A Radar Emitter Recognition Mechanism Based on IFS-Tri-Training Classification Processing

Abstract

Radar Warning Receiver (RWR) is one of the basic pieces of combat equipment necessary for the electromagnetic situational awareness of aircraft in modern operations and requires good rapid performance and accuracy. This paper proposes a data processing flow for radar warning devices based on a hierarchical processing mechanism to address the issue of existing algorithms’ inability to balance real-time and accuracy. In the front-level information processing module, multi-attribute decision-making under intuitionistic fuzzy information (IFS) is used to process radar signals with certain prior knowledge to achieve rapid performance. In the post-level information processing module, an improved tri-training method is used to ensure accurate recognition of signals with low pre-level recognition accuracy. To improve the performance of tri-training in identifying radar emitters, the original algorithm is combined with the modified Hyperbolic Tangent Weight (MHTW) to address the problem of data imbalance in the radar identification problem. Simultaneously, cross entropy is employed to enhance the sample selection mechanism, allowing the algorithm to converge rapidly.

1. Introduction

The identification of radar emitter is a critical component of battlefield situational awareness and threat assessment. If pilots can accurately perceive enemy’s radar emitter information during weal confrontation, they can accurately assess the battlefield situation and make timely confrontation decisions [1]. As electronic countermeasures become more sophisticated, new complex system radars continue to enter the battlefield and gradually take a dominant position. Phased array radars, in particular, with beam cut and parameter random modulation capabilities, are widely used, posing new challenges for identifying radar emitter. Improving the ability of new and complex radar systems, such as phased array radars, to recognize emitter has become a critical issue that alarms must address. As a result, we can only gain an advantage in modern electronic warfare by exploring and supplementing new identification methods [2,3].

Much research on identifying radar emitter is available in the published literature. Among the early methods of identifying radar emitters are the parameter matching method and the expert system method. In the parameter matching method, the signal characteristics obtained by the measurement are directly compared with the known radar database to identify the radar emitter’s attribute information [4,5]. This method has the advantages of being quick to recognize and simple to implement. However, it is overly reliant on prior knowledge and is deficient in reasoning ability. The expert system method constructs inference rules of radar signals recognition based on the experts’ knowledge of radar attributes and then infers and recognizes the radar emitter data. This method possesses some learning and reasoning capabilities [6], but its implementation requires many radar signal parameter examples and experts’ knowledge of radar attributes. As a result, this algorithm has a low recognition efficiency and slow recognition speed.

In recent years, the advantages of machine learning algorithms in recognition have prompted an increasing number of researchers to apply cutting-edge machine learning techniques to the research of radar emitter recognition [7,8]. Currently, neural networks [9,10], Relevance Vector Machine (RVM) [11,12,13], extreme learning machine [14,15,16], weighted-xgboost [17], k-Nearest Neighbor (KNN) [18], and deep learning algorithm [19,20] are widely used in radar emitter recognition. Each machine learning algorithm has distinct advantages and disadvantages. Neural networks are well-suited for pattern recognition classification and functional approximation, but there are issues with local extremum and over (or under) learning. The KNN algorithm is easy to comprehend, but storing known instances requires a significant amount of space. The SVM algorithm is well-suited for solving recognition problems with a small sample size and a high degree of dimension, but when the types and numbers of radar emitters are large, the speed and accuracy of recognition suffer. Weighted-xgboost has a rapid learning rate, but the algorithm is more computationally complex and requires a longer training period. Deep learning, which is a deeper artificial neural network, has a good performance when dealing with large data volumes and multidimensional feature parameters. However, the algorithm’s disadvantage is that the training is time-consuming and requires a large amount of labeled data. In practice, obtaining such a large number of non-cooperative radar emitter data is challenging.

