Miniature unmanned aerial vehicle radio frequency signal monitoring method based on optimized residual neural network

Abstract

The invention discloses a small unmanned aerial vehicle radio-frequency signal monitoring method based on an optimized residual neural network in the technical field of unmanned aerial vehicle radio-frequency signals, and the method comprises the steps: taking a small unmanned aerial vehicle radio-frequency signal as a research object, training an unmanned aerial vehicle radio-frequency signal monitoring model through a cross validation protocol, and employing the grouping one-dimensional convolution. Local features of original one-dimensional signals with multi-scale resolutions are extracted from coarse to fine, a multi-level jump connection structure is built, and a final monitoring model is obtained. Compared with the prior art, the method is not sensitive to environment requirements, can effectively improve monitoring of unmanned aerial vehicles with different distances, different types and different operation modes in a complex environment, and can effectively improve the monitoring accuracy of the unmanned aerial vehicles through a multi-level jumping structure, gradient flow keeping, combination with multi-level pooling and collection of the depth characteristics of the radio frequency signals of the unmanned aerial vehicles. High precision is achieved by improving learning efficiency.

Description

Technical Field

[0001] The present invention belongs to the technical field of unmanned aerial vehicle radio frequency signals, and in particular is a method for monitoring radio frequency signals of small unmanned aerial vehicles based on an optimized residual neural network. Background Art

[0002] Unmanned aerial vehicles (UAVs) are small aircraft without human pilots that can perform tasks through remote control or autonomous flight. They are the result of cutting-edge technology deployed on small and low-cost devices. With the widespread use and popularity of small civilian drones, public safety issues and hidden dangers have become increasingly prominent. The regulation of small drones is crucial to the security of individuals, companies and even the country. Illegal misuse of drones may lead to the leakage of personal privacy, the theft of corporate secrets, and even threaten important national meetings.

[0003] Therefore, supervision and control of drones is a necessary measure to maintain public safety and national security.

[0004] Traditional drone detection and identification methods, such as radar, vision, and acoustics, are not very reliable because they are easily restricted by various environmental factors. Among them, radar detection is affected by factors such as terrain, water vapor, and clutter, visual detection is affected by factors such as light, atmospheric refraction, and weather, acoustic detection is easily affected by factors such as water quality, gas content, and water depth, and infrared detection is affected by meteorological conditions and enemy countermeasures. Compared with traditional drone detection methods, the combination of RF sensing and deep learning provides a solution that can effectively detect and classify drones.

[0005] The task of monitoring drone RF signals requires detecting drones, classifying drones, and identifying drone operation modes. Detecting drones requires judging the presence of drones based on the characteristics of RF signals. Classifying drones requires distinguishing the model of drones. Identifying drone operation modes requires identifying the specific operation modes of drones, such as flying, hovering, etc. It is difficult to achieve the above functions using existing RF sensing. Summary of the invention

[0006] To solve the above problems, the purpose of the present invention is to provide a method for monitoring the radio frequency signals of small drones based on an optimized residual neural network, which can detect the presence of small drones, distinguish their types, and identify their operation modes, and can be used to monitor the presence of drones, the types of drones, and the operation modes of drones. This method can ensure high detection accuracy for the defects of monitoring small drones in complex environments. The radio frequency signals of drones can be monitored intelligently and accurately.

[0007] In order to achieve the above object, the technical solution of the present invention is as follows: A method for monitoring radio frequency signals of a small UAV based on an optimized residual neural network comprises the following steps:

[0008] Step 1, add grouped one-dimensional convolution: build a deep neural network architecture with multiple convolutional layers, each layer is specified by several one-dimensional asymmetric filters of size 1×3 and 1×7, extracting local features of the original one-dimensional signal with multi-scale resolution from coarse to fine;

[0009] Step 2: Build a variation residual block, i.e., a multi-level skip connection structure: Design a multi-level skip connection, combined with a multi-level pool, to adjust the gradient flow and collect more global basic features;

[0010] Step 3, cross-validation protocol evaluation to fine-tune the drone RF signal monitoring model. The cross-validation method is used to randomly divide the dataset into 10 non-overlapping subsets, one of which is used as a validation set for testing, and the rest is used to train the network. This process is repeated 10 times to evaluate the entire dataset, and the overall performance is reflected by calculating the average performance of all subsets.

