UAV signal detection method and system based on deep learning

Abstract

The present invention relates to a drone signal detection method and system based on deep learning. The method collects received signals, obtains a signal time-frequency diagram, and then performs vector accumulation. The signal time-frequency diagrams of two radio frequency channels are conjugate-multiplied and arc-tangented to obtain a time-phase difference spectrum. The spectrum is normalized into a grayscale image and fed into a trained deep learning model for drone target detection. After retaining high-confidence target frames, secondary detection is performed to obtain candidate target frames. Finally, non-maximum suppression processing is performed to obtain the final valid target frame. This makes the algorithm insensitive to pixel mutations. The use of a trained deep learning model also avoids the need to consider the type of convolution kernel to be used, thereby simplifying the detection process, effectively adapting to lower signal-to-noise ratios, and improving the recognition performance of drone image transmission signals under low signal-to-noise ratios.

Description

Technical Field

[0001] The present invention belongs to the technical field of MIMO radar target detection and relates to a drone signal detection method and system based on deep learning. Background Art

[0002] With the increasing use of drone technology, drones are posing increasing challenges to low-altitude security. Consequently, anti-drone systems have emerged. Detecting drone signals is fundamental to their operation, and accurate detection is a crucial and arduous task. Drone detection primarily involves both active and passive devices. Active devices offer the advantage of high precision, but are susceptible to interference in complex terrain, making it difficult to accurately observe the target. Therefore, passive devices are becoming increasingly important.

[0003] Drone control primarily relies on video data transmitted via image transmission signals, especially for high-speed drones. This signal is fed back to the pilot, allowing them to perceive the drone's surroundings. Because the image transmission signal originates from the drone, its detection is essential for locating the target drone. Image transmission signals transmit relatively large amounts of data, have wide signal bandwidths and pulse widths, and are continuous. They are typically modulated using OFDM (Orthogonal Frequency Division Multiplexing), which appears as a relatively regular rectangle on a time-frequency graph.

[0004] Currently, recognition of drone image transmission signals is primarily achieved by binarizing time-frequency plots and then performing edge detection or direct image convolution. Edge detection is prone to failure at low signal-to-noise ratios and can easily introduce interference signals. Selecting a binarization threshold is also difficult. Image convolution requires different convolution kernels, making the algorithm complex to implement. Therefore, improving the recognition performance of drone image transmission signals at low signal-to-noise ratios has become a current technical challenge. Summary of the Invention

[0005] In response to the problems existing in the above-mentioned traditional technologies, the present invention proposes a drone signal detection method based on deep learning and a drone signal detection system based on deep learning, which can improve the recognition performance of drone image transmission signals under low signal-to-noise ratio.

[0006] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions:

[0007] On the one hand, a method for detecting drone signals based on deep learning is provided, comprising the steps of:

[0008] Load the trained deep learning model; the deep learning model is trained using a pre-collected and annotated training dataset;

[0009] After the received signal is collected, it is sampled by ADC, windowed, and short-time Fourier transformed to obtain the signal time-frequency diagram;

[0010] After vector accumulation of the signal time-frequency diagram, the signal time-frequency diagrams of the received signals received by the two RF channels are conjugate multiplied and the arc tangent is calculated to obtain the time-phase difference spectrum;

[0011] The time-phase difference map is converted into a grayscale image and then fed into the deep learning model to obtain a series of target box upper left corner vertex coordinates and lower right corner vertex coordinates, as well as the confidence scores corresponding to the target framed by the target box;

[0012] After retaining the target boxes whose confidence scores are higher than the set confidence threshold, determine the categories of all retained target boxes;

[0013] Traverse the categories of all retained target frames, perform secondary detection on all retained target frames, and obtain a list of candidate target frames corresponding to the categories;

[0014] Perform non-maximum suppression on the candidate target frame list corresponding to the category to obtain the valid target frame corresponding to the category.

[0015] On the other hand, a drone signal detection system based on deep learning is also provided, including:

[0016] The model loading module is used to load the trained deep learning model; the deep learning model is trained using a pre-collected and annotated training dataset;

[0017] The time-frequency conversion module is used to collect the received signal, sample it through ADC, add window and perform short-time Fourier transform to obtain the signal time-frequency diagram;

[0018] A vector spectrum module is used to perform vector accumulation on the signal time-frequency diagram, conjugate multiply the signal time-frequency diagrams of the received signals received by the two RF channels, and calculate the arc tangent to obtain a time-phase difference spectrum;

[0019] The model detection module is used to convert the time-phase difference map into a grayscale image and feed it into the deep learning model to obtain a series of target box upper left corner vertex coordinates and lower right corner vertex coordinates, as well as the confidence score corresponding to the target framed by the target box;

[0020] The target selection module is used to retain the target frames whose confidence scores are higher than the set confidence threshold and then determine the categories of all retained target frames;

[0021] The secondary detection module is used to traverse the categories of all retained target frames, perform secondary detection on all retained target frames, and obtain a list of candidate target frames corresponding to the categories;

[0022] The non-maximum suppression processing module is used to perform non-maximum suppression on the candidate target frame list corresponding to the category to obtain the valid target frame corresponding to the category.