Many researchers improved those algorithms in response to the aforementioned issues. One study empirically decomposed each radar pulse signal, extracted bispectral features, and fed the reduced bispectrum features into the one-dimensional LeNet neural network as the transmitter’s fingerprint features [21]. Some other studies employed different techniques, such as the support vector machine (SVM) radar emitter identification method, which circumvents the slow processing speed of SVM on large sample data based on affinity propagation (AP) clustering [22]; or the hierarchical extreme learning machine (BS+H-ELM), in which the sparse autoencoder (AE) in the H-ELM is used for feature learning after the bispectrum of the radar signal is extracted [23]. The Mutation Gray Wolf Elite Particle Swarm Optimization Balanced eXtreme Gradient Boosting (MGWEPSO-BXGBoost) method, proposed in [24], optimizes the main parameters of BXGBoost using MGWEPSO. The algorithm also makes use of the wolf pack leadership mechanism, elite rules, and mutation thinking to overcome the local optimal solution problem. A two-stage multi-core extreme learning machine (TSMKELM) method to identify specific radar emission sources was proposed in [25], while a multi-feature automatic extraction and fusion method based on deep ensemble learning that uses a convolutional neural network (CNN) to extract and fuse multiple features through a data-driven strategy was introduced in [26]. Moreover, one study [27] analyzed signal features using the short-time Fourier transform (STFT) and k-means algorithm, generated a video spectrogram, and finally used a convolutional neural network (CNN) for automatic recognition based on time-frequency images.

Radar Warning Receiver (RWR) is critical for reconnaissance missions. It places a premium on signal recognition speed and accuracy. The alarm’s existing signal processing flow and algorithm cannot incorporate real-time and accuracy. To meet the index requirements for rapid and accurate identification of radar alarms, this paper proposed a hierarchical processing mechanism for identification. During data processing, the former module leveraged the rapidity of the parameter matching method to perform rapid identification, while the latter module leveraged the high recognition accuracy of the machine learning algorithm to perform accurate identification. Notably, to investigate a novel algorithm for identifying radar emitters, this paper introduced the tri-training algorithm to identify the radar emitter. This algorithm employed different data samples to train three distinct weak classifiers, which exhibited strong robustness and generalization capabilities and can also address the issue of a small sample size. To improve the original tri-training algorithm’s suitability for emitter identification, this paper employed the modified Hyperbolic Tangent Weight (MHTW) and cross-entropy calculation.

2. Hierarchical Processing Model

The basic components of a traditional radar warning device include an antenna, a receiver, a back-end processor, a display, and a recorder [28]. The basic information processing flow is shown in Figure 1. The antenna detects the radar signal in the space and directs it to the receiver, which completes the parameter measurement. The measured parameters form a list of characteristic parameter data in time series. Basic parameters include: Radio Frequency (RF), Time of Arrival (TOA), Pulse Width (PW), Pulse Amplitude (PA), and Angle of Arrinal (AOA). The list of data is sent to the backend processor. The latter performs signal classification, identification, and threat assessment, based on a list of data. The final output is sent to the monitor, which displays the alarm information. The alarm’s existing signal processing flow processes the radar signal sequentially according to the time sequence. Simultaneously, the processing interval is typically 2–3 pulses to maintain the alarm’s real-time nature. If the system cannot sort and identify the signal, it will perform the “unknown” processing. After marking, it will either discard the data immediately or store it in the radar signal database, awaiting analysis and processing by ground data analysts.

Currently, emitter identification is accomplished through apriori information from the emitter database. Due to the widespread use of phased array radars, the ability to change radar parameters is strong, resulting in an increase in the number of unknown signals in the electromagnetic space. Algorithms that rely on a large amount of prior knowledge to identify unknown signals cannot process them correctly. Deep learning algorithms can be used for unknown or aliased signals, but they require a long computing time and cannot achieve real-time performance.

To address the issues inherent in the emitter identification link in the preceding traditional data processing procedures, this paper proposed a hierarchical processing mechanism for radar emitter identification. The model is depicted in Figure 2.

The processing flow of the airborne alarm system based on the hierarchical processing mechanism demonstrates the consideration of both the real-time attribute and the accuracy of alarms. Simultaneously, it possesses technical characteristics such as parallelism, scalability, upgradeability, and data fusion. The specific summary is as follows:

(1)

Parallel reconnaissance. Parallel processing is used in a pre-stage-post-stage fashion, with pre-stage focusing on fast sorting and post-stage focusing on precise sorting, platform model identification, and comprehensive identification of “platform model + working status.” It not only meets the timeliness requirements of reconnaissance warning, it also improves the classification accuracy and identification level of advanced systems and unknown radar sources.

(2)

Scalable. The post-stage sorting and identification are accomplished through a scalable and open architecture. With the advancement of signal processing technology and artificial intelligence algorithms, the background reconnaissance module can be equipped with new signal characteristics and sorting and recognition algorithms.

