first pixel point in the row where the pixel signal image (tx, ty) is extracted, and step represents the step size, which also represents the frequency span of a pixel point.

[0041] (42) Signal end frequency: Calculated by tx, tw and the start coordinates of the frequency domain waterfall diagram. The calculation formula is as follows:

[0042] freq_end=f0+(tx-t0+tw)*step

[0043] Where freq_end represents the end frequency of the signal, f0 represents the start frequency corresponding to this spectrum waterfall diagram, tx represents the pixel coordinate of the upper left corner of the pixel signal image border, t0 represents the coordinate of the first pixel point in the row where the pixel signal image (tx,ty) is extracted, tw represents the width of the pixel signal image border, and step represents the step size, which also represents the frequency span of one pixel point.

[0044] (43) Signal bandwidth and signal center frequency:

[0045] The signal bandwidth is obtained by subtracting the calculated start and end frequencies:

[0046] freq_width=(freq_end-freq_start);

[0047] The signal center frequency is calculated using the following formula:

[0048] freq_mid=(freq_end+freq_start) / 2;

[0049] (44) Signal start time: the matrix time corresponding to the pixel signal image border t0 when the signal starts to pass through;

[0050] (45) Signal end time: the matrix time corresponding to the signal end time (ty+th) row.

[0051] The beneficial effects of the present invention are: after image enhancement through the frequency sweep level matrix, a spectrum waterfall diagram is formed, and the image recognition deep learning algorithm is used to automatically extract the pattern characteristics of radio waves, thereby improving the recognition ability of radio signals in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flowchart of the method of the present invention;

[0053] Figure 2 is a flow chart of the method of the present invention;

[0054] Figure 3 It is a two-dimensional frequency waterfall diagram in a specific example of the present invention. DETAILED DESCRIPTION

[0055] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0056] like Figure 1 、 Figure 2 In the embodiment described, in order to achieve faster and more accurate detection of signals in complex environments, the present invention provides a radio rapid signal recognition method based on color-enhanced image recognition, comprising a radio fixed station monitoring data acquisition and processing module, a signal pattern annotation training module, a radio signal image preprocessing module, and a signal monitoring module. The radio fixed station monitoring data acquisition and processing module is used to acquire radio data and parse radio data frames; the signal pattern annotation training module is used to annotate and train existing signals, annotate the signals in the waterfall chart, and acquire a signal recognition model through deep neural network learning using an image recognition algorithm; the radio signal image preprocessing module is used to generate corresponding waterfall charts by frequency band based on the collected monitoring data through image pseudo-color enhancement; the signal monitoring module is used to perform signal recognition on the preprocessed signal waterfall by calling the trained signal recognition model, extract the corresponding signal, and finally save these signals; specifically, the method comprises the following steps:

[0057] (1) The radio fixed station monitoring data acquisition and processing module issues frequency sweep monitoring tasks to the monitoring equipment through network monitoring. For the collected frequency sweep data, the de-framing algorithm is called to de-frame it and the de-framed data is saved in the form of a frequency sweep level matrix.

[0058] Specifically, the radio fixed station monitoring data acquisition and processing module issues a frequency sweep monitoring task to a monitoring station through the radio networking platform, obtains the corresponding socket frequency sweep data stream, calls the ultrashort wave atomic service protocol to deframe the socket frequency sweep data stream, and obtains the corresponding task information. The deframed data is saved according to the frequency sweep level matrix for subsequent waterfall image generation. The obtained task information includes the following:

[0059] <start_freq,end_freq,step>

[0060] Among them, start_freq represents the starting frequency of the task, end_freq represents the ending frequency of the task, and step represents the step size of this task.

[0061] After the socket frequency sweep data stream is deframed, it is processed according to the following process:

[0062] (11) Obtain the data after de-framing of the socket frequency sweep data stream. If it is the first frame of data received, determine whether its first frequency point is the task start frequency point. If it is not the task start frequency point, discard it until the data with the first frequency point of the frame being the task start frequency point is obtained, and record the time of its data frame;

[0063] (12) Recording the data obtained in step (11), and sequentially obtaining the level values ​​of subsequent frequency points until the level value of the task end frequency point is obtained, thereby forming a one-dimensional complex sequence of level values, representing the spectrum at a certain time point;

[0064] (13) Multiple one-dimensional complex sequences of consecutive time points are stacked in sequence to form a frequency sweep level matrix. In this matrix, the horizontal axis represents the frequency point, the vertical axis represents the time, and the element value corresponds to the level value:

[0065]

[0066] Where tm represents the time of a complete scanning cycle, f i tj Indicates t j Time frequency f i The level value of this matrix also contains the task information of this task:<start_freq,end_freq,step> ;

[0067] (14) The new socket frequency sweep data stream forms a frequency sweep level matrix according to this processing flow.

