High-performance time sequence signal classification model for radio monitoring

Abstract

The invention discloses a high-performance time sequence signal classification model for radio monitoring, and the model comprises a DenseCNN backbone network module which is used for adjusting the number of channels of an input signal feature under the condition that the length and width of the input signal feature are not changed, and then outputting a multi-channel feature map containing the output feature of each convolution unit; the dimension transformation network module is used for converting the multi-channel feature map into a time sequence feature; the time sequence feature extraction network module is used for extracting time sequence features; and the classifier network module is used for carrying out classification prediction on the features extracted by the time sequence feature extraction network module, outputting a classification prediction probability and obtaining classification of signal features of an input model, verification shows that the accuracy of the model is superior to that of a traditional optimal model, and the model has a higher reasoning speed compared with a Bi-LSTM structure.

Description

Technical Field

[0001] The invention relates to the technical field of radio monitoring, and in particular to a high-performance time series signal classification model for radio monitoring. Background Art

[0002] Electromagnetic space involves many key areas, such as mobile phones, broadcasting, satellite communications, navigation, etc. Ensuring the security of electromagnetic space can ensure the stable operation of these communication systems and meet the daily communication needs of the country and the people.

[0003] Electromagnetic space is an indispensable infrastructure in people's daily life, work and social development. At present, digital, networked and intelligent information and communication technologies are developing rapidly, which has had a profound impact on the national economy and social construction, and has promoted the exponential growth of wireless services. However, in the process of rapid development of wireless services, the phenomenon of illegal frequency use has increased sharply, resulting in complex and disorderly challenges in spectrum management. Multiple signals of different sources and natures often share the same spectrum, causing normal communications to be interfered by unknown signals, increasing the bit error rate, and even leading to communication interruption.

[0004] Strengthen supervision and management of illegal frequency usage to ensure people’s communication security and social order, and provide solid protection for the security and development of the country and the people.

[0005] Radio monitoring technology is a technology used to detect, collect and analyze signals in the radio spectrum. It involves the use of specialized equipment and technical means to monitor and record the parameters and characteristics of radio signals to obtain intelligence about communication systems, equipment and activities.

[0006] The working scenarios of radio monitoring mainly include four categories: ground monitoring, ocean monitoring, air monitoring, and space-based monitoring. Ground monitoring includes urban electromagnetic environment monitoring and transportation monitoring, providing protection for the safety of people's livelihood communications and transportation safety.

[0007] Radio monitoring can be summarized into the following six steps: data acquisition, signal detection, signal analysis and identification, spectrum analysis, monitoring reporting and decision support, and spectrum management and intervention. The current problem is that signal analysis and identification are not accurate enough. Therefore, it is necessary to study a high-performance signal feature recognition model to improve the accuracy of signal analysis and identification. Summary of the invention

[0008] The purpose of the present invention is to provide a high-performance time series signal classification model for radio monitoring to solve the problems raised in the above background technology.

[0009] To achieve the above object, the present invention provides the following technical solution: a high-performance time series signal classification model for radio monitoring, comprising: The DenseCNN backbone network module includes several basic convolution units with skip connections, which are used to adjust the number of channels of the input radio signal features while keeping the length and width unchanged, and then output a multi-channel feature map containing the output features of each convolution unit; Dimension transformation network module: used to align the nodes at the same position in the information of different channels of the multi-channel feature map in time, and convert the multi-channel feature map into a time series feature; The time series feature extraction network module includes several stacked Transformer encoders for extracting time series features; The classifier network module includes a two-layer fully connected network, which is used to classify and predict the features extracted by the time series feature extraction network module, output the classification prediction probability, and obtain the classification of the signal features of the input model.

[0010] Preferably, the jump connection of several basic convolutional units is realized by stacking ResBlocks with jump connections, and the DenseCNN backbone network module uses 1×1 convolution to increase or reduce the dimension of feature channels to change the number of output channels.