(3)

Online update capability. For the unknown emitter parameters that recur after the subsequent classification and identification, the data are written into the threat database following recognition by the deep learning algorithm, allowing the threat database to be updated and upgraded online. Additionally, the background sorting and recognition algorithm can also be upgraded after combat for increased efficiency.

(4)

Real-time utilization of reconnaissance data. Compared to the existing alarm system’s processing method of “throwaway” data immediately after processing, the alarm architecture with pre-and post-level memory can enable data reuse and full mining of the information. Signals that cannot be processed by the preceding stage may be forwarded to the subsequent stage for processing and analysis. Additionally, it does not require ground personnel to conduct any post-analysis, which increases the warning device’s combat capability during a war. Post-processing can alleviate the processing load on an airborne electronic reconnaissance system operating in dense pulse flow conditions and improve the system’s adaptability in those conditions.

3. Radar Emitter Recognition Based on IFS

3.1. IFS Multi-Attribute Decision-Making Theory

IFS multi-attribute decision-making is a probabilistic multi-attribute decision-making technique. Its processing process is straightforward and highly automated. Since its decision information is provided in the form of intuitionistic fuzzy numbers or interval intuitionistic fuzzy numbers, it can deal with problems involving uncertainty [29,30,31]. The IFS multi-attribute decision-making algorithm was introduced in this paper to address the problem of radar emitter identification. The degrees of membership and non-membership were used to describe the fuzzy correspondence between the radar emitter signal attributes and individual emitters. The ICP comparison matrix is used to quantify each operating mode’s “pros and cons”.

To assist in the selection of an appropriate membership function that will enhance the decision-making model’s reliability and effectiveness, this section discusses the Gaussian Membership Function [32,33,34]. As decision attributes, Radio Frequency (RF), Pulse Repetition Interval (PRI), Pulse Width (PW), and Duty Cycle (DC) were used. Simultaneously, the membership relationship of the radar signal was calculated by combining the attribute intervals corresponding to various radar operating parameters.

Suppose ${\mu }_{ij}(i=1,2,\cdots ,m;j=1,2,\cdots ,n)$ represents the degree to which the decision target $s_i$ belongs to the object $x_j$. ${\nu }_{ij}$ represents the degree to which the decision objective $s_i$ does not belong to the object $x_j$. The relationship between ${\mu }_{ij}$ and ${\nu }_{ij}$ is ${\mu }_{ij}+{\nu }_{ij}=1$, and its calculation formula is shown in Equation (1) [32].

$${\mu }_{ij}=\exp \left(-\frac{{\left({\theta }_{ix}-\frac{{\theta }_{i\min }+{\theta }_{i\max }}{2}\right)}^2}{2{\lambda }_{\theta }^2}\right)$$

(1)

where ${\lambda }_{\theta }$ is the parameter of the Gaussian membership function, which is set according to the actual situation and is used to ensure the value of the membership degree falls within a reasonable interval. ${\theta }_{ix}$ represents a target $s_j$ characteristic attribute. ${\theta }_{i\min }$ represents the upper boundary value of the attribute, and ${\theta }_{i\max }$ is the lower boundary.

The signal $s_x$ characteristic of a certain radar signal acquired by the alarm is $(P{F}_x,PR{I}_x,P{W}_x,D{C}_x)$. Following Equation (1), it can be concluded that the membership degree of each signal feature corresponding to the radar emitter is ${\mu }_{iRF},{\mu }_{iPRI},{\mu }_{iPW},{\mu }_{iDC}$, and the following intuitionistic fuzzy matrix Q can be constructed as shown in Equation (2).

$$Q=\left(\begin{array}{cccc}\left({\mu }_{1RF},v_{1RF}\right)& \left({\mu }_{1PRI},v_{1PRI}\right)& \left({\mu }_{1PW},v_{1PW}\right)& \left({\mu }_{1DC},v_{1DC}\right)\\ \left({\mu }_{2RF},v_{2RF}\right)& \left({\mu }_{2PRI},v_{2PRI}\right)& \cdots & \left({\mu }_{2DC},v_{2DC}\right)\\ ⋮& ⋮& ⋮& ⋮\\ \left({\mu }_{mRF},v_{mRF}\right)& \left({\mu }_{mPRI},v_{mPRI}\right)& \left({\mu }_{mPW},v_{mPW}\right)& \left({\mu }_{mDC},v_{mDC}\right)\end{array}\right)$$