[0068] (2) The radio signal image preprocessing module reads the frequency sweep level matrix of the radio fixed station monitoring data acquisition and processing module, selects a period of time and frequency band range to combine into a two-dimensional spectrum waterfall diagram, and at the same time enhances the image display by pseudo-enhancement of image color, making it easier to detect characteristic signals;

[0069] Specifically, the radio signal image preprocessing module forms a spectrum waterfall diagram based on the sweep frequency level matrix; the radio signal image preprocessing module starts from the task start time and obtains a 512*512 sub-matrix. If the number of frequency points is not an integer multiple of 512, the last sub-matrix is ​​a sub-matrix with less than 512 columns; the time series needs to meet 512 time points. If the number of rows does not reach 512, wait for the radio fixed station monitoring data acquisition and processing module to accumulate enough data; after the sub-matrix is ​​selected, it is deleted from the original sweep frequency level matrix to avoid processing these data again in the next processing.

[0070] In the submatrix, the element level values ​​are mapped according to the grayscale values ​​to obtain the corresponding grayscale spectrum waterfall diagram. In order to better display the image features for easy recognition, the color enhancement method is continued on the basis of the grayscale image to perform color enhancement to form a spectrum waterfall diagram for subsequent signal recognition.

[0071] In color enhancement, color enhancement is performed based on the characteristics of radio level distribution. The process is as follows:

[0072] (21) The sweep level matrix is ​​quantized in bit depth, and the quantization formula is as follows:

[0073] Graxy(x,y)=(Matrix(x,y)+20)*2 8 / 80

[0074] Where Graxy(x,y) represents the grayscale value of the pixel coordinate (x,y), Matrix(x,y) is the radio frequency sweep level matrix, and 80 represents the signal-to-noise ratio in radio detection;

[0075] (22) According to the characteristics of radio level, a grayscale enhancement algorithm is constructed to perform color enhancement. Based on step (21), a JET mapping function adapted to radio grayscale is constructed, and its mapping function is defined as WGLFC JET (R, G, B), where the mapping relationship between R, G, and B channels is as follows:

[0076]

[0077] (23) The RGB color space after color enhancement is transposed to the HSV color space to enhance the color, depth, and brightness. In order to enhance the color vividness of the image under the HSV model, the V channel is set to the maximum value of 1, and an enhancement coefficient is introduced to solve the problem of low image contrast caused by small frequency level fluctuations due to low device resolution. After color enhancement, the low-contrast image can achieve better adaptive enhancement.

[0078] (3) The signal image annotation training module uses the waterfall chart formed by the historical monitoring task to annotate the signals in the waterfall chart through the signal annotation method, and forms the annotated waterfall chart into a training sample. The training is performed through the YOLO algorithm, and the training sample is divided into a training set and a validation set in a ratio of 2:1. The corresponding paths of the training set and the validation set are configured in the YOLO algorithm. The training is implemented by configuring the training yaml file and selecting the corresponding image segmentation function until the set parameters (confidence, recall rate, precision rate) are met. Finally, a signal recognition model that can be used for image signal recognition reasoning is formed;

[0079] The specific labeling is as follows: the signal image labeling training module uses the labelme tool to label the color-enhanced spectrum waterfall diagram. Taking into account the characteristics of the signal having a fixed working frequency band and continuous existence, and the influence of factors such as signal attenuation, elevation and equipment parameters on each identical signal, the signal is labeled according to each device, and a corresponding signal recognition model is supervised for each device.