[0011] Preferably, the basic convolution unit performs a one-dimensional batch normalization on the input features, then activates them through the ReLU function, and finally outputs them through a one-dimensional convolution layer. The convolution layer ensures that the length of the input features is consistent with the length of the output features by padding zeros at both ends.

[0012] Preferably, the Transformer encoder adds a token for classification to the beginning of the input feature sequence, and adds it to the subsequent learnable position code to obtain a learnable feature. During the network forward calculation, after three Transformer encoder calculations, the vector at this position in the result sequence is taken out as the feature vector of the input classifier network module.

[0013] Preferably, the DenseCNN backbone network module includes 15 basic convolutional units with skip connections, and the signal feature dimension of the input to the first basic convolutional unit is [N×2×1×128], each dimension represents the data BatchSize size, the number of feature channels, the height and width of each channel feature map, and the feature dimension of the output of the DenseCNN backbone network module is [N×242×1×128].

[0014] Preferably, in the Transformer encoder parameter setting, the number of multiple heads is 2 and the Dropout rate is 0.5.

[0015] Preferably, the cross entropy loss is used in the model training process to represent the difference between the probability distribution output by the model and the true probability distribution. The mathematical definition of the cross entropy loss function is: , where N represents the total number of samples and K represents the total number of sample categories. Represents the onehot vector of the i-th sample label in a batch of training samples. This vector can be expressed as y=[y1,…,y k ], when the sample belongs to the kth category, y k =1, otherwise y k =0, Represents the probability vector predicted by the model for the i-th sample, each element y k That is, the probability that the corresponding sample belongs to the kth class.

[0016] Preferably, the model uses accuracy as an evaluation indicator, where the accuracy is the ratio of the number of samples correctly predicted by the model to the total number of samples.

[0017] Preferably, the training data of the model is divided into a training set and a test set in an 8:2 ratio. During the training process, the prediction accuracy of the model on the test set is monitored in real time. When the accuracy stops improving or starts to decrease, the training process is terminated.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: the feature map finally output by the DenseCNN backbone network module in the present model contains the features of the previous layers, the dense connection mode of the basic convolution unit ensures the shallow convolution parameter gradient, and better performance can be obtained with a deeper layer, and the large convolution kernel effectively improves the receptive field, so that the feature richness received by the downstream is higher, and the shallow original features are kept to reach the subsequent classifier through the cross-layer channel cascade mode, and the dimension transformation module reduces the dimension of the features, while minimizing the impact on the accuracy, greatly improving and reducing the computational complexity, and after testing, the model using the Transformer structure has a more obvious performance advantage in the accuracy curve under high signal-to-noise ratio, therefore, the present model uses large convolution kernels and dense connections to obtain multi-scale receptive field features, and uses the self-attention mechanism to extract global features, and compared with the LSTM structure, obtains the ability to accelerate in parallel. Experimental results show that the accuracy of the present model is better than the traditional optimal model, and has a faster reasoning speed than the Bi-LSTM structure, reduces the model parameter amount and floating point number calculation amount, so that it can be applied to resource-constrained systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A schematic diagram of a basic convolution unit in a high-performance time series signal classification model for radio monitoring proposed by the present invention; Figure 2A schematic diagram of converting a multi-channel feature map in a high-performance time series signal classification model for radio monitoring proposed by the present invention into a time series feature; Figure 3 An example audio waveform diagram of a high-performance time series signal classification model for radio monitoring proposed by the present invention. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] The present invention provides a high-performance time series signal classification model for radio monitoring, comprising: The DenseCNN backbone network module includes several basic convolution units with skip connections, which are used to adjust the number of channels of the input radio signal features while keeping the length and width unchanged, and then output a multi-channel feature map containing the output features of each convolution unit. Specifically, Figure 1As shown in Table 1, the input of the DenseCNN backbone network module is a feature with a dimension of [N×2×1×128], where each dimension represents the data BatchSize size, the number of feature channels, and the height and width of each channel feature map. We assume that the input feature dimension of each convolution module is [N×Cin×1×128]. The feature first undergoes a one-dimensional batch normalization, which mainly makes the distribution of each channel of the input feature more stable and accelerates the convergence of the model. Then it is activated by the ReLU function to provide the convolution module with nonlinear representation capabilities. Finally, there is a one-dimensional convolution layer with a large convolution kernel of 1×15 and a sliding step of 1. The convolution layer ensures that the length of the input feature is consistent with the length of the output feature through zero padding at both ends. The number of output channels of all convolution layers in the backbone network is fixed, which we set to 16, which indicates the speed at which the channel of the feature map grows after each layer of densely connected convolution. The output feature map dimension of the basic convolution unit is [N×16×1×128]. It can be seen that the input feature and the output feature only differ in the number of channels, and the length and width of the feature have not changed. The cross-layer connection can be realized in the subsequent channel splicing module. Finally, we get a feature map dimension of [N×(Cin+16)×1×128] after the convolution module is spliced, which will be used as the feature input of the next convolution module. In the DenseCNN backbone network module, we did not use the pooling layer to shorten the width of the feature map to ensure that the feature map finally output by the network can contain the features output by each convolution module. Table 1 shows the changes in the feature dimension of each convolution module output. In the process of signal waveform feature extraction, the features of the shallow convolution represent some basic features of the waveform, and its receptive field is small, while the deep features represent some semantic features, and its receptive field is larger. Through the form of cross-layer connection, the feature map finally output contains receptive field features of multiple scales, so the feature representation ability of the network is richer.