[0023] One of the above technical solutions has the following advantages and beneficial effects:

[0024] The deep learning-based drone signal detection method and system collects received signals, samples them using an ADC, performs windowing, and performs a short-time Fourier transform (SFT) to obtain a signal time-frequency map. This map then undergoes vector accumulation to improve image quality. The time-frequency maps of the signals received by the two RF channels are then conjugate-multiplied and inverse-tangented to produce a time-phase difference map. This map is then normalized to a grayscale image and fed into a trained deep learning model for drone target detection. High-confidence target frames are retained, followed by secondary detection to obtain candidate target frames. Finally, non-maximum suppression is performed to obtain the final valid target frame. By detecting and clustering drone signals on the time-phase difference map, the algorithm becomes insensitive to pixel mutations. Using a trained deep learning model also avoids the need to consider the type of convolution kernel, simplifying the detection process and effectively adapting to low signal-to-noise ratios (e.g., below 10 dB). This reduces false alarm and false alarm rates, improving the recognition performance of drone image transmission signals at low signal-to-noise ratios. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the conventional technology, the following briefly introduces the drawings required for use in the embodiments or the conventional technology descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0026] Figure 1 Flowchart of a method for detecting drone signals based on deep learning in one embodiment;

[0027] Figure 2 A schematic diagram of a process for creating a data set before training in one embodiment;

[0028] Figure 3 The learning rate and loss change curves during training in one embodiment are shown;

[0029] Figure 4 A system framework and architecture flow chart for the method in one embodiment;

[0030] Figure 5 A schematic diagram showing the comparison between a time-frequency diagram and a time-phase difference diagram in one embodiment;

[0031] Figure 61 is a schematic diagram of a specific detection process of a drone signal detection method based on deep learning in one embodiment;

[0032] Figure 7 is a precision-recall curve in one embodiment;

[0033] Figure 8 This is a preliminary inspection result in one embodiment;

[0034] Figure 9 is an Elbow method curve in one embodiment;

[0035] Figure 10 is a clustering result in one embodiment;

[0036] Figure 11 is a secondary detection result in an embodiment;

[0037] Figure 12 The final result in one embodiment;

[0038] Figure 13 2 is a block diagram of a deep learning-based drone signal detection system in one embodiment. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and Examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the description of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0040] It should be noted that the reference to "embodiment" in this document means that the specific features, structures or characteristics described in conjunction with the embodiment may be included in at least one embodiment of the present invention. The presentation of this phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It will be understood by those skilled in the art that the embodiments described herein may be combined with other embodiments. The term "and / or" used in the specification of the present invention and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.

[0041] The following describes the implementation of the present invention in detail with reference to the accompanying drawings in the embodiments of the present invention.

[0042] In one embodiment, Figure 1As shown, a method for detecting drone signals based on deep learning is provided, which may include the following steps S10 to S22:

[0043] S10, loading the trained deep learning model; the deep learning model is trained by using a pre-collected and annotated training data set; wherein, the pre-collected received signal is subjected to a series of processing to obtain a corresponding grayscale image, and then the image transmission signal of the drone is annotated on the grayscale image (i.e., target annotation). The series of grayscale images after annotation are used as a data set to pre-train the deep learning model, so that in the actual detection process, the received signal from the actual detection airspace is collected, and the corresponding grayscale image is obtained after processing and sent to the trained deep learning model for rapid detection.

[0044] S12, after collecting the received signal, sampling it with an ADC, adding a window, and performing a short-time Fourier transform, a signal time-frequency diagram is obtained;

[0045] S14, after performing vector accumulation on the signal time-frequency diagram, conjugate multiplying the signal time-frequency diagrams of the received signals received by the two radio frequency channels and calculating the arc tangent to obtain a time-phase difference spectrum;

[0046] S16, converting the time-phase difference map into a grayscale image and feeding it into a deep learning model to obtain a series of upper left corner vertex coordinates and lower right corner vertex coordinates of the target box, as well as confidence scores corresponding to the target framed by the target box;

[0047] S18, after retaining the target frames whose confidence scores are higher than the set confidence threshold, determining the categories of all retained target frames;

[0048] S20, traverse the categories of all retained target frames, perform secondary detection on all retained target frames, and obtain a list of candidate target frames corresponding to the categories;

[0049] S22, performing non-maximum suppression processing on the candidate target frame list corresponding to the category to obtain a valid target frame corresponding to the category.