[0080] Perform signal labeling for each device, labeling each one by the service frequency band. By labeling the visual border, the annotation information of the image in the background is automatically generated. The annotation information refers to the JSON information of the signal border. The signal outline is identified by the spectrum graph and the corresponding signal label is added. After labeling, it is converted into a corresponding txt file, including the number of the labeled signal, the coordinates of the corresponding contour point of the labeled signal on the graph, the information of the image where the labeled signal is located, and the Base64 code of the image, which facilitates segmentation and recognition according to the corresponding points during model training. The annotation content includes the following information:

[0081]

[0082] Signals are subject to many changes due to environmental influences. Therefore, based on the common characteristics of radio signals, signal annotations are divided into the following four categories:

[0083] Narrow-band continuous signal (0.015, 0.9)

[0084] ●Broadband continuous signal (0.5, 0.9)

[0085] Narrowband transient signal (0.025, 0.5)

[0086] ●Broadband transient signal (0.469, 0.1)

[0087] The (x, y) after the signal represents the percentage of pixels in the image.

[0088] The training process involves dividing and training the dataset based on the labeled images, outputting a signal recognition model for each site, and naming the model using the device's MFID. In the training model, to ensure that the output layer retains higher-resolution features and reduces the probability of "vanishing gradients," the ReLU nonlinear activation function is used to reduce the probability of neuron death during training.

[0089]

[0090] Here, x represents the pixel value of the image.

[0091] (4) The signal monitoring module uses the spectrum waterfall diagram obtained after preprocessing as input. The YOLO algorithm performs image segmentation on the spectrum waterfall diagram according to the trained model based on the category and position of the input image to obtain a concrete image signal. For the image signal, the image features are converted into waveform signal features and time / frequency signal features through pixel coordinates;

[0092] The specific steps for acquiring image signals are as follows: the signal monitoring module loads the corresponding model, takes the pre-processed spectrogram of the signal to be judged as input, and obtains a concrete signal image through the YOLO algorithm. First, for the spectrogram to be judged, the corresponding monitoring station is obtained. The YOLO model trained on the monitoring station is selected based on the MFID corresponding to the monitoring station. The model extracts features from the processed two-dimensional spectrogram to extract information related to the image signal:

[0093] <tx,ty,tw,th,t0>

[0094] Wherein, they represent the coordinates of the upper left point of the pixel signal image (tx, ty), tw represents the width of the pixel signal image border (number of pixels), th represents the height of the pixel signal image border (number of pixels), and t0 represents the coordinates of the first pixel point in the row where the pixel signal image (tx, ty) is extracted.

[0095] In order to convert the acquired image information into actual signal information, the following conversion is required to convert it into signal information in the radio field, specifically:

[0096] (41) Signal start frequency: calculated by tx coordinate and image start coordinate, the calculation formula is as follows

[0097] freq_start=f0+(tx-t0)*step

[0098] Where freq_start represents the starting frequency of the signal, f0 represents the starting frequency corresponding to this spectrum waterfall diagram, tx represents the pixel coordinates of the upper left point of the pixel signal image border, t0 represents the coordinates of the first pixel point in the row where the pixel signal image (tx, ty) is extracted, and step represents the step size, which also represents the frequency span of a pixel point.

[0099] (42) Signal end frequency: Calculated by tx, tw and the start coordinates of the frequency domain waterfall diagram. The calculation formula is as follows:

[0100] freq_end=f0+(tx-t0+tw)*step

[0101] Where freq_end represents the end frequency of the signal, f0 represents the start frequency corresponding to this spectrum waterfall diagram, tx represents the pixel coordinate of the upper left corner of the pixel signal image border, t0 represents the coordinate of the first pixel point in the row where the pixel signal image (tx,ty) is extracted, tw represents the width of the pixel signal image border, and step represents the step size, which also represents the frequency span of one pixel point.

[0102] (43) Signal bandwidth and signal center frequency:

[0103] The signal bandwidth is obtained by subtracting the calculated start and end frequencies:

[0104] freq_width=(freq_end-freq_start)

[0105] The signal center frequency is calculated using the following formula:

[0106] freq_mid=(freq_end+freq_start) / 2

[0107] (44) Signal start time: the matrix time corresponding to the pixel signal image border t0 when the signal starts to pass through;

[0108] (45) Signal end time: the matrix time corresponding to the signal end time (ty+th) row.

[0109] Through this conversion, the image target detection information is converted into radio signal parameters, which can facilitate business personnel to judge the signal characteristic parameters.