[0022] Specifically, in order to change the number of channels of the output features of the DenseCNN backbone network module to meet the needs of the downstream Transformer network, for example, we hope that the number of channels can divide the number of heads to ensure that the feature dimension of each head is the same, or we want to increase or decrease the number of input feature channels to change the computational complexity of the downstream network. In the previous introduction to the backbone network, the number of channels of the final output can also be controlled by changing the growth rate parameter, but we found in the experiment that this control method has many shortcomings. The number of channels of the final output of the backbone network is mainly determined by the growth rate and the number of basic convolutional units. The number of convolutional units corresponds to the model depth. We hope that the model can have enough depth to ensure the ability of nonlinear mapping, and the growth rate is the level of each convolutional unit fusing the information of each channel to obtain a new expression channel. If this value is set too small, it will obviously inhibit the ability to extract model features. If it is set too large, the number of channels of the final output will double. Therefore, we fix the growth rate to a typical value of 16, and the number of channels of the final output of the DenseCNN backbone network module is flexibly changed by the Transition layer we added. The Transition layer includes a 1×1 convolutional layer, which can realize arbitrary changes in the number of output channels while retaining the performance of local features extracted by the DenseCNN backbone network module to the greatest extent.

[0023] Table 1: .

[0024] Dimension transformation network module: It is used to align the nodes at the same position in the information of different channels of the multi-channel feature map in time, and convert the multi-channel feature map into a time series feature, such as Figure 2 As shown in the figure, in the multi-channel feature map, the blocks in the same row represent the feature values ​​of each node in the same channel, and different rows represent different channels of the feature. If the order relationship between the nodes of each channel is not broken up, that is, in the output feature sequence, the vector of each time step is composed of different channels of the original feature, and the division of time steps is still based on the position relationship of each node. This method corresponds well to the characteristics of each sampling point position in the original signal waveform, but the vectors of each dimension in the same time step do not have a unified interpretation, which is equivalent to aligning the nodes at the same position in the information of different channels in time. The downstream network needs to learn this complex information representation.