[0050] The above-mentioned deep learning-based drone signal detection method collects received signals, samples them using an ADC, performs windowing, and performs a short-time Fourier transform (SFT) to obtain a signal time-frequency map. Vector accumulation is then performed to improve image quality. The signal time-frequency maps of the two RF channels are then conjugate-multiplied and inverse-tangented to obtain a time-phase difference map. This map is then normalized to a grayscale image and fed into a trained deep learning model for drone target detection. High-confidence target boxes are retained, followed by secondary detection to obtain candidate target boxes. Finally, non-maximum suppression is performed to obtain the final valid target box. By detecting and clustering drone signals on the time-phase difference map, the algorithm is insensitive to sudden pixel changes. Using a trained deep learning model also avoids the need to consider the type of convolution kernel, simplifying the detection process and effectively adapting to low signal-to-noise ratios (e.g., below 10 dB). This method reduces false alarm and false alarm rates, improving the recognition performance of drone image transmission signals at low signal-to-noise ratios.

[0051] In one embodiment, before the above step S10, the following model training steps may be further included:

[0052] The received signal is sampled by ADC, windowed and short-time Fourier transformed to obtain the signal time-frequency diagram;

[0053] Perform vector accumulation on the signal time-frequency diagram;

[0054] The time-frequency graphs of the received signals from the two RF channels are conjugate multiplied and the arc tangent is calculated to obtain the time-phase difference graph and then converted into a grayscale image;

[0055] The grayscale images corresponding to multiple groups of received signals are used as training data sets for target annotation;

[0056] The training data set after target annotation is sent to the deep learning model for training until the detection error of the deep learning model for the target is lower than the set error threshold, and the trained deep learning model is exported.

[0057] It is understandable that Figure 2 As shown in the figure, during the model training phase, the received signal is first sampled using an ADC (analog-to-digital conversion), windowed (such as a Hamming window), and subjected to a short-time Fourier transform (SFT). This results in a signal time-frequency diagram (in matrix form). The data within the signal time-frequency diagram is complex. The signal time-frequency diagram is two-dimensional and consists of many pixels. Each pixel has a coordinate indicating its location on the signal time-frequency diagram.

[0058] Then, in order to improve the quality of the signal time-frequency diagram, it is necessary to perform vector accumulation on the signal time-frequency diagram; the width of the time-phase difference spectrum is w o , Gao Wei ho ; Vector accumulation means setting a window size. The width of the window w t and high h t , making the width of the time-phase difference spectrum w o Can be w t Divide by integer, the result is w ; High time-phase difference spectrum h o Can be h t Divide by integer, the result is h ; Therefore, the original signal time-frequency diagram can be divided into windows, and summing the data in each window can form a new image of size Signal time-frequency diagram.

[0059] The time-phase difference spectrum is obtained by conjugate multiplication of the signal time-frequency graphs of the received signals of the two RF channels and calculating the inverse tangent. The time-phase difference spectrum is then converted into a grayscale image through normalization. Since the phase difference range is between (-π, π), the following processing is required:

[0060] (1)

[0061] in, is the grayscale image obtained after processing, w , h is the width and height of the time-phase difference spectrum, For coordinates in x , y The phase difference of the point at .

[0062] The grayscale images created by multiple sets of signals are used as training datasets to label the targets of drone image transmission signals.

[0063] The grayscale images of the labeled training data set are sent to the deep learning model for training; the existing deep learning model suitable for target detection of the present invention can be but is not limited to the YOLO model, FasterRCNN model, MaskRCNN model or SSD model. The FasterRCNN model is used as an example in this specification.

