Radar signal open set identification method based on SoftMax information entropy

Abstract

The application provides a radar signal open set identification method based on SoftMax information entropy, research is carried out on the basis of SoftMax probability distribution, and the probability distribution information entropy is calculated for unknown discrimination; the information entropy is calculated based on the known type SoftMax probability distribution of output; the uncertainty degree of the identification result is reflected through the information entropy; whether the uncertainty degree of the identification result is higher than the preset is judged; if yes, the identification result of the radar signal to be identified is updated as an unknown radar signal category and is output; otherwise, the radar signal category with the highest probability value in the identification result is output. Compared with directly using the maximum probability, the threshold discrimination interval is improved, so that the unknown sample can be more easily identified; the application does not need additional network design and training, is simple to realize, has high real-time performance, has good open set performance, can be applied to any network based on the SoftMax layer classification, has wide application range and does not bring too much time consumption, can be used immediately, and has good practicability and real-time performance.

Description

TECHNICAL FIELD

[0001] The present application relates to radar signal recognition technology, in particular to radar signal open set recognition technology based on SoftMax information entropy. BACKGROUND

[0002] In real scenarios, more open and non-static environments will appear some never seen before. Open set recognition is used for the model that has been trained, when the test set appears in the training set, the unknown class data is processed properly, such as identifying as unknown type.

[0003] SoftMax function can convert the feature vector extracted by the network into a probability distribution, which contains the same number of probability values as the class of training data, and each probability value represents the probability of input radar signal sample being judged as the corresponding class. The trained deep learning model can determine the class of new input samples according to the maximum SoftMax probability, but in various practical application scenarios, the new input radar signal to be tested may come from unknown classes that have not been trained, and the role of open set recognition is to determine whether the radar signal belongs to unknown classes. The most basic open set method is to set a threshold for the maximum SoftMax probability, and the sample less than the threshold is unknown class.

[0004] So far, there have been many more advanced open set recognition methods, which can be roughly divided into machine learning based methods, deep discriminative model based methods and generative model based methods. The open set recognition based on machine learning is to improve some traditional machine learning methods such as SVM, sparse representation, nearest neighbor, etc., so that the original method is suitable for open set recognition; The method based on deep discriminative model often improves or replaces the SoftMax layer (usually the last layer of deep learning network, used to convert feature vector into probability distribution, and then the predicted class can be obtained according to the maximum probability), such as OpenMax technology, which uses Weibull distribution to improve SoftMax and increase unknown discrimination ability; The method based on generative model is based on generative adversarial network (GAN), which uses generator to generate false unknown class samples and uses discriminator to perform binary classification learning of known class samples and generated unknown samples. At present, in the field of open set recognition, the methods based on the first two ideas usually need very complex mathematical theory to support, and the implementation in many practical applications is difficult, while the third idea depends on the reliability of network design, and the network parameter adjustment and training is also a lot of work.

[0005] A kind of open set identification method based on maximum SoftMax probability, radar emitter signal is used as identification object, pseudo Wigner-Ville distribution PWVD is used to process signal, time-frequency image sample is generated, is input convolutional neural network CNN and is trained and learned, after training, new time-frequency graph sample is input into the trained model, obtain SoftMax probability distribution, set open set discrimination threshold, the sample with the maximum value of probability distribution less than the threshold is identified as unknown category.Reference: Yang Zhiyuan, Zhang Yuxi, Sun Jinping, etc. Radar emitter signal open set identification method based on CNN [C] / / China Electronics Society Digital Signal Processing Expert Committee. The 14th National DSP Application Technology Academic Conference Proceedings. Beijing University of Aeronautics and Astronautics School of Electronic Information Engineering;Electromagnetic Scattering Key Laboratory, 2022:5.The maximum probability will be a larger value when facing unknown category sample and trained category sample feature similarity is high, which makes it difficult to set open set discrimination threshold.

[0006] An open set identification method based on OpenMax, the output of the full connection layer before the SoftMax layer is processed and Weibull fitting, so that the possibility of new input sample belonging to known category can be judged according to the fitted distribution.Reference: Song Xing. Radar jamming intelligent perception method in complex electromagnetic environment [D]. Xi'an University of Electronic Science and Technology, 2022.The mathematical theory of open set identification method based on OpenMax is complex, and the engineering implementation is difficult, and the additional Weibull fitting training brings more training time overhead.

[0007] At present, the reliability of unknown discrimination directly using maximum SoftMax probability is low in the case of similar samples, and various more advanced open set identification methods are often complex in theory, difficult to implement, and have large training and parameter adjustment time overhead. SUMMARY

[0008] The technical problem to be solved by the present application is to provide a radar signal open set identification method based on a simple and feasible discrimination model idea with strong real-time performance.