[0025] The time series feature extraction network module includes three stacked Transformer encoders for extracting time series features. Specifically, the main parameters set for the Transformer encoder are the number of heads, the output dimension of each head, the Dropout rate, and the parameters of the FeedForward layer. As shown in Table 2, we only give the feature dimension change process and parameter settings of the first Transformer encoder, and the two encoders stacked later are exactly the same. In addition, in the feature dimension, N represents the Batch Size, M represents the number of channels output in the Transition layer, and “@” represents the parallel features output by multiple parallel self-attention mechanism modules.

[0026] Table 2: .

[0027] The classifier network module includes a two-layer fully connected network, which is used to classify and predict the features extracted by the time series feature extraction network module, output the classification prediction probability, and obtain the classification of the signal features of the input model. Specifically, before entering the classifier, the vector of the Token position in the Transformer encoder output needs to be extracted as the input of the classifier. The output of the classifier is the category prediction probability vector of N samples. The dimension of the vector is K, which represents the total number of signal categories.

[0028] Table 3: .

[0029] Regarding model training, cross entropy loss is used during model training to represent the difference between the probability distribution of the model output and the true probability distribution. The mathematical definition of the cross entropy loss function is: , where N represents the total number of samples and K represents the total number of sample categories. Represents the onehot vector of the i-th sample label in a batch of training samples. This vector can be expressed as y=[y1,…,y k ], when the sample belongs to the kth category, y k =1, otherwise y k =0, Represents the probability vector predicted by the model for the i-th sample, each element y k That is, the probability that the corresponding sample belongs to the kth class. At the same time, we use the Adam optimizer

[122] with default parameters to optimize the model to minimize the loss function. During the training process, our Batch Size is set to 320.

[0030] In terms of the evaluation index of the model, we use accuracy as the main metric, which is equal to the proportion of the number of samples correctly predicted by the model to the total number of samples. In the data set, the data under the same modulation mode have different signal-to-noise ratio problems. We believe that the model may be too simplistic to use the average accuracy under the full signal-to-noise ratio as an evaluation index, because in actual tasks, the accuracy under low signal-to-noise ratio is too low and has no practical significance. If the accuracy under high signal-to-noise ratio is sacrificed in exchange for the improvement of the average accuracy under the full signal-to-noise ratio, it is not worth the loss. In short, the accuracy under high signal-to-noise ratio is more important. Therefore, we propose a new metric, namely the average accuracy above 0dB. During training, we rely on the change of the average accuracy above 0dB on the test set as the basis for controlling the early stop of training. In terms of learning rate setting, our model uses 6 sets of learning rates from 1e-4 to 6e-4. Under the 6 sets of learning rates, the model with the highest accuracy above 0dB is selected as the final result of this training.

[0031] The dataset we used for training and testing comes from RadioML2016.10A, which is a popular open source single-carrier modulation classification dataset. The basic information of the dataset is shown in Table 4, and the parameter settings of the dynamic channel used in the dataset are shown in Table 5. This dataset covers a wide range of modulation types, with small data length and sample size, which is suitable for quickly verifying the effectiveness of the model. It is a basic dataset for signal modulation classification.

[0032] Table 4: .

[0033] Table 5: .

[0034] However, from the perspective of data rigor, there are two unreasonable points when the RadioML 2016.10A dataset was generated. First, in the label of the modulation category, the dataset proposer mistakenly marked the modulation mode 4ASK as fourth-order pulse amplitude modulation. We have corrected this in Table 4, but this does not affect the normal training process. It is just that the trained model cannot really be used to identify PAM4 signals. Second, there are some inaccuracies in the signal source used for analog modulation, such as Figure 3 The figure shows a common audio waveform. If we use it as the data source, we need to extract the amplitude part as the source of modulation. If we randomly extract a section as the signal source, it may be a situation where most of the time the signal source is noise. At this time, the output of each analog modulation is also similar to a kind of noise. In this way, there will be some contaminated modulation categories in the data set, which affects the training of the entire data set.

[0035] The software and hardware platform information we used during the model training and testing phases is shown in Table 6. In terms of training data selection, we divided the training set and the test set into an 8:2 ratio. During the training process, we monitored the prediction accuracy of the model on the test set and terminated the training process when the accuracy stopped improving or began to decrease.