[0064] When the target detection error during training is lower than the set error threshold, the trained deep learning model is exported; otherwise, the model parameters of the deep learning model are adjusted and the grayscale images created by multiple sets of signals are returned to the step of annotating the training data set, and retraining is performed until the detection accuracy reaches the standard. Figure 3The figure shows the learning rate and loss curves during training. The target detection error can be evaluated from the perspectives of the model's loss function and detection accuracy. On the one hand, training ends when the loss value is no longer optimized during model training. The loss value can be considered an adaptive threshold (i.e., the loss value will continue to decrease during training, but when the model performance bottleneck is reached, the loss value will fluctuate within a small range or remain at a fixed value). If the loss value does not decrease after several consecutive rounds of training (or after all training rounds are completed), the model is considered to have reached the performance limit of this training and continued training will not improve, thus terminating the training. On the other hand, the error threshold can be determined based on actual needs. For example, but not limited to, when the target detection accuracy exceeds 85% (i.e., the detection error is less than 15%), the model can be determined to be trained and ready for use.

[0065] like Figure 4 As shown, the above-mentioned drone signal detection method based on deep learning includes four process processing links in the detection and clustering stage (i.e., steps S10 to S22), including target detection, clustering, secondary detection and non-maximum suppression (NMS).

[0066] Among them, target detection includes: loading the trained deep learning model; collecting the received signal, sampling the received signal through ADC, adding windows and performing short-time Fourier transform to obtain the signal time-frequency diagram; performing vector accumulation on the signal time-frequency diagram; conjugate multiplication of the signal time-frequency diagrams of the received signals received on the two RF channels and calculating the inverse tangent to obtain the time-phase difference spectrum, such as Figure 5 As shown in the figure, the time-phase difference map is converted into a grayscale image according to formula (1). The grayscale image is fed into the trained deep learning model to obtain a series of upper left corner vertex coordinates and lower right corner vertex coordinates of the target box, as well as the confidence scores corresponding to the target framed by the target box. The target box can be recorded as , the confidence score can be recorded as Each target box The coordinates contain The coordinates of the upper left vertex and The coordinates of the lower right vertex.

[0067] The target frames with confidence scores higher than the set confidence threshold are retained to obtain the target coordinates corresponding to these target frames. The confidence threshold can be determined based on the false alarm and missed alarm requirements to ensure that targets above the threshold are valid targets (reducing the false alarm rate). For example, but not limited to, if a low false alarm rate is required, the confidence threshold can be set to 0.7, otherwise it can be set to 0.3 or other values. This step is achieved through formula (2):

[0068] (2)

[0069] in, where The function returns the number that meets the conditions. is the set confidence threshold. Indicates the target box whose corresponding confidence score is higher than the confidence threshold.

[0070] About clustering: First determine whether clustering is needed. Specifically, consider the horizontal coordinate of the vertex coordinate in the upper left corner of the target box. The starting frequency of the corresponding signal, and the image transmission signal of the drone is OFDM modulated and will not have frequency hopping. The distribution of the starting frequency of the corresponding signal is used to determine whether clustering is needed:

[0071] (3)

[0072] in, N The target frame that is retained The number of target boxes whose confidence scores are higher than the confidence threshold, The horizontal coordinate of the upper left corner of the target box The mean of the distribution of Indicates the The horizontal coordinate of the upper left corner vertex of the target box, is the horizontal axis The standard deviation of the distribution.

[0073] In one embodiment, regarding the above-mentioned step S18, in the process of determining the categories of all retained target frames, if the distribution standard deviation of the horizontal coordinates of the upper left corner vertex coordinates of the target frame is less than the set standard deviation threshold, the number of categories of all retained target frames is determined to be 1 and the secondary detection processing flow is directly entered.

[0074] It can be understood that if the distribution standard deviation is less than the set standard deviation threshold (THS), it can be considered that the target box is concentrated in frequency distribution and has only one category. Clustering is not required, and the secondary detection process is directly entered. The standard deviation threshold can be set according to the accuracy requirements of the target classification. It can generally be set to 20% of the average width of the target box. The smaller the standard deviation threshold, the more likely data that originally belonged to one category will be divided into multiple categories. If the standard deviation threshold is too large, the opposite will happen.

[0075] In one embodiment, regarding the above step S18, in the process of determining the categories of all retained target boxes, if the distribution standard deviation of the horizontal coordinates of the upper left corner vertex coordinates of the target box is not less than the set standard deviation threshold, all retained target boxes are clustered using K-Means classification and Elbow method to determine the number of cluster categories.

[0076] Specifically, if the distribution standard deviation is not less than the set standard deviation threshold, the next step is to determine the number of clusters.