[0009] The technical scheme adopted by the present application to solve the above technical problems is a radar signal open set identification method based on SoftMax information entropy, comprising the following steps:

[0010] Data sample generation step: time-frequency image of radar signal with known model category label is used as training sample to form training set;

[0011] Network training step: use a deep learning network with SoftMax layer as time-frequency image classification network, input the training set into the time-frequency image classification network to complete training;

[0012] The open set identification step: input the time-frequency image of the radar signal to be identified into the time-frequency image classification network which has completed training, and the time-frequency image classification network outputs an identification result, and the identification result is a SoftMax probability distribution of a known model; calculate the information entropy based on the output SoftMax probability distribution of the known model, and the information entropy reflects the uncertainty degree of the identification result, and it is judged whether the uncertainty degree of the identification result is higher than a preset or not, if yes, the identification result of the radar signal to be identified is updated as an unknown radar signal category and is output, otherwise, the radar signal category with the highest probability value in the identification result is output.

[0013] The present application carries out research on the basis of SoftMax probability distribution, calculates the information entropy of the probability distribution for unknown discrimination, and compared with directly using the maximum probability, the threshold discrimination interval is improved, so that the unknown sample can be more easily identified; the present application does not need complex mathematical theory basis, and does not need additional network design and training, and has simple realization, high real-time performance, good open set performance, can be applied to any network based on SoftMax layer classification, wide application range, no excessive time overhead, plug and play, good practicability and real-time performance.

[0014] The present application has the advantages that compared with the identification scheme of the maximum SoftMax probability value, the open set discrimination space can be improved, more unknown samples can be rejected, and the reliability of the SoftMax unknown discrimination method is improved; the mathematical operation complexity is low, easy to implement in engineering; no additional network needs to be built, and the time overhead of parameter adjustment and training is saved. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 It is a radar signal open set identification diagram based on SoftMax information entropy of the present application;

[0016] Figure 2 It is a radar signal time-frequency image of seven known models of the embodiment;

[0017] Figure 3 It is a radar signal time-frequency image of five unknown models of the embodiment;

[0018] Figure 4 It is a SoftMax probability distribution diagram of the known category sample and the unknown category sample of the embodiment;

[0019] Figure 5 It is a diagram of the maximum SoftMax probability and P H distribution of the unknown category of the embodiment;

[0020] Figure 6The open set recognition method based on maximum SoftMax probability and the open set recognition method based on SoftMax information entropy are respectively used for testing set recognition method to obtain the relationship curve between threshold and rejection rate and false recognition rate. DETAILED DESCRIPTION

[0021] As shown in Figure 1 , the flow of the present application includes three parts of data sample generation, network training and open set recognition:

[0022] 1) Data sample generation

[0023] Seven known radar signal models and five unknown radar signal models are prepared, each model including 200 pulse signals, each pulse signal is subjected to short-time Fourier transform, time-frequency image is generated and subjected to grayscale, size scaling and other operations, and finally 224*224 size image samples are obtained, the typical time-frequency image of the known model is shown as Figure 2 , and the unknown model is shown as Figure 3 .

[0024] All the samples of the models are labeled with model names as class labels, and each known model is used as a training set, and the unknown model is used as a test set for subsequent open set recognition test.

[0025] 2) Network training:

[0026] The present application is essentially based on the SoftMax layer, so as long as the deep learning network using the SoftMax layer for classification task can be used as the time-frequency image classification network of the present application for open set recognition. The embodiment uses ResNet18 as the time-frequency image classification network, and ResNet18 is a ResNet containing 18 hidden layers. Since the number of classes used for training is 7, the output size of the last fully connected layer of the ResNet18 network needs to be changed to 7, and a SoftMax layer is added. The classification network outputs the probability value of classification to each class, that is, a SoftMax probability distribution. The training set is input into the classification network, and after multiple rounds of training, the current network model parameters are saved, and the model training is completed.

[0027] 3) Open set recognition:

[0028] The test set sample is input into the trained classification network model, and the classification network model outputs the SoftMax probability distribution of each type. Generally, the SoftMax probability distribution output by the unknown class sample will be smoother than that of the known class. Taking the SoftMax probability distribution of a known class sample and a unknown class sample as an example, as shown in Figure 4The known class samples are most likely to be classified as a certain class, while the unknown class samples are relatively more dispersed in the probability of being classified as each class, because the features of the unknown samples extracted by the network can be different from the features of the known class samples, making the network's class prediction uncertainty high.