[0011] Further, step 1 includes the following method: step 11, input layer: define the number of groups of convolutional layers, the number of filters in each group, and the number of categories for multi-category operation mode classification; step 12, convolutional layer: define the convolutional layer to use a one-dimensional asymmetric filter of 1×3 or 1×7 with a stride of 1 to 3 or 1 to 7; step 13, normalization layer: use a batch normalization layer.

[0012] Further, step 2 includes the following method:

[0013] Step 21, configuration of convolutional layer 1: activated by ELU hidden layer 1, after convolutional layer 1 is activated, it is activated again by ELU hidden layer 2, and finally reaches global average pooling layer 2;

[0014] Step 22, configuration of convolutional layer 2: jump from ELU hidden layer 2 and maximum pooling layer 1 to addition layer 1, then connect to convolutional layer 2 from addition layer 1, activate through ELU hidden layer 3, and finally reach global average pooling layer 5;

[0015] Step 23, configuration of convolutional layer 3: jump from ELU hidden layer 3 and maximum pooling layer 2 to addition layer 2, then connect to convolutional layer 3 from addition layer 2, activate through ELU hidden layer 4, and finally reach global average pooling layer 4;

[0016] Step 24, configuration of convolutional layer 4: jump from ELU hidden layer 4 and maximum pooling layer 3 to addition layer 3, then connect to convolutional layer 4 from addition layer 3, activate through ELU hidden layer 5, and finally reach global average pooling layer 3;

[0017] Step 25, configuration of convolutional layer 5: jump from ELU hidden layer 5 and maximum pooling layer 4 to addition layer 4, then connect to convolutional layer 5 from addition layer 4, activate through ELU hidden layer 6, and finally reach global average pooling layer 6;

[0018] Step 26, configuration of global average pooling layer No. 1: jump from ELU hidden layer No. 6 and maximum pooling layer No. 5 to addition layer No. 5, and directly to global average pooling layer No. 1;

[0019] Step 27, deep cascade layer configuration: the global average pooling layers 1, 2, 3, 4, and 5 are connected in series to reach the deep cascade layer;

[0020] Step 28, output layer configuration: from the deep concatenation layer to the fully connected layer, activated by the softmax activation layer, and finally output the classification result.

[0021] Furthermore, in step 28, the score derived by the fully connected layer is normalized by the exponential function. Converted to the probability distribution of the class; where P j is satisfied 1≤P j ≤0 and The j-th class probability of the input data;

[0022]

[0023] The network assigns each input to one of n mutually exclusive classes based on the output of the softmax function and uses the Laplace transform Calculate the cross entropy loss of multiple classifications, where N represents the number of training signal frames and V ij represents the true value of the i-th input signal associated with the j-th class, and U ij Represents the output result of the network for the jth class of the i-th input signal:

[0024]

[0025] The above scheme has the following beneficial effects: The present invention uses the algorithm of optimizing residual neural network to design a small UAV radio frequency signal monitoring system based on optimized residual neural network. The system takes the radio frequency signal of small UAV as the research object, trains the UAV radio frequency signal monitoring model through the cross-validation protocol, uses grouped one-dimensional convolution to extract the local features of the original one-dimensional signal with multi-scale resolution from coarse to fine, builds a multi-level jump connection structure, and obtains the final monitoring model. Compared with the prior art, this method is insensitive to environmental requirements and can effectively improve the monitoring of UAVs at different distances, different models, and different operating modes in complex environments. Through the multi-level jump structure, the gradient flow is maintained, combined with multi-level pooling, the deep features of the UAV radio frequency signal are collected, and high precision is achieved by improving learning efficiency.

[0026] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 The UAV radio frequency signal monitoring system according to an embodiment of the present invention;

[0028] Figure 2 It is a coarse-to-fine feature extraction structure of an embodiment of the present invention;

[0029] Figure 3 This is a model structure diagram of drone radio frequency signal monitoring based on an optimized residual neural network according to an embodiment of the present invention;

[0030] Figure 4 It is a summary of the model modules of the embodiments of the present invention;

[0031] Figure 5 The drone presence detection result of the embodiment of the present invention;

[0032] Figure 6 The drone type detection result of the embodiment of the present invention;

[0033] Figure 7 The detection result of the operation mode of the drone according to the embodiment of the present invention;

[0034] Figure 8 It is a line chart comparing the detection results of the embodiment of the present invention. DETAILED DESCRIPTION

[0035] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0036] In the description of the present invention, it is necessary to understand that the terms "longitudinal", "lateral", "vertical", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside" and "outside" etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present invention.