(2)

Assuming that $z_{ij}$ represents the “pros-and-cons” relationship between object $x_i$ and $x_j$ in decision-making, the value range of $z_{ij}$ is $[0,1]$. The expression of $z_{ij}$ is shown in Equation (3).

$$z_{ij}=\max \left\{1-\max \left\{\frac{{\pi }_i}{{\pi }_i+{\pi }_j},0\right\},0\right\}$$

(3)

where ${\pi }_i={\displaystyle \prod _{l=1}^n{\left(1-{\mu }_{il}{}^{\prime }\right)}^{{\omega }_l}}-{\displaystyle \prod _{l=1}^n{\left({\nu }_{il}{}^{\prime }\right)}^{{\omega }_l}}$,${\omega }_l$ is the weight of each target attribute. This paper uses intuitionistic fuzzy entropy theory [34,35] to assign attribute weights. The calculation formula of the intuitionistic fuzzy entropy value $W_j$ is shown in Equation (4) [34].

$$W_j=\frac{1}{n}{\displaystyle \sum _{i=1}^n\frac{1-\max \left({\mu }_{ij}{}^{\prime },{\nu }_{ij}{}^{\prime }\right)}{1-\min \left({\mu }_{ij}{}^{\prime },{\nu }_{ij}{}^{\prime }\right)}}$$

(4)

The calculation formula for the weight ${\omega }_l$ of attribute $s_j$ is shown in Equation (5) [35].

$${\omega }_l=\frac{1-W_j}{n-{\displaystyle \sum _{j=1}^n{W}_j}}$$

(5)

The ICP comparison matrix $Z$ can be constructed using the derivation from Equation (5), as shown in Equation (6).

$$Z=\left[\begin{array}{cccc}{z}_{11}& z_{12}& \cdots & z_{1m}\\ z_{21}& z_{22}& \cdots & z_{2m}\\ ⋮& ⋮& \ddots & ⋮\\ z_{m1}& z_{m2}& \cdots & z_{mm}\end{array}\right]$$

(6)

Finally, the value of “pros and cons” ${\alpha }_i$ can be calculated with Equation (7).

$${\alpha }_i=\frac{1}{m\left(m-1\right)}\left({\displaystyle \sum _{l=1}^m{z}_{il}}+\frac{m}{2}-1\right)$$

(7)

The higher the value of $\alpha$, the more confident the algorithm is in the object. The object with the greatest $\alpha$ value is the optimal decision-making object.

3.2. Algorithm Processing Flow

The specific implementation steps of the algorithm are shown in Algorithm 1:

Algorithm 1: IFS Radar Working Pattern Recognition Algorithm

Input: Receiving radar signal Output: Emitter identification information step: Step 1.  Extract signal pulse parameters. Extract parameters such as Radio Frequency (RF), Pulse Repetition Interval (PRI), Pulse Width (PW), and Duty Cycle (DC). Step 2.  Construct the intuitionistic fuzzy matrix. Construct the intuitionistic fuzzy matrix according to the calculated membership of each characteristic parameter and the information stored in the threat database. Step 3.  Assign attribute weights. Measure the importance of attributes in the decision-making process based on the intuit tionistic fuzzy entropy theory and obtain the weight vector. Step 4.  Quantify the “pros and cons” of the radar working mode. Combine the intuitionistic fuzzy matrix of Step 2 and the weight vector of Step 3 to construct the ICP matrix. Then, calculate the “pros-and-cons” value of each working mode. The working mode corresponding to the maximum value is the recognition result.

4. Radar Working Status Recognition based on Improved Tri-Training

The tri-training algorithm is a collaborative training method based on divergence proposed by Zhou and Li [35,36,37]. Compared with collaborative algorithms such as the collaborative regularization method and collaborative EM method, the tri-training algorithm has the advantage of not requiring a sufficiently large number of redundant views in the dataset. The tri-training is a classification algorithm composed of three base classifiers. The classifiers are trained using a variety of perspectives, and the results are obtained via classifier voting. In comparison to other collaborative training methods, which require a significant amount of time to cross-validate the flaws of unlabeled data confidence, the tri-training algorithm is more efficient [38,39,40].