[0036] Table 6: .

[0037] After testing the model, we experimentally verified and analyzed the main improvements in the model, including the impact of large convolution kernels and dense connections on network performance. By comparing the performance of convolution kernels of various sizes in the ResNet model and convolution kernels of various sizes in the DenseCNN model, we demonstrated the effectiveness of the convolution network design in the model of this application; Transformer encoder and feature segmentation, comparing the performance difference between DenseCNN and the model of this application, proving the effectiveness of the Transformer structure in the model of this application; Transition structure and model dimensionality reduction, by changing the channel of Transition output, we obtained two models with different capacity and compared their computational complexity and performance changes; two evaluation indicators, discussing the differences between the model of this application under two accuracy indicators; comparison with other models, analyzing the differences between the model of this application and other models on the RadioML2016.10A dataset, as follows: Regarding the impact of large convolution kernels and dense connections on network performance: In order to verify the effectiveness of the two design methods of large convolution kernel and dense connection in the model of this application, we selected the ResNet structure and the DenseCNN backbone network module in the model of this application for control experiments. In the experiment, each model globally average pooled the channels output by the last layer to obtain a feature vector with the same number of channels, and then passed it through a classifier with the same parameters to obtain the prediction results. Finally, we observed their performance differences under different convolution kernel sizes. We found that in the conventional CNN structure, directly increasing the size of the convolution kernel will not improve performance, but will reduce the accuracy of signal classification. As the size of the convolution kernel increases, the DenseCNN backbone network module maintains performance growth, while the performance of the ResNet network declines. Secondly, the network depths of the three models are different, but whether residual connection or dense connection is used, the performance will increase after the model layers become deeper, which can show that both cross-layer connection methods can effectively expand the model depth.

[0038] Regarding the impact of convolution kernel size and dense connection method on accuracy: There are two reasons why the DenseCNN backbone network module can achieve superior performance. The first is that the feature map output by the model contains the features of the previous layers. In the task of signal waveform classification, the features of each layer are beneficial for classification, rather than relying solely on the final semantic features. The dense connection method in the DenseCNN backbone network module ensures the shallow convolution parameter gradient, so that deeper layers can achieve better performance. The second point is that the large convolution kernel effectively improves the receptive field, making the features received downstream richer. Through the cross-layer channel cascade method, the original features of the shallow layer can be maintained to reach the subsequent classifier, while the addition method of ResNet will destroy the features extracted by the shallow large convolution kernel receptive field.

[0039] About Transformer encoder and feature segmentation: In order to prove that the Transformer encoder structure has better performance than the LSTM structure and is suitable for global feature extraction, this application connects the DenseCNN backbone network module to a two-layer bidirectional LSTM (Bi-LSTM) structure. The number of neurons in the Bi-LSTM hidden node is 128, and the DenseCNN backbone network module does not use the Transition layer to compress the dimension, that is, M=242 in Table 3. At this time, the LSTM structure obtains the same local features as the model of this application, and then observes the performance difference between the two models. The basic performance of the DenseCNN backbone network module is 62% of the full SNR average accuracy and 92% of the average accuracy above 0dB. The results show that the accuracy curve of the model using the Transformer structure has a more obvious performance advantage under high signal-to-noise ratio, and is also greater than other control group models in the range of -10dB to -5dB, but after it is lower than -12dB, the performance of the model will deteriorate. The average accuracy of Bi-LSTM at full SNR is 62.8%, and the average accuracy above 0dB is 92.5%. The performance gain after connecting the LSTM structure to the DenseCNN backbone network module is not obvious. When we use the Transformer structure, the average accuracy at full SNR is 63.49%, and the average accuracy above 0dB is 92.85%, which has reached the highest level on the RadioML2016.10A dataset. In terms of the inference speed of a single sample, the latency of the DenseCNN backbone network module is 40% smaller than that of CNN-BiLSTM.