[0037] In the description of the present invention, unless otherwise specified and limited, it should be noted that the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a mechanical connection or an electrical connection, or it can be the internal connection between two components. It can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to the specific circumstances.

[0038] The following is further described in detail through specific implementation methods:

[0039] A method for monitoring radio frequency signals of small unmanned aerial vehicles based on an optimized residual neural network comprises the following steps:

[0040] Prepare drone RF signal data in advance.

[0041] like Figure 1 As shown, software radio is used to collect the RF signals of the surrounding environment when the small drone does not exist and the RF signals of drones of different models and different operating modes, including the RF signals of the surrounding environment when there is no drone, and record the RF signals of drones in four operating modes: flying image transmission, flying without image transmission, hanging in the air, and only powered on and connected to the controller. In order to adapt to the neural network structure, all data are segmented and added with corresponding labels and serial numbers. If the technical specifications of the collected RF receiver are different, further data normalization processing is required, where x min and x max Represents the minimum and maximum amplitude values ​​in the RF signal data, scaling the value to between [0,1]:

[0042]

[0043] Step 1, add grouped one-dimensional convolution, add grouped one-dimensional convolution: build a deep neural network architecture with multiple convolution layers, each layer is specified by several one-dimensional asymmetric filters of size 1×3 and 1×7, extract local features of the original one-dimensional signal with multi-scale resolution from coarse to fine; including the following methods: Step 11, input layer: define the number of groups of convolution layers, the number of filters in each group, and the number of categories for multi-category operation mode classification; Step 12, convolution layer: define the convolution layer to use a one-dimensional asymmetric filter of 1×3 or 1×7, with a step size of 1 to 3 or 1 to 7; Step 13, normalization layer: use batch normalization layer. Specifically:

[0044] The present invention modifies the network framework of the residual neural network and adds multi-group convolution. Considering the large amount of RF signal data, in order to reduce the number of trainable parameters and computational cost, we perform grouped one-dimensional convolution on the training data. Figure 2 As shown in the figure, a deep neural network architecture with multiple convolutional layers is used, where each convolutional layer is specified by multiple 1×3 and 1×7 one-dimensional asymmetric filters. The number of channels of the conventional convolutional layer filter is always equal to the number of channels of the input of the layer, such as 3×3, 5×5, and 7×7. Specify the input layer as 1×28000×2, the filter stride as [1,7], and since the number of input channels is 2, the number of filter channels is also 2, and group convolution is performed; that is, step 11, define the number of groups of the convolutional layer, the number of filters in each group, and the number of categories for multi-category operation mode classification.

[0045] Convolution operation. Generally, in the spatial domain, the weights W i The convolution kernel and input map M i Calculation is performed at a specific coordinate (x, y). Where C is the convolution scalar and b is the scalar bias:

[0046]

[0047] Batch normalization operation: The convolution of K channels is the sum of the convolution results of K convolution scalars C, and its input is value normalized, using small batches and input channels to calculate the mean μ B and variance where ∈ = 10 -5 is a constant to prevent numerical uncertainty due to very small variances:

[0048]

[0049] The normalization layer handles inputs with zero mean and non-optimal variance by shifting and scaling the normalized values, passing them to the layer following the normalization layer. Once the training process is complete, the normalization layer calculates the mean and variance over the entire training set. During testing, in order to predict the category of the new input, the normalization layer uses the trained mean and variance to activate the normalized value instead of the mean and variance of the mini-batch. In convolutional neural networks for classification, batch normalization is often used between convolutional layers and eLU layers to speed up network training and reduce sensitivity to network initialization. In the formula, k and l represent the scale factor and offset parameters updated during network training, respectively:

[0050]