[0029] Information entropy is an index for measuring the degree of uncertainty of a random variable. The greater the entropy value, the higher the uncertainty of the random variable, and the smaller the uncertainty. Since the SoftMax function essentially gives the probability of a test sample belonging to each class in the training set, the lower the certainty of the network about the class to which the test sample belongs, the closer the probability values of each class output by SoftMax, and each class possibility will exist, which is the reason for the aforementioned smoother probability distribution. The uncertainty of the SoftMax output probability distribution can be calculated using information entropy as follows:

[0030]

[0031]

[0032] where H represents the information entropy value, k is the sample class sequence variable, the lower the information entropy, the more concentrated the SoftMax distribution of the sample, and the higher the confidence of being classified as a known class, N represents the number of classes in the training set, p k is the probability of the current sample being classified as the kth class, q k is the intermediate value, which is calculated by equation (2).

[0033] Optionally, a preset minimum uncertainty tolerance is used as a threshold. If the information entropy H is greater than the threshold, the identification result of the radar signal to be identified is updated as an unknown radar signal class and output, otherwise the radar signal class with the highest probability in the identification result is output.

[0034] Optionally, in order to determine the threshold similar to the maximum SoftMax probability, samples less than the threshold are classified as unknown classes, and the known confidence is measured using the information entropy probability P H

[0035] P H = 1-H (3)

[0036] The preset confidence is used as a threshold. If the information entropy probability P H is less than the threshold, the identification result of the radar signal to be identified is updated as an unknown radar signal class and output, otherwise the radar signal class with the highest probability in the identification result is output.

[0037] Compared with the maximum SoftMax probability of the unknown sample, P​H The value is usually smaller, and 50 unknown samples are input into the trained ResNet18 network to draw the maximum SoftMax probability and P H Distribution as Figure 5 shown, it can be seen that the P H of unknown samples is lower than the maximum SoftMax probability, which will increase the threshold discrimination space and improve the discrimination ability of unknown samples.

[0038] The unknown type rejection rate is used to evaluate the open set performance of the method, as shown in equation (4), which simply means the proportion of unknown types determined as unknown.

[0039]

[0040] Where A or represents the rejection rate, TU is the number of unknown samples determined as unknown types, and FU is the number of unknown samples determined as known types.

[0041] Set different open set thresholds, respectively use the open set recognition method based on the maximum SoftMax probability and the open set recognition method based on the SoftMax information entropy to test set recognition, draw the relationship curve of the threshold and the rejection rate, the false recognition rate of the two methods as Figure 6 . It can be seen from Figure 6 that compared with the open set recognition method based on the maximum SoftMax probability, the open set recognition method based on the SoftMax information entropy can greatly improve the rejection rate of unknown samples, and has better open set performance. Compared with the most basic open set recognition method based on the maximum SoftMax probability, the present application can improve the average rejection rate by about 10% under the same discrimination threshold, and has better open set ability.

Claims

1. A radar signal open set identification method based on SoftMax information entropy, characterized in that, The method comprises the steps of: A data sample generation step: time-frequency images of radar signals with known model category labels are used as training samples to form a training set; A network training step: a deep learning network with a SoftMax layer is used as a time-frequency image classification network, and the training set is input into the time-frequency image classification network to complete training; 3) Open set identification: The time-frequency image of the radar signal to be identified is input into the trained time-frequency image classification network, and the time-frequency image classification network outputs an identification result, which is a SoftMax probability distribution of known models; the information entropy is calculated based on the output SoftMax probability distribution of known models, the uncertainty degree of the identification result is reflected through the information entropy, and it is judged whether the uncertainty degree of the identification result is higher than a preset or not; if yes, the identification result of the radar signal to be identified is updated as an unknown radar signal category and output, otherwise, the radar signal category with the highest probability value in the identification result is output.

2. The method of claim 1, wherein, The specific way to judge whether the uncertainty degree of the identification result is higher than the preset is to compare whether the information entropy H is greater than a preset minimum uncertainty tolerance or not; if yes, it belongs to the case that the uncertainty degree of the identification result is higher than the preset.

3. The method of claim 1, wherein, The specific manner of judging whether the uncertainty degree of the recognition result is higher than the preset is to calculate an information entropy probability P H , P H =1-H, H is the information entropy, and whether the information entropy probability P H is less than the preset confidence is compared. If yes, it is determined that the uncertainty degree of the recognition result is higher than the preset.

4. The method of claim 1, wherein, The calculation method of the information entropy H is as follows: where k is the sample class serial number variable, the lower the information entropy, the more concentrated the SoftMax distribution of the sample, the higher the confidence of being judged as the known class, N represents the number of classes in the training set, p k is the probability of the current sample being judged as the kth class, q k is the intermediate value.

5. The method of claim 1, wherein, The processing method of the time-frequency image of the radar signal is: taking 200 pulse signals of each radar signal, performing short-time Fourier transform on each pulse signal, and then performing grayscale and size scaling operations to obtain a time-frequency image with a size of 224*224.

6. The method of claim 1, wherein, The deep learning network with a SoftMax layer is a ResNet18 network with a SoftMax layer added.

7. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to realize the steps of the method of claim 1.

Images