[0077] Determining the Number of Clusters: Because different drones may transmit signals simultaneously in the time dimension, adding time coordinates to clustering will reduce clustering effectiveness. Therefore, only frequency-dimensional data is used in clustering. K-Means classification (a commonly used unsupervised learning algorithm) and the Elbow method (a method for determining the optimal number of clusters, K, in K-Means clustering) are used to select the appropriate number of clusters. The Elbow method primarily determines the optimal K value based on the sum of squared errors (SSE) of the clusters. SSE is the sum of the squares of the distances from each data point to its cluster center. As the K value increases, the SSE decreases because more clusters make the data points within each cluster more similar, and the distances from the cluster center decrease. However, after a certain K value is reached, the rate of decrease in SSE slows down. The Elbow method plots the SSE against the K value and finds the point on the curve that resembles an elbow. The K value corresponding to this point is the optimal number of clusters.

[0078] Specifically, assuming the number of clusters is , , perform the following iterations: use K-Means clustering to cluster the starting frequency and ending frequency of the target box, that is, initialize k Cluster centers (which can be recorded as Afterwards, the Euclidean distance from each object to each cluster center is calculated:

[0079] (4)

[0080] in, Indicates the object property, Indicates the The cluster center property. Represents the Euclidean distance from the object to the cluster center. The 2 on the cumulative symbol here indicates that the object here has two attributes, namely the horizontal coordinates of the upper left corner vertex coordinates and the horizontal coordinates of the lower right corner vertex coordinates of the target box, corresponding to the starting frequency and ending frequency of the signal. After each classification is completed, the sum of squared errors is calculated. :

[0081] (5)

[0082] along with As the number of samples increases, the sample division becomes finer, and the SSE decreases. The SSE value of the first iteration is the same as that of the The SSE value of the first iteration is significantly reduced compared to the second iteration (e.g., less than the 1 / 10 of the iterations), then stop the iteration, That is the real number of clusters. At this time, the clustering result of K-Means clustering is the real clustering result.

[0083] Secondary detection: Traverse all cluster categories and calculate the mean of the horizontal coordinates of the upper left corner vertex coordinates of all target boxes in each category, the mean of the width of the target box, and the starting frequency and bandwidth of the corresponding signal:

[0084] (6)

[0085] in, For the The mean of the horizontal coordinates of the upper left vertex coordinates of the class, For the The mean width of the target box in the class, For the No. The horizontal coordinate of the upper left corner vertex of the target box, For the No. The horizontal coordinate of the lower right corner vertex of the target box. For the The number of target boxes in the class.

[0086] In one embodiment, the above-mentioned drone signal detection method based on deep learning may further include the following processing steps:

[0087] Traverse the target boxes with confidence scores below the confidence threshold and classify them as valid target boxes if their width deviation is less than the width deviation threshold and the horizontal coordinate deviation of the upper left vertex coordinate is less than the coordinate deviation threshold. The width deviation threshold can be set to ±5% of the width. A higher value results in a higher false alarm rate and a lower missed alarm rate, while a lower value results in a higher missed alarm rate and a lower false alarm rate. The same principle applies to the coordinate deviation threshold.

[0088] It can be understood that the target boxes with confidence scores lower than the confidence threshold are traversed, and the target boxes with smaller width deviation and horizontal coordinate deviation of the upper left vertex coordinate are classified as valid target boxes:

[0089] (7)

[0090] in, For the target boxes whose confidence scores are lower than the confidence threshold, For the k The set of valid target boxes of the class, for The width, for The horizontal coordinate of the upper left vertex coordinate, is the threshold of the width error of the target box, is the threshold of the horizontal coordinate error of the upper left corner vertex coordinate of the target box, is the absolute value function. Specifically, because low-confidence targets may have been directly filtered out in the previous processing, in order to detect targets that may have been missed (i.e., missed alarms) in the model detection results and reduce the missed alarm rate, this embodiment can also re-detect based on the patterns of the already detected targets, as shown in the aforementioned formula (7).

[0091] Non-maximum suppression (NMS): After the second detection, there may be overlapping target frames, and NMS processing is required to remove false alarm signals. First, traverse all categories and calculate the area of ​​all target frames in each category, which can be recorded as :

[0092] (8)

[0093] in, For the No. The coordinates of the upper left corner of the target box, For the No. The coordinates of the lower right corner of the target box.

[0094] Specifically, after sorting the target frames in each category in descending order according to the confidence scores, all candidate frames corresponding to each category are taken as the first candidate frame among all candidate frames of each category. For each category, the first candidate frame among all candidate frames is considered as the valid target frame, and then the overlapping area of ​​the remaining candidate frames and the first candidate frame is calculated, which is recorded as :

[0095] (9)

[0096] in, For the The coordinates of the upper left corner of the overlapping parts of all other remaining candidate boxes of the class and the first candidate box, For the The coordinates of the lower right corner of the overlapping portion of all other remaining candidate boxes of the class with the first candidate box (valid target box). For the The coordinates of the upper left corner vertex of the first candidate box among all the remaining candidate boxes of the class, For the The coordinates of the lower right corner vertex of the first candidate box among all the remaining candidate boxes of the class. For the The coordinates of the upper left corner vertex of all candidate boxes except the first candidate box in the class. For the The coordinates of the lower right corner vertex of all candidate boxes in the class except the first candidate box.