4.1. The Difficulty of Classifying Unlabeled Data Using Cross Entropy

The difficulty of data classification describes how difficult it is for the classifier to classify unlabeled samples correctly. That is, whether the unlabeled samples have apparent characteristics from the perspective of the classifier. This paper introduces the concept of cross entropy to illustrate the difficulty of data classification.

Suppose the unlabeled data are $x_c$, and the average cross entropy of all training sample data in the training sample set of the $i(i=2,3)$ classifier is shown in Equation (8).

$${\widehat{L}}_i\left(x_c\right)=-\frac{1}{{\displaystyle \sum _{j\ne 1}{k}_{ij}}}{\displaystyle \sum _{j\ne 1}{\displaystyle \sum _{k=1}^{k_{ij}}{x}_{ijk}\ln \left(x_c\right)}}$$

(8)

where ${\widehat{L}}_i\left(x_c\right)$ is the average cross entropy of the input data of $x_c$ and the training sample of the $i$th classifier, $x_{ijk}$ is the $k$-th input data of the $i$th sample classified as $j$ by the classifier, and $k_{ij}$ is the training sample of the $i$th classifier classified as $j$ quantity.

Then, the identification difficulty weight $d_{ijk}\left(x_c\right)$ between the unlabeled data input $x_c$ and the labeled data is shown in Equation (9).

$$d_{ijk}\left(x_c\right)=\left\{\begin{array}{cc}{\widehat{L}}_i+x_{ijk}\ln \left(x_c\right)& ,x_{ijk}\ln \left(x_c\right)<{\widehat{L}}_i\\ 0& ,x_{ijk}\ln \left(x_c\right)>{\widehat{L}}_i\end{array}\right.$$

(9)

The average difficulty weight for a single category is determined using Equation (10).

$$D_{ij}=\frac{1}{k_0}{\displaystyle \sum _k^{k_{ij}}{d}_{ijk}}$$

(10)

where $k_0$ is the number of marked data in $d_{ijk}\ne 0$.

The classification difficulty of unlabeled data $x_c$ in the $i$-th classifier according to the obtained difficulty weight is calculated using Equation (11).

$$e=\frac{\underset{j}{\min }\left(D_{ij}\right)}{\underset{j}{\max }\left(D_{ij}\right)}$$

(11)

The tri-training algorithm’s central idea is to cross-validate the unlabeled data $U$ and expand each classifier’s training sample set. The difficulty of the expanded training set is likely to cause the classifier to fail to converge during the expansion process. The classification difficulty of the sample is calculated using cross entropy. If the sample classification is too tricky, it is assumed that this classification may be inaccurate or may encounter noise interference. This result should not be used as a training sample for the first classifier.

4.2. The Use of MHTW to Reduce Data Imbalance

The training samples in the problem of radar working status recognition are the characteristic parameters of various radar models. Indeed, whether they constitute pre-acquired intelligence acquisition or real-time acquisition on the battlefield, the signal data collected by various types of radars suffer from an imbalance problem. To mitigate the impact of data imbalance, this paper proposed an adaptive class MHTW modified by the hyperbolic tangent function. The method for calculating the MHTW is shown in Equation (12).

$$W_i=1+\frac{e^{\sigma K_i}-e^{-\sigma K_i}}{e^{\sigma K_i}+e^{-\sigma K_i}}$$

(12)

where $W_i$ is the weight of the $i$-th class sample, $\sigma$ is the scaling parameter, and $K_i$ is the number parameter of the $i$-th class sample, as shown in Equation (13).

$$K_i=\frac{N_{\max }}{N_i}-1$$

(13)

The constant term combined with the hyperbolic tangent function constrains the upper and lower bounds of MHTW, increasing the model’s stability during the training process. Additionally, $\sigma$ can be adjusted to accommodate datasets with the disparate number of multiple gaps between different classes. This paper plots the change of MHTW value with different $\sigma$ values. As shown in Figure 3.

4.3. Improved Tri-Training Algorithm Flow

This paper improved the tri-training algorithm by utilizing cross-entropy-based unlabeled data classification difficulty. Additionally, a model for recognizing the radar’s operational status was established using the MHTW-tri-training algorithm, as illustrated in Algorithm 2.