[0040] About Transition structure and model dimensionality reduction: Through the previous experiments, it is proved that the model of this application exceeds the CNN-LSTM structure in the modulation classification task, but at this time the computational complexity of the model of this application is relatively high, its parameter volume reaches 2577.2K (unit thousand), and the total floating point operation volume reaches 331.2M (unit million), and the huge amount of calculation exceeds many conventional signal classification models. By analyzing the amount of calculation of each part of the model of this application, as shown in Table 7, we can find that the amount of calculation mainly comes from the Transformer structure. The calculation of the Transformer is mainly to generate the QKV feature matrix. The amount of calculation is proportional to the dimension of the matrix, and the matrix dimension is related to the dimension of the feature vector input to the Transformer, that is, M in Table 3. Therefore, the most direct way to reduce the amount of calculation is to control the size of M, that is, to compress the original features output by the DenseCNN backbone network module and reduce the number of channels, but this compression method may destroy the characteristics of the original shallow features and bring about a reduction in accuracy. In order to compress the DenseCNN backbone network module, we took the following four measures: 1. Reduce M from 242 to 128 through the Transition layer; 2. Reduce the number of hidden layer units in the last FeedForward module in the Transformer Block from 1024 to 128; 3. Remove one Transformer Block; 4. Increase the number of Transformer Block heads from 2 to 8. Through these operations, we get a compressed model. The comparison of this model with the original model in terms of model capacity and accuracy is shown in Table 8. From the perspective of parameter quantity and floating-point calculation quantity, the complexity of the compressed model is about 7 times lower than that of the original model. The main reason is that the complexity of the fully connected computing structure of the Transformer is proportional to the dimension of the input vector. When the input dimension is reduced, the overall cost of the model is also reduced (the input vector length of the Transformer structure is reduced to 1 / 8 of the original length). After compression, the average accuracy of the model at full SNR decreases. This is because after the total amount of model parameters is reduced, the model's ability to fit complex data distribution is also reduced accordingly. This performance degradation can be alleviated by adjusting parameters during model training. In addition, we found that the average accuracy will be slightly improved above 0dB. This may be due to the mutual exclusivity of data distribution under low signal-to-noise ratio and high signal-to-noise ratio in the dataset of this task. The mutual exclusivity of the data means that the model can only learn from one distribution type. As the number of model parameters decreases, the performance deteriorates when fitting fuzzy data under low signal-to-noise ratio, so the accuracy will increase under high signal-to-noise ratio. In terms of the inference speed of the model, the compressed model only needs 30 iterations to reach the optimal effect, and the total training time is less than 40 minutes.

[0041] Table 7:

[0042] Table 8:

[0043] About two evaluation indicators: In previous modulation classification studies, the commonly used evaluation indicator is the average accuracy under full SNR

[128] . However, in our study, we observed that for the model of this application, if the average accuracy under full SNR is continuously improved during training, the model will tend to optimize the accuracy under low SNR. The accuracy below -8dB will increase from 20% to 25%, while the average accuracy under high SNR (above 0dB) will drop from 93% to 91%. Therefore, we use the average accuracy above 0dB to control the progress of training during training, and stop training when the accuracy no longer increases.

[0044] Comparison with other models: We selected three mainstream automatic modulation classification algorithms for comparison, namely DSCLDNN

[105] , ResCNN

[129] and CLDNN

[129] . After comparison, we found that the model of this application achieved the best results. The compressed model achieved the highest classification effect in AM-SSB, but according to our previous analysis of the data set, we believe that at this time, the AM-SSB category includes the noise category, which does not mean that the model's recognition of this category is significantly better than the other two analog modulations.