[0097] Calculate the intersection-over-union ratio of the candidate boxes of each category based on the overlapping area:

[0098] (10)

[0099] in, Indicates the The intersection of the first candidate box in the class and the remaining candidate boxes, Indicates the The area of ​​the first candidate box among all candidate boxes in the class, Indicates the The area of ​​all remaining candidate boxes in the class except the first candidate box.

[0100] For each category, remove all candidate frames whose IoU ratio is greater than the set IoU threshold (since the image transmission signals of a single drone do not overlap, this threshold can be set below 20%) and send the first candidate frame to the valid target frame queue corresponding to the category. If the remaining candidate frames are 0, it means that all candidate frames of the category have been processed, and the valid target frame in the valid target frame queue is the final target frame of the category. If the remaining candidate frames are not 0, then return to the target frame processing step and sort them in order of confidence score from high to low to continue iterating until all candidate frames of the category have been processed.

[0101] In some embodiments, as Figure 6 As shown in the figure, it is a schematic diagram of the specific detection process of the above-mentioned UAV signal detection method based on deep learning. Figure 7 The precision-recall curves shown in Figure 2 show that the recall and accuracy of the deep learning-based drone signal detection method are both over 90%. Figure 8 As shown, Figure 9 is the Elbow method curve, Figure 10 is the clustering result, Figure 11 For the secondary test results, Figure 12 The final result is shown in Figure 1, where the blue box is the image transmission signal of drone target 0, and the red box is the image transmission signal of drone target 1. Drone target 0 and drone target 1 are two different drone targets.

[0102] Traditional edge detection algorithms (such as the Sobel edge detection operator) are more sensitive to pixel mutations. However, due to the presence of noise in the signal time-frequency diagram and the above-mentioned time-phase difference diagram, there will be more pixel mutations, which will cause the traditional edge detection algorithm to fail. The above-mentioned drone signal detection method based on deep learning of the present invention is insensitive to pixel mutations, and therefore does not have such sensitivity issues. It also does not need to consider the type of convolution kernel used like traditional image convolution algorithms. The processing is simpler and can effectively filter out some non-drone signals, with a lower false alarm rate.

[0103] It should be understood that although the above process Figure 1 The steps in the flowchart are shown in the order indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily performed at the same time, but can be performed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0104] In one embodiment, Figure 13As shown, a deep learning-based drone signal detection system 100 is provided, comprising a model loading module 11, a time-frequency conversion module 13, a vector atlas module 15, a model detection module 17, a target selection module 19, a secondary detection module 21, and a non-maximum suppression processing module 23. The model loading module 11 is used to load a trained deep learning model; the deep learning model is trained using a pre-collected and annotated training dataset. The time-frequency conversion module 13 is used to acquire received signals, sample them using an ADC, perform windowing, and perform a short-time Fourier transform to obtain a signal time-frequency map. The vector atlas module 15 is used to perform vector accumulation on the signal time-frequency map, conjugate multiply the signal time-frequency maps of the received signals received by the two RF channels, and calculate the inverse tangent to obtain a time-phase difference map. The model detection module 17 is used to convert the time-phase difference map into a grayscale image and feed it into the deep learning model to obtain a series of target box upper left corner vertex coordinates and lower right corner vertex coordinates, as well as the confidence score corresponding to the target framed by the target box. The target selection module 19 is used to retain target frames with confidence scores above a set confidence threshold and then determine the categories of all retained target frames. The secondary detection module 21 is used to traverse the categories of all retained target frames and perform secondary detection on all retained target frames to obtain a list of candidate target frames corresponding to the categories. The non-maximum suppression processing module 23 is used to perform non-maximum suppression on the list of candidate target frames corresponding to the categories to obtain valid target frames corresponding to the categories.