[0110] Specific examples are as follows:

[0111] A rapid radio signal recognition method based on color-enhanced image recognition was developed for a monitoring facility in a certain province. First, a specific monitoring task was assigned to the monitoring equipment through the networked monitoring work website. The corresponding data stream was obtained and converted into a spectrum graph to complete the image signal recognition. Finally, the image information was converted back into radio signal characteristics to complete the signal recognition. The processing flow is as follows:

[0112] (a) Select a monitoring facility in a province and issue a monitoring task through the networked monitoring workstation website. The parameters are as follows:

[0113] Frequency range Start time End Time Trail Task 1 87M~300M 09:00 18:00 25K

[0114] (b) The system obtains the data stream through the socket and deframes it. After deframing, it forms a sweep level matrix in the form of a complete time scan frame:

[0115] t1 d11 d12 d13 … d1n t2 … … … … … t3 … … … … … … … … … … … tm … … …… … dmn

[0116] Where tm is time, dij is the level value of the corresponding frequency point at the corresponding time;

[0117] (c) Read a 512*512 matrix of corresponding size in the frequency sweep level matrix and convert this matrix into a grayscale image;

[0118] (d) The grayscale image is enhanced by color to form a two-dimensional frequency waterfall diagram, such as Figure 3 As shown;

[0119] (e) To generate a color-enhanced two-dimensional frequency waterfall diagram, call the model trained on the corresponding device to obtain the image information of the corresponding signal. This information includes:

[0120]

[0121] (f) Convert the image information into radio signal features according to the correspondence between pixels and frequencies.

[0122] At this point, a method for rapid radio signal recognition based on color-enhanced image recognition is completed.

[0123] After experimental calculation, the present invention uses traditional digital signal processing methods to convert radio signals into a two-dimensional spectrum waterfall diagram, converts the radio signal recognition problem into target monitoring in the field of image recognition, and converts it into an RGB image with relatively obvious resolution through an enhancement algorithm. Then, by using the advanced achievements of artificial intelligence in the field of images, based on the YOLOv8 model, the radio signal recognition is converted from a monitoring parameter recognition problem to an image target monitoring problem. A signal recognition model for radio signal recognition is constructed, which realizes the conversion of the radio spectrum time domain diagram into a two-dimensional spectrum waterfall diagram, extracts the feature border of the signal based on the two-dimensional waterfall diagram, and finally converts it back into signal parameters. It can effectively identify the visualized radio signal, and the recognition effect is similar to that of the human eye, and the recognition speed can be completed within milliseconds. It brings new ways to intelligent radio signal recognition technology in complex electromagnetic environments in the future, and greatly improves the efficiency of radio business personnel in signal monitoring, identification and discovery.

Claims

1. A method for rapid radio signal recognition using color-enhanced image recognition, characterized in that: The system comprises a radio fixed station monitoring data acquisition and processing module, a signal pattern annotation training module, a radio signal image preprocessing module, and a signal monitoring module. The radio fixed station monitoring data acquisition and processing module is used to realize the acquisition of radio data and the analysis of radio data frames; the signal pattern annotation training module is used to annotate and train existing signals, and acquire a signal recognition model by annotating the signals of the waterfall chart and performing deep neural network learning through an image recognition algorithm; the radio signal image preprocessing module is used to generate corresponding waterfall charts according to frequency bands based on the collected monitoring data through image pseudo-color enhancement; the signal monitoring module is used to perform signal recognition on the preprocessed signal waterfall by calling the trained signal recognition model, extract the corresponding signal, and finally save these signals. The system specifically comprises the following steps: (1) The radio fixed station monitoring data acquisition and processing module issues frequency sweep monitoring tasks to the monitoring equipment through network monitoring. For the collected frequency sweep data, the de-framing algorithm is called to de-frame it and the de-framed data is saved in the form of a frequency sweep level matrix. (2) The radio signal image preprocessing module reads the frequency sweep level matrix of the radio fixed station monitoring data acquisition and processing module, selects a period of time and frequency band range to combine into a two-dimensional spectrum waterfall diagram, and at the same time enhances the image display by pseudo-enhancement of image color, making it easier to detect characteristic signals; (3) The signal image annotation training module uses the waterfall chart formed by historical monitoring tasks to annotate the signals in the waterfall chart through the signal annotation method, and forms the annotated waterfall chart into a training sample. The training is performed through the YOLO algorithm, and the training sample is divided into a training set and a validation set in a ratio of 2:

1. The corresponding paths of the training set and the validation set are configured in the YOLO algorithm. The training is implemented by configuring the training yaml file and selecting the corresponding image segmentation function until the set parameters are met. Finally, a signal recognition model that can be used for image signal recognition reasoning is formed; (4) The signal monitoring module takes the spectrum waterfall diagram obtained after preprocessing as input. The YOLO algorithm performs image segmentation on the spectrum waterfall diagram according to the trained model based on the category and position of the input image to obtain a concrete image signal. For the image signal, the image features are converted into waveform signal features and time / frequency signal features through pixel coordinates.

2. The method for rapid radio signal recognition based on color-enhanced image recognition according to claim 1, wherein: In step (1), specifically: the radio fixed station monitoring data acquisition and processing module issues a frequency sweep monitoring task to a certain monitoring station through the radio networking platform, obtains the corresponding socket frequency sweep data stream, calls the ultrashort wave atomic service protocol to deframe the socket frequency sweep data stream, and obtains the corresponding task information. The deframed data is saved according to the frequency sweep level matrix for subsequent waterfall image generation; the obtained task information includes the following information: <start_freq,end_freq,step> Among them, start_freq represents the starting frequency of the task, end_freq represents the ending frequency of the task, and step represents the step size of this task.

3. The method for rapid radio signal recognition based on color-enhanced image recognition according to claim 2, wherein: After the socket frequency sweep data stream is deframed, it is processed according to the following process: (11) Obtain the data after de-framing of the socket frequency sweep data stream. If it is the first frame of data received, determine whether its first frequency point is the task start frequency point. If it is not the task start frequency point, discard it until the data with the first frequency point of the frame being the task start frequency point is obtained, and record the time of its data frame; (12) Recording the data obtained in step (11), and sequentially obtaining the level values ​​of subsequent frequency points until the level value of the task end frequency point is obtained, thereby forming a one-dimensional complex sequence of level values, representing the spectrum at a certain time point; (13) Multiple one-dimensional complex sequences of consecutive time points are stacked in sequence to form a frequency sweep level matrix. In this matrix, the horizontal axis represents the frequency point, the vertical axis represents the time, and the element value corresponds to the level value: Where tm represents the time of a complete scanning cycle, f i tj Indicates t j Time frequency f i The level value of this matrix also contains the task information of this task:<start_freq,end_freq,step> ; (14) The new socket sweep frequency data stream forms a sweep frequency level matrix according to this processing flow.

4. The method for rapid radio signal recognition based on color-enhanced image recognition according to claim 1, wherein: In step (2), specifically: the radio signal image preprocessing module forms a spectrum waterfall diagram according to the sweep frequency level matrix; the radio signal image preprocessing module starts from the task start time and obtains a sub-matrix of size 512*512. If the number of frequency points is not an integer multiple of 512, the last sub-matrix is ​​a sub-matrix with a column number less than 512; the time series needs to meet 512 time points. If the number of rows does not reach 512, wait for the radio fixed station monitoring data acquisition and processing module to accumulate enough data; after the sub-matrix is ​​selected, delete these sub-matrices from the original sweep frequency level matrix to avoid processing these data again in the next processing; In the submatrix, the element level values ​​are mapped according to the grayscale values ​​to obtain the corresponding grayscale spectrum waterfall diagram. In order to better display the image features for easy recognition, the color enhancement method is continued on the basis of the grayscale image to perform color enhancement to form a spectrum waterfall diagram for subsequent signal recognition.

5. The method for rapid radio signal recognition based on color-enhanced image recognition according to claim 4, wherein: In the color enhancement, color enhancement is performed according to the characteristics of radio level distribution, and the process is as follows: (21) The sweep level matrix is ​​quantized in bit depth, and the quantization formula is as follows: Graxy(x,y)=(Matrix(x,y)+20)*2 8 / 80 Where Graxy(x,y) represents the grayscale value of the pixel coordinate (x,y), Matrix(x,y) is the radio frequency sweep level matrix, and 80 represents the signal-to-noise ratio in radio detection; (22) According to the characteristics of radio level, a grayscale enhancement algorithm is constructed to perform color enhancement. Based on step (21), a JET mapping function adapted to radio grayscale is constructed, and its mapping function is defined as WGLFC JET (R, G, B), where the mapping relationship between R, G, and B channels is as follows: (23) The RGB color space after color enhancement is transposed to the HSV color space to enhance the color, depth, and brightness. In order to enhance the color vividness of the image under the HSV model, the V channel is set to the maximum value of 1, and an enhancement coefficient is introduced to solve the problem of low image contrast caused by small frequency level fluctuations due to low device resolution.