[0045] In summary, this application solves the problems of insufficient performance and slow reasoning speed in common CNN and CNN-LSTM structures, and designs a high-performance time series signal classification model for radio monitoring. It uses large convolution kernels and dense connections to obtain multi-scale receptive field features, and uses the self-attention mechanism to extract global features, and obtains the ability to accelerate in parallel compared with the LSTM structure. Experimental results show that under the RadioML 2016.10A dataset, this model achieves performance that exceeds the traditional optimal model and faster reasoning speed than the Bi-LSTM structure. At the same time, we also propose a compressed version of the model, which reduces the number of model parameters and floating-point calculations, so that it can be applied to resource-constrained systems.

[0046] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A high-performance time series signal classification model for radio monitoring, characterized by: include: The DenseCNN backbone network module includes several basic convolution units with skip connections, which are used to adjust the number of channels of the input radio signal features while keeping the length and width unchanged, and then output a multi-channel feature map containing the output features of each convolution unit; Dimension transformation network module: used to align the nodes at the same position in the information of different channels of the multi-channel feature map in time, and convert the multi-channel feature map into a time series feature; The time series feature extraction network module includes several stacked Transformer encoders for extracting time series features; The classifier network module includes a two-layer fully connected network, which is used to classify and predict the features extracted by the time series feature extraction network module, output the classification prediction probability, and obtain the classification of the signal features of the input model.

2. A high performance time series signal classification model for radio monitoring according to claim 1, characterized in that: The skip connections of several basic convolutional units are realized by stacking ResBlocks with skip connections. The DenseCNN backbone network module uses 1×1 convolution to increase or decrease the dimension of feature channels to change the number of output channels.

3. A high performance time series signal classification model for radio monitoring according to claim 1 or 2, characterized in that: The basic convolution unit performs a one-dimensional batch normalization on the input features, then activates them through the ReLU function, and finally outputs them through a one-dimensional convolutional layer. The convolutional layer ensures that the length of the input features is consistent with the length of the output features by padding them with zeros at both ends.

4. A high performance time series signal classification model for radio monitoring according to claim 1, characterized in that: The Transformer encoder adds a token for classification to the beginning of the input feature sequence, and adds it to the subsequent learnable position code to obtain the learnable feature. During the forward calculation of the network, after three Transformer encoder calculations, the vector at this position in the result sequence is taken out as the feature vector of the input classifier network module.

5. The high performance time series signal classification model for radio monitoring according to claim 1, characterized in that: The DenseCNN backbone network module includes 15 basic convolutional units with skip connections. The signal feature dimension of the first basic convolutional unit is [N×2×1×128]. Each dimension represents the data BatchSize, the number of feature channels, and the height and width of each channel feature map. The feature dimension of the output of the DenseCNN backbone network module is [N×242×1×128].

6. A high performance time series signal classification model for radio monitoring according to claim 1 or 4, characterized in that: In the Transformer encoder parameter setting, the number of heads is 2 and the Dropout rate is 0.

5.

7. A high performance time series signal classification model for radio monitoring according to claim 1, characterized in that: During the training process of the model, cross entropy loss is used to represent the difference between the probability distribution of the model output and the true probability distribution. The mathematical definition of the cross entropy loss function is: , where N represents the total number of samples and K represents the total number of sample categories. Represents the onehot vector of the i-th sample label in a batch of training samples. This vector can be expressed as y=[y1,…,y k ], when the sample belongs to the kth category, y k =1, otherwise y k =0, Represents the probability vector predicted by the model for the i-th sample, each element y k That is, the probability that the corresponding sample belongs to the kth class.

8. A high performance time series signal classification model for radio monitoring according to claim 7, characterized in that: The model uses accuracy as the evaluation indicator, which is the ratio of the number of samples correctly predicted by the model to the total number of samples.

9. A high performance time series signal classification model for radio monitoring according to claim 8, characterized in that: The training data of the model is divided into training set and test set in an 8:2 ratio. During the training process, the prediction accuracy of the model on the test set is monitored in real time. When the accuracy stops improving or starts to decrease, the training process is terminated.

Images