[0051] Modify the ReLU in the traditional residual neural network to eLU hidden layer activation: eLU can achieve faster learning convergence and higher accuracy in this classification task. In the convolutional layer, when 64 filters of size 1×7 and stride [1,7] are used, the width dimension of the output feature map is reduced by seven times, and the depth dimension is the same as the number of filters. Therefore, when the input signal frame after the normalization layer is 1×28000×2, the output unit activated by the first hidden layer is 1×4000×64. Where α=1 is the nonlinear parameter:

[0052]

[0053] Perform convolution grouping: Figure 3 As shown in the figure, there are 1, 2, 3, 4, and 5 convolution layers, the filter size is 1×3, the stride is [1,2], each convolution layer is divided into 8 groups, each group has 8 filters, and the horizontal scale of the feature map output of 1, 2, 3, 4, and 5 convolution layers is {2000, 1000, 500, 250, 125}. After each convolution layer is normalized, it is followed by 2, 3, 4, 5, and 6 eLU activation layers. Compared with traditional residual neural networks, eLU activation layers have more advantages than ReLU activation layers in dealing with the gradient vanishing problem. And under the condition that the filter learning rate meets the general sparse relationship, as the number of non-overlapping filter groups increases, the added grouped convolution method significantly reduces the number of parameters and computational cost.

[0054] According to the above operation methods, group convolution can be realized.

[0055] Step 2: Build a variation residual block, i.e., a multi-level skip connection structure: Design a multi-level skip connection, combined with a multi-level pool, to adjust the gradient flow and collect more global basic features; including the following methods:

[0056] Step 21, configuration of convolutional layer 1: activated by ELU hidden layer 1, after convolutional layer 1 is activated, it is activated again by ELU hidden layer 2, and finally reaches global average pooling layer 2;

[0057] Step 22, configuration of convolutional layer 2: jump from ELU hidden layer 2 and maximum pooling layer 1 to addition layer 1, then connect to convolutional layer 2 from addition layer 1, activate through ELU hidden layer 3, and finally reach global average pooling layer 5;

[0058] Step 23, configuration of convolutional layer 3: jump from ELU hidden layer 3 and maximum pooling layer 2 to addition layer 2, then connect to convolutional layer 3 from addition layer 2, activate through ELU hidden layer 4, and finally reach global average pooling layer 4;

[0059] Step 24, configuration of convolutional layer 4: jump from ELU hidden layer 4 and maximum pooling layer 3 to addition layer 3, then connect to convolutional layer 4 from addition layer 3, activate through ELU hidden layer 5, and finally reach global average pooling layer 3;

[0060] Step 25, configuration of convolutional layer 5: jump from ELU hidden layer 5 and maximum pooling layer 4 to addition layer 4, then connect to convolutional layer 5 from addition layer 4, activate through ELU hidden layer 6, and finally reach global average pooling layer 6;

[0061] Step 26, configuration of global average pooling layer No. 1: jump from ELU hidden layer No. 6 and maximum pooling layer No. 5 to addition layer No. 5, and directly to global average pooling layer No. 1;

[0062] Step 27, deep cascade layer configuration: the global average pooling layers 1, 2, 3, 4, and 5 are connected in series to reach the deep cascade layer;

[0063] Step 28, output layer configuration: from the deep concatenation layer to the fully connected layer, activated by the softmax activation layer, and finally output the classification result.

[0064] Specifically: (1) Adding skip connection processing in the addition layer. Where F represents the feature map, F pi is the most significant feature obtained by the maximum pooling operation, the pool size is (1,2) and the stride is [1,2]. g(i+1) is the feature obtained by the i+1th hidden layer, is the result of element-by-element addition of the first two. The i-th addition layer skip connection can be expressed as:

[0065]

[0066] The element-wise addition layer does not change the feature dimension. The proposed multi-level skip connection allows to retain most of the gradient information in different feature levels. Therefore, when the network becomes deeper, the gradient vanishing or exploding problem can be effectively prevented.

[0067] (2) In the traditional residual network, there is no structure of deep concatenation layer. Adding deep concatenation, such as Figure 4 As shown in the deep cascade module. In order to collect meaningful content from the previous layers, a multi-level global average pooling deep cascade structure is built. The global average pooling layer is set to calculate the mean of each feature map. It is usually located before the fully connected layer to reduce the network size and alleviate overfitting. For example, the global average pooling layer converts the dimension from 1×W×D to 1×1×D by performing average pooling on the channel dimension. The outputs of the 2, 3, 4, 5, and 6 global average pooling layers are first passed to the deep cascade layer to obtain global features. Where Represents the average pooling operations at different pool sizes.