6. The method for rapid radio signal recognition based on color-enhanced image recognition according to claim 1, wherein: In step (3), the labeling is specifically as follows: the signal image labeling training module uses the labelme tool to label the color-enhanced spectrum waterfall diagram, labels the signal according to each device, and monitors the corresponding signal recognition model for each device; the signal labeling work is performed on each device, and each device is labeled one by one according to the service frequency band. By labeling the visual border, the labeling information of the image in the background is automatically generated; The annotation information refers to the JSON information of the signal border. The signal outline is identified by the spectrum graph and the corresponding signal label is added. After annotation, it is converted into a corresponding txt file, including the number of the annotated signal, the coordinates of the corresponding contour point of the annotated signal on the graph, the information of the image where the annotated signal is located, and the Base64 code of the image, which facilitates segmentation and recognition according to corresponding points during model training.

7. The method for rapid radio signal recognition based on color-enhanced image recognition according to claim 1 or 6, wherein: In step (3), the training is as follows: the data set is divided and trained according to the labeled images, and the signal recognition model of each site is output. The model name is named using the MFID code of the device; in the training model, in order to enable the output layer to retain higher resolution features and reduce the probability of gradient disappearance, the ReLU nonlinear activation function is used to reduce the probability of neuron death during training: Here, x represents the pixel value of the image.

8. The method for rapid radio signal recognition based on color-enhanced image recognition according to claim 1, wherein: In step (4), the image signal is obtained as follows: the signal monitoring module loads the corresponding model, takes the pre-processed spectrum waterfall diagram to be judged as input, and obtains the concrete signal image through the YOLO algorithm; first, for the spectrum waterfall diagram to be judged, obtain the monitoring station corresponding to its task, select the YOLO model trained on the monitoring station according to the MFID corresponding to the monitoring station, and perform feature extraction on the processed two-dimensional spectrum waterfall diagram to extract information related to the image signal: <tx,ty,tw,th,t0> Wherein, tx, ty represent the coordinates of the upper left point of the pixel signal image, tw represents the width of the pixel signal image border, th represents the height of the pixel signal image border, and t0 represents the start time of the signal represented by the pixel signal image.

9. The method for rapid radio signal recognition based on color-enhanced image recognition according to claim 8, wherein: In step (4), the image signal is converted into: (41) Signal start frequency: calculated by tx coordinate and image start coordinate, the calculation formula is as follows freq_start=f0+(tx-t0)*step Where freq_start represents the starting frequency of the signal, f0 represents the starting frequency corresponding to this spectrum waterfall diagram, tx represents the pixel coordinates of the upper left point of the pixel signal image border, t0 represents the coordinates of the first pixel point in the row where the pixel signal image (tx, ty) is extracted, and step represents the step size, which also represents the frequency span of a pixel point. (42) Signal end frequency: Calculated by tx, tw and the start coordinates of the frequency domain waterfall diagram. The calculation formula is as follows: freq_end=f0+(tx-t0+tw)*step Where freq_end represents the end frequency of the signal, f0 represents the start frequency corresponding to this spectrum waterfall diagram, tx represents the pixel coordinate of the upper left corner of the pixel signal image border, t0 represents the coordinate of the first pixel point in the row where the pixel signal image (tx,ty) is extracted, tw represents the width of the pixel signal image border, and step represents the step size, which also represents the frequency span of one pixel point. (43) Signal bandwidth and signal center frequency: The signal bandwidth is obtained by subtracting the calculated start and end frequencies: freq_width=(freq_end-freq_start) The signal center frequency is calculated using the following formula: freq_mid=(freq_end+freq_start) / 2 (44) Signal start time: The signal start time passes through the pixel signal image border t0 time; (45) Signal end time: The time corresponding to the signal end time (ty+th) line.

Images