[0105] The deep learning-based drone signal detection system 100 collects received signals, samples them using an ADC, performs windowing, and performs a short-time Fourier transform (SFT) to obtain a signal time-frequency map. This map then undergoes vector accumulation to improve image quality. The time-frequency maps of the signals received by the two RF channels are then conjugate-multiplied and inverse-tangented to obtain a time-phase difference map. This map is then normalized to a grayscale image and fed into a trained deep learning model for drone target detection. High-confidence target frames are retained, followed by secondary detection to obtain candidate target frames. Finally, non-maximum suppression is performed to obtain the final valid target frame. By detecting and clustering drone signals on the time-phase difference map, the algorithm becomes insensitive to pixel mutations. Using a trained deep learning model also avoids the need to consider the type of convolution kernel used, simplifying the detection process and effectively adapting to low signal-to-noise ratios (e.g., below 10 dB). This reduces false alarm and missed alarm rates, improving the recognition performance of drone image transmission signals at low signal-to-noise ratios.

[0106] In one embodiment, in the process of determining the categories of all retained target frames, if the distribution standard deviation of the horizontal coordinates of the upper left corner vertex coordinates of the target frame is less than the set standard deviation threshold, the category of all retained target frames is determined to be 1 and the secondary detection processing flow is directly entered.

[0107] In one embodiment, in the process of determining the categories of all retained target boxes, if the distribution standard deviation of the horizontal coordinates of the upper left corner vertex coordinates of the target box is not less than a set standard deviation threshold, all retained target boxes are clustered using K-Means classification and Elbow method to determine the number of cluster categories.

[0108] In one embodiment, the target selection module 19 is further configured to traverse target frames whose confidence scores are lower than a confidence threshold, and classify target frames whose width deviation is less than the width deviation threshold and whose horizontal coordinate deviation of the upper left corner vertex coordinate is less than the coordinate deviation threshold as valid target frames.

[0109] In one embodiment, the above-mentioned deep learning-based drone signal detection system 100 may further include a model training module, which is used to obtain a signal time-frequency diagram after ADC sampling, windowing, and short-time Fourier transform of the received signal, perform vector accumulation on the signal time-frequency diagram, conjugate multiply the signal time-frequency diagrams of the received signals received by the two RF channels, and calculate the inverse tangent to obtain a time-phase difference map and convert it into a grayscale image. The grayscale images corresponding to the multiple groups of received signals are used as training data sets for target annotation and then sent to the deep learning model for training. When the detection error of the deep learning model for the target is lower than the set error threshold, the trained deep learning model is exported.

[0110] For the specific limitations of the deep learning-based drone signal detection system 100 , please refer to the corresponding limitations of the deep learning-based drone signal detection method above, which will not be repeated here.

[0111] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the above-described method embodiments. Any reference to memory, storage, database, or other media used in the various embodiments provided herein may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus DRAM (RDRAM), and DDR DRAM.

[0112] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0113] The above embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that a person skilled in the art would be able to make numerous modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of the present invention. Therefore, the scope of the present invention shall be determined by the appended claims.

Claims

1. A drone signal detection method based on deep learning, characterized in that: Including steps: Loading a trained deep learning model; the deep learning model is trained using a pre-collected and annotated training dataset; After the received signal is collected, it is sampled by ADC, windowed, and short-time Fourier transformed to obtain the signal time-frequency diagram; After vector accumulation of the signal time-frequency diagram, the signal time-frequency diagrams of the received signals received by the two radio frequency channels are conjugate multiplied and arc tangent is calculated to obtain a time-phase difference spectrum; The time-phase difference map is converted into a grayscale image and then fed into the deep learning model to obtain a series of upper left corner vertex coordinates and lower right corner vertex coordinates of the target box, as well as the confidence scores corresponding to the targets framed by the target box; After retaining the target boxes whose confidence scores are higher than the set confidence threshold, determine the categories of all retained target boxes; Traverse the categories of all retained target frames, perform secondary detection on all retained target frames, and obtain a list of candidate target frames corresponding to the categories; Perform non-maximum suppression on the candidate target frame list corresponding to the category to obtain the valid target frame corresponding to the category; The process of performing vector accumulation on the signal time-frequency diagram includes: setting the window size and summing the data in each window of the original signal time-frequency diagram; setting the window size so that the width of the time-phase difference spectrum is divisible by the width of the window, and the height of the time-phase difference spectrum is divisible by the height of the window.

2. The method for detecting drone signals based on deep learning according to claim 1, characterized in that: In the process of determining the categories of all retained target boxes, if the distribution standard deviation of the horizontal coordinates of the upper left corner vertex coordinates of the target box is less than the set standard deviation threshold, the number of categories of all retained target boxes is determined to be 1 and the secondary detection processing flow is directly entered.