[0068]

[0069] Where F multi represents a deep concatenated stack, where Represented as a join operation, such as Figure 4 The medium depth series module is shown below:

[0070]

[0071] The connection layer aggregates inputs of the same height and width along the depth (or channel) dimension. If each input has 1×1×64 channels, the output has 1×1×320 channels (or depth size), which is the sum of the channels of all inputs. The stacking result is connected to the last jump. The output results of the 5th addition layer are combined to get F final (1×1×384):

[0072]

[0073] F final The signal features of multiple scales are synthesized and then input into a fully connected layer defined by n neurons. The number of neurons in the fully connected layer is determined according to the number of classes in the data set. If a signal data set without drones and a signal data set with drones are used, then n = 2 is used to detect drones; if there are three types of drone signals, then n = 3 is used to classify drones; if there are four operation modes, then n = 4 is used for operation recognition.

[0074] (3) The output layer activation function still uses the softmax layer. The fully connected layer is followed by the softmax layer. This layer uses a normalized exponential function to convert the scores derived by the fully connected layer into Converted to the probability distribution of the class. j is satisfied 1≤P j ≤0 and The j-th class probability of the input data.

[0075]

[0076] The network assigns each input to one of n mutually exclusive classes based on the output of the softmax function and uses the Laplace transform Calculate the cross entropy loss of multiple classifications, where N represents the number of training signal frames and V ij represents the true value of the i-th input signal associated with the j-th class, and U ij Represents the output result of the network for the jth class of the i-th input signal:

[0077]

[0078] Step 3, cross-validation protocol evaluation to fine-tune the drone RF signal monitoring model. The cross-validation method is used to randomly divide the dataset into 10 non-overlapping subsets, one of which is used as a validation set for testing, and the rest is used to train the network. This process is repeated 10 times to evaluate the entire dataset, and the overall performance is reflected by calculating the average performance of all subsets.

[0079] Step 4: Test the detection effect of this system

[0080] (1) During testing, the dataset to be tested is input into the detection model. In the validation, the test set and training set are divided into 3:7. The model performance of three UAV surveillance tasks (i.e., UAV detection, UAV classification, and operation recognition) is benchmarked. Figure 5 , 6 ,7 shows the recognition and detection results when the iteration is 10 times, the initial learning rate is 0.01 and the loss rate is 0.001, where the average recognition accuracy of UAV presence detection is about 99.94%, the average recognition accuracy of UAV type classification is about 86.66%, and the recognition accuracy of UAV operation mode is about 80.13%. Figure 8 It shows that the recognition accuracy changes with the increase of iteration number.

[0081] (2) The present invention compares the test results of the model to confirm the advantages of combining grouped convolution and multi-level skip connections. We compare the model with its three variants, Models A, B, and C. Model A uses standard convolutional layers without merging skip connections and multi-level structures; Model B uses grouped convolutional layers without skip connections but with multi-level structures; Model C uses grouped convolutional layers and no multi-level structures but with skip connections. All networks are trained and evaluated using the same configuration. Compared with Model A using standard convolutional layers, Model B with grouped layers significantly reduces the network size and computational cost, while the performance is slightly reduced. In the comparison between Model C and Model B, the recognition accuracy is improved without increasing the number of parameters and traffic. The grouped convolutional layers are combined with multi-level skip connections. The model proposed in the present invention successfully achieves two goals: low complexity and high accuracy. It is about 70% lower than Model A in terms of complexity, and is about 10% higher than Model B in terms of effective drone classification and operation recognition, but is larger than Model B and Model C in terms of network size because the multi-gap structure of multi-level feature map aggregation increases the number of parameters in the fully connected layer, as shown in Table 1. Table 2 provides the model recognition accuracy obtained after using different hidden layer activation functions while ensuring the same number of iterations. Among them, the use of the eLU activation function generally has a high recognition accuracy, and the operation mode recognition accuracy is lower than that of LeakyReLU.