Algorithm 2: Improved Tri-Training Algorithm

Input: Marked radar signal set $T$, unmarked radar signal set $U$, and scaling parameter $\sigma$ Output: Stable classifier $H$ steps: Step 1.  According to the marked and received unmarked radar emitter signal parameters in the intelligence content, the characteristic parameters used to evaluate the working status of the radar are calculated. The Bootstrap method is used to sample and collect the characteristic data $L$ of the radar emitter in the loaded library, and the different training subsets $L_1,L_2,L_3$ are obtained. Step 2.  Based on the MHTW-BP neural network and three training subsets mentioned in Section 3.2, three different base classifiers are established. Step 3.  Expand the base classifier training sample set. Taking classifier $H_1$ as an example, if $c_{2i}\in C_2$, $c_{3i}\in C_3$ and $c_{2i}=c_{3i}$, then $c_{i2}$ is used as a new labeled training sample to learn from $H_1$. Similarly, classifiers $H_2,H_3$ use the same method to expand their training sample set. Among them, $c_{ij}$ is the classification result of the $i$th unlabeled data by the $j$th classifier. Step 4.  When there are two or more base classifiers to make the same classification of the same unlabeled data, calculate the classification difficulty $e$ of the data according to the method mentioned in Section 3.1. If $e<\delta$, then the data and its majority classification result are taken as the labeled data and added to the training subset of another base classifier. If $e>\delta$, it remains unchanged. Step 5.  The newly received radar information data are added to the unlabeled data sets of the three base classifiers. Step 3 and Step 4 are repeated until the task ends.

5. Simulation Experiment and Analysis

5.1. Simulation Data Generation

The data used to simulate the electromagnetic environment in this paper were collected using electromagnetic environment pulse generation software. The concept of “operational scenario design-scenario data generation-pulse signal simulation” was adopted to generate radar characteristic spectrum datasets. Figure 4 shows the specific method used to generate the dataset discussed in this paper.

Several radar parameters are set in the simulation system. According to different radar types, 5 radar signal data are randomly selected, namely, the early warning radar (radar 5), the guidance radar (radar 1), and the airborne radars (radar 2, radar 4, and radar 3). The range of attributes associated with each radar is listed in

Table 1. Radar working mode attribute range.

Radar Working Mode	$\mathit{R}\mathit{F}/\mathit{M}\mathit{h}\mathit{z}$	$\mathit{P}\mathit{R}\mathit{I}/\mathit{\mu }\mathit{s}$	$\mathit{P}\mathit{W}/\mathit{\mu }\mathit{s}$
Radar 5	[3000, 3600]	[500, 700]	[10, 54]
Radar 1	[5200, 5500]	[300, 350]	[5, 10]
Radar 3	[9800, 11,000]	[25, 40]	[1, 7]
Radar 4	[9600, 9900]	[400, 1500]	[0.5, 2]
Radar 2	[9500, 9900]	[30, 60]	[0.5, 2]

.

In the data generation and combat scenario design, the environmental pulse density is 1 million per second. The combat platform’s participation phase and intervention time in the entire combat process change as the combat process evolves, as does the switch time of its radar. This paper selects 977,190 pulse data generated during 105 s to 140 s in the demonstration process. The data parameter distribution is shown in Figure 5. Among them: radar 1 has a total of 106,827 pulse data; radar 2 has a total of 94,043 data; radar 3 has a total of 128,610 data; radar 4 has a total of 605,650 data; and radar 5 has a total of 20,700 data.

5.2. Radar Emitter Identification Based on IFS Decision

The IFS algorithm can process individual pulses received by the alarm unit without requiring a large amount of training data. To verify the IFS algorithm’s simulation calculation time, this study selected 1000 pulses and calculated the time used for each pulse. Since this paper does not have the conditions for the actual application of the algorithm, the MATLAB simulation platform is used to verify the rapid computing capability of the algorithm. The simulation software environment used in this paper is listed in

Table 2. Software configuration environment.

CPU	Clock Speed	RAM	MATLAB Version
Core(TM) i5-10210U CPU	1.60 GHz	16.0 GB	2019

. The parameter list of 1000 pulse features sequentially counts the time consumed by each pulse through the IFS algorithm. Statistics are obtained using the Monte Carlo method. The running time obtained by the simulation is shown in Figure 6. As shown in the figure, the time required to calculate the single pulse is approximately 30 us, which can achieve the rapid identification recognition and alarm for the single pulse. However, the calculation time of the characteristic parameters for individual pulse data is long, and peaks appear in the graph.