3. The method for detecting drone signals based on deep learning according to claim 2, wherein: In the process of determining the categories of all retained target boxes, if the distribution standard deviation of the horizontal coordinates of the upper left corner vertex coordinates of the target box is not less than the set standard deviation threshold, all retained target boxes are clustered using K-Means classification and Elbow method to determine the number of cluster categories.

4. The method for detecting drone signals based on deep learning according to any one of claims 1 to 3, characterized in that: Also includes the steps: The target frames whose confidence scores are lower than the confidence threshold are traversed, and the target frames whose width deviation is less than the width deviation threshold and whose abscissa deviation of the upper left corner vertex coordinates is less than the coordinate deviation threshold are classified as valid target frames.

5. The method for detecting drone signals based on deep learning according to claim 1, wherein: Before loading the trained deep learning model, the following steps are also included: The received signal is sampled by ADC, windowed and short-time Fourier transformed to obtain the signal time-frequency diagram; Performing vector accumulation on the signal time-frequency diagram; The time-frequency graphs of the received signals from the two RF channels are conjugate multiplied and the arc tangent is calculated to obtain the time-phase difference graph and then converted into a grayscale image; The grayscale images corresponding to multiple groups of received signals are used as training data sets for target annotation; The training data set after target annotation is sent to the deep learning model for training until the detection error of the deep learning model for the target is lower than the set error threshold, and the trained deep learning model is exported.

6. A drone signal detection system based on deep learning, characterized in that: include: Model loading module, used to load trained deep learning models; The deep learning model is trained by using a pre-collected and annotated training data set; The time-frequency conversion module is used to collect the received signal, sample it through ADC, add window and perform short-time Fourier transform to obtain the signal time-frequency diagram; A vector spectrum module is configured to perform vector accumulation on the signal time-frequency graph, conjugate-multiply the signal time-frequency graphs of the received signals received by the two radio frequency channels, and calculate the inverse tangent to obtain a time-phase difference spectrum; wherein the process of performing vector accumulation on the signal time-frequency graph includes: setting a window size and summing the data of the original signal time-frequency graph in each window; setting the window size so that the width of the time-phase difference spectrum is divisible by the width of the window, and the height of the time-phase difference spectrum is divisible by the height of the window; A model detection module is used to convert the time-phase difference map into a grayscale image and then input it into the deep learning model to obtain a series of upper left corner vertex coordinates and lower right corner vertex coordinates of the target box, as well as the confidence scores corresponding to the targets framed by the target box; The target selection module is used to retain the target frames whose confidence scores are higher than the set confidence threshold and then determine the categories of all retained target frames; The secondary detection module is used to traverse the categories of all retained target frames, perform secondary detection on all retained target frames, and obtain a list of candidate target frames corresponding to the categories; The non-maximum suppression processing module is used to perform non-maximum suppression on the candidate target frame list corresponding to the category to obtain the valid target frame corresponding to the category.

7. The deep learning-based drone signal detection system according to claim 6, characterized in that: In the process of determining the categories of all retained target boxes, if the distribution standard deviation of the horizontal coordinates of the upper left corner vertex coordinates of the target box is less than the set standard deviation threshold, the number of categories of all retained target boxes is determined to be 1 and the secondary detection processing flow is directly entered.

8. The deep learning-based drone signal detection system according to claim 7, characterized in that: In the process of determining the categories of all retained target boxes, if the distribution standard deviation of the horizontal coordinates of the upper left corner vertex coordinates of the target box is not less than the set standard deviation threshold, all retained target boxes are clustered using K-Means classification and Elbow method to determine the number of cluster categories.

9. The deep learning-based drone signal detection system according to any one of claims 6 to 8, characterized in that: The target selection module is further used to traverse the target frames whose confidence scores are lower than the confidence threshold, and classify the target frames whose width deviation is less than the width deviation threshold and the horizontal coordinate deviation of the upper left corner vertex coordinate is less than the coordinate deviation threshold as valid target frames.

10. The deep learning-based drone signal detection system according to claim 6, characterized in that: It also includes a model training module, which is used to obtain a signal time-frequency diagram after ADC sampling, windowing and short-time Fourier transform of the received signal. After vector accumulation of the signal time-frequency diagram, the signal time-frequency diagrams of the received signals received by the two RF channels are conjugate multiplied and the inverse tangent is calculated to obtain a time-phase difference map and convert it into a grayscale image. The grayscale images corresponding to multiple groups of received signals are used as training data sets for target annotation and then sent to the deep learning model for training. The trained deep learning model is exported until the detection error of the deep learning model for the target is lower than the set error threshold.

Images