[0082] Table 1

[0083]

[0084]

[0085] Table 2

[0086]

[0087] The above is only an embodiment of the present invention, and the common knowledge such as the known specific structure and / or characteristics in the scheme is not described in detail here. It should be pointed out that for those skilled in the art, several deformations and improvements can be made without departing from the structure of the present invention, which should also be regarded as the protection scope of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A method for monitoring radio frequency signals of small UAVs based on an optimized residual neural network, characterized in that: The following methods are included: Step 1, add grouped one-dimensional convolution: build a deep neural network architecture with multiple convolutional layers, each layer is specified by several one-dimensional asymmetric filters of size 1×3 and 1×7, extracting local features of the original one-dimensional signal with multi-scale resolution from coarse to fine; Step 2: Build a variation residual block, i.e., a multi-level skip connection structure: Design a multi-level skip connection, combined with a multi-level pool, to adjust the gradient flow and collect more global basic features; Step 3, cross-validation protocol evaluation to fine-tune the drone RF signal monitoring model. The cross-validation method is used to randomly divide the dataset into 10 non-overlapping subsets, one of which is used as a validation set for testing, and the rest is used to train the network. This process is repeated 10 times to evaluate the entire dataset, and the overall performance is reflected by calculating the average performance of all subsets.

2. The method for monitoring radio frequency signals of small unmanned aerial vehicles based on optimized residual neural network according to claim 1 is characterized in that: Step 1 includes the following methods: Step 11, input layer: define the number of groups of convolutional layers, the number of filters in each group, and the number of categories for multi-category operation mode classification; Step 12, convolution layer: define the convolution layer to use a 1×3 or 1×7 one-dimensional asymmetric filter with a stride of 1 to 3 or 1 to 7; Step 13, Normalization layer: Use batch normalization layer.

3. The method for monitoring radio frequency signals of small unmanned aerial vehicles based on optimized residual neural network according to claim 2 is characterized in that: Step 2 includes the following methods: Step 21, configuration of convolutional layer 1: activated by ELU hidden layer 1, after convolutional layer 1 is activated, it is activated again by ELU hidden layer 2, and finally reaches global average pooling layer 2; Step 22, configuration of convolutional layer 2: jump from ELU hidden layer 2 and maximum pooling layer 1 to addition layer 1, then connect to convolutional layer 2 from addition layer 1, activate through ELU hidden layer 3, and finally reach global average pooling layer 5; Step 23, configuration of convolutional layer 3: jump from ELU hidden layer 3 and maximum pooling layer 2 to addition layer 2, then connect to convolutional layer 3 from addition layer 2, activate through ELU hidden layer 4, and finally reach global average pooling layer 4; Step 24, configuration of convolutional layer 4: jump from ELU hidden layer 4 and maximum pooling layer 3 to addition layer 3, then connect to convolutional layer 4 from addition layer 3, activate through ELU hidden layer 5, and finally reach global average pooling layer 3; Step 25, configuration of convolutional layer 5: jump from ELU hidden layer 5 and maximum pooling layer 4 to addition layer 4, then connect to convolutional layer 5 from addition layer 4, activate through ELU hidden layer 6, and finally reach global average pooling layer 6; Step 26, configuration of global average pooling layer No. 1: jump from ELU hidden layer No. 6 and maximum pooling layer No. 5 to addition layer No. 5, and directly to global average pooling layer No. 1; Step 27, deep cascade layer configuration: the global average pooling layers 1, 2, 3, 4, and 5 are connected in series to reach the deep cascade layer; Step 28, output layer configuration: from the deep concatenation layer to the fully connected layer, activated by the softmax activation layer, and finally output the classification result.

4. The method for monitoring radio frequency signals of small unmanned aerial vehicles based on an optimized residual neural network according to claim 3 is characterized in that: In step 28, the score derived by the fully connected layer is normalized by the exponential function Converted to the probability distribution of the class; where P j is satisfied 1≤P j ≤0 and The j-th class probability of the input data; The network assigns each input to one of n mutually exclusive classes based on the output of the softmax function and uses the Laplace transform Calculate the cross entropy loss of multiple classifications, where N represents the number of training signal frames and V ij represents the true value of the i-th input signal associated with the j-th class, and U ij Represents the output result of the network for the jth class of the i-th input signal:

Images