In this paper, the data and radar parameters of the electromagnetic environment pulse generation software were fed into the IFS decision-making algorithm. The recognition outcome is depicted in Figure 7.

It can be seen from the simulation diagram that the radar 1 and radar 5 can be accurately identified by the IFS-based identification algorithm. The IFS algorithm has low recognition accuracy for radars 2, 3, and 4. Radar 4 and the Radar 2 have serious aliasing recognition errors, and the two are often identified as each other’s models. Radar 4 and Radar 2 appear in the recognition results of Radar 3. To reflect the IFS recognition results for each radar model, this study calculated recognition statistics, as shown in

Table 3. Recognition accuracy for radar.

Radar Number	Radar 1	Radar 2	Radar 3	Radar 4	Radar 5
Recognition results	99.91%	51.45%	84.87%	53.23%	99.51%

.

As be seen from the data, Radars 1 and 5 have a recognition accuracy of greater than 99%. Radar 3 has an approximate recognition accuracy of 85%, while Radars 2 and 4 have merely around 50% recognition accuracy. To better understand the reasons for the identification results, this paper compared the identified result data with the original parameter data. Radar 1 and Radar 5 recognition errors are primarily caused by inaccuracies in raw data collected during reconnaissance and measurement. The operating parameters of Radars 2, 3, and 4 are similar, as are the parameters of the detected signal, which is prone to error when making IFS decision-making recognition. This recognition error is most noticeable in datasets where the operating parameters of Radars 2 and 4 are nearly identical.

5.3. Recognition based on the Improved Tri-Training Algorithm

To address the low-recognition accuracy of Radars 2, 3, and 4, the data were transmitted to the post-processing module, in which an improved tri-training algorithm was adopted for radar source identification. To evaluate the proposed algorithm more accurately, this paper compared accuracy (Accuracy, ACC), average recall rate (mean Average Recall, mAR) and average precision (mean Average Precision, mAP), as summarized in

Table 4. Radar recognition evaluation index.

Algorithm	ACC/%	mAR/%	mAP/%
Tri-Training	88.91	71.45	87.87
BP	79.89	51.10	58.88
SVM	84.86	66.17	63.14

.

As shown in the table, the improved tri-training algorithm recognizes objects with an recognition accuracy of 88.91%, an increase of 38% compared to the IFS decision recognition algorithm used in the pre-processing module. Additionally, the improved tri-training algorithm has advantages over the BP and the SVM algorithms. This paper employed the Monte Carlo method to loop 200 times to verify the performance of the tri-training algorithm better. Figure 8 illustrates the simulation results. Specifically, Figure 8b–d depict the simulation results of Radars 2, 3, and 4, respectively, while Figure 8a shows the total recognition results of the three radars.

As illustrated in the figure, the improved tri-training algorithm is more accurate at recognizing radar emitters than the BP and the SVM algorithms. However, the simulation results vary significantly, particularly when recognizing Radars 2 and 4. The primary reason is that there is uncertainty in the sub-dataset selection, which substantially impacts the accuracy of each training.

6. Conclusions

Recognizing radar emitter is critical for accurately assessing the degree of a radar threat. This paper examined and resolved the need for rapid and accurate identification of the alarm device. Additionally, novel methods for improving the recognition of radar emitter were investigated. The hierarchical processing mechanism’s signal processing flow was proposed. The front stage employs the IFS decision-making algorithm, which can recognize and alarm single pulses. The latter stage employed tri-training to determine the source of collected radar signal. The MHTW was used to improve the algorithm’s ability to deal with unbalanced data, making it more suitable for emitter identification. Simultaneously, cross-entropy calculations were used to enhance the algorithm’s capacity for dataset selection. Based on the simulation results, it was found that the IFS algorithm was capable of recognizing radar models without significant overlap in operating parameters. The enhanced tri-training algorithm enabled the recognition of overlapped radar signals. The correct rate of recognition, on the other hand, was constrained by the initial selection of the sub-dataset. In future research, the initial selection of improved subsets can be examined to improve the stability and accuracy of the tri-training recognition accuracy.